A Parkinsonian Digital Biomarker Learned as an Anomaly Deep Generative Representation

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Parkinson's Disease (PD) is a prevalent neurodegenerative disorder affecting movement, with gait alterations being key diagnostic indicators.
• Current gait analysis relies on subjective expert observation, leading to potential diagnostic inaccuracies (up to 24% false positives).
• Computational tools offer objective alternatives, but often require extensive, balanced datasets impractical for clinical settings.

Purpose of the Study:

• To develop a self-supervised learning framework for discovering gait-motion patterns in Parkinson's Disease.
• To create a digital biomarker for objective PD diagnosis using gait analysis from video data.
• To overcome the limitations of expert-based diagnosis and data-intensive machine learning models.

Main Methods:

• A self-supervised generative representation model was employed, utilizing video reconstruction as a pretext task.
• An anomaly detection framework was integrated to identify gait deviations.
• A hidden embedding gait descriptor was extracted as a digital biomarker.

Main Results:

• The digital biomarker effectively distinguished between Parkinson's Disease patients and control subjects.
• The model achieved a high classification accuracy, with an Area Under the Curve (AUC) of 99.4%.
• The approach was trained solely on control population data, demonstrating robust generalization.

Conclusions:

• The proposed self-supervised method provides an objective and accurate digital biomarker for Parkinson's Disease diagnosis.
• This approach minimizes subjectivity in gait analysis, offering a valuable tool for clinical application.
• The technique holds potential for early and reliable detection of PD through gait pattern analysis.