Evaluation of insulin sensitivity temporal prediction by using quantile regression combined with neural network model.

Abstract

Stress-induced hyperglycemia, a pathologically high blood glucose level, is a frequent complication in intensive care units. Blood glucose (BG) level control is crucial but challenging due to patient variability. The Stochastic TARgeted (STAR) protocol is clinically used for blood glucose control, which uses the current and predicted future patient insulin sensitivity (SI) parameter to assess BG outcomes of alternative treatment options. Neural network (NN) models using quantile regression (QR) have enhanced SI prediction performance. However, remains a challenge in determining the optimal NN configuration to best predict SI. This study aims to find the NN configuration yielding the highest prediction accuracy to improve the STAR protocol and explores the behaviour of the QR method in predicting the percentiles of a non-Gaussian multi-mode distribution of a physiological parameter. Alternative NN architectures combined with QR were implemented and trained on a large dataset comprising 1,897 patients collected between 2011 and 2023 using five-fold cross-validation ensuring model robustness. Prediction performance was evaluated among NN configurations and compared using case-specific metrics across the global SI domain as well as within subdomains to analyse the models' local performance. Outcomes indicate QR applied to simpler NN, consisting of one-hidden layer with four neurons, achieves a best prediction performance at a minimum network size. Using more complex NN did not improve the prediction performance significantly. However, at long prediction horizons, no compact network demonstrated improved outcomes. A more general methodological outcome of the study is that QR-based prediction does not need to be combined with complex NN to achieve the best prediction performance. The QR-based method was found to be appropriate for the SI prediction problem in short-term predictions which may improve the STAR protocol's clinical outcomes. Overall, the study provides a generalisable, empirical approach to network configuration optimisation for similar problems.