KINGSTON, R.I. – November 17, 2025 – Depression, a prevalent mental health condition, affects more than 10% of the U.S. population, or approximately 35 million people. This number has skyrocketed significantly in the aftermath of the COVID-19 pandemic.
Despite affecting millions of Americans, tools that help identify those at risk remain outdated and inadequate in an increasingly complex mental health landscape.
Historically, assessment of depression has relied on the Patient Health Questionnaire-9. This is a self-administered survey that, although widely used, has been shown to have limitations in early detection and nuanced assessment.
But Tingting Chao, assistant professor of business analytics and artificial intelligence at the University of Rhode Island’s School of Business, is confronting the limitations of traditional screening from a new perspective.
Zhao recently published research in IEEE Transactions on Affective Computing suggesting that medical professionals should use artificial intelligence as a diagnostic aid.
“We were looking to determine whether a person develops depression,” Zhao said.
Zhao’s research utilizes text data from three different forms of communication: clinical interview transcripts, SMS text messages, and typed responses to open-ended questions.
“Our motivation was to determine whether we could develop machine learning methods to accurately identify people who may be affected by depression,” Zhao said.
The algorithm Zhao employed is XGBoost, an advanced machine learning framework that builds decision trees and examines the output to detect complex patterns. In this context, identify linguistic indicators such as specific words, expressions, and emotional tone that may indicate a tendency toward depression.
“We put a lot of wood together,” Zhao said.
By applying machine learning to a text dataset, Zhao generated PHQ-9 scores and identified text markers that indicate symptoms of depression. Her findings revealed that the algorithm predicted early signs of depression with remarkable accuracy.
Of the participants whose clinical interviews were analyzed, 41% showed signs of depression. That percentage increased to 46% for typed responses and jumped to 61% for those whose SMS communications were evaluated.
“Ultimately, we’re going to use all of these different markers and features in combination to reach our conclusions,” Zhao said.
Zhao’s model also revealed specific linguistic cues that correlate with depressive tendencies.
In clinical interview transcripts, negative emotions were an early predictor of someone’s likelihood of developing depression. The model identified “love” and “communication” as the most important features in predicting underlying depressive symptoms in the input replies. Although these words generally convey positive connotations, their frequent use in written responses was statistically associated with underlying psychological distress. This suggests that people experiencing depression are expressing a strong desire for love, connection, or understanding, which may manifest linguistically through such words.
Zhao says the future applications are limitless. She claims that this information can be collected using mobile apps. However, she says clinically certified experts should always be involved in treatment and data review.
“Ultimately, we hope this will be something that doctors can use as part of their pre-clinical diagnostic screening,” Zhao said.
