Pose-based tremor type and level analysis for Parkinson's disease from video

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Current Parkinson's disease (PD) diagnosis relies on clinical examination, with accuracy varying (73-84%) and depending on assessor experience.
• An automated, interpretable system is needed to enhance the robustness of PD diagnostic decisions.

Purpose of the Study:

• To develop and validate a deep learning-based system (SPA-PTA) for classifying Parkinson's tremor (PT) and estimating its severity.
• To support early detection of PD using consumer-grade videos.

Main Methods:

• Utilized a novel attention module with a lightweight pyramidal channel-squeezing-fusion architecture for PT analysis.
• Input data consisted of consumer-grade videos of front-facing individuals.
• System performance was evaluated using leave-one-out cross-validation on PT classification and severity estimation tasks.

Main Results:

• Achieved 91.3% accuracy and 80.0% F1-score for classifying PT versus non-PT.
• Attained 76.4% accuracy and 76.7% F1-score in the multiclass tremor severity rating task.

Conclusions:

• The SPA-PTA system provides cost-effective PT classification and severity estimation, serving as an early warning for undiagnosed PD patients.
• Offers a potential solution for PD diagnosis support in resource-limited settings.