Uncovering Latent Biomechanical Signatures in Parkinson's Gait using Explainable AI and Synthetic Profiling
This flagship project aims to identify, explain, and simulate latent biomechanical phenotypes in Parkinson’s gait using clustering, generative models, and explainable AI. We move beyond classical classification by exploring hidden structures in movement patterns and creating a synthetic simulator to empower clinical research, training, and screening.
To discover biomechanical phenotypes in Parkinson's patients, explain their key features, and generate realistic synthetic patient profiles using ctGAN, with potential applications in:
- Clinical education
- Disease monitoring
- Hypothesis generation
- Anomaly detection in digital gait biomarkers
| Module | Functionality | Output |
|---|---|---|
data_prep/ |
Data cleaning, feature selection, imputation, scaling | Clean dataset (no identifiers or prodromes) |
cluster_engine/ |
Clustering (KMeans, HDBSCAN, silhouette/PCA/UMAP tuning) | 3–6 biomechanical phenotypes |
xai_explain/ |
SHAP + SHAPSet visualizations per cluster | Signature plots for clinical interpretation |
gan_generator/ |
Cluster-wise ctGAN training | Realistic synthetic gait profiles |
anomaly_detector/ |
Outlier scoring for new patients | Biomechanical anomaly index |
webapp_predictor/ |
Upload/inject patient → assign cluster → show XAI + simulate via GAN | Full Streamlit interface for clinical users |
docs/ |
README, clinical guide, case studies, GIFs, deployment tips | High-impact public GitHub profile |
- Clustering: KMeans, HDBSCAN, PCA, UMAP
- Explainability: SHAP, SHAPSet plot (
shapiq), beeswarm, heatmaps - Generative Modeling:
ctGAN(via SDV library) - Web Application: Streamlit (optionally Django + React)
- Packaging: Modular Python structure under
src/parkinsonbiomech/
| Dimension | What Makes It Special |
|---|---|
| Innovation | Use of GAN to generate biomechanical gait phenotypes |
| Explainability | Transparent cluster interpretation via SHAP and SHAPSet |
| Usability | Relevant for clinicians, therapists, researchers, students |
| Open Science | Fully open-source, modular, cite-ready |
| Web-Ready | Interactive app ready for deployments, interviews, and grants |
| Reusability | Packaged logic, testable code, and clear separation of concerns |
- Finalize data cleaning + numeric conversion
- Proceed to clustering via
02_clustering.ipynb - Train ctGAN models per biomechanical cluster
- Integrate SHAP explainability for simulated patients
- Develop the interactive web interface
- Clinical researchers interested in biomechanical phenotyping
- Data scientists working on interpretable models in healthcare
- Developers building medical education tools with synthetic data
- Students seeking advanced AI + neuroscience integration projects
Open to collaborations! Want to add transfer learning? Test GAN explainability? Deploy with Docker? Get in touch or open a PR.
This project integrates explainable AI and generative AI techniques into biomechanics and gait analysis, with a focus on Parkinson's disease.