Machine learning-based identification of a minimal wearable sensor for gait assessment in vestibular schwannoma

Abstract

Vestibular schwannoma (VS) impairs balance and gait, often leading to substantial functional limitations that are incompletely captured by standard clinical assessments. While wearable inertial sensors and machine learning offer promise for objective gait analysis, clinical translation is limited by uncertainty regarding optimal sensor placement, task selection, and interpretability. We recorded six-dimensional kinematics from 32 individuals with unilateral vestibular schwannoma and 32 age-matched healthy controls as they performed the ten tasks of the Functional Gait Assessment (FGA). Adapting a previously developed deep learning framework, we systematically evaluated all combinations of ten gait tasks and six sensor locations, yielding 60 task–sensor models. Models were trained and tested using subject-level leave-one-out cross-validation. To support clinical interpretation, we derived a continuous kinematic score reflecting the similarity of an individual’s gait patterns to those observed in vestibular pathology. Model performance depended strongly on task and sensor selection. Gait with eyes closed (FGA8) emerged as the most informative task, and wrist-mounted sensors consistently outperformed head and trunk sensors across tasks. A wrist-worn sensor during gait with eyes closed provided discrimination between VS and control participants with overall accuracies ranging from 60 to 83%, while substantially reducing instrumentation burden. The resulting kinematic score significantly differentiated groups, and model accuracies correlated with established clinical measures including the Dizziness Handicap Inventory. These findings demonstrate that machine learning can be used not only to classify gait pathology, but to identify a minimal, clinically deployable wearable configuration for objective gait assessment in vestibular schwannoma. The proposed wrist-based kinematic score offers a continuous, interpretable metric that complements existing clinical assessments and supports longitudinal monitoring across in-clinic and remote rehabilitation settings. This work advances the development of quantitative, scalable digital biomarkers for precision rehabilitation in vestibular disorders.