A Novel Activation Maximization-based Approach for Insight into Electrophysiology Classifiers

Abstract

Spectral analysis remains a hallmark approach for gaining insight into electrophysiology modalities like electroencephalography (EEG). As the field of deep learning has progressed, more studies have begun to train deep learning classifiers on raw EEG data, which presents unique problems for explainability. A growing number of studies have presented explainability approaches that provide insight into the spectral features learned by deep learning classifiers. However, existing approaches only attribute importance to different frequency bands. Most of the methods cannot provide insight into the actual spectral values or the relationship between spectral features that models have learned. Here, we present a novel adaptation of activation maximization for electrophysiology time-series that generates samples that indicate the features learned by classifiers by optimizing their spectral content. We evaluate our approach within the context of EEG sleep stage classification with a convolutional neural network, and we find that our approach is able to identify spectral patterns known to be associated with each sleep stage. We also find surprising results suggesting that our classifier may have prioritized the use of eye and motion artifact when identifying Awake samples. Our approach is the first adaptation of activation maximization to the domain of raw electrophysiology classification. Additionally, our approach has implications for explaining any classifier trained on highly dynamic, long time-series.

A Systematic Approach for Explaining Time and Frequency Features Extracted by CNNs from Raw EEG Data

“…( 2) There are multiple large publicly available datasets within the domain of sleep stage classification that help with reproducibility of analyses (29-31). ( 3) Multiple studies have already presented explainability methods for the domain of sleep stage classification, which will make it easier to compare our findings with results from previous studies (17)(18)(19)(20)32).…”

“…However, when datasets consist of thousands of samples and there is no way to combine insights from the perturbation of each sample, that approach does not provide useful global conclusions on the nature of the time-domain features extracted. Another study used activation maximization to optimize the spectral content of a sample (18). While the method does yield a sample in the time domain that maximizes activation for a particular class, it does not provide insight into the relative importance of different time domain features.…”

“…However, they were only applied to networks trained on samples that were around 30 time points long, and EEG samples can be hundreds to thousands of time points long. Additionally, a previous study showed that these methods do not generalize well to sample lengths relevant for EEG analysis (18). It should be noted that existing explainability methods can be applied to some forms of electrophysiology like electrocardiograms (ECG) that have regularly repeated waveforms (25,26).…”

“…Several studies have presented methods that use activation maximization to identify spectral features that maximize activation of the CNN (17)(18)(19). Two studies examined the effect of sinusoids at different frequencies upon activations of nodes in the early layers of the classifier (17) (19), and the remaining study varied the spectral representation of a sample until it maximized the activation of the final output node (18). Other existing studies involve perturbing canonical frequency bands of samples and examining the effect upon the predictions (20) (21) or performance of a classifier (22).…”

“…While these methods are very innovative, they still inherently restrict the feature space to frequency features. Several studies have presented methods that use activation maximization to identify spectral features that maximize activation of the CNN (17)(18)(19). Two studies examined the effect of sinusoids at different frequencies upon activations of nodes in the early layers of the classifier (17) (19), and the remaining study varied the spectral representation of a sample until it maximized the activation of the final output node (18).…”