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

Area of Science:

• Intensive Care Medicine
• Computational Biology
• Machine Learning

Background:

• Stress-induced hyperglycemia is a common complication in ICUs, necessitating precise blood glucose (BG) control.
• Patient variability makes BG management challenging, despite the use of protocols like Stochastic TARgeted (STAR).
• Insulin sensitivity (SI) prediction is key to the STAR protocol's effectiveness.

Purpose of the Study:

• To identify the optimal neural network (NN) configuration for enhanced SI prediction accuracy.
• To improve the performance of the STAR protocol through better SI forecasting.
• To explore quantile regression (QR) for predicting percentiles of non-Gaussian physiological data.

Main Methods:

• Implemented and trained various NN architectures with QR on a dataset of 1,897 patients (2011-2023).
• Utilized five-fold cross-validation for robust model evaluation.
• Assessed prediction performance using case-specific metrics across global and local SI domains.

Main Results:

• A simple NN with one hidden layer and four neurons, combined with QR, achieved optimal SI prediction performance.
• More complex NN architectures did not yield significant improvements in prediction accuracy.
• QR-based SI prediction does not require complex NNs for effective performance, especially for short-term horizons.

Conclusions:

• QR-based SI prediction is suitable for short-term forecasting, potentially enhancing clinical outcomes of the STAR protocol.
• The study offers a generalizable method for optimizing NN configurations in similar predictive tasks.
• Empirical validation supports the use of simpler NN models for efficient SI prediction.