A Model Visualization-based Approach for Insight into Waveforms and Spectra Learned by CNNs.

Area of Science:

• Computational Neuroscience
• Machine Learning in Medicine
• Signal Processing

Background:

• Deep learning, particularly CNNs, is increasingly used for electrophysiology analysis.
• Explainability of CNNs using raw time-series data, like EEG, remains a significant challenge.
• Existing methods for explaining CNNs in electrophysiology have limited utility for unique waveforms.

Purpose of the Study:

• To present a novel model visualization-based approach for analyzing CNN filters in electrophysiology.
• To extract explainable information from EEG waveforms learned by CNNs.
• To provide insights into learned spectral features alongside waveform analysis.

Main Methods:

• Analysis of filters in the first convolutional layer of a CNN.
• Application of the method to automated sleep stage classification using EEG data.
• Identification of filter subgroups, their spectral properties, and relative importance.

Main Results:

• Demonstrated the viability of the visualization approach in sleep stage classification.
• Identified 3 subgroups of filters with distinct spectral characteristics.
• Uncovered unique EEG waveforms learned by the classifier crucial for performance.

Conclusions:

• The proposed method offers a significant advancement in the explainability of electrophysiology classifiers.
• This approach can aid in the development and validation of clinical time-series classifiers.
• Provides valuable insights into learned features for future electrophysiology studies.