Objective: This study aims to find a digital behavioral marker Parkinson’s disease (PD) in activities of daily living via machine learning (ML) analysis of 3D body and hand poses.

Background: Understanding activities of daily living (ADL) for PD is crucial, as ADL difficulties reduce quality of life. Nevertheless, previous ADL research has focused on the MDS-UPDRS II questionnaire, which is subjective and self-reported. While recent ML models have enabled quantitative analysis of poses in PD patients, few researches have used pose analysis to study changes in ADLs. [1,2]

Method: Sequential drinking activity was recorded simultaneously using six cameras from 10 PD patients and 5 healthy subjects. We used the Three-dimensional (3D) Understanding and Learning of Impairments in Parkinson’s disease (TULIP) system from our previous study [3] to extract 2D and 3D poses from multi-camera videos. For each subject, pose sequences were segmented into drinking epochs using 3D coordinates of the write. We defined global features from the 3D body and hand poses, including angles, distances, speeds, accelerations, postural changes. Features showing a high degree of correlation with each other were condensed. A random forest model was trained using selected features to make a classification between PD and healthy subjects. Leave-One-Subject-Out (LOSO) paradigm was used for the internal validation (Figure 1), and the data recorded from the Neurology Clinic (37 PD patients) was used for the external validation.

Results: We found six statistical important features as shown in Figure 2 and Figure 3. Using these features, our model performed 86.7% of classification accuracy in 15 subjects and 83% in the external validation. (F1 score=0.9, AUC=0.87 (Figure 4))

Conclusion: Our paper proposed the explainable ML model for PD patient’s ADL analysis. 3D poses were used to quantify movements, and our work suggested a potential way to detect PD using daily activities.

Daily Activity Analysis Pipeline

3D Hand and Arm Keypoints Configuration

Important 3D Features

Classification Results

References: [1] Lawrence, Blake J., et al. “Activities of daily living, depression, and quality of life in Parkinson’s disease.” PloS one 9.7 (2014): e102294.
[2] Bouça-Machado, Raquel, et al. “Measurement tools to assess activities of daily living in patients with Parkinson’s disease: A systematic review.” Frontiers in neuroscience 16 (2022): 945398.
[3] Kim, Kyungdo, et al. “TULIP: Multi-Camera 3D Precision Assessment of Parkinson’s Disease.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/daily-activity-analysis-of-parkinsons-disease-using-machine-learning-with-3d-poses/