Data Augmentation Effects on Highly Imbalanced EEG Datasets for Automatic Detection of Photoparoxysmal Responses

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Photosensitivity is a neurological disorder causing epileptic discharges (Photoparoxysmal Responses - PPRs) triggered by visual stimuli.
• Diagnosis involves Intermittent Photic Stimulation (IPS) and Electroencephalogram (EEG) monitoring, with manual PPR identification by specialists.
• EEG recordings for photosensitive epilepsy often present highly imbalanced datasets due to infrequent PPR occurrences.

Purpose of the Study:

• To improve the automatic detection of Photoparoxysmal Responses (PPRs) in photosensitive epilepsy.
• To address the challenge of imbalanced datasets in EEG recordings for photosensitive epilepsy.
• To evaluate the effectiveness of Data Augmentation (DA) in enhancing machine learning model performance for PPR detection.

Main Methods:

• Applied Data Augmentation (DA) techniques to generate synthetic PPR segments from existing EEG data.
• Utilized K-Nearest Neighbors and a One-Hidden-Dense-Layer Neural Network to assess the impact of DA.
• Compared model performance before and after the DA stage, including metrics like Accuracy, Specificity, and Sensitivity.

Main Results:

• Data Augmentation significantly improved the balance of the EEG dataset.
• Machine learning models demonstrated enhanced robustness and generalization capabilities after DA.
• Performance improvements of approximately 20% in Accuracy and Specificity were observed, with no loss in Sensitivity.

Conclusions:

• Data Augmentation is an effective strategy for improving automated PPR detection in photosensitive epilepsy.
• DA contributes to more robust and generalizable machine learning models for neurological disorder diagnosis.
• The findings suggest a promising approach for enhancing diagnostic accuracy in clinical settings, as demonstrated in ongoing trials at Burgos University Hospital.