A Combination of Decision Trees with Machine Learning Ensembles for Blood Glucose Level Predictions

Abstract

Abstract Researchers have developed many patient-specific physiological and machine learning (ML) models to predict blood glucose levels (BGLs) of Type-1 diabetes (T1D) patients. However, less is known on how traditional ensemble and non-ensemble ML algorithms can be combined to predict BGLs. This research’s primary objective is to evaluate various ensemble ML models for generalized BGL prediction and evaluate their novel combination with the decision tree (DCT) models. Twenty-four-hour data of 40 patients at 15-min intervals were generated using the automated insulin dosage advisor (AIDA) simulator. The decision tree (DCT), random forest (RF), extra trees (EXT), gradient boost (GBoost), Adaboost (ABoost), and bagging models were evaluated on the data. A new two-stage model using decision tree (DCT) and Adaboost (DCT-ABoost) was developed, where the predictions from the DCT model were fed as an extra input to the ABoost model for the final BGL prediction. The results revealed that the DCT-ABoost model outperformed the traditional models (DCT, RF, EXT, GBoost, ABoost, and bagging) and other new two-stage models (DCT-EXT, DCT-Bagging, DCT-RF, and DCT-GBoost) designed in this research. This research highlights the utility of developing new multi-stage models for generalized BGL prediction of T1D patients.KeywordsDiabetesTime-seriesMachine learningDecision treeExtra treeEnsembleRandom forestExtra treesAdaBoostGradient boostBaggingCascading