Remote AI Screening for Parkinson's Disease: A Multimodal, Cross-Setting Validation Study
Remote AI Screening for Parkinson's Disease: A Multimodal, Cross-Setting Validation Study
Where did the research take place?
The study site has not been established. Author addresses may differ from where the research occurred.
Publication status: preprint
A plain-language reading has not been prepared for this paper yet.
Original abstract
Abstract PARK is a web-based artificial intelligence (AI) tool for remote screening of Parkinson's disease (PD) using video and audio recordings of speech, facial expression, and motor tasks performed via webcam. The study draws on data from 1,865 participants across diverse global demographics and recording environments, including supervised clinical settings and unsupervised home use. On three independent test sets (n=389; 188 with PD), PARK achieved strong predictive performance on classifying individuals with and without PD, with accuracy ranging from 80.2% to 80.6% and area under the receiver operating characteristic curve (AUROC) from 0.85 to 0.87. When evaluated on 30 randomly selected individuals, PARK's assessments were 83.7% accurate and showed high agreement with three movement disorder specialists. The model generalized well across age, sex, and ethnic subgroups and incorporated mechanisms to withhold uncertain predictions to support safe use in unsupervised settings. Usability studies confirmed high participant satisfaction and preference for remote screening. These findings support the potential of PARK as an accessible, scalable, and clinically aligned tool to identify individuals with PD when access to traditional healthcare settings is scarce.