Fairness-Constrained Temporal Feature Learning Algorithm for Cross-Population Glucose Prediction and Adherence Risk Stratification Using Continuous Glucose Monitoring Data

Abstract

Continuous glucose monitoring (CGM) data provides granular temporal information that enables individualized diabetes management, yet existing glucose prediction algorithms are predominantly trained and evaluated on small, homogeneous cohorts, raising concerns about cross-population generalizability and demographic equity. This study proposes a fairness-constrained temporal feature learning algorithm that integrates patch-based Transformer encoding with adversarial debiasing to improve cross-population glucose prediction accuracy and adherence risk stratification. Using four publicly available CGM benchmark datasets encompassing 243 participants across diverse diabetes types and demographic backgrounds, we evaluate the proposed approach against five baseline algorithms across 30-minute and 60-minute prediction horizons. Subgroup-stratified analysis reveals that the fairness-constrained approach reduces the maximum inter-group RMSE disparity from 4.83 mg/dL to 1.97 mg/dL at the 30-minute horizon while maintaining competitive overall prediction accuracy (RMSE: 18.26 mg/dL). CGM wear-time gap features extracted by the temporal encoder achieve an AUC of 0.817 for 7-day adherence risk prediction. These findings demonstrate that incorporating fairness constraints into CGM temporal feature learning can mitigate demographic performance disparities without substantial accuracy trade-offs, supporting more equitable data-driven diabetes care aligned with national chronic disease reduction strategies.