Contactless Upper-Limb Bradykinesia Monitoring for Parkinson's Disease via Semantic-Aware mmWave Sensing in Daily Life

Abstract

Upper-limb bradykinesia, a cardinal symptom of Parkinson's disease (PD), significantly disrupts daily activities with slowness and reduced amplitude of voluntary movement. Current clinical assessments are episodic and subjective, and fail to capture real-world symptom fluctuations, while wearable systems suffer from low long-term adherence and vision-based solutions raise privacy concerns. We introduce mmBrady, the first contactless system that enables continuous, privacy-preserving estimation of upper-limb bradykinesia in daily environments using mmWave radar. To address core challenges including coarse clinical supervision, noisy superimposed signals, and heterogeneous daily activities, we propose three key modules: (1) a decoupled strategy using cross-modal dense supervision for kinematic extractors; (2) a hierarchical framework modeling spatial joint synergy to disentangle overlapping reflections; and (3) a semantic-aware multi-instance regressor that aggregates motion-stage-specific observations to distinguish pathological slowness from normal pauses. Evaluated on 42 PD patients across 8 real-world environments (clinic, nursing home, and homes), mmBrady achieves a high correlation of 0.73 and a low MAE of 1.59 with clinician-anchored MDS-UPDRS bradykinesia scores, which is even better than state-of-the-art single-wearable solutions but with no user burden. We will release the first contactless bradykinesia dataset to advance in-home neurological healthcare. We envision that mmBrady paves the way for longitudinal, low-burden PD monitoring to support timely medication adjustments and improve patient quality of life.