Introduction
Parkinson’s Disease (PD) assessment currently relies on the Unified Parkinson’s Disease Rating Scale (UPDRS), a process where clinicians observe a patient and assign scores from 0 to 4. This method is inherently subjective and prone to high inter-rater variability; disagreements between experts can be as high as 53%. While machine learning offers a path toward objective "digital biomarkers," progress has been hindered by a lack of high-quality data.
Existing PD video datasets are typically limited to single-camera (2D) recordings covering only one or two activities. Because human movement is fundamentally three-dimensional, 2D projections lose critical information due to perspective and depth ambiguities. This makes it difficult to capture the fine-scale motor pathologies characteristic of PD. The TULIP dataset addresses these gaps by providing the first multi-camera, markerless 3D video resource for high-precision clinical assessment.
The TULIP Dataset
The primary contribution of this research is the TULIP dataset (Three-dimensional Understanding and Learning of Impairments in Parkinson’s disease). TULIP significantly expands the scope and technical quality of available PD data.
As shown in the comparison above, TULIP triples the standard frame rate to 80 fps and utilizes six synchronized cameras. While previous studies focused on 1-3 activities, TULIP covers 25 different UPDRS motor tasks, including finger tapping, gait, seated stepping, and "global spontaneity." The dataset includes 29 distinct UPDRS Part III scores per subject, provided by three clinical experts.
Methodology
The researchers developed a custom multi-camera system to record 15 subjects (10 PD patients, 5 healthy controls) in a 6.3 x 3-meter arena.
To ensure clinical scalability, the team employed a markerless pose estimation pipeline rather than cumbersome wearable sensors. They used MediaPipe for hand tracking and MMPose for body tracking to extract 2D keypoints. These points were then triangulated into 3D space using a Direct Linear Transformation (DLT) algorithm, followed by interpolation and smoothing.
From these 3D trajectories, they extracted spatiotemporal kinematic features such as angular speed, amplitude, and freezing events. A suite of classifiers (SVM, Random Forest, XGBoost, and MLP) was trained using a Leave-One-Subject-Out (LOSO) validation scheme. To handle clinician variability, the researchers incorporated "bias matrices" to weight labels from the three different expert neurologists.
Results and Discussion
The study demonstrates a clear "3D advantage." In tasks like finger tapping, 3D models consistently outperformed 2D versions because they are not fooled by the angle of the hand relative to a single camera.
Key results include:
- Diagnosis Accuracy: For finger tapping, a Multi-Layer Perceptron (MLP) using 3D features achieved an F1 score of 0.87 for distinguishing PD from healthy controls.
- Score Prediction: The 3D MLP achieved an F1 of 0.69 in predicting specific 0-4 UPDRS scores, significantly higher than the 0.57 F1 achieved by 2D-based models.
- Gait Assessment: A 3D Random Forest model achieved a 0.72 F1 score, actually outperforming two of the three human clinicians when compared against the consensus "gold standard" labels.
- Precision: The 3D tracking error was remarkably low—approximately 21mm for finger tapping and 56mm for gait—validating the effectiveness of the markerless approach.
Conclusion and Future Work
The TULIP dataset proves that off-the-shelf computer vision tools are now precise enough for clinical-grade kinematic analysis without requiring patients to wear specialized motion-capture gear. While the current study is limited by a small sample size of 15 subjects, it sets a new benchmark for objective PD assessment.
Future directions include using unsupervised learning to discover new movement signatures more sensitive than the 0-4 UPDRS scale, adapting the system for smaller clinic rooms, and expanding the dataset to include other movement disorders like Essential Tremor to test differential diagnostic capabilities.
Summary updated