A Systematic Approach for Explaining Time and Frequency Features Extracted by Convolutional Neural Networks From Raw Electroencephalography Data

Abstract

In recent years, the use of convolutional neural networks (CNNs) for raw resting-state electroencephalography (EEG) analysis has grown increasingly common. However, relative to earlier machine learning and deep learning methods with manually extracted features, CNNs for raw EEG analysis present unique problems for explainability. As such, a growing group of methods have been developed that provide insight into the spectral features learned by CNNs. However, spectral power is not the only important form of information within EEG, and the capacity to understand the roles of specific multispectral waveforms identified by CNNs could be very helpful. In this study, we present a novel model visualization-based approach that adapts the traditional CNN architecture to increase interpretability and combines that inherent interpretability with a systematic evaluation of the model via a series of novel explainability methods. Our approach evaluates the importance of spectrally distinct first-layer clusters of filters before examining the contributions of identified waveforms and spectra to cluster importance. We evaluate our approach within the context of automated sleep stage classification and find that, for the most part, our explainability results are highly consistent with clinical guidelines. Our approach is the first to systematically evaluate both waveform and spectral feature importance in CNNs trained on resting-state EEG data.

Improving Multichannel Raw Electroencephalography-based Diagnosis of Major Depressive Disorder via Transfer Learning with Single Channel Sleep Stage Data*

“…We adapted the 1D-CNN architecture initially developed in [19] and later used in [8], [20]. The architecture was first developed to work with much longer samples that might contain multiple sleep stages.…”

“…We lastly computed the percent change in test performance following the perturbation of each cluster of filters. This analysis is adapted from [8] and provides an important point of comparison for our novel filter activation analysis by relating it to existing approaches.…”

“…While these methods can offer improved performance relative to traditional machine learning methods applied to extracted EEG features [6], [7], they also make explainability more challenging. As such, a growing number of studies have developed EEG-specific explainability methods [8]- [13]. However, while these methods do provide important insights, they are mainly feature importance approaches.…”

“…Many studies use methods that involve data perturbation [1], [3], [12], which can lead to out of distribution samples that provide poor explanations [18]. One alternative approach that has provided useful insights is model visualization (i.e., structuring first layer filters such that characteristics of features can be understood) [8], [9], [19], [20].…”

“…As recently taxonomized in [4], existing EEG explainability methods generally fall into one of five categories based on the types of features that they identify. They can provide insight into the importance of key (1) frequency bands [8]- [13], (2) waveforms [2], [8], [9], [17], [18], (3) channels [3], [5], [19], (4) spatial interactions [4], and (5) spatio-spectral interactions [4], [5], though the third through fifth categories are restricted to multi-channel EEG data. Many studies use methods that involve data perturbation [1], [3], [12], which can lead to outof-distribution samples that yield poor explanations [20].…”