Fairness in EEG-Based Depression Detection

As artificial intelligence becomes increasingly integrated into mental health diagnostics, ensuring
fairness and transparency is critical—particularly in sensitive applications such as depression detection.
This project explores fairness-aware deep learning using raw EEG signals from the MODMA (Multi-
modal Open Dataset for Mental-disorder Analysis) dataset [1]. Three deep neural architectures are
developed and evaluated: a Convolutional Neural Network (CNN) as a baseline model, a CNN combined
with Long Short-Term Memory (CNN-LSTM), and a CNN combined with Gated Recurrent Units and
Attention mechanism (CNN-GRU-Attention). To address potential biases, five mitigation techniques are
applied across three stages: preprocessing, in-processing, and post-processing. The models are evaluated
using both performance metrics (accuracy, precision, recall, F1-score) and fairness metrics (disparate
impact, statistical parity, equal opportunity, average odds, and equalized accuracy).