Label-Free Parkinson's Disease Screening from Face and Voice through Mechanistic Interpretability

2026-08-09

AI Summary

This study addresses the scarcity of labeled data and privacy constraints in Parkinson’s disease screening by proposing an unsupervised, cross-modal framework that operates without disease labels. The approach leverages frozen pre-trained Vision Transformer (for facial expressions) and HuBERT (for speech) models, guided by synthetic dysarthria-induced contrastive activation directions and alignment principles to enable interpretable anomaly detection. It fuses k-nearest neighbor anomaly scores with Contrastive Activation Addition (CAA) for enhanced performance. Evaluated on the YouTubePD benchmark, the fused model achieves an AUROC of 0.802 and a negative predictive value of 0.95, substantially outperforming unimodal baselines and demonstrating both methodological efficacy and clinical promise.

Abstract

Parkinson's disease (PD) is the second most common neurodegenerative disorder. Typical machine learning screening methods require PD labels, but the available data is limited by privacy concerns and the need for expert annotation. We propose a label-free face-plus-voice PD screen built entirely on frozen pretrained encoders--a face-expression Vision Transformer and HuBERT--in which no PD label touches any fit; the reference is training controls only. The voice modality uses a synthetic-dysarthria contrastive activation addition (CAA) direction built from time-stretch and breathy degradation of healthy speech; the face modality uses a k-nearest-neighbor anomaly score to the control embedding cluster. We introduce the alignment principle, a post-hoc analysis showing that a synthetic-degradation CAA detector works when the cosine similarity between the synthetic and real disease directions exceeds zero. Measured on the YouTubePD benchmark, this cosine is +0.37 for voice (CAA works, AUROC 0.765) and -0.48 for face (CAA fails; anomaly succeeds, AUROC 0.751). Equal-weight late fusion reaches AUROC 0.802 (95% CI [0.70,0.89]) with NPV 0.95, supporting a rule-out triage interpretation. An overfitting audit shows the voice detector transfers cleanly, while the face-side--and thus fused--AUROC is potentially optimistic pending external validation.

Problem

• Parkinson's disease
• label-free screening
• face and voice
• neurodegenerative disorder
• privacy constraints

Innovation

• label-free screening
• contrastive activation addition
• mechanistic interpretability
• frozen pretrained encoders
• synthetic dysarthria