Blood Glucose Anomaly Correction and Personalized Short-Term Prediction Based on Continuous Glucose Monitoring Data

Diabetes, as a chronic metabolic disease, relies on precise blood glucose monitoring and management for treatment. The quality of continuous glucose monitoring (CGM) data and prediction accuracy are critical for achieving closed-loop control in an automated insulin delivery system (AIDS). To address challenges such as sensor anomalies and physiological delays in CGM data, this study proposes an attention-based framework for glucose anomaly detection and short-term prediction. First, we design an anomaly detection and correction module that integrates a multiperiod spatiotemporal attention mechanism with a dual-task joint learning framework to simultaneously optimize anomaly identification and interpolation. In addition, a blood glucose prediction module based on a recurrent self-attention mechanism and transfer learning is proposed. Experimental results show that the proposed anomaly detection model achieves an accuracy of 96.4%, with Pearson correlation coefficients (PearsonCCs) and intragroup correlation coefficients between interpolation results and actual values reaching 0.854 and 0.846, respectively. On the Yuwell dataset, the glucose prediction module attains average root mean square errors (RMSEs) of 6.27, 10.40, and 14.31 mg/dL at 15, 30, and 45 min. The average clinical safety zone coverage in Clarke error grid analysis reaches 99.52%. On the public OhioT1DM dataset, the average clinical safety zone coverage reaches 98.92%. Our approach significantly outperforms some existing methods in detection sensitivity, prediction accuracy, and clinical applicability, providing a robust and practical solution for diabetes management.