Subject-Independent Hypoglycemia Warnings via Multi-Horizon Transformers and Causal Hysteresis (Guardian Pro)

Vedant Thakkar

Abstract: Hypoglycemia prevention in diabetes management is hindered by a critical trade-off between sensitivity and temporal precision. While tabular baselines like XGBoost achieve high discrimination, they frequently exhibit "temporal smearing," triggering alarms several hours before an event occurs by over-relying on broad diurnal trends. We present Guardian Pro, a subject-independent warning system combining a multi-horizon Transformer with a Causal Hysteresis filter to provide actionable "Day 1" alerts for new patients. Evaluated on the Shanghai Diabetes dataset (N=112, comprising predominantly T2DM patients), Guardian Pro achieves statistical parity with gradient-boosted baselines in point-wise discrimination (AUROC 0.985) but demonstrates superior clinical utility. Specifically, Guardian Pro improves event-based recall to 92.5% (vs. 77.3% for XGBoost) and tightens the average lead time to 128 minutes, providing a precise warning window compared to the diffuse, clinically non-actionable warnings of the tabular baseline (~200 minutes). Crucially, we isolate the impact of model architecture by evaluating a "Fair XGBoost" baseline trained on flattened raw temporal sequences. While this baseline achieved competitive temporal precision (137.5 minutes), it suffered a severe drop in safety, detecting only 38.3% of hypoglycemic events (vs. Guardian Pro's 92.5%) despite high AUROC (0.979). Furthermore, a multi-task variant of Guardian Pro generalizes effectively across horizons, achieving 0.9957 AUROC at the 10-minute horizon and 0.9760 AUROC at the 40-minute horizon on the held-out test cohort. [span_3](start_span)These results indicate that attention-based architectures, coupled with causal hysteresis, play a crucial role in resolving the trade-off between early warning and patient safety in subject-independent CGM monitoring.