Scientists at Shenzhen University in China have developed an artificial intelligence model that predicts major depressive disorder up to four years in advance.
The computational biomarker analyzes how the human brain processes facial expressions in order to identify neural activity patterns associated with a future risk of depression. Major depressive disorder is a common mental health condition affecting more than 332 million people worldwide, according to World Health Organization figures cited in the study.
Led by Lu Han, an assistant professor at Shenzhen University's School of Artificial Intelligence, the research team published its findings this month in the peer-reviewed journal Science Advances. Shenzhen University is a public research institution based in Guangdong province in China.
Brain scanning across European youth
The research team built its model using data from IMAGEN, a longitudinal study tracking adolescent health and brain development across several European countries. Participants underwent functional magnetic resonance imaging, or fMRI, scans at 19 years old while viewing human faces displaying a range of emotional expressions.
Alongside the neuroimaging scans, the 19-year-old participants completed standardized questionnaires to evaluate their mental health symptoms. Researchers monitored the participants until age 23 to observe which individuals subsequently developed clinical depressive symptoms.
Speaking to the news outlet Global Times, Lu explained that individuals without depression easily distinguish emotional changes in facial expressions and respond appropriately, such as reacting with kindness to a smile. In contrast, individuals affected by depression tend to assume that others are angry with them.
Deep learning and facial threat signals
Because of this perceptual tendency, the investigators focused their analysis on angry expressions, treating them as a social threat signal linked to the interpersonal difficulties characteristic of depression. The team constructed a deep learning model designed to emulate how the human brain processes visual input and encodes abstract emotional concepts such as anger.
The results demonstrated that 19-year-olds whose neural responses to facial expressions were biased toward negative emotions or negative memories had the highest likelihood of developing a form of depression by age 23. Functional magnetic resonance imaging measures neural activity by detecting blood flow changes, allowing researchers to observe real-time brain responses to visual stimuli.
The team also identified a connection between the computational biomarker and a specific genetic variant designated as rs11123030. Lu indicated that this correlation suggests inherited genetic factors influence how an individual perceives and processes the emotional expressions of others.
Clinical testing and disorder specificity
To verify the accuracy of the biomarker, the researchers tested the algorithm against data from Stratify, a separate clinical trial comparing depression with other psychiatric conditions such as alcoholism, binge eating, and drug abuse.
The Stratify trial evaluated a cohort of more than 400 adolescents experiencing various mental health conditions. Out of that group, abnormal biomarker values appeared exclusively in the 134 participants who held a formal clinical diagnosis of major depressive disorder.
Patients displaying symptoms of addiction, anorexia, or bulimia recorded biomarker levels similar to those seen in participants with no signs of mental health issues. Lu noted that these comparative results support the specific diagnostic capability of the artificial intelligence model for major depression.
Environmental factors and future applications
Lu emphasized that the biomarker provides predictive insights independent of family stress and socioeconomic circumstances, though it does not replace those environmental risk factors. He stated that depression is not driven by a single cause, but rather emerges from complex interactions between genetic factors, brain development, and life experiences.
Before the tool can achieve widespread clinical application, Lu noted that the findings must be validated in middle-aged adult populations and across diverse ethnic groups. Current clinical testing has focused primarily on adolescent cohorts.
Looking further ahead, Lu suggested the study could advance robotic perception systems by helping machines mimic human emotional interpretation more accurately. He described this direction as a step toward embodied artificial intelligence capable of recognizing subtle emotional states through facial cues.
