Wearable Inertial Phenotyping of Freezing of Gait in Parkinson’s Disease: Biomechanical Signatures, Subject-Independent Classification, and Clinical Associations

This is the authors' abstract. We don't add key points for this paper.

Freezing of gait (FoG) in Parkinson’s disease is clinically heterogeneous, yet wearable-sensor studies predominantly focus on binary FoG detection. We investigated whether right-ankle inertial signals can characterize clinically meaningful FoG manifestations within a three-stage conditional framework. Accelerometer and gyroscope data from 22 participants in the FoG-STAR dataset were segmented into 1 s windows. Three subject-independent classification stages were evaluated separately: locomotor non-FoG versus FoG, kinetic versus akinetic FoG, and shuffling versus trembling. Nine interpretable biomechanical features were compared with Logistic Regression, RBF-SVM, Random Forest, and a raw temporal 1D-CNN using frozen participant-independent partitions. The CNN yielded a mean Stage-1 balanced accuracy (0.877 ± 0.034), whereas for Stage-2 kinetic–akinetic discrimination, Logistic Regression and the CNN yielded mean balanced accuracies of 0.730 ± 0.044 and 0.727 ± 0.065, respectively. Stage-3 performance was lower and more variable (Logistic Regression, 0.666 ± 0.133). Nested leave-one-subject-out analysis restricted to mixed-phenotype participants further highlighted subject-level heterogeneity. Kinetic FoG showed greater residual movement amplitude and movement-rate characteristics than akinesia. Protocol-observed locomotor FoG burden was positively associated with FoG-Q (ρ=0.536) and MDS-UPDRS III (ρ=0.785). Wearable inertial signals therefore showed clinically interpretable differences among FoG manifestations, while finer phenotype discrimination remained heterogeneous and requires external validation in independent cohorts.