Frozen Stage-2 Base and Cascade Predictions and Verification Code for DKA Risk Modeling

This record provides frozen Stage-2 Base and Cascade predictions and self-contained Python verification code for paired performance comparisons in diabetic ketoacidosis (DKA) risk modeling.The partitions contain 447 records with 133 composite events, 867 records with 247 events, and 68 records with 12 events, respectively.

The code evaluates AUC, average precision, Brier score, calibration intercept and slope, fixed-threshold classification, empirical ROC/PR coordinates, calibration bins, decision curves and paired Cascade-minus-Base differences with patient-level bootstrap intervals. Models, probabilities, calibrators and thresholds are held fixed. This deposit does not provide original EHR extraction or model-development code and does not establish clinical effectiveness or consistent cascade superiority.

Directory reconstruction after downloading the individual files: keep the unprefixed files in one working directory. For each filename beginning with expected__, create an expected subdirectory, remove the expected__ prefix and place the file there. For each filename beginning with verification_output__, create a verification_output subdirectory, remove that prefix and place the file there. Prefixes only encode the original directory structure.

To run: install Python 3.10 or newer, run python -m pip install -r requirements.txt, then run python stage2_verification.py. The script uses relative paths, generates verification_output and compares its calculations with the frozen references. The documented test environment was Python 3.12.14, NumPy 2.3.5 and pandas 3.0.1. Bootstrap sampling uses 1,000 paired resamples per partition with seed 2603502 plus the partition size. The tolerance is 1e-10 for general metrics and 1e-5 for independently solved calibration diagnostics.