Evaluation of interpretability for deep learning algorithms in EEG emotion recognition: A case study in autism.

Current models on Explainable Artificial Intelligence (XAI) have shown a lack of reliability when evaluating feature-relevance for deep neural biomarker classifiers. The inclusion of reliable saliency-maps for obtaining trustworthy and interpretable neural activity is still insufficiently mature for practical applications. These limitations impede the development of clinical applications of Deep Learning. To address, these limitations we propose the RemOve-And-Retrain (ROAR) algorithm which supports the recovery of highly relevant features from any pre-trained deep neural network. In this study we evaluated the ROAR methodology and algorithm for the Face Emotion Recognition (FER) task, which is clinically applicable in the study of Autism Spectrum Disorder (ASD). We trained a Convolutional Neural Network (CNN) from electroencephalography (EEG) signals and assessed the relevance of FER-elicited EEG features from individuals diagnosed with and without ASD. Specifically, we compared the ROAR reliability from well-known relevance maps such as Layer-Wise Relevance Propagation, PatternNet, Pattern-Attribution, and Smooth-Grad Squared. This study is the first to bridge previous neuroscience and ASD research findings to feature-relevance calculation for EEG-based emotion recognition with CNN in typically-development (TD) and in ASD individuals.