Related Experiment Video
Updated: Jul 13, 2026
07:42
A Data-Driven Approach to Quantifying Immune States in Sepsis
Published on: February 7, 2025
Performance of Machine Learning Models for Sepsis and Stroke Detection Using EMS Data
Lawrence H Brown1,2, Remle P Crowe3, Oleksandr Ivanov4
1Dell Medical School, University of Texas at Austin, Austin, Texas.
Prehospital Emergency Care
|April 13, 2026
Summary
Machine learning models can identify sepsis and stroke in emergency medical services (EMS) data, improving early detection. This study shows the feasibility of applying hospital-trained models to prehospital electronic health records for better patient outcomes.
Area of Science:
- Emergency Medicine
- Data Science
- Machine Learning
Background:
- Early recognition of sepsis and stroke by emergency medical services (EMS) is crucial for improving patient triage, treatment, and outcomes.
- Machine learning (ML) offers potential to enhance EMS sepsis and stroke screening by analyzing complex patterns in electronic health record (EHR) data.
- Research applying ML to EMS data for these conditions remains limited.
Purpose of the Study:
- To evaluate the feasibility and performance of ML models, originally designed for hospital emergency department (ED) triage data, when applied to EMS EHR data for sepsis and stroke detection.
Main Methods:
- Retrospective analysis of linked EMS and ED records for adult patients transported to a single hospital.
- Hospital ML models were adapted to EMS EHR data, incorporating vital signs, chief complaints, and impressions, with ED physician diagnoses as the reference standard.
- Analyses included different prediction aggregation methods (majority vs. any prediction) and incorporation of EMS free-text narratives.
Main Results:
- The primary ML model ('majority of predictions') achieved 68% sensitivity and 86% specificity for sepsis, and 71% sensitivity and 94% specificity for stroke.
- The 'any prediction' approach increased sensitivity but decreased specificity.
- Incorporating free-text narratives further enhanced sensitivity at the cost of specificity.
Conclusions:
- Applying ML models trained on ED data to prehospital EMS data is feasible for early sepsis and stroke identification.
- Future research should focus on retraining models with EMS-specific data, prospective validation, and developing real-world implementation strategies.