Background: Gait impairments are among the most prevalent and disabling symptoms in Parkinson's Disease (PD), featuring complex and highly heterogeneous manifestations.
Methods: We propose a deep learning-based framework to assess gait impairments using smartphone-recorded videos. This framework demonstrated high proficiency in predicting PD severity.
Results: The model achieved 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%.
Parkinson's disease (PD) is the second most prevalent progressive neurodegenerative disorder, affecting more than 10 million people worldwide. Gait impairments are among the most common and disabling symptoms in PD, characterized by intricate underlying mechanisms and substantial individual variation in clinical manifestations.
Current clinical rating scales like UPDRS have low sensitivity, inherent subjectivity, and dependency on clinical specialists, limiting their utility in routine assessment of gait impairments.
Our framework uses smartphone-recorded videos and deep learning to provide objective, precise assessment of gait impairments, enabling home-based monitoring.
We extract both traditional clinically used motion markers and discover novel digital biomarkers sensitive to disease progression and medication response.
We developed a novel Siamese contrastive network architecture that efficiently extracts precise spatiotemporal motion characteristics of the entire body joints from gait videos recorded with a single smartphone.
Our architecture fuses gait videos recorded from both left and right lateral perspectives during shuttle walks, ensuring accurate identification of lateral motion characteristics.
We extract body skeletons from smartphone videos using DWPose, with modifications to correct misdetections of leg keypoints for accurate feature extraction.
Our model analyzes personalized joint impacts on gait impairment severity over walking time, enabling extraction of both traditional and novel digital biomarkers.
The model demonstrated high accuracy in predicting gait impairment severity, with performance comparable to clinical experts (F1 score: 0.806).
Our framework accurately identified medication-induced changes in gait impairments with 73.68% precision, including subtle responses undetectable by UPDRS.
We identified novel digital biomarkers sensitive to disease progression and medication response, such as linear velocities and accelerations of skeletal joints.
The model's performance was validated against the consensus of three clinical specialists, demonstrating alignment with expert assessments and the ability to handle inter-rater variability.
We developed an online assessment system (FAGI-PD) that enables home-based, objective, routine assessments of gait impairments in Parkinson's disease using smartphone-recorded videos.
Enables frequent, convenient assessment of gait impairments in home settings, reducing the need for clinical visits.
Provides immediate assessment results and trends over time for patients and clinicians.
Tracks disease progression and treatment response over time, enabling personalized therapy adjustments.
Code for data preprocessing, model training, and feature extraction available at Code Ocean
Access Code