Detection of CVA and TIA at Emergency Department Triage: Developing a Prediction Machine Learning Model
2Hospital Management, Sheba Medical Center, Israel
Background and Purpose: Early identification of patients with CVA/TIA at emergency department (ED) triage can expedite CT evaluation. We aimed to develop a prediction model for CVA/TIA during triage.
Materials and methods: We retrieved data for consecutive adult patients (1/2012 to 12/2018) who performed head CT in our ED.
We obtained the following triage variables: chief complaint recorded by a triage nurse, demographics, vital signs, pain score, emergency severity index (ESI), comorbidities, and home medications. Previous ED visits and hospitalizations were also computed. CVA/TIA diagnoses were verified using ICD9 coding.
We used a gradient boosting model to predict CVA/TIA at triage time. Included variables were used as input for the model. We trained the model on years 2012-2017 data and tested on year 2018 data.
We evaluated the AUCs of single variables and the full model to predict CVA/TIA. We used Youden’s index to find the model’s optimal sensitivity and specificity.
Results: The study cohort included 93,202 patients who underwent head CT in our ED. Among those, 7,280 patients (7.8%) were finally diagnosed with CVA/TIA.
Variables with higher AUC included chief complaint (0.84), pain score (0.67), age (0.63), and systolic blood pressure (0.62).
The full model showed an AUC of 0.87 (95% CI: 0.86 - 0.88) for detecting CVA/TIA at triage. The model showed a sensitivity of 84.1% and specificity of 75.0% for detecting CVA/TIA.
Conclusion: Nurse staff recorded chief complaint has high accuracy for CVA/TIA. A machine-learning model may further expedite head CT scans for CVA/TIA at triage.