| Abstract [eng] |
Reliable short-horizon continuous glucose monitoring (CGM) forecasting is essential to prevent hypoglycemia, yet most data-driven models provide neither physiological awareness nor actionable safety guaranties. We introduce a trustworthy forecasting framework that unifies continuous-time neural dynamics, calibration-based conformal risk control (empirically validated in held-out subjects), and tail-risk-aware adaptation. First, a Neural Controlled Differential Equation (Neural CDE) encoder models irregular, multi-channel streams (CGM, insulin, carbohydrates, activity, time-of-day). A physiology regularizer anchors latent dynamics to plausible glucose responses (e.g., post-prandial rise, insulin-mediated decay) and penalizes simulator-inconsistent trends. Second, Conformal Risk Control calibrates prediction sets and provides an actionable deferral policy when uncertainty or low-glucose risk is high, achieving prediction sets calibrated to a nominal coverage target and empirically validated under subject-disjoint evaluation. Third, a risk-sensitive policy (CVaR-optimized PPO) adapts predictions to minimize tail errors, explicitly weighting under-predictions near hypoglycemia. We evaluated in BrisT1D and OhioT1DM with patient-wise splits and performed in-silico stress tests in the UVA/Padova simulator (meal/bolus perturbations, basal overrides, sensor noise). Across 30/60-minutes horizons, our method achieves MAE=3.0–3.9 mg/dL, RMSE=8.5–9.8 mg/dL, and MARD=2.1–2.7%, outperforming strong baselines by 10–14 mg/dL RMSE and 10–13 mg/dL MAE on average. The risk control is calibrated at a nominal target of 90%, achieving 90.5 to 91.5% coverage with a average set width of 18-20 mg/dL and 5% to 6% deferral. Clinically, the rates of the SEG/Parkes A/B-zone increase by +10–14 pp, and event-centric analyzes show PR-AUC=0.92–0.95 and ROC-AUC=0.94–0.96 with earlier warnings (9–11 min median lead-time) and a 35–45% reduction in hypoglycemia under-prediction near the threshold. To our knowledge, this is the first end-to-end framework combining physiology-constrained Neural CDEs, conformal risk control with deferral, and CVaR-PPO adaptation for CGM forecasting, delivering verifiable safety properties alongside state-of-the-art accuracy. |
