Wearable sensor–derived gait signatures for Parkinson's disease: an exploratory diagnostic study integrating dimensionality reduction and machine learning

BackgroundParkinson's disease (PD) is characterized by progressive motor dysfunction, in which gait disturbances often emerge early yet remain underutilized in diagnostic frameworks. Advances in wearable sensor technology enable high-resolution quantification of gait dynamics, but the clinical translation of such high-dimensional data is limited by redundancy and model overfitting.ObjectiveTo determine the exploratory diagnostic efficacy of synthesizing wearable sensor-derived gait parameters, cognitive scores, and blood biomarkers utilizing a dual-stage dimensionality reduction framework coupled with logistic regression machine learning for the precise identification of PD.MethodsIn this cross-sectional study, demographic characteristics, blood biomarkers, cognitive assessments, and wearable sensor–derived gait parameters were collected from patients with PD and matched controls. Following initial group comparisons, least absolute shrinkage and selection operator (LASSO) regression was applied to identify candidate predictors from multimodal variables. Principal component analysis (PCA) was subsequently performed to reduce dimensionality and extract latent structures. Logistic regression models were constructed using both the original LASSO-selected variables and PCA-derived components. Model performance was assessed using 10-fold cross-validation, with discrimination quantified by the area under the receiver operating characteristic curve (AUC) and accuracy.ResultsBaseline characteristics were comparable between groups except for age. LASSO identified 22 potential predictors spanning demographic, biochemical, cognitive, and gait domains. PCA further condensed these variables into six principal components, explaining 62.983% of the total variance. Using a rigorous nested cross-validation framework on the primary cohort, the PCA-based model achieved an internal mean AUC of 0.949 with an accuracy of 0.827. Furthermore, when evaluated on a completely independent external cohort (N = 36) to ensure generalizability, the finalized PCA-based model demonstrated robust diagnostic performance, yielding an AUC of 0.891 and an accuracy of 0.750. Both approaches showed robust discriminative capacity, with dimensionality reduction enhancing overall efficiency and stability.ConclusionIntegrating wearable gait signatures with LASSO-PCA dimensionality reduction and logistic regression yields reliable, interpretable PD classification, with PCA-condensed features demonstrating enhanced diagnostic efficiency. These findings warrant external validation in multicentre, longitudinal cohorts to advance non-invasive, sensor-based PD screening.