The need for speed: using nystagmus velocity profiles and machine learning models to separate canalithiasis BPV from its mimics.

PURPOSE: Separating BPV from other positional nystagmus types can be challenging. We examined the utility of nystagmus slow-phase velocity (SPV) profiles when seeking to separate canalithiasis benign positional vertigo (BPV) from its mimics.

METHODS

231 eye-videos canalithiasis BPV were compared against 245 non-BPV positional nystagmus recordings whose diagnoses included: vestibular migraine, posterior fossa tumors, and cerebellar ataxia. A custom-written pupil-tracker generated 2D eye-position traces and plotted SPV as a function of time. Nystagmus onset latency, 50% rise time, peak velocity, peak latency, time to 50%, and 95% decay of SPV (T50 and T95) were recorded. Machine learning (ML) models were developed to separate BPV from non-BPV positional nystagmus.

RESULTS

Median nystagmus onset, 50% rise time, peak velocity, and peak latency were 0.789 s, 0.488 s, 36.530°/s, and 2.487 s for BPV and 0.321 s, 0 s, 9.306°/s, and 4.418 s for non-BPV groups. BPV demonstrated a paroxysmal profile with T50 and T95 of 3.062 s and 10.616 s; the non-BPV group had a T50 of 19.02 s and did not decay to 95% by 1000 s. All six metrics demonstrated significant differences in the dominant plane. Among the ML models tested, CatBoost achieved the highest performance with an accuracy, sensitivity, and specificity of 93.3, 92.7, and 93.9%. Statistical methods separated these two groups with an accuracy, a sensitivity, and a specificity of 87.4, 94.8, and 80.4%.

CONCLUSION

Machine learning models, when applied to SPV profiles, exceeded statistical methods and may prove useful when developing video-nystagmography-based diagnostic aids for non-experts.