Evaluation of Interpretability for Deep Learning algorithms in EEG\n Emotion Recognition: A case study in Autism

Abstract

Current models on Explainable Artificial Intelligence (XAI) have shown an\nevident and quantified lack of reliability for measuring feature-relevance when\nstatistically entangled features are proposed for training deep classifiers.\nThere has been an increase in the application of Deep Learning in clinical\ntrials to predict early diagnosis of neuro-developmental disorders, such as\nAutism Spectrum Disorder (ASD). However, the inclusion of more reliable\nsaliency-maps to obtain more trustworthy and interpretable metrics using neural\nactivity features is still insufficiently mature for practical applications in\ndiagnostics or clinical trials. Moreover, in ASD research the inclusion of deep\nclassifiers that use neural measures to predict viewed facial emotions is\nrelatively unexplored. Therefore, in this study we propose the evaluation of a\nConvolutional Neural Network (CNN) for electroencephalography (EEG)-based\nfacial emotion recognition decoding complemented with a novel\nRemOve-And-Retrain (ROAR) methodology to recover highly relevant features used\nin the classifier. Specifically, we compare well-known relevance maps such as\nLayer-Wise Relevance Propagation (LRP), PatternNet, Pattern-Attribution, and\nSmooth-Grad Squared. This study is the first to consolidate a more transparent\nfeature-relevance calculation for a successful EEG-based facial emotion\nrecognition using a within-subject-trained CNN in typically-developed and ASD\nindividuals.\n