MIMIC-TRIAGE: An Explainable Gradient-Boosted Framework for Early ICU Risk Prediction and Resource Optimization

Abstract

Manual clinical judgment or rule-based protocols are still widely used in patient triage for healthcare institutions, resulting in a lack of early identification of critically ill patients. The vast majority of current machine learning (ML) models are used to build a prediction for outcomes after the patient has been admitted (post-hoc prediction, e.g., mortality or length of stay) rather than for prioritizing admission at the moment of admission. This proposed work is an explainable, data-driven framework that is coupled with a rule-based clinical decision interface for early patient risk stratification, leveraging structured Electronic Health Record (EHR) data, referred to as "Medical Information Mart for Intensive Care (MIMIC-TRIAGE)". The system leverages demographic data, admission information, and several clinical features derived from 129,327 patient encounters to address the critical need for real-time interpretability in the early admission workflow. It features an interpretable LightGBM model and generates extremely reliable patient level risk scores with feature engineering. A rule-based assistant provides Shapley Additive Explanations (SHAP) values which allow for clinicians to ask about patient risk factors in real-time. These scores are further refined to categorize patients into different high, medium and low-risk groups. This system classifies patients most likely to need admission to the intensive care unit (ICU) and uses an aggregated risk profile to estimate the overall demand for the ICU. Experimental results on the MIMIC-IV dataset show that the ICU recall rate is 85%, while the accuracy, F1-score, precision, and Area Under the Curve (AUC) are 61%, 40%, 26% and 78%, respectively. This approach aims to increase recall, minimize the risk of missing important cases and facilitate prompt clinical action and optimization of ICU resources.