Developing a stroke alert trigger for clinical decision support at emergency triage using machine learning.

Area of Science:

• Emergency Medicine
• Neurology
• Data Science

Background:

• Acute stroke requires immediate emergency department (ED) assessment and treatment.
• Timely identification and activation of stroke systems are crucial for effective stroke management.
• Balancing false negatives and false positives in stroke detection is essential to avoid treatment delays and reduce neurologist burden.

Purpose of the Study:

• To develop and evaluate a stroke-alert trigger for identifying suspected stroke patients at ED triage.
• To compare the performance of rule-based algorithms with machine learning techniques for stroke detection.

Main Methods:

• Utilized clinical features from 1361 ED patients presenting within 12 hours of symptom onset.
• Compared rule-based algorithms (Face Arm Speech Test, Balance Eyes FAST) with six machine learning models (including Logistic Regression, Random Forest) using resampling techniques.
• Employed Logistic Regression to identify key predictive features for stroke suspicion.

Main Results:

• Machine learning models outperformed rule-based algorithms, achieving higher Area Under the Precision-Recall Curve (AUPRC) values.
• Top ML models, particularly Logistic Regression with SMOTE or class weighting, demonstrated strong predictive performance (AUPRC up to 0.787).
• Key features for stroke alert included presenting complaint, triage level, age, diastolic blood pressure, body temperature, and pulse rate.

Conclusions:

• Machine learning techniques significantly enhance the accuracy of prediction models for identifying suspected acute stroke patients.
• These ML models can be integrated into electronic triage systems to provide clinical decision support.
• Improved ED triage for stroke can lead to faster treatment and better patient outcomes.