KINGSTON, R.I. – Nov. 17, 2025 – Depression, a pervasive mental health condition, affects more than 10% of the U.S. population, or roughly 35 million people. That figure has surged markedly in the aftermath of the COVID-19 pandemic.
Despite impacting millions of Americans, the tools that help identify those at risk remain outdated and insufficient for an increasingly complex mental health landscape.
Historically, depression assessments have relied on the Patient Health Questionnaire-9, a self-administered survey that, while widely used, has shown limitations in early detection and nuanced assessment.
But Tingting Zhao, an assistant professor in business analytics and artificial intelligence in the University of Rhode Island’s College of Business, is confronting the limitations of conventional screening through a new lens.
Zhao recently published a study in the journal IEEE Transactions on Affective Computing suggesting that medical professionals should leverage artificial intelligence as a supplementary diagnostic instrument.
“We were looking at trying to determine whether a person is going to develop depression or not,” said Zhao.
Zhao’s research draws on text data from three distinct 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 a machine learning method to accurately identify people who may be impacted by depression,” said Zhao.
The algorithm Zhao employed is XGBoost, a sophisticated machine learning framework that builds decision trees and looks at the outputs to detect intricate patterns. In this context, it identifies linguistic indicators—such as specific words, expressions, or emotional tones that may signal depressive tendencies.
“We put a lot of trees together,” said Zhao.
By applying machine learning to the textual datasets, Zhao generated PHQ-9 scores and identified textual markers indicative of depressive symptoms. Her findings revealed that the algorithm predicted early signs of depression with notable precision.
Among the participants whose clinical interviews were analyzed, 41% exhibited depressive indicators. That proportion increased to 46% for typed responses and surged to 61% among those whose SMS communications were evaluated.
“In the end, we are going to use all these different markers and features together to try and reach a conclusion,” said Zhao.
Zhao’s model also uncovered specific linguistic cues that correlated with depressive tendencies.
In clinical interview transcripts, negative emotions were an early predictor of someone potentially suffering from depression. In typed replies, the model identified “love” and “communication” as features with the highest importance in predicting potential depressive symptoms. Although these words generally convey positive meaning, their frequent use in written replies was statistically associated with underlying emotional distress. This suggests that individuals experiencing depression might express a stronger desire for affection, connection, or understanding, which manifests linguistically through such words.
Zhao says that future applications are limitless. She argues that this information could potentially be collected by using a mobile app. However, she says there should always be a clinically certified professional involved in treatment and reviewing the data.
“At the end of day, we hope this could be something that physicians can use as part of their screenings before their clinical diagnosis,” said Zhao.