Predicting blood glucose levels using machine learning techniques with metaheuristic optimisers

Abstract

Introduction: Persons with Type-1 diabetes need to continuously monitor their blood glucose level to remain within a healthy range. Using machine learning techniques researchers can predict blood glucose values with the benefit of providing the patient with future blood glucose values with the aim of primitively taking action. The focus of this study was to investigate the use of metaheuristic optimisers to strategically tune the hyperparameter configuration of these machine learners in the context of blood glucose prediction using the OhioT1DM dataset with the aim of improving the predictive performance of the machine learners. Research questions: i) What is the degree of improvement when using a metaheuristic approach over a completely random search given the same search space? ii) How can the computation be carried out in a shorter time, what are possible ways of distributing the workload among several machines? Methodology: A few machine learners namely the MLP, RNN and XGBoost were implemented for the prediction of blood glucose level. Moreover, two metaheuristic optimisers, the genetic algorithm and particle swarm optimisation, and random search were used to perform hyperparameter optimisation. The experimentation was run three times to obtain an average of the performance. Due to the increased computation load in running multiple runs a Spark cluster running on EC2 instances was considered to reduce the computation time. Results & evaluation: The results obtained from the experimentation give an indication that for the context of the ohioT1DM dataset and configurations set, the metaheuristic optimisers consistently provide a slightly better predictive performance when given enough iterations. Conclusion: This study demonstrated that the use of metaheuristic optimisers in the context of blood glucose prediction when using the OhioT1DM dataset can provide improved results over random search. It is noted that using such techniques significantly increased computational load.