Machine Learning Fairness for Depression Detection using EEG Data

TL;DR

This work tackles fairness in EEG-based depression detection by framing it as a binary classification problem with gender as the sensitive attribute and evaluating five bias-mitigation strategies across pre-, in-, and post-processing stages. It employs three DL architectures (Deep-Asymmetry CNN, GTSAN, and 1D-CNN-LSTM) on three EEG datasets (Mumtaz, MODMA, Rest) and uses both prediction metrics ($Acc$, $P$, $F1$) and fairness metrics ($M_{SP}$, $M_{EOpp}$, $M_{EOdd}$, $M_{EAcc}$) with practical bounds for assessment. The findings reveal dataset and algorithmic biases across datasets, with mitigation methods showing heterogeneous effects and often failing to address all fairness notions, underscoring the need for multiple fairness metrics and dataset-aware strategies in high-stakes EEG-based depression detection. The study highlights the importance of fairness-aware evaluation and motivates further research into robust bias mitigation tailored to imbalanced, sensitive-domain EEG data. Overall, the work provides a foundational analysis of bias in EEG depression detection and points to the necessity of comprehensive fairness frameworks in clinical decision-support contexts.

Abstract

This paper presents the very first attempt to evaluate machine learning fairness for depression detection using electroencephalogram (EEG) data. We conduct experiments using different deep learning architectures such as Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Unit (GRU) networks across three EEG datasets: Mumtaz, MODMA and Rest. We employ five different bias mitigation strategies at the pre-, in- and post-processing stages and evaluate their effectiveness. Our experimental results show that bias exists in existing EEG datasets and algorithms for depression detection, and different bias mitigation methods address bias at different levels across different fairness measures.

Figures (1)

• Figure 1: Simplified pipeline: highlighted parts indicate how bias is mitigated at the pre-, in- and post-processing stages.