Development and Validation of Machine Learning Models to Optimize Imaging and Referrals for Dizziness in the...

Area of Science:

• Emergency Medicine
• Artificial Intelligence
• Clinical Decision Support

Background:

• Dizziness and vertigo are common emergency department (ED) complaints, often leading to extensive testing due to the lack of reliable risk stratification tools.
• A small percentage (2%-5%) of these patients receive a serious diagnosis like stroke or TIA, indicating potential for overtreatment.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting serious diagnoses in ED patients presenting with dizziness or vertigo.
• To assess the potential of ML models to improve diagnostic accuracy and reduce resource utilization compared to existing methods.

Main Methods:

• A multicenter cohort study of 6637 ED patients with dizziness, vertigo, or imbalance was conducted.
• Four ML models were trained and validated to predict serious diagnoses (stroke, TIA, dissection, tumor) within 30 days.
• Model performance was evaluated using AUC and diagnostic accuracy metrics, and compared to the Sudbury Vertigo Risk score.

Main Results:

• All developed ML models demonstrated strong predictive performance, with AUCs ranging from 0.92 to 0.97.
• The LASSO logistic regression model showed excellent discrimination (AUC: 0.97) with high sensitivity (97%) and specificity (91%).
• Hypothetical application of ML models suggested significant reductions in CT utilization (53%-85%) and referrals (11%-73%).

Conclusions:

• Machine learning models show comparable or superior discrimination to the Sudbury Vertigo Risk score for identifying serious diagnoses in dizzy ED patients.
• These ML tools have the potential to enhance diagnostic specificity and decrease unnecessary healthcare resource utilization.
• External validation of these promising ML models is recommended before widespread clinical implementation.