Hand tremor detection in videos with cluttered background using neural network based approaches

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Hand tremor detection is crucial for diagnosing and monitoring neurodegenerative diseases like Parkinson's.
• Current methods often rely on wearable sensors or are sensitive to environmental factors.
• A robust, non-invasive video-based approach is needed for real-world applications.

Purpose of the Study:

• To investigate the accuracy of advanced neural network architectures for automatic hand tremor detection in videos.
• To develop an algorithm capable of detecting hand tremors even with cluttered backgrounds.
• To enable the use of videos recorded in non-research settings for tremor analysis.

Main Methods:

• Examined various neural network architectures and feature sets for tremor detection.
• Compared different combinations of features and classification models.
• Utilized cross-validation to assess model prediction accuracy.

Main Results:

• The highest classification accuracy for tremor detection (vs. non-tremor) reached 80.6%.
• This optimal performance was achieved using a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model.
• Features based on frequency and amplitude changes proved most effective.

Conclusions:

• Advanced neural networks, particularly CNN-LSTM, show high potential for accurate, non-invasive hand tremor detection.
• The developed algorithm can effectively detect tremors in videos with cluttered backgrounds.
• This method facilitates tremor analysis in diverse, real-world environments, aiding clinical diagnosis and research.