An Explainable Deep Learning Approach for Multimodal Electrophysiology Classification

In recent years, more biomedical studies have begun to use multimodal data to improve model performance. As such, there is a need for improved multimodal explainability methods. Many studies involving multimodal explainability have used ablation approaches. Ablation requires the modification of input data, which may create out-of-distribution samples and may not always offer a correct explanation. We propose using an alternative gradient-based feature attribution approach, called layer-wise relevance propagation (LRP), to help explain multimodal models. To demonstrate the feasibility of the approach, we selected automated sleep stage classification as our use-case and trained a 1-D convolutional neural network (CNN) with electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG) data. We applied LRP to explain the relative importance of each modality to the classification of different sleep stages. Our results showed that across all samples, EEG was most important, followed by EOG, and EMG. For individual sleep stages, EEG and EOG had higher relevance for classifying awake and non-rapid eye movement 1 (NREM1). EOG was most important for classifying REM, and EEG was most relevant for classifying NREM2-NREM3. Also, LRP gave consistent levels of importance to each modality for correctly classified samples across folds and inconsistent levels of importance for incorrectly classified samples. Our results demonstrate the additional insight that gradient-based approaches can provide relative to ablation methods and highlight their feasibility for explaining multimodal electrophysiology classifiers.