Automated limb motor assessment in Parkinson’s disease via a time-frequency state-space model

Assessment of limb motor function is central to the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS), but current practice relies on subjective visual inspection. Existing deep learning approaches remain limited in characterizing the non-stationary and nonlinear dynamics of pathological extremity movement, including bradykinesia and intermittent tremor. We developed TF-Mamba, a time-frequency dynamics-aware framework for automated and standardized scoring of finger tapping, hand movements, and leg agility based on expert-consensus MDS-UPDRS ratings. The framework integrates a spatiotemporal feature backbone with a physical signal dynamics module that decomposes nonlinear joint trajectories in the time-frequency domain. The model was trained using data from 1479 patients with Parkinson’s disease at four Chinese centers and independently validated in 450 patients from three centers. For five-level severity scoring, TF-Mamba achieved an AUC of up to 0.90 in external validation and outperformed state-of-the-art baselines by 8–10%. Energy Entropy, Amplitude Coefficient of Variation, and Frequency Stability were associated with clinical severity (all p < 0.001). In a reader study, the model reduced evaluation time by 60–70% and improved the agreement of junior physicians’ ratings with expert consensus. Collectively, TF-Mamba provides an interpretable and scalable framework for MDS-UPDRS limb motor assessment.