Ranking Features Distinguish Vestibular Migraine with Positional Vertigo from Benign Paroxysmal Positional Vertigo Using Machine Learning

Objectives: Vestibular migraine with positional vertigo (VMwPV) can mimic benign paroxysmal positional vertigo (BPPV), leading to diagnostic challenges. This study aimed to compare clinical and nystagmus-related features between VMwPV and BPPV patients and to establish a differential diagnostic model using machine learning. Methods: 101 VMwPV and 219 BPPV patients were enrolled. The nystagmus-related features included spontaneous nystagmus (SN) and parameters during positional testing: latency, pure vertical or vertical-component positional nystagmus (PN), PN lasting <1 min, and the maximum slow phase velocity (SPVmax) of PN. Vertigo/dizziness during positional testing was also included. Using multiple machine learning algorithms, diagnostic models were constructed, and the optimal model was selected. The Shapley Additive Explanations (SHAP) analysis was applied to rank the feature importance. Results: (1) Compared with BPPV, VMwPV patients were more likely to have SN and pure vertical/vertical-component PN but less likely to have latency, PN lasting <1 min, and provoked vertigo/dizziness; the SPVmax of PN was lower in VMwPV than in BPPV. (2) No statistically significant differences in area under the curve (AUC) were observed among the five models. Logistic Regression was selected as the optimal model based on its superior interpretability. (3) Vertigo/dizziness during positional testing and PN lasting <1 min were ranked as the top two most important features based on SHAP analysis. These findings should be interpreted with caution due to group selection and incorporation bias. Conclusions: Apart from the subjective provoked vertigo/dizziness, other objective variables such as PN lasting <1 min may have potential value in differentiating VMwPV and BPPV, which require multicenter studies and prospective external validation.