A Bayesian Latent-Threshold Framework for Fall Occurrence in Parkinson’s Disease

Falls in Parkinson’s disease (PD) are commonly modeled as direct outcomes of isolated gait features, although wearable measurements are noisy indicators of broader motor-control processes. We propose a Bayesian latent-threshold framework in which fall occurrence is modeled as the probabilistic manifestation of an unobserved gait instability construct. In 269 individuals with Parkinson’s disease performing a 30-m walk with a single L5-mounted inertial sensor, trunk-derived features were grouped into lower-trunk kinematics, rhythmicity–recurrence, and neuromotor complexity domains. Within-domain principal component analysis provided three domain-specific axes, which were integrated in a Bayesian hierarchical measurement model to estimate subject-specific latent instability with posterior uncertainty. Fall occurrence was then modeled using a Bayesian probit threshold formulation. Latent instability showed a positive association with fall occurrence (posterior probability β > 0 = 0.964), and the inferred threshold (τ ≈ 0.32) identified a transition region where small changes in instability produced the largest changes in fall probability. Counterfactual contrasts showed baseline-dependent risk reductions, maximal near the threshold. This framework reframes wearable fall-risk modeling from feature-level prediction toward uncertainty-aware latent digital biomarker inference.