Using Deep Learning for Task and Tremor Type Classification in People with Parkinson's Disease.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Rehabilitation Technology

Background:

• Hand tremor is a primary symptom of Parkinson's disease (PD), significantly impacting daily activities.
• Current wearable devices aim to suppress tremor but struggle to differentiate from voluntary motion.
• Advanced algorithms are needed for precise identification of tremor and voluntary movements.

Purpose of the Study:

• To design algorithms for automatic identification of tremor type and voluntary motions using only surface electromyography (sEMG) data.
• To develop a system capable of distinguishing between different types of Parkinson's disease tremor and voluntary actions.
• To enhance the development of assistive technologies for individuals with Parkinson's disease.

Main Methods:

• Implementation of a bidirectional long short-term memory (BiLSTM) algorithm utilizing sEMG data.
• Automated hyperparameter selection for the BiLSTM model using a regularized evolutionary algorithm.
• Classification of tasks and tremor types in individuals with Parkinson's disease.

Main Results:

• The BiLSTM model achieved 84±8% accuracy in classifying tasks performed by individuals with PD.
• Tremor classification accuracy reached 88±5% using the developed sEMG-based models.
• Both classification models significantly outperformed chance levels (20% for task, 33% for tremor).

Conclusions:

• sEMG signals alone are sufficient for accurately identifying distinct tasks and tremor types in Parkinson's disease.
• The developed algorithms demonstrate the potential for creating more effective tremor-suppressing wearable devices.
• This research paves the way for improved non-invasive monitoring and management of Parkinson's disease symptoms.