PULSAR: Graph based Positive Unlabeled Learning with Multi Stream Adaptive Convolutions for Parkinson's Disease Recognition

TL;DR

PULSAR tackles the challenge of scalable Parkinson's disease screening from accessible webcam data by framing the task as a Positive Unlabeled learning problem. It combines a four-stream spatio-temporal graph neural network with adaptive graph convolution to capture complex hand motion patterns during finger-tapping, leveraging bone, velocity, and acceleration information in addition to joint coordinates. The model demonstrates strong performance on a limited dataset, achieving an accuracy of 80.95% on the validation set and 71.29% on an independent test, with statistically significant improvements over several baselines. This approach offers a practical, privacy-conscious screening tool that could extend PD assessment to underserved regions, with potential applicability to other movement disorders after scaling up data and addressing medication effects.

Abstract

Parkinson's disease (PD) is a neuro-degenerative disorder that affects movement, speech, and coordination. Timely diagnosis and treatment can improve the quality of life for PD patients. However, access to clinical diagnosis is limited in low and middle income countries (LMICs). Therefore, development of automated screening tools for PD can have a huge social impact, particularly in the public health sector. In this paper, we present PULSAR, a novel method to screen for PD from webcam-recorded videos of the finger-tapping task from the Movement Disorder Society - Unified Parkinson's Disease Rating Scale (MDS-UPDRS). PULSAR is trained and evaluated on data collected from 382 participants (183 self-reported as PD patients). We used an adaptive graph convolutional neural network to dynamically learn the spatio temporal graph edges specific to the finger-tapping task. We enhanced this idea with a multi stream adaptive convolution model to learn features from different modalities of data critical to detect PD, such as relative location of the finger joints, velocity and acceleration of tapping. As the labels of the videos are self-reported, there could be cases of undiagnosed PD in the non-PD labeled samples. We leveraged the idea of Positive Unlabeled (PU) Learning that does not need labeled negative data. Our experiments show clear benefit of modeling the problem in this way. PULSAR achieved 80.95% accuracy in validation set and a mean accuracy of 71.29% (2.49% standard deviation) in independent test, despite being trained with limited amount of data. This is specially promising as labeled data is scarce in health care sector. We hope PULSAR will make PD screening more accessible to everyone. The proposed techniques could be extended for assessment of other movement disorders, such as ataxia, and Huntington's disease.

Figures (8)

• Figure 1: Overview of the PD screening pipeline. A participant can perform finger-tapping task in front of a computer webcam. A hand tracking model is used to locate the key points of the hand. A spatio temporal graph is constructed specifically for the finger-tapping task. Four different feature streams (joint, bone, velocity and acceleration) are generated and fed to the proposed PULSAR model for prediction.
• Figure 2: Demonstration of the described data cleaning process.
• Figure 3: Spatio Temporal Graph Construction. (a) The spatial connections of the finger and wrist joints. Natural joint connections are denoted by solid lines. The first type of augmented edge is denoted by dashed lines, the second type by dotted lines, and the third by a blue dashed arrowhead. (b) The temporal connections are between the same joints in the successive frames.
• Figure 4: Feature representation of multi stream network of PULSAR
• Figure 5: Box and whisker plot for accuracy, macro F1 and AUROC of PULSAR and baseline models.