Edge-Based Deep Learning for Population-Specific Blood Glucose Level Prediction in Type 1 Diabetes Mellitus

Abstract

This study evaluates a population-based deep learning (DL) approach to predict blood glucose levels (BGL) in patients with Type 1 Diabetes Mellitus (T1DM) using edge computing. Eight DL architectures, namely Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Neural Hierarchical Interpolation for Time Series Forecasting (NHITS), Kolmogorov-Arnold Network (KAN), Temporal Convolutional Network (TCN), Bidirectional Temporal Convolutional Network (BiTCN), Temporal Fusion Transformer (TFT), and Deep Non-Parametric Time Series Model (DeepNPTS), were optimized using Optuna Auto-Tuning and tested on the ShanghaiT1DM dataset, which consists of 16 subjects with a continuous glucose monitoring (CGM) sampling interval of 15 minutes. Two input configurations (univariate and multivariate) were assessed at prediction horizons (PHs) of 30 and 60 minutes through cloud-based simulations in Google Colab and deployment on an NVIDIA Jetson Orin Nano edge device. Performance is evaluated using MAE, MAPE, and RMSE for predictive performance, along with inference time, latency, throughput, and power consumption for deployment efficiency. In both PHs, KAN and TFT achieve superior predictive performance compared to the LSTM baseline and other models, with minimal differences observed between univariate and multivariate configurations. Crucially, KAN uniquely achieves the highest throughput (4.05 preds./s) at the lowest power consumption (7.64 W), making it the first architecture jointly characterized for population-based BGL prediction and uncompressed edge deployment, with improved predictive performance on edge over cloud at the 60-minute horizon, unseen in other architectures. These characteristics position KAN as a strong candidate for efficient on-device inference. Furthermore, with consistent power consumption of approximately 8 W across all eight models, the edge device implementation yielded predictive performance comparable to cloud-based simulations. These findings confirm that population-specific, edge-based BGL prediction is a feasible and practical solution for contemporary T1DM management.