Machine Learning Models Predict Serious Dizziness Causes in the ED with AUC up to 0.97

In a multicenter cohort of 6637 ED dizziness patients, ML models predicted stroke/TIA/dissection/tumor with AUC up to 0.97 (sensitivity 97%, specificity 91%), with hypothetical CT use reductions of 53-85%. External validation needed.

• Retrospective multicenter cohort of 6637 patients (mean age 78, 57.8% female); 3.3% had a serious diagnosis. ML models (LASSO, XGBoost, etc.) compared with the Sudbury Vertigo Risk Score.
• At a 5% threshold, sensitivities ranged 53-97% and specificities 84-96% across models; confidence intervals overlapped substantially, limiting model selection.
• Study limitation: retrospective design, no external validation, and the resource reduction analysis is hypothetical—real-world impact remains uncertain.

RadPigeon summaries are original and for information only. They are not clinical advice.