Real-time Hypoglycemia Risk Management: A Calibrated and Interpretable HMM Approach for Continuous Glucose Monitoring

Abstract

Objective: This study enhances clinical interpretability and calibration of real-time 60-minute hypoglycemia risk from Continuous Glucose Monitoring. It reframes Hidden Markov Model risk as a posterior-weighted probability of glucose ≤ 70mg/dL, compared with logistic regression and a stacked ensemble. Methods: We developed an interpretable HMM P70 model for comparison of its performance with regularized logistic regression and a stacked ensemble. Evaluation, using a multi-patient CGM cohort, discrimination, calibration and net benefit Results: HMM P70 showed significantly improved calibration and lower Brier scores. However, logistic regression and ensemble models had improved discrimination, net clinical benefit, and earlier alert. The best combined performance was provided by the ensemble model. Conclusion: HMM P70 is a well-calibrated, interpretable real-time risk signal. It is complementary to more discriminative logistic regression models, which will help improve understanding and safe deployment of hypoglycemia risk tools.