Comparison of Three Artificial Intelligence Methods for Predicting 90% Quantile Interval of Future Insulin Sensitivity of Intensive Care Patients

Abstract

Three alternative artificial intelligence-based insulin sensitivity prediction methods are compared in this study. Insulin sensitivity prediction is an essential step in calculating the optimal treatment options in model-based glycemic control protocol of insulin-dependent intensive care patients. The prediction methods must predict not only the expected value of the insulin sensitivity for a given time horizon but also the 90% confidence interval making the prediction problem more specific compared to the common prediction problems. All of the proposed prediction methods - proposed in our previous publications - use different neural network models: a classification deep neural network model, a Mixture Density Network based model, and a Quantile regression based model. The patent data set used for the development and accuracy assessment is from 3 clinical ICU cohorts, including 820 treatment episodes of 606 patients and 68,631 hours of treatment. To evaluate the efficacy of the prediction in the context of clinical requirements, three metrics are used Success rate, Interval ratio, and I-Score are applied.