Abstract
Detecting subtle balance and coordination impairments in individuals with vestibular schwannomas (VS) remains a challenge, particularly when traditional clinical assessments appear normal. Focusing on patients prior to surgical intervention, this study offers a unique opportunity to capture early motor adaptations before permanent vestibular loss. We developed a deep learning-based classifier to distinguish VS patients from age-matched healthy controls using kinematic data collected during standardized gait tasks. Participants performed a short-duration (< 10 s) straight-path walk at a normal pace (an item from the Functional Gait Assessment) and a 30-second walk with intermittent eye closure. Six inertial measurement units (IMUs) were placed at different body locations, and models were trained using data from each sensor independently. We employed a convolutional neural network (CNN) tailored to this clinical application. The classifier achieved a maximum accuracy of 0.74 for controls and 0.71 for VS patients. Notably, the model detected early-stage compensatory movement patterns-even in the absence of clinical score differences-by extracting features from wrist and trunk sensors, regions not emphasized in standard evaluations. Performance improved with dataset size, particularly with more subjects, highlighting the value of broader data collection in clinical machine learning. Pretraining on external datasets that included both healthy and pathological subjects further improved accuracy, underscoring the importance of dataset diversity. These findings demonstrate the potential of deep learning to detect functional impairment when clinical scoring falls short and provide practical guidance for sensor placement, dataset design, and model training in vestibular and balance disorders.