Consistency of Feature Importance Algorithms for Interpretable EEG Abnormality Detection.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Machine learning advances offer potential for automated electroencephalography (EEG) abnormality detection.
• Interpretable models provide control but often underperform complex deep learning methods.
• Bridging the gap between performance and interpretability in EEG analysis is crucial.

Purpose of the Study:

• To investigate the feasibility of combining deep learning with interpretable EEG features.
• To evaluate the performance and interpretability of a hybrid approach for EEG abnormality detection.
• To assess the consistency of different feature importance attribution methods in a deep learning context.

Main Methods:

• Developed a convolutional neural network (CNN) using multi-channel EEG frequency bands as input.
• Applied four feature importance attribution methods: Layer-wise Relevance Propagation (LRP), DeepLIFT, Integrated Gradients (IG), and Guided GradCAM.
• Analyzed the consistency and reliability of feature attributions across the chosen methods.

Main Results:

• The study demonstrated a feasible approach to integrate deep learning with interpretable EEG inputs.
• Layer-wise Relevance Propagation (LRP), DeepLIFT, and Integrated Gradients (IG) showed consistent feature importance attributions.
• Guided GradCAM produced deviating attributions compared to the other three methods.

Conclusions:

• Combining deep learning with interpretable EEG features offers a promising direction for abnormality detection.
• The consistency among LRP, DeepLIFT, and IG highlights their potential for robust medical interpretation.
• Employing a suite of feature importance methods is recommended to ensure the reliability of deep learning-based medical interpretations.