Objective: To investigate whether deep neural networks can learn meaningful representations of gait in Parkinson’s disease (PD), and to probe whether these representations group PD subjects into distinct gait classes.

Background: PD patients often differ in their disease trajectories and exhibit heterogeneous responses to pharmacological and surgical treatments. If disease progression or treatment response patterns were associated with specific motor symptoms, one could develop new strategies for personalized medicine. Among motor symptoms, PD patients can exhibit different gait characteristics, not all of which are captured by the MDS-UPDRS scale. We hypothesized that PD gait patterns could be grouped into classes, which we tested using a machine learning approach.

Method: We utilized a dataset of walk test videos collected from 15 subjects and scored for gait impairment by clinicians. The videos were collected from multiple angles, allowing us to track over time the 3D positions of 26 body joints on each subject. These 3D pose traces were segmented into individual stride sequences and then normalized to 100 frames using linear interpolation. A variational autoencoder (VAE) network was trained to learn, without human input, 128 spatiotemporal features that best summarized the gait data across the cohort. These features were then input for a random forest classifier to distinguish clinician-assessed healthy from impaired gait profiles.

Results: The VAE was trained on 389 stride samples from all 15 subjects. We visualized the extracted 128 features using a dimensionality reduction method (t-SNE). We observed healthy subjects and PD subjects were well separated [Fig. 1a] and multiple subjects clustered together, providing preliminary evidence for the existence of gait classes [Fig. 1b]. The extracted features were then used to train a random forest classifier to predict healthy vs. PD gait, achieving 80% classification accuracy, surpassing the 73% accuracy reported by traditional feature-based machine learning methods [1]. In the future, these results should be validated over a larger PD dataset.

Conclusion: Our pilot results demonstrate the potential for future explorations of PD gait classes and any associations they might have with disease and treatment properties.

Figure 1

References: [1]. Kim K, Lyu S, Mantri S, Dunn TW. TULIP: Multi-camera 3D Precision Assessment of Parkinson’s Disease. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2024.

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/unsupervised-analysis-and-clustering-of-3d-parkinsonian-gait/