Interpretable polar-flow decomposition for torsional nystagmus lateralization in posterior-canal BPPV

Objective Torsional nystagmus during the Dix–Hallpike maneuver represents a cardinal vestibulo-ocular reflex (VOR) sign in posterior-canal benign paroxysmal positional vertigo (pc-BPPV). This study develops an interpretable framework for automated classification of torsional nystagmus direction and temporal localization of abnormal events from clinical infrared recordings. Methods Infrared Dix–Hallpike videos were processed via You Only Look Once version 8 nano (YOLOv8n) pupil localization, dynamic iris masking, and Recurrent All-Pairs Field Transforms (RAFT) optical flow. Flow vectors were projected into pupil-centered polar coordinates to extract tangential and radial motion features. A bidirectional gated recurrent unit with attention (BiGRU-attention) network performed segment-level three-class classification with sliding-window patient-level inference. Results Five-fold cross-validation yielded segment-level accuracy of 94.19% ± 0.88% with balanced F1-scores across all three classes. On an independent 45-patient long-video test set, patient-level accuracy reached 97.78% at 60 frames per second (fps) and 100% (45/45) at 30 fps. Temporal localization achieved mean intersection-over-union (IoU) of 0.899 and 0.896, with hit rates of 96.8 and 100%, respectively. Conclusion The polar-flow framework enables physiologically interpretable and robust classification of pc-BPPV-related torsional-upbeat nystagmus. By explicitly decomposing eye movements into tangential and radial components, it provides reliable lateralization while demonstrating the potential for clinically meaningful temporal localization of abnormal nystagmus events.