Attentive Prototype Learning with Wearable Sensor Mutual Information for Fall Risk Stratification of Parkinson's Patients.

Parkinson's disease (PD), with its rising global prevalence, poses severe risks from falls and motor impairments. Current fall risk assessments rely heavily on subjective clinical evaluations, underscoring the need for quantitative methods. In this exploratory study, wearable inertial and photoelectric sensors attached to the limbs and trunk were used to objectively collect biomechanical movement data during standardized MDS-UPDRS motor assessments. Leveraging the clinically validated correlation between Hoehn-Yahr (H-Y) staging and fall risk, we propose a data-driven framework to quantify risk. Mutual information (MI) analysis links biomechanical features to H-Y stages, generating a weighted Fall FRS (FRS). Machine learning validation was further performed to preliminarily evaluate the discriminative capability of the proposed FRS in stratifying patients by risk severity. Based on a cohort of 92 PD patients, experimental results on the independent test set showed that incorporation of the FRS improved classification accuracy from 50.00% to 82.14%, while the macro-average AUC increased from 0.698 to 0.907. These findings suggest that wearable sensor-based biomechanical assessment may provide useful quantitative information for exploratory fall-risk stratification in PD patients.