# Sobel neural network for EEG-based major depressive disorder screening

## TL;DR

This paper introduces the Sobel Network, a new neural architecture that improves accuracy in detecting major depressive disorder using EEG data.

## Contribution

The Sobel Network integrates Sobel-inspired operations within convolutional layers for end-to-end learning of EEG features specific to depression.

## Key findings

- The Sobel Network achieved 98.67% accuracy in distinguishing MDD patients from healthy controls.
- It outperformed other deep learning models in sensitivity (99.18%) and specificity (98.10%).
- The model was tested on a dataset of 34 MDD patients and 30 healthy controls from a public EEG repository.

## Abstract

Early and objective screening for major depressive disorder (MDD) is crucial, with electroencephalography (EEG) offering significant potential. However, developing accurate automated tools requires architectures adept at capturing subtle, discriminative spatiotemporal features in EEG signals. This paper introduces the Sobel Network, a novel neural architecture designed specifically for EEG-based MDD screening, namely, identifying MDD patients from healthy controls (HC). Unlike approaches using Sobel operators solely for preprocessing, the Sobel Network integrates Sobel-inspired operations intrinsically within its convolutional layers, enabling end-to-end learning of features emphasizing gradient patterns and edge-like information highly relevant to depression biomarkers in EEG. We evaluate the Sobel Network on a publicly available EEG dataset from the Hospital of Universiti Sains Malaysia (HUSM). This dataset comprises 34 patients diagnosed with MDD (17 men; mean age, 40.3 ± 12.9 years) and 30 healthy controls (HC; 21 men; mean age, 38.2 ± 15.6 years). The results demonstrate that the proposed architecture significantly outperforms other deep learning models in key metrics including accuracy (achieving 98.67%), sensitivity (99.18%), and specificity (98.10%). The Sobel Network presents a promising avenue to improve the accuracy and robustness of automated EEG-based depression screening tools, offering practical impact for clinical decision support.