TL;DR
This study evaluates the effectiveness of several popular pose-estimation libraries (like OpenPose and MediaPipe) in measuring joint range of motion for patients with suspected Ehlers-Danlos Syndrome (EDS). The researchers found that while current models are promising for screening large joints like elbows and shoulders, specialized "finetuned" models are necessary to accurately track smaller, hypermobile joints like the fifth finger.
What Problem Does This Paper Address?
Diagnosing Ehlers-Danlos Syndrome (EDS) and Generalized Hypermobility Spectrum Disorders (G-HSD) relies heavily on the Beighton exam—a series of maneuvers to test if joints move beyond the typical range of motion. However, manual assessment using a goniometer (a physical protractor for joints) is notoriously inconsistent between different clinicians, especially in non-specialized primary care settings.
While computer vision (CV) and pose estimation offer a path toward objective, automated screening, there is a major hurdle: most AI models are trained on datasets of people performing "normal" activities. Hypermobile joints represent "edge cases" that these models rarely see during training. This paper investigates whether off-the-shelf CV libraries can actually handle the extreme angles found in this clinical population.
Methods
The researchers recorded 97 adults at a specialized EDS clinic performing various joint movements using a standard mobile phone. They then compared three measurement methods:
- Clinical Gold Standard: Manual measurements taken by a clinician using a digital goniometer.
- Manual Video Annotation: Researchers manually clicked on joint keypoints in video frames to establish the "best possible" CV performance.
- Automated Pose Estimation: Five body-tracking libraries (AlphaPose, Detectron, MediaPipe-Pose, MoveNet-Thunder, OpenPose) and two hand-tracking libraries (AlphaPose, MediaPipe-Hands) were used to automatically calculate joint angles.
The "Secret Sauce": Because standard hand-tracking libraries performed poorly on the fifth finger, the team developed a custom domain-specific model. They used the DeepLabCut framework to finetune a MobileNet-v2 backbone specifically on their hypermobile dataset.
Datasets
- Clinical Dataset: 97 adults (89 female, mean age 35) with suspected EDS/G-HSD. This provided a unique benchmark of extreme joint angles.
- Pre-training Data: The open-source libraries used standard benchmarks like COCO (Common Objects in Context) and Halpe-Full Body.
- Custom Training: For the fifth-finger model, the researchers used an 80/20 train-test split of their own manually annotated video data.
Key Insights and Ideas
- The Generalization Gap is Real: Standard models struggle with hypermobility. It’s not just that the joints move "more"; it's that the visual configuration of a hyperextended limb often confuses models trained on typical human geometry.
- No "One Library to Rule Them All": Performance varied significantly by joint. Detectron was the top performer for elbows, while MoveNet-Thunder (a model designed for speed and mobile use) actually proved most robust for shoulders and knees.
- Small Joints Require Specialized Models: Off-the-shelf hand libraries were almost useless for measuring fifth-finger hyperextension. However, the finetuned custom model achieved a high correlation ($rho = 0.786$) with clinical measurements, proving that domain-specific training is the key to clinical utility.
- Framing Matters More Than You Think: The ankle measurements failed across the board. The researchers realized that by framing the whole body to help the AI "see" the person, they lost the detail and the specific anatomical plane needed to measure ankle dorsiflexion accurately.
Key Results
The study used Spearman’s correlation ($rho$) to compare automated results to clinical goniometry:
- Elbow: Detectron led with $rho = 0.722$.
- Shoulder & Knee: MoveNet-Thunder performed best ($rho = 0.632$ and $0.608$, respectively).
- Fifth Finger: Standard libraries failed ($rho \approx 0.2$), but the finetuned custom model hit $rho = 0.786$.
- Ankle: No significant correlation was found for any automated library, highlighting a need for better recording protocols.
Limitations and Future Work
The primary limitation was the 2D nature of the video. Pose estimation is sensitive to the camera angle; if the limb isn't perfectly parallel to the lens, the calculated angle will be off. While the team tried 3D pose estimation, preliminary results were poor, likely due to the lack of 3D training data for hypermobile poses.
Future Directions: * Ankle-Specific Protocols: Moving from "whole-body" framing to "lower-leg-only" videos to improve ankle tracking. * Screening Thresholds: The next step is to move from measuring angles to classifying whether a joint is "hypermobile" or "normal" to assist in clinical triage. * Diversity: The study population was 92% female. Future work needs to validate these models across a more diverse range of body types and genders.
Summary updated