Automatic Recognition of Epileptiform EEG Abnormalities.

Area of Science:

• Neuroscience
• Medical Informatics
• Machine Learning

Background:

• Long-term electroencephalogram (EEG) examinations are crucial for diagnosing conditions like epilepsy.
• Current automated abnormality detection methods often lack generalizability due to reliance on single datasets.
• Improving the efficiency and reliability of EEG analysis is essential for clinical practice.

Purpose of the Study:

• To develop and evaluate a machine learning solution for detecting interictal abnormal EEG segments.
• To assess the generalizability of the developed classifier across different publicly available datasets.
• To identify factors influencing the performance of automated EEG abnormality detection.

Main Methods:

• A machine learning classifier was developed and optimized using the TUH Abnormal EEG Corpus.
• The classifier was subsequently re-trained and tested on various combinations of public EEG datasets.
• Performance was evaluated based on accuracy across different training and testing data splits.

Main Results:

• Internal results on datasets were comparable to the state-of-the-art in EEG abnormality detection.
• Cross-dataset training and testing yielded accuracies ranging from 67.51% to 99.50%.
• Lower accuracy was observed when training data was highly preprocessed and limited in size.

Conclusions:

• Machine learning offers an efficient approach to automated detection of abnormal EEG segments.
• Training and testing on diverse datasets are critical for robust generalization of EEG analysis tools.
• Data preprocessing and dataset size significantly impact the performance of automated EEG interpretation models.