Machine Learning with Imbalanced EEG Datasets using Outlier-based Sampling.

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy affects over 65 million people globally, necessitating advanced seizure detection.
• Implantable devices use machine learning to detect and treat seizures via electrical stimulation.
• Training data imbalance, with rare seizures (<1%), hinders algorithm performance.

Purpose of the Study:

• To develop an improved pre-processing method for imbalanced training data in epilepsy seizure detection.
• To address the performance drawbacks of conventional sampling techniques like down-sampling and up-sampling.

Main Methods:

• Proposed an outlier-based sampling method to reduce the majority class (normal activity).
• Utilized Exponentially Decaying Memory Signal Energy (EDMSE) features with Isolation Forests and ANOVA for outlier detection.
• Compared the outlier-based method with conventional techniques using KNN and Logistic Regression classifiers.

Main Results:

• Achieved approximately 2% higher accuracy compared to conventional methods.
• Reduced false positives by approximately 38%.
• Demonstrated a latency reduction of approximately 3 seconds.

Conclusions:

• The outlier-based sampling method effectively addresses data imbalance in epilepsy seizure detection.
• This novel approach enhances classifier performance, reduces false alarms, and lowers detection latency.
• The findings support the development of more reliable implantable therapeutic devices for epilepsy.