Toward Robust Patient-Level Parkinson’s Disease Classification: Evaluating Stability Across Participant Compositions Using Wearable-Sensor and Questionnaire Information

Open Access
Article
Conference Proceedings
Authors: Ruka Hashimoto, Masaomi Kimura
Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized byheterogeneous motor symptoms that vary across individuals and over time. Wearableinertial sensors provide a low-burden means of capturing motor characteristics, but reliableclassification requires consideration of performance variability across participantcompositions. Building on our previous temporal–frequency integrated deep learningframework, this study evaluated the robustness of patient-level PD classification andexamined whether complementary questionnaire information improves classificationstability. Using the Parkinson’s Disease Smartwatch dataset, participants with PD andhealthy controls (HC) were evaluated by stratified five-fold participant-level cross-validation.Movement signals were classified using a model integrating temporal and wavelet-basedfrequency features, and segment-level probabilities were averaged to obtain patient-levelpredictions. A 30-item non-motor symptom questionnaire was independently classifiedusing logistic regression, and movement-based and questionnaire-based outputs wereintegrated with equal weighting to obtain probabilities of PD. The movement-based modelachieved an accuracy of 0.854±0.029, while the questionnaire-based classifier achieved0.864±0.020. Integration increased accuracy to 0.927±0.027. In particular, Recall(HC)improved from 0.791±0.107 for movement-only classification to 0.896±0.036 afterintegration, indicating reduced fold-level variability, while the change in PD classificationperformance was comparatively modest. Evaluations using alternative participant partitionseeds showed similar overall tendencies without substantial increases in variability. Theseresults suggest that integrating complementary motor and non-motor information canimprove both classification performance and stability across participant compositions,supporting more robust patient-level PD classification.

Keywords: Parkinson’s Disease, Wearable Sensors, Deep Learning, Patient-Level Classification, Multimodal Integration

DOI: 10.54941/ahfe1008440

Cite this paper
Downloads
0
Visits
1
Download PDF

More from this volume

← A Task Selection Method for Alzheimer's Disease Patient Prediction to Reduce Patients’ BurdenArtificial Intelligence-Assisted Rapid On-site Evaluation: A Human Factors-Informed Patient Safety Framework for Cytopathology →
View all articles in Human Factors in Design, Engineering, and Computing →