1 Why bioLeak

bioLeak is a leakage-aware modeling toolkit for biomedical and machine-learning analyses. Its purpose is to prevent and diagnose information leakage across resampling workflows where training and evaluation data are not truly independent because samples share subjects, batches, studies, or time.

Standard workflows are often insufficient. Random, row-wise cross-validation assumes samples are independent. Global preprocessing (imputation, scaling, feature selection) done before resampling lets test-fold information shape the training process. These choices inflate performance and can lead to incorrect biomarker discovery, misleading clinical signals, or models that fail in deployment.

Data leakage means any direct or indirect use of evaluation data in training or feature engineering. In biomedical data, leakage commonly appears as:

  • repeated measurements from the same patient split across folds
  • batch or study effects that align with the outcome
  • duplicates or near-duplicates across train and test
  • time-series lookahead or random splits that use future information
  • preprocessing that uses all samples instead of training-only statistics

bioLeak addresses these issues with leakage-aware splitting, guarded preprocessing that is fit only on training data, and audit diagnostics that surface overlaps, confounding, and duplicates.

2 Guided workflow

The sections below walk through a leakage-aware workflow from data setup to audits. Each step includes a leaky failure mode and a corrected alternative.

2.1 Example data

library(bioLeak)

set.seed(123)
n <- 160
subject <- rep(seq_len(40), each = 4)
batch <- sample(paste0("B", 1:6), n, replace = TRUE)
study <- sample(paste0("S", 1:4), n, replace = TRUE)
time <- seq_len(n)

x1 <- rnorm(n)
x2 <- rnorm(n)
x3 <- rnorm(n)
linpred <- 0.7 * x1 - 0.4 * x2 + 0.2 * x3 + rnorm(n, sd = 0.5)
p <- stats::plogis(linpred)
outcome <- factor(ifelse(runif(n) < p, "case", "control"),
                  levels = c("control", "case"))

df <- data.frame(
  subject = subject,
  batch = batch,
  study = study,
  time = time,
  outcome = outcome,
  x1 = x1,
  x2 = x2,
  x3 = x3
)

df_leaky <- within(df, {
  leak_subject <- ave(as.numeric(outcome == "case"), subject, FUN = mean)
  leak_batch <- ave(as.numeric(outcome == "case"), batch, FUN = mean)
  leak_global <- mean(as.numeric(outcome == "case"))
})

df_time <- df
df_time$leak_future <- c(as.numeric(df_time$outcome == "case")[-1], 0)
predictors <- c("x1", "x2", "x3")


# Example data (first 6 rows)
head(df)
#>   subject batch study time outcome          x1         x2          x3
#> 1       1    B3    S4    1 control  0.21444531  1.0149432 -0.09736927
#> 2       1    B6    S3    2    case -0.32468591 -1.9927485  0.21615254
#> 3       1    B3    S1    3    case  0.09458353 -0.4272793  0.88246516
#> 4       1    B2    S3    4 control -0.89536336  0.1166373  0.20559750
#> 5       2    B2    S4    5 control -1.31080153 -0.8932076 -0.61643584
#> 6       2    B6    S2    6    case  1.99721338  0.3339029 -0.73479925

# Outcome class counts
as.data.frame(table(df$outcome))
#>      Var1 Freq
#> 1 control   91
#> 2    case   69

The table preview displays the metadata columns (subject, batch, study, time), the binary outcome, and three numeric predictors (x1, x2, x3).

The class count table shows the baseline prevalence. Stratified splits try to preserve these proportions (for grouped modes, stratification is applied at the group level).

2.2 Create leakage-aware splits with make_split_plan()

make_split_plan() is the foundation. It returns a LeakSplits object with explicit train/test indices (or compact fold assignments) and metadata. For data.frame inputs, a unique row_id column is added automatically, so group = "row_id" is the explicit way to request sample-wise CV. It assumes that the grouping columns you provide are complete and that samples sharing a group must not cross folds. Grouped stratification uses the majority class per group and is ignored for study_loocv, time_series, and survival outcomes. Misuse to avoid:

  • using group = "row_id" (sample-wise CV) when subjects repeat
  • using the wrong grouping column (for example batch instead of subject)
  • using time-series CV with unsorted or missing time values
  • relying on stratification when the outcome is missing, single-class at the group level, or a time/event outcome
  • assuming time_series will always return exactly v folds

Leaky example: row-wise CV when subjects repeat

leaky_splits <- make_split_plan(
  df,
  outcome = "outcome",
  mode = "subject_grouped",
  group = "row_id",
  v = 5,
  repeats = 1,
  stratify = TRUE
)

cat("Leaky splits summary (sample-wise CV):\n")
#> Leaky splits summary (sample-wise CV):
leaky_splits
#> LeakSplits object (mode = subject_grouped, v = 5, repeats = 1)
#> Outcome: outcome | Stratified: TRUE | Nested: FALSE
#> ------------------------------------------------------
#>   fold repeat_id train_n test_n
#> 1    1         1     127     33
#> 2    2         1     128     32
#> 3    3         1     128     32
#> 4    4         1     128     32
#> 5    5         1     129     31
#> ------------------------------------------------------
#> Total folds: 5 | Hash: b1c98ff810948d48352dc8426a350983

The printed LeakSplits summary reports the split mode, number of folds, and fold-level training and test set sizes.

Because group = "row_id", each sample is treated as its own group. As a result, repeated samples from the same subject may appear in both the training and test sets, which introduces information leakage.

Leakage-safe alternative: group by subject

safe_splits <- make_split_plan(
  df,
  outcome = "outcome",
  mode = "subject_grouped",
  group = "subject",
  v = 5,
  repeats = 1,
  stratify = TRUE,
  seed = 10
)

cat("Leakage-safe splits summary (subject-grouped CV):\n")
#> Leakage-safe splits summary (subject-grouped CV):
safe_splits
#> LeakSplits object (mode = subject_grouped, v = 5, repeats = 1)
#> Outcome: outcome | Stratified: TRUE | Nested: FALSE
#> ------------------------------------------------------
#>   fold repeat_id train_n test_n
#> 1    1         1     124     36
#> 2    2         1     128     32
#> 3    3         1     128     32
#> 4    4         1     128     32
#> 5    5         1     132     28
#> ------------------------------------------------------
#> Total folds: 5 | Hash: 0b297788da70728ebdae0ff2f83ee0d9

Here, each subject is confined to a single fold. Test set sizes are multiples of the number of samples per subject, confirming that subjects were not split across folds.

The fold sizes remain comparable because stratify = TRUE balances outcome proportions across folds while respecting subject-level boundaries.

Other leakage-aware modes

batch_splits <- make_split_plan(
  df,
  outcome = "outcome",
  mode = "batch_blocked",
  batch = "batch",
  v = 4,
  stratify = TRUE
)

cat("Batch-blocked splits summary:\n")
#> Batch-blocked splits summary:
batch_splits
#> LeakSplits object (mode = batch_blocked, v = 4, repeats = 1)
#> Outcome: outcome | Stratified: TRUE | Nested: FALSE
#> ------------------------------------------------------
#>   fold repeat_id train_n test_n
#> 1    1         1      73     87
#> 2    2         1     138     22
#> 3    3         1     132     28
#> 4    4         1     137     23
#> ------------------------------------------------------
#> Total folds: 4 | Hash: eeacae466015baafa3b81a09cf6150fd

study_splits <- make_split_plan(
  df,
  outcome = "outcome",
  mode = "study_loocv",
  study = "study"
)

cat("Study leave-one-out splits summary:\n")
#> Study leave-one-out splits summary:
study_splits
#> LeakSplits object (mode = study_loocv, v = 5, repeats = 1)
#> Outcome: outcome | Stratified: FALSE | Nested: FALSE
#> ------------------------------------------------------
#>   fold repeat_id train_n test_n
#> 1    1         1      96     64
#> 2    2         1     121     39
#> 3    3         1     131     29
#> 4    4         1     132     28
#> ------------------------------------------------------
#> Total folds: 4 | Hash: 08feeefcf0e5ceb839fe3f4f198dfab6

time_splits <- make_split_plan(
  df,
  outcome = "outcome",
  mode = "time_series",
  time = "time",
  v = 4,
  horizon = 2
)

cat("Time-series splits summary:\n")
#> Time-series splits summary:
time_splits
#> LeakSplits object (mode = time_series, v = 4, repeats = 1)
#> Outcome: outcome | Stratified: FALSE | Nested: FALSE
#> ------------------------------------------------------
#>   fold repeat_id train_n test_n
#> 1    2         1      39     40
#> 2    3         1      79     40
#> 3    4         1     119     40
#> ------------------------------------------------------
#> Total folds: 3 | Hash: a73baf6d3a817601acb7e15d8028b441

nested_splits <- make_split_plan(
  df,
  outcome = "outcome",
  mode = "subject_grouped",
  group = "subject",
  v = 3,
  nested = TRUE,
  stratify = TRUE
)

cat("Nested CV splits summary:\n")
#> Nested CV splits summary:
nested_splits
#> LeakSplits object (mode = subject_grouped, v = 3, repeats = 1)
#> Outcome: outcome | Stratified: TRUE | Nested: TRUE
#> ------------------------------------------------------
#>   fold repeat_id train_n test_n
#> 1    1         1     104     56
#> 2    2         1     104     56
#> 3    3         1     112     48
#> ------------------------------------------------------
#> Total folds: 3 | Hash: 97e5fbd9cdcc675432bef43384a02337

Each summary reports the number of folds and the fold sizes for the corresponding split strategy.

  • Batch & Study: The test set sizes vary because batches/studies are not evenly sized. study_loocv uses one study per fold, so v and repeats are ignored.

  • Time series: The training set size increases across folds while the test set size follows the block size. Early blocks with no prior training data (or too-small test windows) are skipped, so you may see fewer than v folds.

Assumptions and intent by mode

  • subject_grouped: keeps all samples from a subject together
  • batch_blocked: keeps batches/centers together (leave-one-group-out when v is at least the number of batches; otherwise multiple batches per fold)
  • study_loocv: holds out each study in turn for external validation
  • time_series: rolling-origin splits with an optional horizon to prevent lookahead

Use repeats for repeated CV in grouped/batch modes (it is ignored for study_loocv and time_series), stratify = TRUE to balance outcome proportions at the group level when applicable, and nested = TRUE to attach inner folds (one repeat). For large datasets, progress = TRUE reports progress; storing explicit indices can be memory intensive.

2.3 Scalability

Handling Large Datasets (Compact Mode)

For large datasets (e.g., \(N > 50,000\)) with many repeats, storing explicit integer indices for every fold can consume gigabytes of memory. Use compact = TRUE to store a lightweight vector of fold assignments instead. fit_resample() will automatically reconstruct the indices on the fly. Compact mode is not supported when nested = TRUE (it falls back to full indices). For time_series compact splits, the time column must be present in the stored split metadata so folds can be reconstructed.

# Efficient storage for large N
big_splits <- make_split_plan(
  df,
  outcome = "outcome",
  mode = "subject_grouped",
  group = "subject",
  v = 5,
  compact = TRUE  # <--- Saves memory
)

cat("Compact-mode splits summary:\n")
#> Compact-mode splits summary:
big_splits
#> LeakSplits object (mode = subject_grouped, v = 5, repeats = 1)
#> Outcome: outcome | Stratified: FALSE | Nested: FALSE
#> ------------------------------------------------------
#>   fold repeat_id train_n test_n
#> 1    1         1     128     32
#> 2    2         1     128     32
#> 3    3         1     128     32
#> 4    4         1     128     32
#> 5    5         1     128     32
#> ------------------------------------------------------
#> Total folds: 5 | Hash: 0c2f23413d312968be7b81c660011fc9
cat(sprintf("Compact storage enabled: %s\n", big_splits@info$compact))
#> Compact storage enabled: TRUE

The summary is identical to a regular split, but the underlying storage is a compact fold-assignment vector. Use the compact flag to confirm the memory- saving mode is active.

2.4 Guarded preprocessing and imputation

bioLeak uses guarded preprocessing to prevent leakage from global imputation, scaling, and feature selection. The low-level API is:

  • .guard_fit() to fit preprocessing on training data only
  • predict_guard() to apply the trained guard to new data
  • .guard_ensure_levels() to align factor levels across train/test
  • impute_guarded() as a convenience wrapper for leakage-safe imputation

.guard_fit() fits a pipeline that can winsorize, impute, normalize, filter, and select features. It one-hot encodes non-numeric columns and carries factor levels forward to new data. Assumptions and misuse to avoid:

  • fs = "ttest" requires a binary outcome with enough samples per class
  • fs = "lasso" requires glmnet
  • impute = "mice" is not supported for guarded prediction
  • impute = "none" with missing values will trigger a median fallback and missingness indicators to avoid errors

Supported preprocessing options include imputation (median, knn, missForest, none), normalization (zscore, robust, none), filtering by variance or IQR (optionally min_keep), and feature selection (ttest, lasso, pca). Winsorization is controlled via impute$winsor and impute$winsor_k in guarded steps; in impute_guarded(), use winsor and winsor_thresh. impute_guarded() returns a LeakImpute object with guarded data, diagnostics, and the fitted guard state.

Leaky example: global scaling before CV

df_leaky_scaled <- df
df_leaky_scaled[predictors] <- scale(df_leaky_scaled[predictors])
scaled_summary <- data.frame(
  feature = predictors,
  mean = colMeans(df_leaky_scaled[predictors]),
  sd = apply(df_leaky_scaled[predictors], 2, stats::sd)
)
scaled_summary$mean <- round(scaled_summary$mean, 3)
scaled_summary$sd <- round(scaled_summary$sd, 3)

# Leaky global scaling: means ~0 and SDs ~1 computed on all samples
scaled_summary
#>    feature mean sd
#> x1      x1    0  1
#> x2      x2    0  1
#> x3      x3    0  1

The summary shows that scaling used the full dataset, so test-fold statistics influenced the training transformation. This violates the train/test separation and can bias performance estimates.

Leakage-safe alternative: fit preprocessing on training only

fold1 <- safe_splits@indices[[1]]
train_x <- df[fold1$train, predictors]
test_x <- df[fold1$test, predictors]

guard <- .guard_fit(
  X = train_x,
  y = df$outcome[fold1$train],
  steps = list(
    impute = list(method = "median", winsor = TRUE),
    normalize = list(method = "zscore"),
    filter = list(var_thresh = 0, iqr_thresh = 0),
    fs = list(method = "none")
  ),
  task = "binomial"
)

train_x_guarded <- predict_guard(guard, train_x)
test_x_guarded <- predict_guard(guard, test_x)

cat("GuardFit object:\n")
#> GuardFit object:
guard
#> Guarded preprocessing pipeline
#>  - Imputation: median
#>  - Normalization: zscore
#>  - Filter: var>=0, iqr>=0
#>  - Feature selection: none
#>  - Output features: 3
cat("\nGuardFit summary:\n")
#> 
#> GuardFit summary:
summary(guard)
#> GuardFit summary
#>  n_train=124, p_out=3, removed_by_filter=0
#>  Steps:
#>  step                 action
#>     1            winsor: k=5
#>     2   impute: none (no NA)
#>     3      normalize: zscore
#>     4 filter: var>=0, iqr>=0
#>     5               fs: none
#>  hash: bacf825a215c48894b993150135d3249

# Guarded training data (first 6 rows)
head(train_x_guarded)
#>            x1           x2          x3
#> 5  -1.4287254 -0.846746034 -0.58336326
#> 6   1.9282279  0.352168157 -0.70450573
#> 7   0.5110636  0.427913734 -0.08735148
#> 8  -1.3683145 -0.006339485  0.36484181
#> 9  -0.7187394 -2.383299658 -1.01654516
#> 10 -1.3015499  2.538309224 -0.14109035

# Guarded test data (first 6 rows)
head(test_x_guarded)
#>              x1         x2          x3
#> 1   0.119085787  1.0175596 -0.05210949
#> 2  -0.428021276 -1.9210219  0.26877352
#> 3  -0.002549215 -0.3915238  0.95073054
#> 4  -1.007141255  0.1398948  0.25797066
#> 25 -1.105722181  0.5243484  2.03908390
#> 26  1.601954989 -0.5499985  0.86726497

The GuardFit object records the preprocessing steps and the number of features retained after filtering. The summary reports how many features were removed and the preprocessing audit trail. The guarded train/test previews show that missing values were imputed and (if requested) scaled using training-only statistics; the test data never influences these values.

Factor level guard (advanced)

raw_levels <- data.frame(
  site = c("A", "B", "B"),
  status = c("yes", "no", "yes"),
  stringsAsFactors = FALSE
)

level_state <- .guard_ensure_levels(raw_levels)

# Aligned factor data with consistent levels
level_state$data
#>   site status
#> 1    A    yes
#> 2    B     no
#> 3    B    yes

# Levels map
level_state$levels
#> $site
#> [1] "A" "B"
#> 
#> $status
#> [1] "no"  "yes"

The returned data keeps factor levels consistent across folds, while the levels list records the training-time levels (including any dummy levels added to preserve one-hot columns). Use these when transforming new data to avoid misaligned model matrices.

Leaky example: imputation using train and test together

train <- data.frame(a = c(1, 2, NA, 4), b = c(NA, 1, 1, 0))
test <- data.frame(a = c(NA, 5), b = c(1, NA))

all_median <- vapply(rbind(train, test),
                     function(col) median(col, na.rm = TRUE),
                     numeric(1))
train_leaky <- as.data.frame(Map(function(col, m) { col[is.na(col)] <- m; col },
                                 train, all_median))
test_leaky <- as.data.frame(Map(function(col, m) { col[is.na(col)] <- m; col },
                                test, all_median))

# Leaky medians computed on train + test
data.frame(feature = names(all_median), median = all_median)
#>   feature median
#> a       a      3
#> b       b      1

# Leaky-imputed training data
train_leaky
#>   a b
#> 1 1 1
#> 2 2 1
#> 3 3 1
#> 4 4 0

# Leaky-imputed test data
test_leaky
#>   a b
#> 1 3 1
#> 2 5 1

The medians above were computed using both train and test data, so the test set influences the imputation values. This is a classic leakage pathway because test information directly alters the training features.

Leakage-safe alternative: impute_guarded()

imp <- impute_guarded(
  train = train,
  test = test,
  method = "median",
  winsor = FALSE
)

# Guarded-imputed training data
imp$train
#>   a b
#> 1 1 1
#> 2 2 1
#> 3 2 1
#> 4 4 0

# Guarded-imputed test data
imp$test
#>   a b
#> 1 2 1
#> 2 5 1

Here, the imputation statistics are computed from the training set only. The missing value in the test set (column a) is replaced by 2, which is the training median. In contrast, in the leaky example above, it was replaced by 3, the global median.

This confirms that the test set is transformed using fixed values learned from the training data, preserving a clean separation between training and evaluation. The LeakImpute object also contains missingness diagnostics in imp$summary$diagnostics, and guarded outputs use the same encoding as .guard_fit() (categorical variables are one-hot encoded). Use vars to impute only a subset of columns when needed.

2.5 Fit and resample with fit_resample()

fit_resample() combines leakage-aware splits with guarded preprocessing and model fitting. It supports:

  • built-in learners (glmnet, ranger)
  • parsnip model specifications or workflows::workflow objects (recommended)
  • recipes::recipe preprocessing (prepped on training folds only)
  • rsample resamples (rset/rsplit) via splits, with split_cols to map metadata
  • custom learners via custom_learners (legacy/advanced)
  • multiple metrics, class weights, positive class override, and yardstick metrics

Assumptions and misuse to avoid:

  • outcome must be binary or multiclass (factor) for classification, numeric for regression, or a survival outcome (a Surv column or time/event columns)
  • positive_class must be a single value that exists in the outcome levels
  • class_weights only applies to binomial/multiclass tasks; workflows must handle weights internally
  • if you supply a matrix without metadata, you must remove leakage columns yourself (group, batch, study, time)

Use learner_args to pass model-specific arguments. For custom learners (advanced), you can supply separate fit and predict argument lists. Set parallel = TRUE to use future.apply for fold-level parallelism when available. When learner is a parsnip specification, learner_args and custom_learners are ignored. When learner is a workflow, preprocess and learner_args are ignored. When a recipe or workflow is used, the built-in guarded preprocessing list is bypassed, so ensure the recipe/workflow itself is leakage-safe.

Metrics used in this vignette:

  • AUC: area under the ROC curve (0.5 = random, 1.0 = perfect; higher is better)
  • PR AUC: area under the precision-recall curve (0 to 1; higher is better)
  • Accuracy: fraction of correct predictions (0 to 1; higher is better)
  • RMSE: root mean squared error (0 to infinity; lower is better)
  • Macro F1: average F1 across classes (multiclass; higher is better)
  • Log loss: cross-entropy for probabilities (lower is better)
  • C-index: concordance for regression/survival (higher is better)

Always report the mean and variability across folds, not a single fold value. When using yardstick metrics, the positive class is the second factor level; set positive_class to control this.

Parsnip model specification (recommended)

spec <- parsnip::logistic_reg(mode = "classification") |>
  parsnip::set_engine("glm")

This spec uses base R glm under the hood and does not require extra model packages. Use it in the examples below; custom learners are covered in the Advanced section.

Leaky example: leaky features and row-wise splits

fit_leaky <- fit_resample(
  df_leaky,
  outcome = "outcome",
  splits = leaky_splits,
  learner = spec,
  metrics = c("auc", "accuracy"),
  preprocess = list(
    impute = list(method = "median"),
    normalize = list(method = "zscore"),
    filter = list(var_thresh = 0),
    fs = list(method = "none")
  )
)
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%

cat("Leaky fit summary:\n")
#> Leaky fit summary:
summary(fit_leaky)
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: binomial
#> Outcome: outcome
#> Positive class: case
#> Learners: logistic_reg/glm
#> Total folds: 5
#> Refit performed: Yes
#> Hash: ha16bdf5a
#> 
#> Cross-validated metrics (mean ± SD):
#>            learner auc_mean auc_sd accuracy_mean accuracy_sd
#> 1 logistic_reg/glm     0.75  0.059         0.687       0.054
#> 
#> Audit overview:
#>  fold n_train n_test          learner features_final
#>     1     127     33 logistic_reg/glm             17
#>     2     128     32 logistic_reg/glm             17
#>     3     128     32 logistic_reg/glm             17
#>     4     128     32 logistic_reg/glm             17
#>     5     129     31 logistic_reg/glm             17
metrics_leaky <- as.data.frame(fit_leaky@metric_summary)
num_cols <- vapply(metrics_leaky, is.numeric, logical(1))
metrics_leaky[num_cols] <- lapply(metrics_leaky[num_cols], round, digits = 3)

# Leaky fit: mean and SD of metrics across folds
metrics_leaky
#>            learner auc_mean auc_sd accuracy_mean accuracy_sd
#> 1 logistic_reg/glm     0.75  0.059         0.687       0.054

The summary reports cross-validated performance. AUC ranges from 0.5 (random) to 1.0 (perfect), while accuracy is the fraction of correct predictions. Because the splits are leaky, these metrics can be artificially high and should not be used for reporting.

Leakage-safe alternative: grouped splits and clean predictors

fit_safe <- fit_resample(
  df,
  outcome = "outcome",
  splits = safe_splits,
  learner = spec,
  metrics = c("auc", "accuracy"),
  preprocess = list(
    impute = list(method = "median"),
    normalize = list(method = "zscore"),
    filter = list(var_thresh = 0),
    fs = list(method = "none")
  ),
  positive_class = "case",
  class_weights = c(control = 1, case = 1),
  refit = TRUE
)
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%

cat("Leakage-safe fit summary:\n")
#> Leakage-safe fit summary:
summary(fit_safe)
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: binomial
#> Outcome: outcome
#> Positive class: case
#> Learners: logistic_reg/glm
#> Total folds: 5
#> Refit performed: Yes
#> Hash: hba29ffa5
#> 
#> Cross-validated metrics (mean ± SD):
#>            learner auc_mean auc_sd accuracy_mean accuracy_sd
#> 1 logistic_reg/glm    0.713  0.103         0.619       0.126
#> 
#> Audit overview:
#>  fold n_train n_test          learner features_final
#>     1     124     36 logistic_reg/glm             13
#>     2     128     32 logistic_reg/glm             13
#>     3     128     32 logistic_reg/glm             13
#>     4     128     32 logistic_reg/glm             13
#>     5     132     28 logistic_reg/glm             13
metrics_safe <- as.data.frame(fit_safe@metric_summary)
num_cols <- vapply(metrics_safe, is.numeric, logical(1))
metrics_safe[num_cols] <- lapply(metrics_safe[num_cols], round, digits = 3)

# Leakage-safe fit: mean and SD of metrics across folds
metrics_safe
#>            learner auc_mean auc_sd accuracy_mean accuracy_sd
#> 1 logistic_reg/glm    0.713  0.103         0.619       0.126

# Per-fold metrics (first 6 rows)
head(fit_safe@metrics)
#>   fold          learner       auc  accuracy
#> 1    1 logistic_reg/glm 0.7956656 0.6666667
#> 2    2 logistic_reg/glm 0.8178138 0.7500000
#> 3    3 logistic_reg/glm 0.6078431 0.5625000
#> 4    4 logistic_reg/glm 0.7460317 0.6875000
#> 5    5 logistic_reg/glm 0.6000000 0.4285714

fit_resample() returns a LeakFit object. Use summary(fit_safe) for a compact report and inspect fit_safe@metrics and fit_safe@predictions for details. When refit = TRUE, the final model and preprocessing state are stored in fit_safe@info$final_model and fit_safe@info$final_preprocess.

The mean +/- SD table is the primary performance summary to report, while the per-fold metrics reveal variability and potential instability across folds.

By default, fit_resample() stores refit inputs in fit_safe@info$perm_refit_spec to enable audit_leakage(perm_refit = "auto"). Set store_refit_data = FALSE to reduce memory and provide perm_refit_spec manually when you want refit-based permutations. When multiple learners are passed, refit = TRUE refits only the first learner; set refit = FALSE if you do not need a final model.

The examples above use a parsnip model specification. To swap in other models, replace spec with another parsnip spec (see the gradient boosting example below).

Multiclass classification (optional)

if (requireNamespace("ranger", quietly = TRUE)) {
  set.seed(11)
  df_multi <- df
  df_multi$outcome3 <- factor(sample(c("A", "B", "C"),
                                     nrow(df_multi), replace = TRUE))

  multi_splits <- make_split_plan(
    df_multi,
    outcome = "outcome3",
    mode = "subject_grouped",
    group = "subject",
    v = 4,
    stratify = TRUE,
    seed = 11
  )

  fit_multi <- fit_resample(
    df_multi,
    outcome = "outcome3",
    splits = multi_splits,
    learner = "ranger",
    metrics = c("accuracy", "macro_f1", "log_loss"),
    refit = FALSE
  )

  cat("Multiclass fit summary:\n")
  summary(fit_multi)
} else {
  cat("ranger not installed; skipping multiclass example.\n")
}
#>   |                                                                              |                                                                      |   0%  |                                                                              |==================                                                    |  25%  |                                                                              |===================================                                   |  50%  |                                                                              |====================================================                  |  75%  |                                                                              |======================================================================| 100%
#> Multiclass fit summary:
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: multiclass
#> Outcome: outcome3
#> Learners: ranger
#> Total folds: 4
#> Refit performed: No
#> Hash: hc88c6471
#> 
#> Cross-validated metrics (mean ± SD):
#>   learner accuracy_mean accuracy_sd macro_f1_mean macro_f1_sd log_loss_mean
#> 1  ranger         0.288       0.075         0.279       0.069          1.11
#>   log_loss_sd
#> 1       0.013
#> 
#> Audit overview:
#>  fold n_train n_test learner features_final
#>     1     116     44  ranger             14
#>     2     120     40  ranger             14
#>     3     120     40  ranger             14
#>     4     124     36  ranger             14

Multiclass fits compute accuracy, macro-F1, and log loss when class probabilities are available. positive_class is ignored for multiclass tasks.

Survival outcomes (optional)

if (requireNamespace("survival", quietly = TRUE) &&
    requireNamespace("glmnet", quietly = TRUE)) {
  set.seed(12)
  df_surv <- df
  df_surv$time_to_event <- rexp(nrow(df_surv), rate = 0.1)
  df_surv$event <- rbinom(nrow(df_surv), 1, 0.7)

  surv_splits <- make_split_plan(
    df_surv,
    outcome = c("time_to_event", "event"),
    mode = "subject_grouped",
    group = "subject",
    v = 4,
    stratify = FALSE,
    seed = 12
  )

  fit_surv <- fit_resample(
    df_surv,
    outcome = c("time_to_event", "event"),
    splits = surv_splits,
    learner = "glmnet",
    metrics = "cindex",
    refit = FALSE
  )

  cat("Survival fit summary:\n")
  summary(fit_surv)
} else {
  cat("survival or glmnet not installed; skipping survival example.\n")
}
#>   |                                                                              |                                                                      |   0%  |                                                                              |==================                                                    |  25%  |                                                                              |===================================                                   |  50%  |                                                                              |====================================================                  |  75%  |                                                                              |======================================================================| 100%
#> Survival fit summary:
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: survival
#> Outcome: time_to_event
#>  Outcome: event
#> Learners: glmnet
#> Total folds: 4
#> Refit performed: No
#> Hash: h041adef3
#> 
#> Cross-validated metrics (mean ± SD):
#>   learner cindex_mean cindex_sd
#> 1  glmnet         0.5         0
#> 
#> Audit overview:
#>  fold n_train n_test learner features_final
#>     1     120     40  glmnet             14
#>     2     120     40  glmnet             14
#>     3     120     40  glmnet             14
#>     4     120     40  glmnet             14

For survival tasks, supply outcome = c(time_col, event_col) or a Surv column; stratify is ignored for time/event outcomes and class_weights are not used.

Passing learner-specific arguments (optional)

if (requireNamespace("glmnet", quietly = TRUE)) {
  fit_glmnet <- fit_resample(
    df,
    outcome = "outcome",
    splits = safe_splits,
    learner = "glmnet",
    metrics = "auc",
    learner_args = list(glmnet = list(alpha = 0.5)),
    preprocess = list(
      impute = list(method = "median"),
      normalize = list(method = "zscore"),
      filter = list(var_thresh = 0),
      fs = list(method = "none")
    )
  )
  cat("GLMNET summary with learner-specific arguments:\n")
  summary(fit_glmnet)
} else {
  cat("glmnet not installed; skipping learner_args example.\n")
}
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#> GLMNET summary with learner-specific arguments:
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: binomial
#> Outcome: outcome
#> Positive class: case
#> Learners: glmnet
#> Total folds: 5
#> Refit performed: Yes
#> Hash: hba29ffa5
#> 
#> Cross-validated metrics (mean ± SD):
#>   learner auc_mean auc_sd
#> 1  glmnet    0.713  0.088
#> 
#> Audit overview:
#>  fold n_train n_test learner features_final
#>     1     124     36  glmnet             13
#>     2     128     32  glmnet             13
#>     3     128     32  glmnet             13
#>     4     128     32  glmnet             13
#>     5     132     28  glmnet             13

This summary reflects the same guarded preprocessing but a different model configuration (here, alpha = 0.5). Use the mean +/- SD metrics to compare learner settings under identical splits.

SummarizedExperiment inputs (optional)

if (requireNamespace("SummarizedExperiment", quietly = TRUE)) {
  se <- SummarizedExperiment::SummarizedExperiment(
    assays = list(counts = t(as.matrix(df[, predictors]))),
    colData = df[, c("subject", "batch", "study", "time", "outcome"), drop = FALSE]
  )

  se_splits <- make_split_plan(
    se,
    outcome = "outcome",
    mode = "subject_grouped",
    group = "subject",
    v = 3
  )

  se_fit <- fit_resample(
    se,
    outcome = "outcome",
    splits = se_splits,
    learner = spec,
    metrics = "auc"
  )
  cat("SummarizedExperiment fit summary:\n")
  summary(se_fit)
} else {
  cat("SummarizedExperiment not installed; skipping SE example.\n")
}
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#> SummarizedExperiment fit summary:
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: binomial
#> Outcome: outcome
#> Positive class: case
#> Learners: logistic_reg/glm
#> Total folds: 3
#> Refit performed: Yes
#> Hash: h3f97c442
#> 
#> Cross-validated metrics (mean ± SD):
#>            learner auc_mean auc_sd
#> 1 logistic_reg/glm    0.673  0.062
#> 
#> Audit overview:
#>  fold n_train n_test          learner features_final
#>     1     104     56 logistic_reg/glm              3
#>     2     108     52 logistic_reg/glm              3
#>     3     108     52 logistic_reg/glm              3

The summary is identical in structure to the data.frame case because fit_resample() extracts predictors and metadata from the SummarizedExperiment object.

Note that features_final here reflects only the assay predictors (x1, x2, x3), because the assay was constructed without metadata columns.

2.6 Tidymodels interoperability

bioLeak integrates with rsample, recipes, workflows, and yardstick:

  • splits can be an rsample rset/rsplit; use split_cols (default "auto") or bioLeak_mode / bioLeak_perm_mode attributes to map group/batch/study/time metadata.
  • preprocess can be a recipes::recipe (prepped on training folds only).
  • learner can be a workflows::workflow.
  • metrics can be a yardstick::metric_set.
  • as_rsample() converts LeakSplits to an rsample object.
library(bioLeak)
library(parsnip)
library(recipes)
library(yardstick)

set.seed(123)
N <- 60
df <- data.frame(
  subject = factor(rep(paste0("S", 1:20), length.out = N)), 
  outcome = factor(rep(c("ClassA", "ClassB"), length.out = N)),
  x1 = rnorm(N),
  x2 = rnorm(N),
  x3 = rnorm(N)
)

spec <- logistic_reg() |> set_engine("glm")

# Use bioLeak's native split planner to avoid conversion errors
set.seed(13)

# Use make_split_plan instead of rsample::group_vfold_cv
# This creates a subject-grouped CV directly compatible with fit_resample
splits <- make_split_plan(
  df, 
  outcome = "outcome", 
  mode = "subject_grouped", 
  group = "subject", 
  v = 3
)

rec <- recipes::recipe(outcome ~ x1 + x2 + x3, data = df) |>
  recipes::step_impute_median(recipes::all_numeric_predictors()) |>
  recipes::step_normalize(recipes::all_numeric_predictors())

metrics_set <- yardstick::metric_set(yardstick::roc_auc, yardstick::accuracy)

fit_rs <- fit_resample(
  df,
  outcome = "outcome",
  splits = splits,
  learner = spec,
  preprocess = rec,
  metrics = metrics_set,
  refit = FALSE
)
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%

if (exists("as_rsample", where = asNamespace("bioLeak"), mode = "function")) {
    rs_export <- as_rsample(fit_rs@splits, data = df)
    print(rs_export)
}
#> # Manual resampling 
#> # A tibble: 3 × 2
#>   splits          id   
#>   <list>          <chr>
#> 1 <split [39/21]> Fold1
#> 2 <split [39/21]> Fold2
#> 3 <split [42/18]> Fold3

Use split_cols to ensure split-defining metadata are dropped from predictors when using rsample inputs.

if (requireNamespace("workflows", quietly = TRUE)) {
  wf <- workflows::workflow() |>
    workflows::add_model(spec) |>
    workflows::add_formula(outcome ~ x1 + x2 + x3)

  fit_wf <- fit_resample(
    df,
    outcome = "outcome",
    splits = safe_splits,
    learner = wf,
    metrics = "auc",
    refit = FALSE
  )

  cat("Workflow fit summary:\n")
  summary(fit_wf)
} else {
  cat("workflows not installed; skipping workflow example.\n")
}
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#> Workflow fit summary:
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: binomial
#> Outcome: outcome
#> Positive class: ClassB
#> Learners: workflow_1
#> Total folds: 5
#> Refit performed: No
#> Hash: hba29ffa5
#> 
#> Cross-validated metrics (mean ± SD):
#>      learner auc_mean auc_sd
#> 1 workflow_1    0.696  0.055
#> 
#> Audit overview:
#>  fold n_train n_test    learner features_final
#>     1     124     36 workflow_1              3
#>     2     128     32 workflow_1              3
#>     3     128     32 workflow_1              3
#>     4     128     32 workflow_1              3
#>     5     132     28 workflow_1              3

When auditing rsample-based fits, pass perm_mode = "subject_grouped" (or set attr(rs, "bioLeak_perm_mode") <- "subject_grouped") so restricted permutations respect the intended split design.

2.7 Nested tuning with tune_resample()

tune_resample() runs leakage-aware nested tuning with tidymodels. It expects outer splits that include inner folds (use nested = TRUE in make_split_plan()), and a parsnip model spec or workflow. Survival tasks are not yet supported. Use inner_v, inner_repeats, and inner_seed to control inner resampling when inner folds are not precomputed.

library(bioLeak)
library(parsnip)
library(recipes)
library(tune)
library(glmnet)

# --- 1. Create Data  ---
set.seed(123)
N <- 60
df <- data.frame(
  subject = factor(rep(paste0("S", 1:20), length.out = N)), 
  # FIX: Use sample() to randomize outcome and avoid perfect subject-confounding
  outcome = factor(sample(c("ClassA", "ClassB"), N, replace = TRUE)),
  x1 = rnorm(N),
  x2 = rnorm(N),
  x3 = rnorm(N)
)

# --- 2. Generate Nested Splits ---
set.seed(1)
nested_splits <- make_split_plan(
  df, 
  outcome = "outcome", 
  mode = "subject_grouped", 
  group = "subject", 
  v = 3, 
  nested = TRUE  
)

# --- 3. Define Recipe & Model ---
rec <- recipes::recipe(outcome ~ x1 + x2 + x3, data = df) |>
  recipes::step_impute_median(recipes::all_numeric_predictors()) |>
  recipes::step_normalize(recipes::all_numeric_predictors())

spec_tune <- parsnip::logistic_reg(
  penalty = tune::tune(),
  mixture = 1,
  mode = "classification"
) |>
  parsnip::set_engine("glmnet")

# --- 4. Run Tuning ---
tuned <- tune_resample(
  df,
  outcome = "outcome",
  splits = nested_splits,
  learner = spec_tune,
  preprocess = rec,
  inner_v = 2,
  grid = 3,
  metrics = c("auc", "accuracy"),
  seed = 14
)
#>   |                                                                              |                                                                      |   0%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |======================================================================| 100%
  
summary(tuned)
#> 
#> ============================
#>  bioLeak Tuning Summary
#> ============================
#> 
#> Task: binomial
#> Outcome: outcome
#> Positive class: ClassB
#> Tuning Grid: 3
#> Selection Rule: best (Metric: roc_auc)
#> Outer Folds: 3
#> 
#> Outer Loop Metrics (mean ± SD):
#>               learner accuracy_mean accuracy_sd roc_auc_mean roc_auc_sd
#> 1 logistic_reg/glmnet         0.585       0.053        0.546      0.173
#> 
#> Best Parameters (First 5 Folds):
#>  penalty fold
#>        0    1
#>        0    2
#>        0    3

2.8 Advanced: Using Gradient Boosting with Parsnip

bioLeak natively supports tidymodels specifications. You can pass a parsnip model specification directly to fit_resample(). This allows you to use state-of-the-art algorithms like XGBoost, LightGBM, or SVMs while ensuring all preprocessing remains guarded.

if (requireNamespace("parsnip", quietly = TRUE) &&
    requireNamespace("xgboost", quietly = TRUE) &&
    requireNamespace("recipes", quietly = TRUE)) {
  
  # 1. Define the model spec
  xgb_spec <- parsnip::boost_tree(
    mode = "classification",
    trees = 100,
    tree_depth = 6,
    learn_rate = 0.01
  ) |>
    parsnip::set_engine("xgboost")
  
  # 2. Define a Recipe (CORRECTED)
  # FIX: Define the recipe on data WITHOUT 'subject'.
  # This matches the data bioLeak passes (which has subject removed).
  df_for_rec <- df[, !names(df) %in% "subject"]
  
  rec_xgb <- recipes::recipe(outcome ~ ., data = df_for_rec) |>
    recipes::step_dummy(recipes::all_nominal_predictors()) |>
    recipes::step_impute_median(recipes::all_numeric_predictors())
  # Note: No need for step_rm(subject) because it's already gone!
  
  # 3. Fit
  fit_xgb <- fit_resample(
    df,
    outcome = "outcome",
    splits = nested_splits,
    learner = xgb_spec,
    metrics = "auc",
    preprocess = rec_xgb 
  )
  
  cat("XGBoost parsnip fit summary:\n")
  print(summary(fit_xgb))
} else {
  cat("parsnip/xgboost/recipes not installed.\n")
}
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#> XGBoost parsnip fit summary:
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: binomial
#> Outcome: outcome
#> Positive class: ClassB
#> Learners: boost_tree/xgboost
#> Total folds: 3
#> Refit performed: Yes
#> Hash: h825b7ff0
#> 
#> Cross-validated metrics (mean ± SD):
#>              learner auc_mean auc_sd
#> 1 boost_tree/xgboost    0.616  0.058
#> 
#> Audit overview:
#>  fold n_train n_test            learner features_final
#>     1      39     21 boost_tree/xgboost              3
#>     2      39     21 boost_tree/xgboost              3
#>     3      42     18 boost_tree/xgboost              3
#> 
#>              learner  auc_mean     auc_sd
#> 1 boost_tree/xgboost 0.6157407 0.05800909

The summary reports cross-validated AUC for a non-linear gradient boosting model. Use the mean \(\pm\) SD table to compare against baseline models, and confirm that any gains do not coincide with leakage signals in the audit diagnostics.

2.9 Advanced: Custom learners

Custom learners are used when a model is not available as a parsnip specification or when a lightweight wrapper around base R is needed.

Each custom learner must define fit and predict components. The fit function must accept x, y, task, and weights as inputs, while the predict function must accept object and newdata.

custom_learners <- list(
  glm = list(
    fit = function(x, y, task, weights, ...) {
      df_fit <- data.frame(y = y, x, check.names = FALSE)
      stats::glm(y ~ ., data = df_fit,
                 family = stats::binomial(), weights = weights)
    },
    predict = function(object, newdata, task, ...) {
      as.numeric(stats::predict(object, newdata = as.data.frame(newdata),
                                type = "response"))
    }
  )
)

cat("Custom learner names:\n")
#> Custom learner names:
names(custom_learners)
#> [1] "glm"
cat("Custom learner components (fit/predict):\n")
#> Custom learner components (fit/predict):
lapply(custom_learners, names)
#> $glm
#> [1] "fit"     "predict"

This output confirms that each custom learner provides both a fit and a predict function.

Custom learners can be used with fit_resample() as follows:

fit_resample(
  ...,
  learner = "glm",
  custom_learners = custom_learners
)

2.10 Visual diagnostics

bioLeak includes plotting helpers for quick diagnostic checks:

  • plot_fold_balance() shows class balance per fold.
  • plot_overlap_checks() highlights train/test overlap for a metadata column.
  • plot_time_acf() checks autocorrelation in time-series predictions.
  • plot_calibration() shows reliability of binomial probabilities.
  • plot_confounder_sensitivity() summarizes performance by batch/study strata.

Tabular helpers are also available:

  • calibration_summary() returns calibration curve data and ECE/MCE/Brier metrics.
  • confounder_sensitivity() returns per-stratum performance summaries.

Misuse to avoid

  • Plotting without predictions (fits with no successful folds).
  • Using plot_overlap_checks() when the column is not present in the split metadata.
  • Using plot_time_acf() without a time column or for non-temporal data.
  • Using plot_calibration() outside binomial tasks.
  • Using plot_confounder_sensitivity() with a metric unsupported by the task.

For classification, plot_fold_balance() uses the positive class recorded in fit@info$positive_class, or defaults to the second factor level if this is not set. For multiclass tasks, it shows per-class counts without a proportion line.

if (requireNamespace("ggplot2", quietly = TRUE)) {
  plot_fold_balance(fit_safe)
} else {
  cat("ggplot2 not installed; skipping fold balance plot.\n")
}

The bar chart shows the counts of Positives (blue) and Negatives (tan) in each fold. The dashed blue line represents the proportion of the positive class.

In a well-stratified fit, this line remains relatively stable across folds. Large fluctuations indicate poor stratification, which can lead to unstable fold-level performance estimates.

if (requireNamespace("ggplot2", quietly = TRUE)) {
  plot_calibration(fit_safe, bins = 10)
} else {
  cat("ggplot2 not installed; skipping calibration plot.\n")
}

Calibration curves compare predicted probabilities to observed event rates. Large deviations from the diagonal indicate miscalibration. Use min_bin_n to suppress tiny bins and learner = to select a model when multiple learners are stored in the fit.

if (requireNamespace("ggplot2", quietly = TRUE)) {
  plot_confounder_sensitivity(fit_safe, confounders = c("batch", "study"),
                              metric = "auc", min_n = 3)
} else {
  cat("ggplot2 not installed; skipping confounder sensitivity plot.\n")
}

Confounder sensitivity plots highlight whether performance varies substantially across batch or study strata.

cal <- calibration_summary(fit_safe, bins = 10, min_bin_n = 5)
conf_tbl <- confounder_sensitivity(fit_safe,
                                   confounders = c("batch", "study"),
                                   metric = "auc",
                                   min_n = 3)

# Calibration metrics
cal$metrics
#>     n bins min_bin_n        ece       mce     brier
#> 1 160   10         5 0.06014931 0.6653969 0.2335112

# Confounder sensitivity table (first 6 rows)
head(conf_tbl)
#>   confounder level metric direction     value  n positive_rate
#> 1      batch    B1    auc    higher 0.7569444 26     0.3076923
#> 2      batch    B2    auc    higher 0.7348485 23     0.4782609
#> 3      batch    B3    auc    higher 0.5538462 28     0.4642857
#> 4      batch    B4    auc    higher 0.7152778 24     0.5000000
#> 5      batch    B5    auc    higher 0.5446429 22     0.3636364
#> 6      batch    B6    auc    higher 0.6764706 37     0.4594595
if (requireNamespace("ggplot2", quietly = TRUE)) {
  plot_overlap_checks(fit_leaky, column = "subject")
  plot_overlap_checks(fit_safe, column = "subject")
} else {
  cat("ggplot2 not installed; skipping overlap plots.\n")
}

The overlap plots compare the number of unique subjects appearing in the training and test sets for each fold.

Top plot (fit_leaky)
The red line shows high overlap counts, accompanied by the explicit “WARNING: Overlaps detected!” annotation. This confirms that the same subjects appear in both the training and test sets, indicating information leakage.

Bottom plot (fit_safe)
The overlap line remains flat at zero across all folds. This confirms that subject_grouped splitting successfully keeps subjects isolated, ensuring that no subject appears in both sets simultaneously.

2.11 Audit leakage with audit_leakage()

audit_leakage() combines four diagnostics in one object:

  • permutation gap: tests whether model signal is non-random
  • batch or study association with folds (chi-square and Cramers V)
  • target leakage scan: feature-wise outcome association to flag proxy columns
  • duplicate and near-duplicate detection across train/test (by default) using feature similarity

Assumptions and misuse to avoid:

  • coldata (or stored split metadata) must align with prediction IDs; rownames or a row_id column help alignment
  • X_ref must align to prediction IDs (rownames, row_id, or matching order)
  • plot_perm_distribution() requires return_perm = TRUE
  • target_threshold should be set high enough to flag only strong proxies
  • target leakage scan includes univariate associations plus an optional multivariate/interaction check (target_scan_multivariate, *_components, *_interactions, *_B); proxies outside X_ref can still pass undetected
  • for rsample splits, set perm_mode (or bioLeak_perm_mode) so restricted permutations respect the intended design
  • for time-series audits, ensure a valid time column and set time_block / block_len as needed

Interpretation Note: The label-permutation test refits models by default when refit inputs are available (perm_refit = "auto" with store_refit_data = TRUE) and B <= perm_refit_auto_max. Otherwise it keeps predictions fixed, so the p-value quantifies prediction-label association rather than a full refit null. A large gap indicates a non-random association. To determine if that signal is real or leaked, check the Batch Association (confounding), Target Leakage Scan (proxy features), and Duplicate Detection (memorization) tables. Use perm_refit = FALSE to force fixed-prediction shuffles or perm_refit = TRUE with perm_refit_spec to always refit.

Use feature_space = "rank" to compare samples by rank profiles, and sim_method (cosine or pearson) to control similarity. For large n, nn_k and max_pairs limit duplicate searches. Use duplicate_scope = "all" to include within-fold duplicates (data-quality checks). Use ci_method = "if" (influence-function) or ci_method = "bootstrap" with boot_B to obtain a confidence interval for the permutation gap; include_z controls whether a z-score is reported. Set perm_stratify = "auto" to stratify permutations only when outcomes support it.

Leakage-aware audit

X_ref <- df[, predictors]
X_ref[c(1, 5), ] <- X_ref[1, ]

audit <- audit_leakage(
  fit_safe,
  metric = "auc",
  B = 20,
  perm_stratify = TRUE,
  batch_cols = c("batch", "study"),
  X_ref = X_ref,
  sim_method = "cosine",
  sim_threshold = 0.995,
  return_perm = TRUE
)
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%

cat("Leakage audit summary:\n")
#> Leakage audit summary:
summary(audit)
#> 
#> ==============================
#>  bioLeak Leakage Audit Summary
#> ==============================
#> 
#> Task: binomial | Outcome: outcome | Splitting mode: subject_grouped | Positive class: case
#> Hash: hba29ffa5 | Folds: 5 | Repeats: 1
#> 
#> Label-Permutation Association Test:
#>   Method: refit per permutation (auto)
#>   Observed metric: 0.656
#>   Permuted mean ± SD: 0.544 ± 0.032
#>   Gap: 0.112 (larger gap = stronger non-random signal)
#>   This test does NOT diagnose information leakage. Use the Batch Association,
#>   Target Leakage Scan, and Duplicate Detection sections to check for leakage.
#> 
#> Batch / Study Association:
#>   batch (repeat 1): χ² = 17.043 (df = 20.000), p = 0.650
#>   study (repeat 1): χ² = 9.487 (df = 12.000), p = 0.661
#> 
#> Target Leakage Scan: not available.
#> 
#> Multivariate Target Scan: not available.
#> 
#> No near-duplicates detected.
#> 
#> Interpretation:
#>   ✓ Strong non-random signal.
if (!is.null(audit@permutation_gap) && nrow(audit@permutation_gap) > 0) {
  # Permutation significance results
  audit@permutation_gap
}
#>   metric_obs perm_mean  perm_sd      gap        z  p_value n_perm
#> 1   0.656474   0.54449 0.031651 0.111984 2.611722 0.047619     20
if (!is.null(audit@batch_assoc) && nrow(audit@batch_assoc) > 0) {
  # Batch/study association with folds (Cramer's V)
  audit@batch_assoc
} else {
  cat("No batch or study associations detected.\n")
}
#>       batch_col repeat_id      stat df     pval  cramer_v
#> batch     batch         1 17.043088 20 0.650174 0.1631865
#> study     study         1  9.487022 12 0.660865 0.1405867
if (!is.null(audit@target_assoc) && nrow(audit@target_assoc) > 0) {
  # Top features by target association score
  head(audit@target_assoc)
} else {
  cat("No target leakage scan results available.\n")
}
#> No target leakage scan results available.
if (!is.null(audit@duplicates) && nrow(audit@duplicates) > 0) {
  # Top duplicate/near-duplicate pairs by similarity
  head(audit@duplicates)
} else {
  cat("No near-duplicates detected.\n")
}
#> No near-duplicates detected.

The permutation table reports the observed metric, the mean under random label permutation, the gap (difference), and a permutation p-value. For metrics where higher values indicate better performance, larger gaps reflect stronger non-random signal.

The batch association table reports chi-square statistics and Cramer’s V. Large p-values and small V values indicate that folds are not aligned with batch or study labels (which is the desired outcome when these should be independent).

The target scan table lists features with the strongest associations with the outcome. For numeric features, the score is \(|\mathrm{AUC} - 0.5| \times 2\) for classification or the absolute correlation for regression. For categorical features, the score is Cramér’s V or eta-squared. Scores closer to 1 indicate stronger outcome association.

The duplicate table lists pairs of samples with near-identical profiles that cross train/test folds (by default). In this setup, the artificially duplicated pair (X_ref[c(1, 5), ] <- X_ref[1, ]) should appear near the top of the list. Use duplicate_scope = "all" to include within-fold duplicates.

if (requireNamespace("ggplot2", quietly = TRUE)) {
  plot_perm_distribution(audit)
} else {
  cat("ggplot2 not installed; skipping permutation plot.\n")
}

The histogram shows the null distribution (gray bars) of the performance metric under random label permutation.

The blue dashed line represents the average performance of a random model (the permuted mean). The red solid line represents the observed performance of the fitted model.

A genuine signal is indicated when the red line lies well to the right of the gray distribution; overlap suggests weak or unstable signal.

Audit per learner

if (requireNamespace("ranger", quietly = TRUE) && 
    requireNamespace("parsnip", quietly = TRUE)) {
    
    # 1. Define specs
    # Standard GLM (no tuning)
    spec_glm <- parsnip::logistic_reg() |> 
        parsnip::set_engine("glm")
    
    # Random Forest
    spec_rf <- parsnip::rand_forest(
        mode = "classification",
        trees = 100
    ) |>
        parsnip::set_engine("ranger")
    
    # 2. Fit with CORRECT splits
    fit_multi <- fit_resample(
        df,
        outcome = "outcome",
        splits = nested_splits,  # <--- FIX: Use the valid split object
        learner = list(glm = spec_glm, rf = spec_rf),
        metrics = "auc"
    )
    
    # 3. Run the audit
    audits <- audit_leakage_by_learner(fit_multi, metric = "auc", B = 20)
    cat("Per-learner audit summary:\n")
    print(audits)
    
} else {
    cat("ranger/parsnip not installed.\n")
}
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#> Per-learner audit summary:
#> LeakAuditList with 2 learners
#>  learner metric_obs perm_mean    gap p_value      z n_perm batch_p_min
#>      glm      0.537     0.609 -0.071   0.857 -0.800     20          NA
#>       rf      0.559     0.586 -0.027   0.667 -0.308     20          NA
#>  batch_col_min_p batch_v_max batch_col_max_v dup_pairs dup_threshold
#>             <NA>          NA            <NA>        NA         0.995
#>             <NA>          NA            <NA>        NA         0.995
#>  dup_max_sim
#>           NA
#>           NA

This example uses ranger for the random forest specification; if it is not installed, the code chunk is skipped.

Use parallel_learners = TRUE to audit learners concurrently when future.apply is available.

The printed table summarizes each learner’s observed metric, permutation gap, p-value, and key batch/duplicate summaries. Use it to compare signal strength across models while checking for leakage risks. Pass learners = to audit a subset, or mc.cores = to control parallel workers.

HTML audit report

audit_report() accepts either a LeakAudit or a LeakFit object. When a LeakFit is provided, it first runs audit_leakage() and then forwards any additional arguments to the audit step. If multiple learners were fit, pass learner = via ... to select one. Use open = TRUE to open the report in a browser after rendering.

if (requireNamespace("rmarkdown", quietly = TRUE) && rmarkdown::pandoc_available()) {
  report_path <- audit_report(audit, output_dir = ".")
  cat("HTML report written to:\n", report_path, "\n")
} else {
  cat("rmarkdown or pandoc not available; skipping audit report rendering.\n")
}

The report path points to a standalone HTML file containing the same audit tables and plots, suitable for sharing with collaborators or archiving as a quality control record.

2.12 Time-series leakage checks

Time-series data require special handling. Random splits can leak information from the future into the past. Use mode = "time_series" with a prediction horizon, and audit with block permutations. Choose time_block = "circular" or "stationary"; when block_len is NULL, the audit uses a default block length (~10% of the test block size, minimum 5).

Leaky example: lookahead feature

time_splits <- make_split_plan(
  df_time,
  outcome = "outcome",
  mode = "time_series",
  time = "time",
  v = 4,
  horizon = 1
)

cat("Time-series splits summary:\n")
#> Time-series splits summary:
time_splits
#> LeakSplits object (mode = time_series, v = 4, repeats = 1)
#> Outcome: outcome | Stratified: FALSE | Nested: FALSE
#> ------------------------------------------------------
#>   fold repeat_id train_n test_n
#> 1    2         1      40     40
#> 2    3         1      80     40
#> 3    4         1     120     40
#> ------------------------------------------------------
#> Total folds: 3 | Hash: 412182f8014315f1879c89587fd76f1e

fit_time_leaky <- fit_resample(
  df_time,
  outcome = "outcome",
  splits = time_splits,
  learner = spec,
  metrics = "auc"
)
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%

cat("Time-series leaky fit summary:\n")
#> Time-series leaky fit summary:
summary(fit_time_leaky)
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: binomial
#> Outcome: outcome
#> Positive class: case
#> Learners: logistic_reg/glm
#> Total folds: 3
#> Refit performed: Yes
#> Hash: h3b18276c
#> 
#> Cross-validated metrics (mean ± SD):
#>            learner auc_mean auc_sd
#> 1 logistic_reg/glm    0.661  0.021
#> 
#> Audit overview:
#>  fold n_train n_test          learner features_final
#>     1      40     40 logistic_reg/glm             14
#>     2      80     40 logistic_reg/glm             14
#>     3     120     40 logistic_reg/glm             14

Time-series splitting trains on growing windows, so performance can differ from standard CV simply because early folds have smaller training sets. Regardless of the score, the presence of the leak_future feature makes this estimate methodologically invalid.

Leakage-safe alternative: remove lookahead and audit with blocks

df_time_safe <- df_time
df_time_safe$leak_future <- NULL

fit_time_safe <- fit_resample(
  df_time_safe,
  outcome = "outcome",
  splits = time_splits,
  learner = spec,
  metrics = "auc"
)
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%

cat("Time-series safe fit summary:\n")
#> Time-series safe fit summary:
summary(fit_time_safe)
#> 
#> ===========================
#>  bioLeak Model Fit Summary
#> ===========================
#> 
#> Task: binomial
#> Outcome: outcome
#> Positive class: case
#> Learners: logistic_reg/glm
#> Total folds: 3
#> Refit performed: Yes
#> Hash: h3b18276c
#> 
#> Cross-validated metrics (mean ± SD):
#>            learner auc_mean auc_sd
#> 1 logistic_reg/glm    0.661  0.021
#> 
#> Audit overview:
#>  fold n_train n_test          learner features_final
#>     1      40     40 logistic_reg/glm             13
#>     2      80     40 logistic_reg/glm             13
#>     3     120     40 logistic_reg/glm             13

audit_time <- audit_leakage(
  fit_time_safe,
  metric = "auc",
  B = 20,
  time_block = "stationary",
  block_len = 5
)
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |=======================                                               |  33%  |                                                                              |===============================================                       |  67%  |                                                                              |======================================================================| 100%

cat("Time-series leakage audit summary:\n")
#> Time-series leakage audit summary:
summary(audit_time)
#> 
#> ==============================
#>  bioLeak Leakage Audit Summary
#> ==============================
#> 
#> Task: binomial | Outcome: outcome | Splitting mode: time_series | Positive class: case
#> Hash: h3b18276c | Folds: 3 | Repeats: 1
#> 
#> Label-Permutation Association Test:
#>   Method: refit per permutation (auto)
#>   Observed metric: 0.645
#>   Permuted mean ± SD: 0.529 ± 0.032
#>   Gap: 0.116 (larger gap = stronger non-random signal)
#>   This test does NOT diagnose information leakage. Use the Batch Association,
#>   Target Leakage Scan, and Duplicate Detection sections to check for leakage.
#> 
#> Batch / Study Association:
#>   batch (repeat 1): χ² = 8.237 (df = 10.000), p = 0.606
#>   study (repeat 1): χ² = 2.383 (df = 6.000), p = 0.881
#> 
#> Target Leakage Scan: not available.
#> 
#> Multivariate Target Scan: not available.
#> 
#> No near-duplicates detected.
#> 
#> Interpretation:
#>   ✓ Strong non-random signal.
if (!is.null(audit_time@permutation_gap) && nrow(audit_time@permutation_gap) > 0) {
  # Time-series permutation significance results
  audit_time@permutation_gap
}
#>   metric_obs perm_mean perm_sd      gap       z  p_value n_perm
#> 1   0.645362  0.529196 0.03175 0.116166 2.22862 0.047619     20

if (requireNamespace("ggplot2", quietly = TRUE)) {
  plot_time_acf(fit_time_safe, lag.max = 20)
} else {
  cat("ggplot2 not installed; skipping ACF plot.\n")
}

The safe fit summary provides the leakage-resistant performance estimate. Compare leaky vs. safe fits to gauge inflation risk; features_final should drop when the lookahead feature is removed.

The ACF plot shows the autocorrelation of out-of-fold predictions by lag. The dashed red lines represent the 95% confidence interval for white noise; large bars outside the bands indicate residual temporal dependence. plot_time_acf() requires numeric predictions and time metadata aligned to the fit.

3 Parallel Processing

bioLeak uses the future framework for parallelism (Windows, macOS, Linux). fit_resample(), audit_leakage(), audit_leakage_by_learner(), and simulate_leakage_suite() honor the active plan when parallel = TRUE (or parallel_learners = TRUE).

library(future)

# Use multiple cores (works on all OS)
plan(multisession, workers = 4)

# Run a heavy simulation
sim <- simulate_leakage_suite(..., parallel = TRUE)

# Parallel folds or audits
# fit_resample(..., parallel = TRUE)
# audit_leakage(..., parallel = TRUE)
# audit_leakage_by_learner(..., parallel_learners = TRUE)

# Return to sequential processing
plan(sequential)

This chunk only configures the parallel plan, so it does not produce a result object. Use it as a template before running compute-heavy functions.

3.1 Simulation suite

simulate_leakage_suite() runs Monte Carlo simulations that inject specific leakage mechanisms and then evaluates detection with the audit pipeline. It is useful for validating your leakage checks before applying them to real data. Available leakage types include none, subject_overlap, batch_confounded, peek_norm, and lookahead. Use signal_strength, prevalence, rho, K, repeats, horizon, and preprocess to control the simulation. The output is a LeakSimResults data frame with one row per seed. Metrics are selected automatically based on outcome type (AUC for binary classification).

if (requireNamespace("glmnet", quietly = TRUE)) {
  sim <- simulate_leakage_suite(
    n = 80,
    p = 6,
    mode = "subject_grouped",
    learner = "glmnet",
    leakage = "subject_overlap",
    seeds = 1:2,
    B = 20,
    parallel = FALSE,
    signal_strength = 1
  )
  
  # Simulation results (first 6 rows)
  head(sim)
} else {
  cat("glmnet not installed; skipping simulation suite example.\n")
}
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%
#>   |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   |                                                                              |                                                                      |   0%  |                                                                              |==============                                                        |  20%  |                                                                              |============================                                          |  40%  |                                                                              |==========================================                            |  60%  |                                                                              |========================================================              |  80%  |                                                                              |======================================================================| 100%
#>   seed metric_obs      gap  p_value         leakage            mode
#> 1    1   0.828914 0.238567 0.047619 subject_overlap subject_grouped
#> 2    2   0.825397 0.210306 0.047619 subject_overlap subject_grouped

Each row corresponds to one simulation seed.

  • metric_obs is the observed performance (AUC for this simulation). Higher values suggest stronger apparent signal, which can be inflated by leakage.

  • gap and p_value indicate that the leaked signal is statistically distinguishable from random noise.

Use metric_obs to gauge the magnitude of the leakage effect, and gap to assess detection sensitivity.

3.2 Objects and summaries

bioLeak uses S4 classes and list-like results to capture provenance and diagnostics:

  • LeakSplits: returned by make_split_plan(), printed with show()
  • LeakFit: returned by fit_resample(), summarized with summary()
  • LeakAudit: returned by audit_leakage(), summarized with summary()
  • LeakAuditList: returned by audit_leakage_by_learner(), printed directly
  • GuardFit: returned by .guard_fit(), printed and summarized
  • LeakImpute: returned by impute_guarded() (guarded data + diagnostics)
  • LeakTune: returned by tune_resample() (nested tuning summaries)
  • LeakSimResults: returned by simulate_leakage_suite() (per-seed results)

These objects store hashes and metadata that help reproduce and audit the workflow. Inspect their slots (@metrics, @predictions, @permutation_gap, @duplicates) to drill into specific leakage signals.