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

Abstract

Background: Dizziness and vertigo are common emergency department (ED) presentations, but only 2%-5% receive a serious diagnosis, such as stroke or transient ischemic attack (TIA). Due to the lack of reliable validated prediction tools, many undergo unnecessary imaging and consultations, highlighting the need for improved risk stratification.

Objective: To develop machine learning (ML) models that predict serious diagnoses in ED patients presenting with dizziness or vertigo.

Methods: This multicenter cohort study included 6637 ED patients with dizziness, vertigo, or imbalance from September 2014-December 2022. The primary outcome was a serious diagnosis-stroke, TIA, vertebral artery dissection, or brain tumor-within 30 days, adjudicated by a blinded committee. Data were split 80/20 into training and test sets. Four ML models (decision tree, LASSO logistic regression, random forest, XGBoost) were trained on 17 variables using 5-fold cross-validation and evaluated alongside the Sudbury Vertigo Risk score. Performance was assessed using area under the curve (AUC) and diagnostic accuracy measures. Computed tomography (CT) and referral rates were hypothetically compared pre- and post-model application.

Results: Among 6637 patients (mean age 78.1; 57.8% female), 3.3% had a serious diagnosis. All ML models demonstrated strong discrimination, with AUCs ranging from 0.92 to 0.97. At a 5% predicted probability threshold, sensitivities ranged from 53%-97% and specificities from 84% to 96%. Logistic regression with LASSO demonstrated a favorable balance between discrimination (AUC: 0.97, sensitivity: 97% and specificity: 91%), although confidence intervals overlapped substantially across models. In a hypothetical model-based analysis, ML-guided classification corresponded to projected reductions in CT utilization and referrals ranging from 53%-85% and 11%-73%, respectively.

Conclusions: Select ML models demonstrated discrimination comparable to the Sudbury Vertigo Risk Score while potentially improving specificity and reducing projected resource utilization. These tools show promise, but external validation is needed.