Abstract
Gait impairments are among the most prevalent and disabling symptoms in Parkinson's Disease (PD), featuring complex and highly heterogeneous manifestations. Here, we propose a deep learning-based framework to assess gait impairments using smartphone-recorded videos. This framework demonstrated high proficiency in predicting PD severity, with a micro-average area under the receiver operating characteristic curve (AUC) of 0.87 and an F1 score of 0.806, comparable to the average performance of three clinical specialists. Additionally, it effectively discerned the comprehensive efficacy of medications on gait impairments with a precision of 73.68%. In particular, it demonstrated the ability to discriminate medication-induced fine-granular gait changes beyond the resolution of the Unified Parkinson's Disease Rating Scale (UPDRS). Furthermore, our interpretable framework enabled the extraction of traditional clinically used motion markers and the discovery of novel digital biomarkers sensitive to disease progression and medication response. The findings underscore its great potential for efficiently assessing disease progression in both clinical and home settings, as well as evaluating disease-modifying effects in clinical trials to promote personalized therapies.