Social Contextualization of Datasets for Mental Health AI: a Review of Gender-linked Sociotechnical Misalignments.

This study examines the relationship between gender, social context, and mental health detection. We rebuilt our dataset from The Distress Analysis Interview Corpus Wizard-of-Oz (DAIC_WOZ) and applied a mixed methodology by first coding the interview responses qualitatively and then analyzing the coded responses quantitatively. Our research findings revealed that: first, the relationship between mental health scores and the responses depends on socially contextual factors. Second, gender, due to its social-contextual implications, differentially links three types of responses (e.g., on family ties, personality type, and travel habits) to a mental health state. We believe gender and social context together do have a certain connection with an individual’s mental health. We used our findings to discuss the study implications and highlight the importance of including gender-related social contexts along with gender information in training datasets to help create better AI decision-making for detecting mental health disorders.