Review of machine learning models in short- and long-term glucose forecasting and hypoglycemia classification

Type 1 Diabetes (T1D) is a chronic autoimmune disorder that requires lifelong insulin therapy. A common side effect is hypoglycemia, characterized by decreased blood glucose levels (BGL) below 70 mg/dL. Diabetes care can be optimized using machine learning (ML) models that can predict and alert patients to potential glycemic abnormalities. The ML models can be classified into regression-based, in which glucose levels are forecasted, and classification-based, in which adverse events are classified. This review analyzes the performance of ML models applied to T1D and compares these in terms of short- and long-term prediction horizons (PHs), defined as 15–120 min and 3 to more than 24 h, respectively. This review investigates: 1) How much in advance can glucose values or a hypoglycemic event be accurately predicted? 2) Which ML methods have the best performance? 3) Which factors impact the performance? And 4) Does personalization increase performance? The results indicate that 1) a PH of up to 1 h provides the best results. 2) Conventional ML methods yield the best results for classification and deep learning (DL) for regression. A single model cannot adequately classify across multiple PHs. 3) The model performance is influenced by multivariate datasets and the input sequence length (ISL). 4) Finally, personal data enhances performance, but due to limited data quality, population-based models are preferred. • Systematic review of glucose forecasting and hypoglycemia classification. • Short- (15–120 min) and long-term (2 to ≥24h) prediction horizons are considered. • Machine learning methods yield best results for hypoglycemia classification. • Ensemble learning methods yield best results for glucose forecasting. • Personal, psychological and physiological data improve performance.