Data-driven Temporal Evolution Pattern Mining of Blood Glucose for Type 1 Diabetes —— A Comparison Study

The temporal evolution patterns of blood glucose for Type 1 diabetes patients are often non-stationary and highly nonlinear due to uncertain daily activities. In this work, we first conduct a comprehensive study on different performance aspects of various glucose prediction models, including mixing mechanism-based, transformer-based, recurrent neural network, convolutional kernel-based and other classical models. Moreover, we investigate the model accuracy, the generalization capability and the performance stability of different models. Then a comprehensive performance evaluation and comparison has been carried out based on the OhioT1DM dataset, which comprises various time-varying individual features from 12 patients with Type 1 diabetes over 8 weeks. Compared to other models, the TSMixer model demonstrates outstanding performances in both prediction accuracy and model stability, while the PatchMixer model achieves best generalization ability. The findings of this study provide empirical insights and valuable guidance for adequate personalized prediction models, and further facilitate proactive decisions about diabetes management in response to different lifestyles.