Time-Aware Multimodal Sensor Fusion for Blood Glucose Forecasting in Non-Diabetic Adults

Abstract

Glucose dysregulation in non-diabetic individuals can precede metabolic disease onset by years, yet most glucose prediction models target diabetic populations with structured insulin records. This article presents Patch Time-Aware Cross-Attention (Patch-TACA), a multimodal transformer-based framework for long-term blood glucose forecasting in healthy individuals. The approach combines patch-based self-supervised pretraining with a time-aware cross-attention mechanism to integrate continuous glucose monitoring (CGM) data with physiological and behavioral measures, such as accelerometry, heart rate, electrodermal activity, and meal macronutrients, without requiring explicit resampling of irregularly sampled signals. Patch-TACA is assessed using data from twelve healthy participants across three laboratory sessions. It achieves an RMSE of 14.26 $\pm$ 3.48 mg/dL at the 90-minute prediction horizon, outperforming GlySim and Gluformer baselines. Moreover, employing a self-supervised method for pre-training reduced RMSE by 7.9% compared to training with random weight initialization, and hyperglycemia prediction accuracy reached 93.9%. These results demonstrate that multimodal sensor fusion with self-supervised learning enables accurate long-horizon glucose forecasting in non-diabetic populations, supporting proactive metabolic health monitoring before clinical disease onset.