Modelling type 2 diabetic patients’ glucose metabolism for real-time predictive healthcare service

Abstract

Continuous glucose monitoring (CGM) sensors and other smart wearables offer valuable opportunities for advancing diabetes disease self-management. However, whether patients may take advantage of CGM data, take preventative behavior, and achieve improved glucose self-management result under natural living conditions still needs to be explored. Human’ s capacity of time series data processing in glucose self-management scenario needs to be evaluated. Accordingly, this study aims to (1) examine whether wearing a CGM sensor alone is effective for glucose self-management and explore key obstacles if any, and (2) develop a high-performance predictive model for real-time glucose forecasting to facilitate self-management. Thirty Type 2 Diabetes (T2D) patients were recruited to collect 13–14 days of glucose, dietary, exercise, and medication data in real-world settings using CGM sensors and smart bands, supplemented by self-reported loggings. The results show no significant improvement in patients' mean glucose levels. An attention-enhanced long short-term memory (ALSTM) glucose prediction model was developed and validated using data from 20 out of the 30 participants, achieving high predictive accuracy for 30/60-min prediction accuracy (errors <5 %). When applied to the other 10 participants, the model combined with deep transfer learning gained high prediction accuracy (errors <6.4 % and 4.7 %, respectively), and may enable early-stage predictions for new patients with limited data. Further experiment with 60 graduate students showed that human’ s predicting accuracy cannot compete with the model and is insufficient for self-management. The proposed approach holds promise for future predictive interventions, enabling timely and personalized glucose management strategies for T2D patients.