Neural Architecture Search for Ultra-low Memory Blood Glucose Forecasting on the Edge

Abstract

Predictive modeling of blood glucose levels of patients is a key enabler for automated management of diabetes. However, such models are typically executed on battery-operated and resource-constrained edge devices, which limits the deployment feasibility of complex machine learning methods. To alleviate this challenge, this paper proposes a neural architecture search formulation tailored to the problem of blood glucose forecasting. This approach exploits joint Bayesian optimization of task performance and computational cost, thereby producing highly optimized architectures. Experimental results show that the resulting models attain state-of-the-art performance in predicting occurrences of future hyperglycemia and hypoglycemia on the OhioT1DM dataset, with accuracies of 88.28% and 79.44%, respectively. Moreover, the proposed models achieve state-of-the-art performance in blood glucose level estimation for all dataset-prediction horizon combinations, while providing improvements up to 130× and 1000× for memory and processing costs, respectively.