A Taxonomy of Explainability Methods for Deep Learning Models Applied to Raw Electroencephalograms

1647 

Abstract Submission 

Charles Ellis¹, Robyn Miller¹, Vince Calhoun²

¹Center for Translational Research in Neuroimaging and Data Science (TReNDS), Atlanta, GA, ²Center for Translational Research in Neuroimaging and Data Science (TReNDS), Decatur, GA

Charles Ellis  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Robyn Miller, PhD  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Vince Calhoun  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Decatur, GA

Deep learning is being increasingly applied to raw electroencephalograph (EEG) data for automated neuropsychiatric disorder diagnosis. Unfortunately, while these methods can boost model performance, they can also reduce explainability, which has led to the development of a variety of explainability (XAI) methods uniquely adapted to raw EEG data. In this study, we taxonomize existing approaches (Figure 1) and present a new XAI approach.
Existing raw EEG XAI approaches generally fall within the categories of spatial (i.e., identifying key channels) (Pathak et al., 2021), spectral (i.e., identifying key frequency bands) (Nahmias and Kontson, 2020), or temporal (i.e., identifying key waveforms) (Ellis et al., 2022) explainability approaches, and among each of these categories can be found both local (i.e., providing insight into the classification of an individual sample) and global (i.e., providing insight into how a model generally behaves) explainability approaches. The categories can be further subdivided based upon the characteristics of the approaches (e.g., perturbation (Nahmias and Kontson, 2020), activation maximization (Ellis et al., 2021b), or gradient-based methods (Ellis et al., 2021a)).
However, while many approaches have been developed for identifying EEG spatial, spectral, and temporal importance, it is also important to identify interactions uncovered by models between channels and frequency bands, so in this study, we develop a model for major depressive disorder (MDD) diagnosis and present a novel raw EEG XAI approach that identifies spatiospectral interactions between frequency bands in a given channel and other channels.

We used 19 channels of resting-state EEG data from 30 individuals with MDD (MDDs) and 28 healthy controls (HCs). The data is available at: https://figshare.com/articles/dataset/EEG_Data_New/4244171. We segmented the data into 25-second samples and downsampled to 200 Hz. We trained a 1-dimensional convolutional neural network with 10-fold subject-wise cross-validation to classify between the two classes.
In our XAI analysis, we applied layer-wise relevance propagation (LRP) (Bach et al., 2015) with an αβ-relevance rule (α=1, β=0) to identify the relative importance of each channel. We then converted the data to the frequency domain, successively zeroed out canonical frequency bands (δ, θ, α, β, γlow, γhigh) in each channel, converted the perturbed data back to the time domain, reapplied LRP, and applied t-tests with false discovery rate correction (α=0.001) to identify significant interactions between frequency bands and channels (i.e., whether a channel's relevance changed after a frequency band was perturbed in another channel). We analyzed the model from the fold with the highest test performance and performed separate analyses for MDDs, HCs, and all samples.

Our model performance was relatively high, particularly when compared to other papers that used robust subject-wise cross-validation approaches. The mean and standard deviation of our model accuracy, balanced accuracy, sensitivity, and specificity were 85.00 ± 7.82, 86.52 ± 8.14, 90.33 ± 11.68, and 82.70 ± 16.82, respectively.
As shown in Figure 2, for HCs, our model identified widespread interactions between parietal electrodes and electrodes in most other regions. In contrast, for MDDs, our model identified interactions primarily between frontal and central electrodes and electrodes in other regions, which fits with other MDD studies that have identified frontal connectivity effects (Movahed et al., 2021).

In this study, we presented a taxonomy of existing EEG XAI approaches and further expanded it by contributing a new XAI approach that identifies spatiospectral interactions between different EEG frequency bands and electrodes. We used our approach within the context of automated MDD diagnosis and identified patterns comparable to those identified in traditional connectivity-based analyses in MDD.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Classification and Predictive Modeling ²

EEG/MEG Modeling and Analysis ¹

Computational Neuroscience

Data analysis

Electroencephaolography (EEG)

Machine Learning

Psychiatric Disorders