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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Current Explainable Artificial Intelligence (XAI) models lack reliability in assessing feature relevance for deep neural biomarker classifiers.
• The maturity of saliency maps for interpreting neural activity is insufficient for practical clinical applications, hindering Deep Learning development.
• This gap limits the clinical translation of advanced AI techniques in neuroscience.

Purpose of the Study:

• To introduce and evaluate the RemOve-And-Retrain (ROAR) algorithm for recovering highly relevant features from pre-trained deep neural networks.
• To assess the clinical applicability of ROAR in Face Emotion Recognition (FER) using electroencephalography (EEG) signals for Autism Spectrum Disorder (ASD) research.
• To compare the reliability of ROAR against established relevance mapping techniques.

Main Methods:

• Trained a Convolutional Neural Network (CNN) using EEG signals for Face Emotion Recognition (FER).
• Applied the RemOve-And-Retrain (ROAR) algorithm to identify and recover salient EEG features.
• Compared ROAR's feature relevance assessment with Layer-Wise Relevance Propagation, PatternNet, Pattern-Attribution, and Smooth-Grad Squared.
• Evaluated feature relevance in both typically developing (TD) and ASD individuals.

Main Results:

• The ROAR algorithm demonstrated effectiveness in recovering highly relevant EEG features for FER.
• Reliability of feature relevance was enhanced compared to traditional methods when applied to EEG data.
• The study successfully bridged neuroscience findings with feature relevance calculation for EEG-based emotion recognition.
• Identified distinct EEG feature relevance patterns between TD and ASD individuals during FER tasks.

Conclusions:

• The ROAR algorithm offers a reliable method for feature relevance recovery in deep neural networks for clinical applications.
• This approach significantly improves the interpretability and trustworthiness of AI models analyzing neural activity.
• The findings support the potential of EEG-based FER using ROAR for advancing ASD diagnosis and understanding.
• This study pioneers the integration of advanced XAI techniques with neuroimaging for emotion recognition in ASD.