Towards Interpretable Machine Learning in EEG Analysis.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Automatic detection of abnormalities in electroencephalography (EEG) is crucial for neurological disorder diagnosis.
• Interpretable machine learning models are needed for clinical adoption and trust.

Purpose of the Study:

• To dissect a machine learning model for EEG abnormality detection to understand the contribution of each component.
• To evaluate the trade-off between model complexity, accuracy, and interpretability.

Main Methods:

• A machine learning model comprising several shallow artificial neural networks aggregated via voting was developed.
• The model was systematically simplified to assess the impact of each component on classification accuracy.
• Key features, such as relative wavelet energy, were extracted and analyzed.

Main Results:

• The most successful model setup achieved an 81% classification accuracy.
• Stepwise simplification led to a decrease in accuracy.
• A simplified model using only relative wavelet energy achieved 75% accuracy, significantly above the random guess baseline of 54%.

Conclusions:

• It is feasible to build simple, interpretable machine learning models for EEG abnormality detection.
• These models can achieve accuracy scores comparable to state-of-the-art, complex methods.
• Interpretability is achievable without substantial compromise in classification performance.