Advancing blood glucose prediction with neural architecture search and deep reinforcement learning for type 1 diabetics

For individuals with Type-1 diabetes mellitus, accurate prediction of future blood glucose values is crucial to aid its regulation with insulin administration, tailored to the individual’s specific needs. The authors propose a novel approach for the integration of a neural architecture search framework with deep reinforcement learning to autonomously generate and train architectures, optimized for each subject over model size and analytical prediction performance, for the blood glucose prediction task in individuals with Type-1 diabetes. The authors evaluate the proposed approach on the OhioT1DM dataset, which includes blood glucose monitoring records at 5-min intervals over 8 weeks for 12 patients with Type-1 diabetes mellitus. Prior work focused on predicting blood glucose levels in 30 and 45-min prediction horizons, equivalent to 6 and 9 data points, respectively. Compared to the previously achieved best error, the proposed method demonstrates improvements of 18.4 % and 22.5 % on average for mean absolute error in the 30-min and 45-min prediction horizons, respectively, through the proposed deep reinforcement learning framework. Using the deep reinforcement learning framework, the best-case and worst-case analytical performance measured over root mean square error and mean absolute error was obtained for subject ID 570 and subject ID 584, respectively. Models optimized for performance on the prediction task and model size were obtained after implementing neural architecture search in conjunction with deep reinforcement learning on these two extreme cases. The authors demonstrate improvements of 4.8 % using Long Short Term Memory-based architectures and 5.7 % with Gated Recurrent Units-based architectures for patient ID 570 on the analytical prediction performance by integrating neural architecture search with deep reinforcement learning framework. The patient with the lowest performance (ID 584) on the deep reinforcement learning method had an even greater performance boost, with improvements of 10.0 % and 12.6 % observed for the Long Short-Term Memory and Gated Recurrent Units, respectively. The subject-specific optimized models over performance and model size from the neural architecture search in conjunction with deep reinforcement learning had a reduction in model size which ranged from 20 to 150 times compared to the model obtained using only the deep reinforcement learning method. The smaller size, indicating a reduction in model complexity in terms of the number of trainable network parameters, was achieved without a loss in the prediction performance.