Development of an Explainable Deep Learning-Based Decision Support System for Blood Glucose Levels Forecasting in Type 1 Diabetes Using Edge Computing

Maintaining blood glucose levels within the euglycemic range to avoid hypo/hyperglycemic events is a very important and complex challenge for people with Type 1 Diabetes. Current solutions employ complex deep learning models requiring substantial computational resources, often necessitating cloud-based processing, raising privacy concerns and significant CO2 emissions. Furthermore, the 'black-box' nature of these neural networks obscures the reasoning behind their predictions, impeding user trust and understanding. This paper introduces a decision support system leveraging a Long Short-Term Memory (LSTM) neural network for glycemic forecasting that utilizes solely Continuous Glucose Monitoring (CGM) data, using edge-computing for overcoming some of the above limitations. We present a streamlined model with an architecture optimized for edge devices, ensuring data privacy and reducing CO2 emissions by eliminating the need for transmitting and storing more than needed sensitive data. The system encompasses a full communication pipeline from CGM data acquisition and transmission to on-device prediction, culminating in the display of results for patient consultation. Our findings show that the model's performance, with an average Root Mean Square Error of 14.22 mg/dL and Clarke Error Grid Analysis with 91.95% of predictions in zone A and 7.05% in zone B for a prediction horizon of 30 min, is on par with current high-standard models. Crucially, we integrate a renowned explainability algorithm to elucidate the predictive process, offering valuable insights into the model's decision-making framework. This transparency aims to bolster user confidence and facilitate a deeper understanding of the predictive outcomes, marking a significant advancement in personalized diabetes management.