ONLINE RESOURCE 1 - COMPLETE NUMERICAL RESULTS

article_title: Auditing imbalanced hospital readmission prediction: patient separation, resampling contamination and probability calibration
journal: International Journal of Data Science and Analytics
authors: Mohammad Rabiul Hossen Chowdhury; Shomi Aktar
corresponding_author_affiliation: Department of Mathematics, The University of Alabama, Tuscaloosa, Alabama, USA
corresponding_author_email: saktar@ua.edu

The 16 CSV files reproduce all tabulated values from former appendix pages 28-49, retaining the displayed numerical precision. No new analyses or rounding changes were introduced. Metadata columns appear after the original table columns. Each file is a rectangular CSV with one header row. The literal -- denotes undefined values; it is not zero.

This supplement reports the saved summary results underlying the corrected analysis. It includes all primary configurations and operating points, all grouped validation folds, all bootstrap intervals, all subgroup rows, both operating points for the seven-model extension, the risk-map cells, all grouped SHAP values, and the MLP stopping history. Patient/encounter-level prediction arrays and threshold-curve coordinates are available as CSV files from the corresponding author upon reasonable request rather than printed as thousands of rows.

Abbreviations are AUC (ROC-AUC), AP (average precision), BS (Brier score), Rec (recall), PPV (precision), FPR (false-positive rate), Acc (accuracy), and TP/FP/FN/TN (confusion counts). Configuration suffixes N, W, R, and S denote no resampling, class weighting, random oversampling, and SMOTENC. For operating points, R0.5 denotes raw scores at 0.5, C0.5 calibrated scores at 0.5, and C10 calibrated scores at the threshold selected on calibration patients under the FPR constraint. Counts are encounters unless labeled patients. Precision is assigned zero when no positive predictions are made, matching the saved analysis convention; this value does not represent an estimable positive predictive value. Other undefined metrics are shown as ``--''. Subgroup intervals are conditional and exploratory; no multiplicity adjustment is implied.

TABLES A.12-A.13
These tables retain the paired sensitivity intervals and the repeat-level resampling audit. The corresponding full fold records and sensitivity predictions are available as CSV files from the corresponding author upon reasonable request. Intervals are experiment minus the corresponding original full-feature baseline, conditional on fitted models; repeated-split SDs are not bootstrap confidence intervals.

TABLES A.14-A.16
These discharge-sensitivity rows use the retained evaluation set (11,610 encounters; 1,348 events). H denotes the historical fitted model restricted to eligible evaluation encounters, E denotes refitting after exclusions, and D denotes that refit without discharge disposition. All thresholds are calibration-selected. The 95 percent intervals condition on fitted models and use 1,000 patient-cluster draws. Delta intervals compare E or D with H for the same algorithm and the same evaluation encounters.

FILE INDEX
Table_A01.csv (former Table A.1; 24 data rows): All 24 primary model/score/operating-point records; each uses 11,919 evaluation encounters and 1,355 positive outcomes.
Columns: Model, Op, t, AUC, AP, BS, Rec, PPV, FPR, Acc, TP, FP, FN, TN

Table_A02.csv (former Table A.2; 24 data rows): All grouped fitting-partition cross-validation folds. Probabilities are raw; thresholded metrics use 0.5. Each fold fits its own preprocessing and resampling.
Columns: Model, Fold, n, Events, AUC, AP, BS, Rec, PPV, FPR, Acc, TP, FP, FN, TN

Table_A03.csv (former Table A.3; 45 data rows): All conditional patient-cluster bootstrap intervals, including paired XGB-N minus LR-N contrasts. All intervals are 95 percent percentile intervals.
Columns: Model/contrast, Metric, Lower, Upper, Replicates

Table_A04.csv (former Table A.4; 128 data rows): All subgroup performance summaries for the eight primary configurations at their calibration-selected thresholds.
Columns: Model, Field, Group, n, Events, t, AUC, AP, BS, Rec, PPV, FPR, Acc

Table_A05.csv (former Table A.5; 128 data rows): Confusion counts corresponding to every subgroup summary. Small-group results require cautious interpretation.
Columns: Model, Field, Group, TP, FP, FN, TN

Table_A06.csv (former Table A.6; 96 data rows): Complete subgroup patient-cluster intervals for the two unresampled baselines. FNR is false-negative rate; recall intervals in the figures equal one minus the reversed FNR bounds.
Columns: Model, Field, Group, n, Patients, Events, Metric, Estimate, Lower, Upper, Valid, Attempted

Table_A07.csv (former Table A.7; 14 data rows): All seven-model extension metrics at both operating points. Evaluation size is 11,919 encounters for every row.
Columns: Model, Op, Fit n, t, AUC, AP, BS, Rec, PPV, FPR, Acc, F1

Table_A08.csv (former Table A.8; 14 data rows): Confusion counts for all seven-model extension records.
Columns: Model, Op, TP, FP, FN, TN

Table_A09.csv (former Table A.9; 16 data rows): All risk-map cells. Shown indicates whether the cell meets the 100-encounter display threshold. Rates are encounter-based proportions.
Columns: Stay bin, Prior bin, n, Patients, Events, Observed, Mean LR, Mean XGB, Shown

Table_A10.csv (former Table A.10; 43 data rows): All grouped XGB SHAP values from the 2,000-encounter explanation sample, in raw log-odds units.
Columns: Original feature, Mean absolute grouped SHAP

Table_A11.csv (former Table A.11; 15 data rows): Complete MLP inner stopping history. Epoch five minimizes the observed stopping loss under the implemented improvement rule; a fresh model is refitted for five epochs.
Columns: Epoch, Grouped validation log loss

Table_A12.csv (former Table A.12; 14 data rows): All paired sensitivity intervals; 95 percent patient-cluster percentile intervals.
Columns: Experiment, Metric, Lower, Upper, Replicates

Table_A13.csv (former Table A.13; 240 data rows): All 240 repeat-level records from the controlled donor-pool audit. Shared-patient encounters count test encounters belonging to patients represented in outer training.
Columns: Repeat, Split, Model, Arm, CV AUC, Test AUC, Test n, Shared-patient encounters

Table_A14.csv (former Table A.14; 6 data rows): Complete discharge-eligibility metrics.
Columns: Model, t, AUC, AP, BS, Rec, PPV, FPR, Acc, TP, FP, FN, TN

Table_A15.csv (former Table A.15; 30 data rows): Conditional absolute and paired-difference intervals for discharge sensitivity.
Columns: Model, Metric, Contrast, Lower, Upper

Table_A16.csv (former Table A.16; 85 data rows): All grouped mean absolute XGB SHAP contributions in the discharge-eligibility refits (raw log-odds).
Columns: Model, Feature, Mean absolute contribution

