PhysioCGM Re-evaluation: Leakage-Controlled Splits, Predictions, Results, and Code for AI-Based Noninvasive Hypoglycemia Detection

Data and code for: Abu Owida, H.; Alsarhan, A.; Olamat, A.; Banat, M.; Ozel, P. Artificial Intelligence for Noninvasive Hypoglycemia Detection From Multimodal Wearable Signals (manuscript submitted to Computers, 2026). The study re-evaluates the published multimodal deep learning pipeline of the PhysioCGM dataset (ECG, PPG and EDA networks with late fusion; Quamer et al., Sci. Data 2025) under four evaluation protocols: the original random five-fold segment split, random folds with a separate validation set, day-blocked folds (unseen days, 60-min gap) and leave-one-participant-out (unseen participants). It adds leakage baselines (absolute recording time, time of day, summary physiology, nearest neighbour), a proximity analysis, a pre-specified feature-based detector, robustness analyses (unfitted fusion, fixed-epoch training, three training runs with different random seeds under the unseen-day protocol, glucose inclusion range) and an external check on the D1NAMO dataset, including the scores of an ECG model trained on all eight PhysioCGM participants. Contents. (1) The training, validation and test assignment of every 5-min segment under each protocol; (2) segment-level test predictions and per-fold metrics of every model and run, including the revision runs (folder revision/); (3) statistical results and the source data of all figures and tables; (4) analysis code (MIT licence), including a patch to the public PhysioCGM implementation (github.com/PSI-TAMU/PhysioCGM, commit 6940f1d) and the scripts for the new protocols and analyses. A README gives the run order and software versions. Source data. PhysioCGM (doi:10.6084/m9.figshare.28136294.v1, CC0 1.0) and D1NAMO (doi:10.5281/zenodo.5651217, CC BY-SA 4.0). Raw data are not redistributed.