A Reproducible Clinical Decision-Support Suite on MIMIC-IV

A research team developed a comprehensive suite of 26 clinical artificial-intelligence models on the MIMIC-IV database covering four areas: predicting deterioration and outcomes in intensive-care units, emergency-department triage, electrocardiogram interpretation, and clinical natural-language processing. The models employ various techniques including gradient-boosted trees and deep neural networks. Results achieve high accuracy: AUROC 0.884 for ICU mortality prediction, 0.830 for acute kidney injury, and 0.875 for emergency-department-to-ICU admission prediction. Deep learning on raw ECG signals improved myocardial-infarction detection by 0.142 AUROC compared to traditional methods. The model trained on MIMIC-IV data successfully generalized to another US dataset with 199,133 stays with only a 0.044 AUROC performance drop. All models are integrated into the zMed Critical Care application with clinical decision-support tools.