Interpretable SincNet-based Deep Learning for Emotion Recognition from EEG brain activity.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Deep learning models show promise in medicine but often lack interpretability, limiting their use in clinical decision support.
• Autism Spectrum Disorder (ASD) is associated with distinct neural oscillatory activity patterns.
• Interpretable AI is crucial for understanding and validating medical applications.

Purpose of the Study:

• To introduce and evaluate SincNet, an interpretable deep learning technique, for analyzing neural activity in individuals with ASD.
• To develop a novel SincNet-based neural network for emotion detection in ASD patients using electroencephalogram (EEG) signals.
• To investigate the interpretability of SincNet by examining the learned filters for EEG spectrum analysis.

Main Methods:

• Utilized SincNet, a convolutional neural network employing trainable sinc-functions for customized band-pass filter learning.
• Applied SincNet to analyze EEG signals from individuals with ASD to detect emotions.
• Inspected the learned filters of SincNet to identify relevant EEG frequency bands for emotion prediction.

Main Results:

• The SincNet model automatically identified suppression in the high-alpha (9-13 Hz) and beta (13-30 Hz) frequency bands in individuals with ASD.
• These findings align with existing neuroscience research linking alpha and beta band suppression to behavioral deficits in ASD.
• The interpretable nature of SincNet allowed for clear identification of the spectral features used in emotion recognition.

Conclusions:

• SincNet offers an interpretable deep learning approach for analyzing neural activity in ASD.
• The model successfully detects emotions in ASD patients using EEG, highlighting specific spectral patterns.
• SincNet achieves high performance in emotion recognition without compromising interpretability, making it suitable for medical decision support.