Wrist sensor-based tremor severity quantification in Parkinson's disease using convolutional neural network

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Tremor is a primary symptom of Parkinson's disease (PD), necessitating accurate severity assessment for effective treatment.
• Current methods for tremor assessment may lack the precision required for continuous monitoring in daily life.

Purpose of the Study:

• To develop and evaluate a novel tremor assessment system utilizing a convolutional neural network (CNN) for differentiating Parkinson's disease symptom severity.
• To enhance the precision of monitoring Parkinson's disease tremor symptoms in real-world settings.

Main Methods:

• Tremor signals were collected from 92 Parkinson's disease patients using a custom wearable device (SNUMAP) with accelerometer and gyroscope sensors.
• Data were transformed into the frequency domain and converted into 2D images for training a CNN model.
• The CNN architecture was trained by convolving tremor signal images with kernels and compared against existing machine learning algorithms.

Main Results:

• The proposed CNN-based system achieved high accuracy in differentiating tremor severity.
• Performance metrics included an accuracy of 0.85 and a linear weighted kappa of 0.85.
• The CNN model demonstrated superior performance compared to previously studied machine learning algorithms.

Conclusions:

• The developed CNN system offers a precise and potentially more accessible method for monitoring Parkinson's disease tremor severity.
• This approach could facilitate more accurate treatment adjustments and improved patient management in daily life.
• Wearable sensor data combined with advanced AI presents a promising avenue for objective neurological symptom assessment.