Wearable Noninvasive Blood Glucose Monitoring via ECG: A Multi-Attention Hierarchical Feature Fusion and Fuzzy Integral Approach

Continuous glucose monitoring (CGM) is an advanced technology with significant potential to improve diabetes management. However, its widespread adoption remains limited due to the high cost and the invasive nature of current measurement techniques. Changes in blood glucose (BG) levels stimulate the sympathetic and parasympathetic systems, leading to variations in the electrocardiogram (ECG). Hence, we proposed an artificial intelligence (AI)-driven, noninvasive BG monitoring approach based on ECG signals, addressing the need for affordable and continuous glucose tracking. Firstly, we applied multi-attention-based feature extraction with channel and spatial aspects to highlight important features and suppress dispensable ones. Secondly, we proposed multi-hierarchical feature fusion to gain the inter-layer second-order feature and enhance their representation capability. Lastly, fuzzy integral was adopted to couple different BG estimation values for decision-level fusion. To verify the feasibility of our approach, we collected a total of 103 days of ECG data with concomitant BG values from 21 participants and conducted a 10-fold cross-validation. Evaluation results revealed that our approach demonstrated competitive performance compared with several existing models. The root-mean-square error was 26.57 mg/dL, and the mean absolute relative difference was 13.39%. In addition, the zone A + B of Clarke and Parkes error grid analysis were 98.50% and 99.41%, respectively. Therefore, these results support the technical feasibility of implementing AI-driven, ECG-based BG monitoring in consumer electronics.