Blood glucose prediction in type 1 diabetes using deep learning on the edge

Real-time blood glucose (BG) prediction can enhance decision support systems for insulin dosing such as bolus calculators and closed-loop systems for insulin delivery. Deep learning has been proven to achieve state-of-the-art performance in BG prediction. However, it is usually seen as a very computationally expensive approach, hence difficult to implement in wearable medical devices such as transmitters in continuous glucose monitoring (CGM) systems. In this work, we introduce a novel deep learning framework to predict BG levels with the edge inference on a microcontroller unit embedded in a low-power system. By using glucose measurements from a CGM sensor and a recurrent neural network that builds on long-short term memory, the personalized models achieves state-of-the-art performance on a clinical data set obtained from 12 subjects with T1D. In particular, the proposed method achieves an average root mean square error of 19.10 ± 2.04 for a 30-minute prediction horizon (PH) and 32.61 ± 3.45 for a 60-minute PH with high clinical accuracy. Notably, the framework has been optimized to achieve a minimum use of hardware resources (34KB FLASH and 1KB SRAM) as well as an execution time of 22 ms for low power operations (8 µW). The presented system has the potential to be implemented in wearable medical devices for diabetes care (CGM and insulin pumps) and to be integrated within an Internet of Things platform.