Learning Ordinal Latent Representations for Subject-Independent Parkinsonian Rigidity Estimation

Quantifying Parkinsonian rigidity from wearable sensors is hindered by high subject variability and the tendency of deep learning models to ignore the ordinal nature of disease severity, producing entangled latent representations and unstable subject-independent predictions. This work proposes a hybrid architecture that enforces ordinal geometry in the latent space, augmenting a CORAL ordinal regression head with center-loss regularization to map increasing severity onto distinct, ordered regions of the embedding. The model was trained on IMU data from 147 individuals in the WearGait-PD dataset and compared against a CORAL-only baseline under strict Leave-One-Subject-Out (LOSO) and Group K-Fold cross-validation. Under LOSO, the hybrid achieved mean Quadratic Weighted Kappa (QWK) $\kappa=0.18$ (best fold $\kappa=0.42$); under Group K-Fold, mean QWK reached $\kappa=0.24$ with no catastrophic failures. Subject-independent estimation remains inherently difficult, yet the hybrid substantially improved worst-case reliability: negative-Kappa folds fell from 8 to 5, and worst-case Kappa improved by 64\% ($-0.093 \rightarrow -0.033$). It also established a strong, stable monotonic relationship between embedding magnitude and severity (Spearman $\rho=0.67 \pm 0.09$, versus a near-zero, unstable baseline $\rho=0.10 \pm 0.56$), and improved the latent separation ratio by 30\%. Warm-up trajectory analysis across all five Group K-Fold folds confirms this dynamic is systematic rather than idiosyncratic. These results show that aligning internal model geometry with the ordinal structure of disease severity improves prediction stability and clinical safety. Whether equivalent structure can be induced from reduced-sensor configurations suited to resource-constrained settings remains an open question for future work.