Strategies of Multi-Step-ahead Forecasting for Blood Glucose Level using LSTM Neural Networks: A Comparative Study

Predicting the blood glucose level (BGL) is crucial for self-management of Diabetes. In general, a BGL prediction is done based on the previous measurements of BGL, which can be taken either (manually) by using sticks or (automatically) by using continuous glucose monitoring (CGM) devices. To allow the diabetic patients to take appropriate actions, the BGL predictions should be done ahead of time; thus a multi-step ahead prediction is suitable. Therefore, many Multi-Step-ahead Forecasting (MSF) strategies have been developed and evaluated, and can be categorized in five types: Recursive, Direct, MIMO (for Multiple Input Multiple Output), DirMO (combining Direct and MIMO) and DirRec (combining Direct and Recursive). However, none of them is known to be the best strategy in all contexts. The present study aims at: 1) reviewing the MSF strategies, and 2) determining the best strategy to fit with a LSTM Neural Network model. Hence, we evaluated and compared in terms of two performance criteria: Root-Mean-Square Error (RMSE) and Mean Absolute Error (MAE), the five MSF strategies using a LSTM Neural Network with an horizon of 30 minutes. The results show that there is no strategy that significantly outperformed others when using the Wilcoxon statistical test. However, when using the Sum Ranking Differences method, HMO is the best strategy for both RMSE and MAE criteria.