Abstract
For people with Type 1 diabetes (T1D), accurate blood glucose (BG) forecasting is crucial for the effective delivery of insulin by Artificial Pancreas (AP) systems. Deep learning frameworks like Long Short-Term-Memory (LSTM) have been widely used to predict BG using continuous glucose monitor (CGM) data. However, these methods usually require large amounts of training data for personalized forecasts. Moreover, individuals with diabetes exhibit diverse glucose variability (GV), resulting in varying forecast accuracy. To address these limitations, we propose a novel deep learning framework: Incrementally Retrained Stacked LSTM (IS-LSTM). This approach gradually adapts to individualsâ data and employs parameter-transfer for efficiency. We compare our method to three benchmarks using two CGM datasets from individuals with T1D: OpenAPS and Replace-BG. On both datasets, our approach significantly reduces root mean square error compared to the state of the art (Stacked LSTM): from 14.55 to 10.23 mg/dL (OpenAPS) and 17.15 to 13.41 mg/dL (Replace-BG) at 30-minute Prediction Horizon (PH). Clarke error grid analysis demonstrates clinical feasibility with at least 98.81% and 97.25% of predictions within the clinically safe zone at 30- and 60-minute PHs. Further, we demonstrate the effectiveness of our method in cold-start scenarios, which helps new CGM users obtain accurate predictions.