AUTOMATED NYSTAGMUS DETECTION FOR BPPV USING YOLOV11s SEGMENTATION AND LSTM CLASSIFICATION

Muhammad Hafiz Azhari, Martin Ferguson-Pell, Erfan Hajiparvaneh, Jade Rosenberger, codi Issac

Keywords: Vestibular, Benign Paroxysmal Positional Vertigo, Computer Vision, long short term memory, Artificial Intelligence, Object Detection, Machine Learning

TL;DR: Automated nystagmus detection using YOLOv11s segmentation and LSTM classification achieved 82% accuracy, showing promise for faster, objective BPPV diagnosis.

Abstract: Benign Paroxysmal Positional Vertigo (BPPV) is a leading cause of vertigo, marked by abnormal eye movements (nystagmus) during diagnostic maneuvers. Diagnosis currently relies on manual clinical interpretation, which can be subjective and time-intensive. In this study, we explored machine learning approaches for automated BPPV classification using eye movement data. A total of 2,500 VNG recordings were collected at ISAAC Physiotherapy (Edmonton, Canada). From these, frame segments were annotated for nystagmus, and YOLOv11s-based segmentation was applied to extract position, velocity, and shape features, which were then converted into 100-step time series. An LSTM classifier was trained with data augmentation (Gaussian noise injection, time warping, random masking, and axis flipping) to distinguish between normal eye movement, horizontal nystagmus, and vertical nystagmus. Results demonstrate the feasibility of leveraging machine learning for automated BPPV detection and classification.

Submission Number: 51