Clinical Scorecard: Analysis of Tremor Types and Severity in Parkinson's Disease Using Video-Based Pose Assessment

At a Glance

Key Highlights

• Proposed a novel open-source video-based deep learning system (SPA-PTA) for PT classification and tremor severity estimation
• System uses Eulerian video magnification to enhance subtle tremors and AlphaPose for accurate 2D upper body pose extraction
• Achieved high performance: 91.3% accuracy and 80.0% F1-score in PT classification; 76.4% accuracy and 76.7% F1-score in tremor rating classification

Guideline-Based Recommendations

Diagnosis

• Use clinical assessment combined with video-based pose analysis for improved PD tremor detection
• Ensure video frame rate is at least twice the highest tremor frequency (3–7 Hz) to meet Nyquist limits for valid tremor analysis
• Focus on upper body (hands and arms) pose features as PT predominantly affects these regions

Management

• Employ non-intrusive, cost-effective video-based systems to assist early PD diagnosis and tremor severity monitoring
• Utilize deep learning models with spatial attention mechanisms to interpret joint-wise relevancy in tremor features

Monitoring & Follow-up

• Apply Eulerian video magnification to enhance subtle tremors for better feature extraction during follow-up assessments
• Normalize pose data to reduce bias from video differences and ensure consistent tremor evaluation

Risks

• Be aware of limitations in clinical diagnostic accuracy (73–84%) due to subjective expert opinions
• Consider potential noise and artifacts in video data; use signal processing methods like EVM to mitigate

Patient & Prescribing Data

Individuals with suspected or confirmed Parkinson’s Disease exhibiting tremor symptoms

Video-based pose assessment can support early diagnosis and tremor severity classification, potentially guiding treatment decisions

Clinical Best Practices

• Use consumer-grade cameras to capture videos with sufficient frame rate (≥30 Hz) for tremor analysis
• Apply state-of-the-art pose estimation algorithms (AlphaPose) focusing on upper body keypoints excluding head for privacy
• Incorporate graph neural networks with spatial attention and pyramidal channel-squeezing–fusion architecture for robust tremor classification
• Validate video suitability for tremor analysis by checking Nyquist frequency limits before processing
• Normalize pose data by aligning key joints to a global origin to reduce variability across recordings

References

• Global Parkinson's Disease prevalence
• Dopaminergic neuron loss in PD
• Clinical assessment and diagnosis of PD
• Diagnostic accuracy in PD
• Neuroimaging in PD diagnosis
• CSF biomarkers in PD
• Speech-based PD diagnosis limitations
• Gait disturbance in early PD
• Prevalence of tremor in early-onset PD
• Wearable sensor data for hand tremor
• Preliminary work on tremor-type classification
• AlphaPose algorithm for pose estimation
• Eulerian Video Magnification method
• Nyquist frequency limits for tremor analysis
• Typical frequency range of Parkinson’s Tremor
• Limitations of 3D pose estimation for tremor analysis
• Localization of Parkinson’s Tremor to upper body
• Graph neural networks for relational data