Validation of remote multimodal AI screening for Parkinson disease across diverse settings

Timely detection of Parkinson's disease (PD) remains limited by reliance on in-person neurological evaluations that are often costly and geographically inaccessible. To address these barriers, we develop PARK (Parkinson's Analysis with Remote Kinetic-tasks) - a web-based artificial intelligence (AI) tool that screens for PD using short webcam recordings of facial expression, motor, and speech tasks.

Across eight independent studies (n = 1,865 participants; 670 with PD), participants completed three standardized tasks (smile mimicry, finger tapping, and pangram utterance) via webcam. Task-specific neural networks estimate PD risk and uncertainty, which are integrated through an uncertainty-calibrated fusion model (UFNet). Model performance is evaluated on one internal and two external test sets representing supervised and unsupervised real-world environments. Three movement disorder specialists also reviewed videos from 30 participants to benchmark clinical agreement of the PARK tool. User experience is assessed through structured surveys containing open-ended or multiple-choice questions.

PARK achieves accuracies of 80.2-80.6% and AUROC of 0.85-0.87 across all evaluation cohorts, with 83.3-86.5% sensitivity and 71.2-78.4% specificity. Predictive performance remains stable across sex, age, and ethnicity. Agreement with clinician judgments reaches Cohen's κ = 0.59. Uncertainty estimates reflect diagnostic confidence, and performance declines at high-uncertainty levels. Usability is rated highly (System Usability Scale  > 70) in both supervised and unsupervised settings, with low perceived risk and strong user preference for remote screening.

PARK demonstrates promising accuracy and favorable user acceptance for remote PD screening, highlighting its potential as an accessible, equitable, and uncertainty-aware tool for neurological assessment when traditional care is challenging to obtain.