AI reveals race-based differences in the expression of depression
New study shows that language-based predictive models for detecting depression severity may generalize well to white Americans, but not to their Black counterparts.
Source: Getty Images
By - Researchers from the University of Pennsylvania, Philadelphia, and the National Institute on Drug Abuse (NIDA) have demonstrated that artificial intelligence (AI) models designed to predict depression severity using language found in social media posts perform significantly better on white Americans than they do on Black individuals, according to a study published in Proceedings of the National Academy of Sciences.
The research team noted that depression and an individual’s use of language are correlated, highlighting the potential utility of predictive models able to accurately use language data to detect the condition.
However, the researchers further indicated that language use varies based on demographic features like gender and age. Despite this, there is limited research examining whether and how the relationship between depression and language may be impacted by race.
To close this research gap, the team assessed the effect of race on the depression-language association by using AI to analyze relevant language features – negative emotions and first-person pronouns – in a matched cohort of 868 English-speaking Black and white Americans with self-reported depression.
The language in social media posts has also been shown to provide useful insights into users’ mental health, leading the researchers to analyze participants’ Facebook posts.
Multiple models were trained for the task, either on Facebook language used by white participants or their Black peers.
Models trained on data from white participants showed strong predictive performance when tested on the white subgroup. However, when the same models were trained using data from Black participants, they performed poorly when tested on the Black subgroup, demonstrating only slightly better performance when applied to the white cohort.
The analysis further revealed that Black and white people with depression used different language when expressing their thoughts on Facebook.
In white participants, depression severity was correlated with increased use of first-person singular pronouns such as “I,” “me” and “my.” However, this association was absent for Black participants. Similarly, white people tended to use more language describing feelings of belongingness, self-deprecation, self-criticism, despair and being an anxious outsider as depression severity increased, but this association was also not present for Black participants.
The research team hypothesized that these results suggest that depression may manifest differently in language for some Black people; factors like speech rate or tone, rather than word selection, may be more relevant.
The researchers underscored that their findings necessitate further exploration into the effects of depression on natural language across diverse populations in order to enhance prediction models and clinical care.
“Our research represents a step forward in building more inclusive language models. We must make sure that AI models incorporate everyone's voice to make technology fair for everyone,” said senior author Brenda Curtis, PhD, MsPH, chief of the Technology and Translational Research Unit in the Translational Addiction Medicine Branch at NIDA’s Intramural Research Program in a news release. “Paying attention to the racial nuances in how mental health is expressed lets medical professionals better understand when an individual needs help and provide more personalized interventions.”
This research is one of a host of efforts seeking to apply advanced analytics to improve mental healthcare.
Researchers from Weill Cornell Medicine, Columbia University, UC Berkeley School of Public Health, the University of Hong Kong and the University of Kentucky recently used analytics to identify five preventable suicide risk profiles.
Using information related to toxicology, factors precipitating death, means of death and other data, the researchers found that mental health disorders alone; comorbid mental health and substance use disorders; polysubstance use; crisis, alcohol-related, and intimate partner problems; and physical health problems formed the basis for distinct suicide risk classes that could be used to inform prevention strategies.
