Benchmarking Early Deterioration Prediction Across Hospital-Rich and MCI-Like Emergency Triage Under Constrained Sensing

A team of researchers including KMA Solaiman has introduced a novel, clinically grounded benchmarking framework for evaluating AI models that predict early patient deterioration in emergency departments. Published on arXiv, the paper addresses a critical flaw in existing research: most models are tested using data that wouldn't be available during the crucial initial triage assessment. The authors created a realistic benchmark by comparing two scenarios—a 'hospital-rich' setting with full data and a 'MCI-like' (Mass Casualty Incident) setting restricted to only the vital signs available within a patient's first hour of presentation. This approach prevents data leakage and provides a more accurate measure of a model's real-world utility when seconds count and information is scarce.

The study, which utilized a deduplicated patient cohort from the MIMIC-IV-ED database, yielded a surprising and practical finding. Across multiple machine learning modeling approaches, predictive performance declined only modestly when models were limited to basic physiological measurements. Structured ablation studies pinpointed respiratory rate and oxygenation measures as the most influential signals for early risk stratification. The research demonstrates that AI models exhibit 'graceful degradation,' maintaining stable performance even as available sensor data is reduced. This work provides a essential tool for developing and validating triage decision-support systems that are actually deployable in field hospitals, disaster zones, or under-resourced clinics, moving AI from theoretical benchmarks to practical clinical impact.