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

Area of Science:

• Artificial Intelligence
• Mental Health Assessment
• Computational Linguistics

Background:

• Speech-based AI models show promise for detecting depression and Post-traumatic stress disorder (PTSD).
• Existing models often exhibit gender bias, leading to disparities in prediction accuracy and fairness.
• Addressing demographic biases is crucial for equitable AI applications in mental healthcare.

Purpose of the Study:

• To develop and evaluate a domain adversarial training approach to mitigate gender bias in speech-based depression and PTSD detection.
• To integrate gender information as distinct domains into a pretrained speech foundation model.
• To assess the impact of this approach on detection performance using the E-DAIC dataset.

Main Methods:

• Domain adversarial training was employed, treating different genders as separate domains.
• Gender information was integrated into a pretrained speech foundation model.
• The model's effectiveness was validated on the E-DAIC dataset.

Main Results:

• The proposed domain adversarial method significantly improved detection performance compared to the baseline.
• An increase in F1-score of up to 13.29 percentage points was observed.
• The results demonstrate the effectiveness of addressing gender disparities in AI mental health tools.

Conclusions:

• Domain adversarial training is an effective strategy for mitigating gender bias in speech-based mental health AI.
• The improved fairness enhances the clinical applicability and trustworthiness of these AI models.
• Addressing demographic disparities is essential for developing equitable and reliable AI for mental health assessment.