Epileptic States Recognition Using Transfer Learning.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signal recognition is crucial for epilepsy diagnosis and assessment.
• Conventional methods often suffer from unsatisfactory accuracy due to inconsistent data distributions between training and testing sets.
• Addressing data distribution discrepancies is vital for robust epilepsy detection algorithms.

Purpose of the Study:

• To develop an effective transfer learning method for EEG signal recognition that overcomes data distribution challenges.
• To improve the accuracy, sensitivity, and specificity of epileptic state recognition.
• To enable knowledge transfer from training data to testing data with differing distributions.

Main Methods:

• Utilized a cross-domain mean joint approximation embedding (CMJAE) transductive transfer learning approach.
• Combined subspace learning and joint distribution adaptation to address marginal and conditional distribution discrepancies.
• Implemented a method for knowledge transfer by measuring distribution differences between training and testing datasets.

Main Results:

• Achieved an average recognition accuracy of 97.5% for different epileptic states.
• Demonstrated high sensitivity (94.3%) and specificity (92.7%) in recognizing epileptic states.
• Outperformed conventional machine learning and deep learning methods in recognition tasks.

Conclusions:

• The CMJAE transfer learning method offers a promising strategy for enhancing epileptic state recognition algorithms.
• This approach effectively learns models from differently distributed data with low computational cost.
• The findings suggest a significant advancement in automated EEG-based epilepsy diagnosis and assessment.