Temporal Convolutional Networks Involving Multi-Patient Approach for Blood Glucose Level Predictions

Abstract

Data-driven techniques like neural networks (ANNs) have often been studied for predicting blood glucose levels (BGLs) of diabetes patients. However, the application of temporal convolutional networks (TCNs) is less known. Moreover, much of the existing literature contains patient-specific BGL models. The objective of this research is to evaluate TCNs for generalized BGL prediction of type-1 diabetes (T1D) patients and compare different calibration methods for selecting optimal hyperparameters for the TCNs. The results obtained by TCNs for BGL prediction using different calibration methods are also compared with artificial neural networks (ANNs). The ANNs were designed to have nearly the same number of trainable parameters as the TCNs. Twenty-four-hour time-series of forty T1D patients at fifteen-minutes intervals were generated using the Automated Insulin Dosage Advisor (AIDA) simulator. Thirty-two patients were chosen at random for model training, and the remaining eight patients were used for model testing. Three calibration methods based on training data, validation data, and k-fold cross-validation were used for hyperparameter tuning. The results showed that the second calibration method of randomly splitting the training data into training and validation datasets, and using the validation results for selecting optimal hyperparameters was the most appropriate. The TCN model with four temporal convolution layers followed by a dense layer obtained the best results. This research highlights the utility of TCNs for generalized BGL prediction.