Objective: This study aimed to classify Parkinson’s disease (PD) severity subtypes by integrating objective multimodal data with machine learning (ML) techniques. We applied unsupervised clustering to differentiate PD subtypes, assessed associations with Unified Parkinson’s Disease Rating Scale (UPDRS) scores, and identified digital biomarkers using logistic regression models.
Background: Classifying and predicting PD severity based on observable motor symptoms is challenging due to the heterogeneity of PD, varying severity levels, and the limitations of traditional clinical assessments. A data-driven approach leveraging multimodal data and ML can enhance the accuracy of PD subtype classification. [1].
Method: We analyzed multimodal data from 102 individuals with PD, including clinical characteristics, physical function, lifestyle factors, and gait parameters collected from motion analysis systems and wearable sensors. Clustering was performed using the K-means algorithm, and mutual information analysis evaluated associations with UPDRS scores. Key features were identified through logistic regression with recursive feature elimination to assess classification performance.
Results: Three PD severity subtypes were identified, ranging from mild (Cluster 1) to severe (Cluster 3). Significant associations emerged between multimodal data and UPDRS scores, with gait parameters from wearable sensors demonstrating the highest predictive value. A model utilizing bilateral ankle gait signals achieved perfect classification (AUC = 1.0), effectively distinguishing between mild and severe PD subtypes [Figure 1]. These findings suggest that gait asymmetry is a critical marker of PD severity.
Conclusion: Sensor-based gait features hold significant potential as digital biomarkers for PD severity classification, disease progression monitoring, and personalized treatment strategies [2].
Figure1
References: [1] Birkenbihl, C., Ahmad, A., Massat, N. J., Raschka, T., Avbersek, A., Downey, P., … & Fröhlich, H. (2023). Artificial intelligence-based clustering and characterization of Parkinson’s disease trajectories. Scientific Reports, 13(1), 2897.
[2] Sotirakis, C., Su, Z., Brzezicki, M. A., Conway, N., Tarassenko, L., FitzGerald, J. J., & Antoniades, C. A. (2023). Identification of motor progression in Parkinson’s disease using wearable sensors and machine learning. npj Parkinson’s Disease, 9(1), 142.
To cite this abstract in AMA style:
H. Park, C. Youm, H. Choi, J. Hwang, M. Kim. Clustering and Identification of Parkinson’s Disease Severity Subtypes Using Multimodal Data and Machine Learning Approaches [abstract]. Mov Disord. 2025; 40 (suppl 1). https://www.mdsabstracts.org/abstract/clustering-and-identification-of-parkinsons-disease-severity-subtypes-using-multimodal-data-and-machine-learning-approaches/. Accessed October 5, 2025.« Back to 2025 International Congress
MDS Abstracts - https://www.mdsabstracts.org/abstract/clustering-and-identification-of-parkinsons-disease-severity-subtypes-using-multimodal-data-and-machine-learning-approaches/