Domain Adversarial Training for Mitigating Gender Bias in Speech-based Mental Health Detection

TL;DR

This work addresses gender bias in speech-based mental health detection by applying domain adversarial training (DAT) that treats gender as a domain alongside a pretrained speech foundation model. The method optimizes a joint objective, $\mathcal{L}_{Final} = \mathcal{L}_{Mental} + \lambda \mathcal{L}_{Dis}$, to promote domain-invariant representations while predicting depression or PTSD on the E-DAIC dataset. Experiments show substantial improvements in F1 scores, up to 13.29 percentage points, and reduced gender disparities across both tasks, demonstrating improved cross-gender generalization. The results underscore the importance of bias-aware learning in AI-driven medical diagnostics and point to future work on additional demographic factors and larger datasets to further enhance fairness and clinical applicability.

Abstract

Speech-based AI models are emerging as powerful tools for detecting depression and the presence of Post-traumatic stress disorder (PTSD), offering a non-invasive and cost-effective way to assess mental health. However, these models often struggle with gender bias, which can lead to unfair and inaccurate predictions. In this study, our study addresses this issue by introducing a domain adversarial training approach that explicitly considers gender differences in speech-based depression and PTSD detection. Specifically, we treat different genders as distinct domains and integrate this information into a pretrained speech foundation model. We then validate its effectiveness on the E-DAIC dataset to assess its impact on performance. Experimental results show that our method notably improves detection performance, increasing the F1-score by up to 13.29 percentage points compared to the baseline. This highlights the importance of addressing demographic disparities in AI-driven mental health assessment.

Figures (2)

• Figure 1: Illustration of fine-tuning and the proposed domain adversarial training architectures designed to mitigate gender bias in speech-based mental health detection.
• Figure 2: T-SNE results of wav2vec2 fine-tuning and proposed DAT on E-DAIC test set for gender labels.