Multivariate Long-Term Forecasting of T1DM: A Hybrid Econometric Model-Based Approach

Abstract

AbstractThere has been a huge effort for enhanced forecasting of blood glucose value for the development of accurate alarm systems that detect instances of hypoglycemia and hyperglycemia beforehand. In this paper, the blood glucose forecasting is carried out using model-based econometric method. Multivariate long-term forecasting machine learning algorithms, namely ARIMAX, ANN and SVM are employed for blood glucose prediction on T1DM dataset. The algorithms are implemented on two different T1DM datasets of patients from varying age, gender, nationality and race. The performance statistics of each of these algorithms is then paralleled upon. The data preparation event in the proposed blood glucose perdition model divides the independent variables into physiological and psychological variables. These variables are then keyed into the novel physiological, psychological and hybrid econometric time series model proposed for blood glucose forecasting. This time series model training is based on a unique time-flag method, where the 24 h time zone is divided into seven distinct time zones termed as time-flag. Considering hybrid econometric model, the performance metric of ANN was better in comparison with other algorithms. ANN exhibited MAPE of 3.70, 6.89, 14.86, 21.49, 13.37 (in mg/dL) for 30 min, 1 h, 2 h, 3 h and 24 h prediction horizon. The results obtained are clearly indicative that multivariate algorithms can be significantly accurate for long-term prediction of blood glucose value.KeywordsARIMAXANNBlood glucose predictionEconometric modelHybrid modelLong-term predictionMultivariateMAPEPhysiological modelPsychological modelRMSESVMT1DMTheil’s UUnivariate