Research map / 3D vision / 3D generation
Care-PD: A Multi-Site Anonymized Clinical Dataset for Parkinson’s Disease Gait Assessment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Care-PD, a multi-site anonymized SMPL-mesh gait dataset for Parkinson's disease from 9 cohorts, and benchmarks encoders and gait features on UPDRS-gait severity prediction and motion pretext tasks.
Problem
Objective PD gait assessment is limited by small, single-site, unstandardized motion datasets lacking clinical labels; there is no large, public, clinically annotated 3D gait dataset for PD.
Approach
Harmonize RGB video and optical motion-capture gait recordings from 9 PD cohorts into anonymized 30 Hz SMPL mesh sequences, and benchmark frozen motion encoders with lightweight probes against handcrafted gait features under within-dataset, cross-dataset, LODO and MIDA protocols, plus pretraining/fine-tuning on the data for 2D-to-3D lifting and 3D reconstruction.
Key results
- Pretraining/fine-tuning MotionAGFormer on Care-PD reduces MPJPE from 60.7 to 7.5 mm and raises UPDRS-gait F1 from 48.1 to 65.1.
- MoMask fine-tuned on Care-PD improves MPJPE from 22.5 to 8.7 mm and F1 from 41.4 to 62.7.
- Under multi-dataset in-domain adaptation (MIDA), most backbones exceed 0.69 macro-F1 on BMClab, with MixSTE reaching 0.78 versus 0.67 in within-dataset LOSO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune MotionAGFormer on Care-PD instead of evaluating zero-shot | MotionAGFormer pretrained on H3.6M · Care-PD pooled test splits | MPJPE: 60.7 -> 7.5 | Helped |
| Fine-tune MotionAGFormer on Care-PD instead of evaluating zero-shot | MotionAGFormer pretrained on H3.6M · Care-PD pooled test splits | F1-score: 48.1 -> 65.1 | Helped |
| Fine-tune MoMask on Care-PD instead of evaluating zero-shot | MoMask pretrained on HumanML3D · Care-PD pooled test splits | MPJPE: 22.5 -> 8.7 | Helped |
| Fine-tune MoMask on Care-PD instead of evaluating zero-shot | MoMask pretrained on HumanML3D · Care-PD pooled test splits | F1-score: 41.4 -> 62.7 | Helped |
| Use Care-PD rather than healthy-gait fine-tuning (ablation) | MotionAGFormer pretrained on H3.6M · Care-PD pooled test splits | MPJPE: 29.8 -> 7.5 | Helped |
| Use Care-PD rather than healthy-gait fine-tuning (ablation) | MoMask pretrained on HumanML3D · Care-PD pooled test splits | F1-score: 40.6 -> 62.7 | Helped |
| Add external H3.6M pretraining before training on Care-PD (ablation) | MotionAGFormer · Care-PD pooled test splits | MPJPE: 9.0 -> 7.5 | Helped |
| Add external HumanML3D pretraining before training on Care-PD (ablation) | MoMask · Care-PD pooled test splits | MPJPE: 9.6 -> 8.7 | Helped |
| Add multi-dataset source training and target-cohort fine-tuning (MIDA) rather than LODO | MixSTE · BMClab | Macro-F1: 0.67 to 0.78 | Helped |
| Evaluate without rare UPDRS-gait class 3 (ablation) | Encoder-based severity models · Within- and cross-dataset evaluation | Macro-F1 | Helped |
Surprising findings
- Fine-tuning on healthy gait gives much smaller gains than fine-tuning on pathological Care-PD gait, indicating the improvement stems from pathological kinematics rather than walking alone.
- The handcrafted-feature Random Forest baseline remains competitive on T-SDU-PD in the 3-class setup (0.49) but deteriorates on other cohorts.
Limitations
- Cross-dataset performance drops substantially (0.2 to 0.4 F1), indicating persistent domain gaps.
- The 3DGait cohort is small (90 walks, 43 participants), making LOSO evaluation unstable and results on it low.
- UPDRS-gait score 3 is rare and appears in only two datasets, limiting evaluation of severe impairment.
Experimental setup
Datasets: PD-GaM BMClab T-SDU-PD 3DGait KUL-DT-T DNE E-LC T-SDU T-LTC H3.6M HumanML3D TOAGA
Benchmarks: Care-PD
Metrics: macro F1 (labels 0-3) macro F1 (labels 0-2) MPJPE PA-MPJPE acceleration error
Baselines: Random Forest on handcrafted gait features POTR MixSTE PoseFormerV2 MotionBERT MotionCLIP
Builds on: SMPL WHAM SparseFusion MotionAGFormer MoMask
Closest papers by meaning
- TULIP: Multi-camera 3D Precision Assessment of Parkinson's Disease (CVPR 2024 · similarity 0.90)
- Accessible, At-Home Detection of Parkinson’s Disease via Multi-Task Video Analysis (AAAI 2025 · similarity 0.87)
- Breaking Data Silos in Parkinson’s Disease Diagnosis: An Adaptive Federated Learning Approach for Privacy-Preserving Facial Expression Analysis (AAAI 2025 · similarity 0.85)
- Cross-Covariate Gait Recognition: A Benchmark (AAAI 2024 · similarity 0.84)
- ReMAP: Neural Model Reprogramming with Network Inversion and Retrieval-Augmented Mapping for Adaptive Motion Forecasting (NeurIPS 2024 · similarity 0.83)
- VidSole: A Multimodal Dataset for Joint Kinetics Quantification and Disease Detection with Deep Learning (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.