Background and Aims: Accurate prediction of glucose levels in patients with type 1 diabetes mellitus (T1DM) is critical both for their glycemic control and for the development of closed-loop systems. Methods: In this study, we utilized real-life, retrospective, continuous glucose monitoring data from 141 T1DM patients (9,083 connection days, 1,592,506 glucose measurements) and in silico data generated by the UVA/Padova T1DM simulator to evaluate different computational methods for glucose prediction. We evaluated the performance of the models using both measures of numerical accuracy, measured by the root mean square error, and clinical accuracy, measured by the percentage of time in each of the Clarke error grid (CEG) zones, and compared the predictions done by autoregressive (AR) models, tree-based methods, artificial neural networks, and a novel model that we devised and optimized for this task. Results: Our novel model, constructed on real-life data, achieved clinical accuracy of 99.3% and 95.8% in predicting the glucose level 30 and 60 min ahead, respectively, and reduced the percentage of glucose predictions in zones C-E of the CEG by 60.6% and 38.4% in these prediction horizons, compared with a standard AR model. The model was superior to all other models across all age groups and achieved higher clinical accuracy in subgroups of patients with high glucose variability and greater time spent in hypoglycemia. Compared with real-life data, when evaluated on in silico data, the model had a higher clinical and numerical accuracy. Conclusions: A model that optimizes for CEG zones may significantly improve clinical accuracy and clinical outcomes of treatment decisions in T1DM patients. Results obtained from simulated data may overestimate the performance of models on real-life data.
Keywords: Artificial neural network; Clinical accuracy; Continuous glucose monitoring; Glucose prediction; Type 1 diabetes.