To evaluate the ability of nystagmus velocity profiles to differentiate between BPV from BPV mimics by using both inferential statistics and machine-learning models

BPV is common, correctable and often undertreated especially in ER and primary care. Differentiating BPV from other forms of positional nystagmus (central positional nystagmus of vestibular migraine) can be challenging. We hypothesized that video nystagmography coupled with nystagmus slow-phase velocity (SPV) profiling can help correct identification of BPV nystagmus.  

We used 476 expert-tagged BPV and non BPV central positional nystagmus videos to generate 2D eye position traces. Horizontal and vertical nystagmus slow phase velocities were plotted as a function of time to create SPV profiles. Nystagmus onset latency, 50% rise time, peak SPV, peak latency, time to decay by 50 and 95% (T50 and T95) were measured.  Machine learning algorithms were applied to separate BPV from non BPV nystagmus

The best metric for separating BPV and non-BPV positional nystagmus was the 95% nystagmus decay time (T95) which identified BPV nystagmus with an accuracy, sensitivity and specificity of 87.4%, 94.8% and 80.4%. When we applied supervised machine learning algorithms, to the nystagmus velocity profile, the best performing model (CatBoost) achieved an accuracy, sensitivity and specificity of 93.3, 92.7 and 93.9%.

Nystagmus slow-phase velocity profiles help differentiate BPV from its mimics. Machine learning algorithms further improve the accuracy of SPV profile-based BPV detection. These algorithms could be incorporated into nystagmus recording systems to assist non-expert physicians to correctly identify and treat BPV in the front-line.