Towards Interpretation for Blood Glucose Level Prediction Models

Abstract

Patients with diabetes need to manage their blood glucose (BG) level to prevent diabetic complications such as retinopathy and cardiovascular diseases. We developed tree-based machine learning (ML) and deep learning (DL) models with continuous glucose monitoring (CGM) data points to improve the BG management. We extracted 20 CGM time series from 20 virtual patients with type 1 diabetes generat-ed by UVA/Padova Type 1 Diabetes Metabolic Simulator, set 12 CGM data points as input, and the CGM data point after 30-min prediction horizon as output. The long short-term memory showed the lowest average root mean squared error (17.37 mg/dL) and mean absolute percentage error (8.33 %). In the clinical analysis, the deep neural network showed the highest percentage in region A (92.53 %) of Clarke error grid analysis (CEGA) and all models had the high percentage in region A and B (> 99 %) of CEGA. Then, we analyzed each model’s feature importance and found that the models exhib-ited different feature importance. We believe that the presented method will help to manage BG levels of patients with diabetes and to interpret the BG level predictive models.