# Reducing Emergency Diagnostic Uncertainty with TRACE: Triage and Risk Assessment via Cost Estimation

## TL;DR

TRACE is a machine-learning framework that improves emergency triage and diagnosis by combining risk assessment and patient similarity to reduce diagnostic uncertainty.

## Contribution

TRACE introduces a novel framework that integrates expected-value calculations and patient similarity metrics for triage and diagnosis in emergency medicine.

## Key findings

- TRACE-T improved triage prediction accuracy with random forest models (from 0.605 to 0.705) and reduced root mean square error.
- TRACE-Dx achieved high similarity scores (93.3% and 92.5%) in predicted diagnoses compared to actual outcomes.
- The framework generated five ranked diagnoses per patient, showing strong alignment with real clinical outcomes.

## Abstract

Diagnostic uncertainty significantly impacts patient safety in emergency medicine, leading to missed diagnoses and severe harm. Current predictive models primarily emphasize diagnostic likelihood without explicitly addressing potential clinical harm from errors. We propose Triage and Risk Assessment via Cost Estimation (TRACE), a machine-learning framework that incorporates expected-value calculations, defined as the probability-weighted estimate of clinical harm, and patient similarity metrics to address both diagnostic accuracy and risk assessment.

Using the Medical Information Mart for Intensive Care IV - Emergency Department dataset, we developed TRACE, comprising two modules: the expected value-powered triage index (TRACE-T), which calculates expected patient acuity from vital signs and chief complaints, and the patient similarity diagnosis engine (TRACE-Dx), which predicts diagnoses by identifying historically similar patients and weighing their outcomes by clinical harm. We assessed TRACE-T’s predictive performance, our primary outcome, using decision trees, random forests, and Lasso (least absolute shrinkage and selection operator) regression. The TRACE-Dx predictions, our secondary outcome, were evaluated through string matching (comparing diagnostic text) and sentence embedding similarity (comparing diagnostic phrases).

Our final analysis included a total of 2,501 patients from the dataset, due to requirements for diagnosis-string cleaning and computational demands of similarity calculations. Within this subset, TRACE-T significantly improved triage prediction accuracy, with the random forest classifier’s accuracy increasing from 0.605 to 0.705 (P = .04) and demonstrating a notable reduction in root mean square error from 0.635 to 0.541 (P < .001). The decision tree model improved from 0.467 to 0.593 (P = .78) but did not reach statistical significance. The TRACE-Dx generated five expected value-ranked predicted diagnoses per encounter (12,505 predictions across 2,501 patients) and achieved average sentence embedding and string match similarities of 93.3% (95% CI, 92.7–94.0%) and 92.5% (95% CI, 90.7–94.3%), respectively, indicating strong alignment with actual outcomes.

Expected value-based clinical harm modeling with patient similarity scoring enhances triage accuracy and diagnostic prediction in emergency care. Triage and Risk Assessment via Cost Estimation provides interpretable, actionable insights that could be incorporated into real-time clinical workflows as decision-support tools to reduce diagnostic uncertainty and improve patient outcomes.