Fairness-Aware Few-Shot Learning for Audio-Visual Stress Detection

The introduction of FairM2S marks a significant step in addressing gender bias in AI-driven stress detection, a critical issue for equitable mental healthcare. Existing models often demonstrate bias, particularly in scenarios with limited data. FairM2S counters this by employing a fairness-aware meta-learning framework that integrates Equalized Odds constraints during training and adaptation. This innovative approach not only achieved an impressive accuracy of 78.1% but also significantly reduced the Equal Opportunity metric to 0.06, indicating substantial fairness improvements. Additionally, the release of the SAVSD dataset, which includes smartphone-captured audio-visual data annotated for gender, provides a valuable resource for ongoing research into fairness in AI. Together, these contributions position FairM2S as a leading method for scalable and equitable few-shot stress detection, highlighting the importance of fairness in advancing mental health AI solutions.