AI-driven multimodal triage system for early ICU risk prediction in emergency departments

Abstract

Artificial Intelligence has been redefining the healthcare industry through the use of complex algorithms and large-scale clinical data to identify predictive patterns that can reduce human error and optimize clinical decision-making compared to traditional triage methods. Traditional triage systems such as CTAS, ESI, and MTS primarily rely on subjective clinical judgment and are designed to assess immediate acuity scale rather than predict short-term clinical deterioration. In this thesis, a calibrated multimodal AI-driven triage decision-support framework is proposed for predicting ICU admission within 24 hours of emergency department presentation. To predict high risk patients, structured triage variables, free-text clinical complaints, and short-horizon temporal physiological trends are extracted from large-scale electronic health record datasets, including MIMIC-IV and eICU. Risk estimation is performed using calibrated gradient-boosted tree models, text-based classifiers, and temporal sequence-based neural modeling to produce interpretable and operationally reliable predictions. In emergency settings, calibration techniques ensure that predicted probabilities correspond closely to true outcome likelihoods, enabling safer prioritization under uncertainty. Model performance is evaluated using both discrimination and calibration metrics, reflecting predictive accuracy and probabilistic reliability under imbalance. To assess operational relevance, a novel capacity-aware simulation framework is developed to compare traditional triage ordering with AI-based prioritization under constrained resources. Results indicate that the proposed system improves early identification of high-risk patients and supports more effective prioritization decisions in emergency care.