Accessible, At-Home Detection of Parkinson's Disease via Multi-task Video Analysis

TL;DR

Parkinson's disease often goes undiagnosed due to limited access to care. The authors present a scalable at-home PD screening framework leveraging three remote tasks (finger-tapping, smile, and pangram speech) and a large multi-task video dataset. A novel Uncertainty-calibrated Fusion Network (UFNet) fuses task-specific features through calibrated self-attention that accounts for per-task uncertainty, while enabling withholding of uncertain predictions to protect patients. The approach yields high accuracy and AUROC, outperforms single-task and baseline fusion methods, and demonstrates robustness across demographics with potential for broad, resource-efficient deployment. The work also provides a substantial, privacy-conscious dataset of extracted features to spur further research in remote PD screening and health AI.

Abstract

Limited accessibility to neurological care leads to underdiagnosed Parkinson's Disease (PD), preventing early intervention. Existing AI-based PD detection methods primarily focus on unimodal analysis of motor or speech tasks, overlooking the multifaceted nature of the disease. To address this, we introduce a large-scale, multi-task video dataset consisting of 1102 sessions (each containing videos of finger tapping, facial expression, and speech tasks captured via webcam) from 845 participants (272 with PD). We propose a novel Uncertainty-calibrated Fusion Network (UFNet) that leverages this multimodal data to enhance diagnostic accuracy. UFNet employs independent task-specific networks, trained with Monte Carlo Dropout for uncertainty quantification, followed by self-attended fusion of features, with attention weights dynamically adjusted based on task-specific uncertainties. To ensure patient-centered evaluation, the participants were randomly split into three sets: 60% for training, 20% for model selection, and 20% for final performance evaluation. UFNet significantly outperformed single-task models in terms of accuracy, area under the ROC curve (AUROC), and sensitivity while maintaining non-inferior specificity. Withholding uncertain predictions further boosted the performance, achieving 88.0+-0.3%$ accuracy, 93.0+-0.2% AUROC, 79.3+-0.9% sensitivity, and 92.6+-0.3% specificity, at the expense of not being able to predict for 2.3+-0.3% data (+- denotes 95% confidence interval). Further analysis suggests that the trained model does not exhibit any detectable bias across sex and ethnic subgroups and is most effective for individuals aged between 50 and 80. Requiring only a webcam and microphone, our approach facilitates accessible home-based PD screening, especially in regions with limited healthcare resources.

Figures (3)

• Figure 1: An illustrative overview of the proposed model. Task-specific models are shallow neural networks trained with task-specific features and MC dropout. Standard deviations in multiple rounds of inference obtained from the task-specific models are used to calibrate the attention scores when fusing task-specific features. Finally, another shallow neural network is trained to differentiate individuals with and without PD and withhold prediction when the model is uncertain.
• Figure 2: Misclassification rate of the best UFNet model across demographic subgroups. The analysis is done on the test set participants. Error bars demonstrate 95% CIs.
• Figure 3: Effect of decision threshold. For this plot, we used the best performing UFNet model with 30 different random seeds. Shaded regions show the 95% confidence intervals.