Diagnosis of Acute Central Dizziness With Simple Clinical Information Using Machine Learning

Abstract

Background: Acute dizziness is a common symptom among patients visiting emergency medical centers. Extensive neurological examinations aimed at delineating the cause of dizziness often require experience and specialized training. We tried to diagnose central dizziness by machine learning using only basic clinical information. Methods: Patients were enrolled who had visited an emergency medical center with acute dizziness and underwent diffusion-weighted imaging. The enrolled patients were dichotomized as either having central (with a corresponding central lesion) or non-central dizziness. We obtained patient demographics, risk factors, vital signs, and presentation (non-whirling type dizziness or vertigo). Various machine learning algorithms were used to predict central dizziness. The area under the receiver operating characteristic curve (AUROC) was measured to evaluate diagnostic accuracy. The SHapley Additive exPlanations (SHAP) value was used to explain the importance of each factor. Results: Of the 4,481 visits, 414 (9.2%) were determined as central dizziness. Central dizziness patients were more often older and male and had more risk factors and higher systolic blood pressure. They also presented more frequently with non-whirling type dizziness (79 vs. 54.4%) than non-central dizziness. Catboost model showed the highest AUROC (0.738) with a 94.4% sensitivity and 31.9% specificity in the test set (n = 1,317). The SHAP value was highest for previous stroke presence (mean; 0.74), followed by male (0.33), presentation as non-whirling type dizziness (0.30), and age (0.25). Conclusions: Machine learning is feasible for classifying central dizziness using demographics, risk factors, vital signs, and clinical dizziness presentation, which are obtainable at the triage.