A recent study has revealed that artificial intelligence (AI) models analyzing social media posts may accurately detect signals of depression in white Americans but not in their Black counterparts. The findings underscore the importance of including diverse racial and ethnic groups in training AI models for healthcare-related tasks.
Published in PNAS (Proceedings of the National Academy of Sciences), the study utilized an AI tool to analyze language in social media posts from 868 volunteers, including equal numbers of Black and white adults. Participants shared similar characteristics such as age and gender, and all completed a validated questionnaire used to screen for depression by healthcare providers.
The results showed that while certain language patterns, such as frequent use of first-person pronouns and self-deprecating terms, were associated with depression in white individuals, these associations did not hold true for Black individuals. Study co-author Sharath Chandra Guntuku from the Center for Insights to Outcomes at Penn Medicine highlighted the surprise at these discrepancies, indicating that previous findings did not universally apply.
Guntuku emphasized that while social media data cannot diagnose depression, it could aid in risk assessment for individuals or groups. Earlier research by the team analyzed language in social media posts to assess communities’ mental health during the COVID-19 pandemic.
Brenda Curtis from the U.S. National Institute on Drug Abuse at the National Institutes of Health, who collaborated on the study, noted that language indicating depression in social media posts has shown potential in predicting outcomes for patients with substance abuse disorders, such as treatment dropout and relapse.
The study’s findings underscore the need for AI models to be trained on diverse datasets to ensure accurate and equitable outcomes, particularly in healthcare applications where racial disparities in health outcomes are prevalent.
