Using machine learning to identify Parkinson's disease severity subtypes with multimodal data

Abstract

Background: Classifying and predicting Parkinson's disease (PD) is challenging because of its diverse subtypes based on severity levels. Currently, identifying objective biomarkers associated with disease severity that can distinguish PD subtypes in clinical trials is necessary. This study aims to address the clinical applicability and heterogeneity of PD using PD severity subtypes classification and digital biomarker development by combining objective multimodal data with machine learning (ML) approaches.

Methods: We analyzed datasets that combine clinical characteristics, physical function and lifestyle data, gait parameters in motion analysis systems, and wearable sensors collected from persons with PD (n = 102) to perform clustering for subtype classification.

Results: We identified three PD severity subtypes, each exhibiting different patterns of clinical severity, with the severity increasing as it progressed from clusters 1 to 3. We found significant mutual information between all/single modalities and the unified PD rating scale scores, identifying potential modalities with high feature importance using ML. Among all modalities, the principal components of gait parameters derived from wearable sensors were identified as the most associated indicators of PD severity. A model utilizing the first principal component of the left and right ankle achieved perfect classification with an area under the curve of 1.0, accurately distinguishing clinically severe subtypes from mild subtypes of PD. These findings suggest that gait features in both ankles can reflect asymmetry factors associated with PD severity subtypes, which contributes to high classification performance.

Conclusions: Digital biomarkers obtained from wearable sensors attached bilaterally to body segments demonstrate potential for classifying PD severity subtypes and tracking disease progression. Our findings emphasized the clinical value of sensor-based gait analysis in PD management, which suggested its integration into personalized monitoring systems and therapeutic interventions for persons with PD.

Keywords: Clustering; Digital biomarker; Machine learning; Multimodal data; Parkinson's disease; Severity subtype.

Fig. 1

Workflow to classify PD severity subtypes and develop digital biomarkers

Fig. 2

Heatmap of feature importance based on multimodal data with demographics, PFL, and gait parameters. The relative importance score of each feature was determined using feature importances from the random forest or LASSO model. The random forest attribute “feature_importances_” enabled determining the contribution of each feature toward reducing the impurity of nodes in the forest. Higher values indicated more important features. The LASSO attribute “coef_” represented the coefficients of the linear model, where larger coefficients (absolute values) indicated more important features. PFL, physical function and lifestyle; GP_Motion, gait parameters' principal components (PCs) in the motion analysis system; GP_Sensors, gait parameters' PCs in wearable sensors; and UPDRS, unified Parkinson’s disease rating scale

Fig. 3

Identified digital biomarkers associated with MDS-UPDRS scores for distinguishing three PD severity subtypes. Representative variables were selected based on logistic regression results. The AUC for each ROC curve was calculated, and the mean and standard deviation of the AUCs were obtained