Abstract Nystagmus is a key indicator of vestibular disorders, including benign paroxysmal positional vertigo (BPPV). Accurate diagnosis of BPPV is essential, as it is treatable with specific bedside maneuvers that lead to rapid symptom resolution, thereby improving patient outcomes and reducing unnecessary treatments. In clinical practice, identification of positional nystagmus relies on eliciting and interpreting eye movements during provocative maneuvers, with or without video nystagmography (VNG). This process can be subjective and difficult to standardize when signals are subtle, noisy, or temporally variable. We present DSF-BPPVNet, a delay-aware neural architecture for BPPV classification from VNG traces. The model combines temporal convolution, delayed-state feedback, and residual refinement to support classification from temporally structured eye-movement signals. The model was evaluated on 3,111 VNG traces from 705 patients using 5-fold cross-validation and compared with established deep-learning baselines. In the patient-independent setting, DSF-BPPVNet achieved the strongest overall performance among the evaluated models, with an F1-score of 0.819 ± 0.020. Explainability analyses were also performed to characterize model attribution patterns and temporal weighting behavior.
Chaturvedi et al. (Mon,) studied this question.