Related Experiment Video
Updated: Aug 12, 2025
09:52
Setting Up a Stroke Team Algorithm and Conducting Simulation-based Training in the Emergency Department - A Practical Guide
Published on: January 15, 2017
17.3K
A Machine Learning Approach to Support Urgent Stroke Triage Using Administrative Data and Social Determinants of
Min Chen1, Xuan Tan2, Rema Padman3
1Department of Information Systems & Business Analytics, College of Business, Florida International University, Miami, FL, United States.
Journal of Medical Internet Research
|January 30, 2023
Summary
This study developed a machine learning (ML) stroke prediction model using readily available patient data, significantly improving diagnostic accuracy, especially in underserved areas. The algorithm effectively identifies stroke, reducing misdiagnosis of stroke mimics.
Area of Science:
- Medical Informatics
- Public Health
- Machine Learning in Healthcare
Background:
- Timely stroke diagnosis and triage are critical for effective management.
- Existing machine learning (ML) methods often rely on detailed clinical data not always available during initial patient triage, particularly in rural or underserved settings.
- This limitation highlights the need for stroke prediction tools utilizing more accessible data.
Purpose of the Study:
- To develop an ML stroke prediction algorithm using data available at hospital presentation.
- To evaluate the added value of social determinants of health (SDoH) in enhancing stroke prediction.
- To create a tool that assists in the early detection and triage of stroke patients.
Main Methods:
- Retrospective analysis of Florida hospital records (2012-2014) merged with American Community Survey SDoH data.
- Development of stroke and stroke mimic cohorts using a case-control design.
- Comparison of ML models (gradient boosting machine, random forest) against logistic regression, utilizing TreeSHAP for model interpretability.
Main Results:
- The ML model achieved high sensitivity, with a false-negative rate below 4% for stroke mimics.
- ML classifiers outperformed logistic regression across various data inputs.
- Key predictors included age, chronic conditions, and primary payer; individual-level SDoH significantly boosted predictive performance (AUC from 0.694 to 0.823).
Conclusions:
- A sensitive stroke prediction model was developed using routinely collected, widely available patient data.
- The algorithm is suitable for resource-limited hospitals lacking advanced diagnostic tools.
- Incorporating individual-level SDoH data substantially improved stroke prediction accuracy.
