AI Model Spots Depression Risk in Gay and Bisexual Men Using Minority Stress Signals

Men who have sex with men carry a substantially elevated burden of depression compared with the general population, yet the screening tools used to detect it were almost never built for them. A new study from Chongqing Medical University, published in BMC Psychology, shows that an interpretable machine learning model can identify which individuals face the highest risk of depressive symptoms by weaving together the specific psychosocial stressors that sexual minorities encounter. The work, led by Xinjing Liu and colleagues, offers a template for moving depression screening beyond generic questionnaires toward population-specific, data-driven risk assessment.

The research team recruited 1,369 men who have sex with men through an online cross-sectional survey conducted between November and December 2022, using a non-probability sampling approach. Rather than relying on demographic variables alone, the investigators grounded their candidate risk factors in the Minority Stress Model, a framework that holds that sexual minorities face chronically elevated stress from stigma, prejudice, and identity-related pressures that the general population does not experience. The collected variables spanned sociodemographic characteristics, MSM-related stigma, mental resilience, social support, and MSM sexual identity, with depressive symptoms serving as the primary outcome the models were trained to predict.

Methodologically, the study took a deliberately systematic approach to model construction. The researchers combined three feature selection methods with ten different machine learning algorithms, building a grid of candidate models whose performance could be compared head to head. Feature selection techniques such as the Least Absolute Shrinkage and Selection Operator, known as LASSO, help prune the variable space by penalizing weak predictors, which reduces overfitting and improves generalizability. Among all the algorithms evaluated, LightGBM, a gradient boosting framework that builds an ensemble of decision trees in a leaf-wise growth pattern, emerged as the clear winner, achieving an area under the receiver operating characteristic curve, or AUC, of 0.945 and outperforming its rivals across multiple evaluation metrics.

An AUC of 0.945 is a striking figure in clinical prediction. The AUC measures the probability that the model ranks a randomly chosen depressed individual as higher risk than a randomly chosen non-depressed one, with 0.5 representing chance-level performance and 1.0 representing perfect discrimination. Values above 0.9 are generally considered excellent, and they are rare in mental health prediction, where outcomes are shaped by a tangle of biological, social, and psychological influences. The authors attribute much of this performance to the integration of multidimensional features that generic screening instruments ignore, particularly the minority-specific stressors that the standard tools were never designed to capture.

Crucially, the team did not stop at prediction accuracy. Because black-box models are difficult to deploy in clinical settings where clinicians need to understand why a patient was flagged, the researchers applied SHAP analysis, or SHapley Additive exPlanations, to interpret the contribution of each risk factor to the model’s output. SHAP values, borrowed from cooperative game theory, allocate the prediction of each individual case fairly across the input features, revealing both the direction and magnitude of each variable’s influence. This interpretability layer showed that mental resilience, identity superiority, experiences of stigma, and internalized homophobia were among the key drivers of depression risk in the model.

The prominence of these particular factors is scientifically meaningful. Internalized homophobia, the direction of negative societal attitudes toward one’s own sexual identity, and experienced stigma are classic distal and proximal stressors in minority stress theory, while low mental resilience acts as an amplifier that weakens the capacity to buffer those stressors. Identity superiority, a dimension of how individuals evaluate their own sexual identity relative to others, adds a less commonly measured but apparently informative signal. Together, the findings suggest that the psychological machinery linking stigma to depression can be quantified and ranked, giving researchers and clinicians a data-driven map of which levers matter most.

To translate the model into something usable at the point of care, the investigators converted the predicted probabilities of the optimal model into risk scores, which were then used for risk stratification. The stratification analysis revealed that individuals placed in the high-risk group had a significantly higher prevalence of depressive symptoms than those in the low-risk group, a difference that was statistically robust at P < 0.001. This step matters because raw probabilities are hard to act on, whereas discrete risk tiers map naturally onto clinical workflows, allowing scarce mental health resources to be concentrated on the individuals most likely to benefit from early intervention.

The team went one step further and integrated the prediction results of the optimal model into web-based applications that support probability estimation and interactive decision-making. Visualization tools were employed to demonstrate the model’s classification threshold and its potential practical utility, making the underlying logic transparent to users rather than hidden inside an opaque algorithm. Such tools could, in principle, allow a counselor or public health worker to enter a client’s profile and receive an interpretable risk estimate along with the specific factors pushing that estimate up or down, which is precisely the kind of actionable output that interpretable machine learning is meant to deliver.

The study’s context is important for understanding its motivation. In China, the mental health burden among men who have sex with men remains underestimated, and there has been a notable lack of precise assessment approaches that integrate multidimensional psychosocial factors. Conventional depression screens, calibrated on general-population samples, systematically miss the minority-specific pathways to depression, which means that high-risk individuals can go unidentified until symptoms become severe. By building the feature set directly from the Minority Stress Model, the researchers aimed to close that gap and to demonstrate that population-specific modeling can outperform one-size-fits-all instruments.

The authors conclude that machine learning models can effectively identify MSM individuals at high risk for depressive symptoms by integrating multidimensional features, and that risk stratification informed by model-derived probabilities, together with web-based applications, could assist with earlier recognition and more tailored support strategies. If validated in longitudinal and more diverse samples, this approach could contribute to improving the personalization of mental health screening for sexual minority populations, shifting the field from reactive diagnosis toward proactive, targeted prevention.

Subject of Research: Interpretable machine learning prediction of depression risk among men who have sex with men using the Minority Stress Model

Article Title: Interpretable machine learning for identifying high-risk depression among men who have sex with men: an online cross-sectional study based on the Sexual Minority Stress Model

Article References: Liu, X., Chen, H., Liu, J., Zhou, T., Lin, B., & Zhong, X. (2026). Interpretable machine learning for identifying high-risk depression among men who have sex with men: an online cross-sectional study based on the Sexual Minority Stress Model. BMC Psychology. https://doi.org/10.1186/s40359-026-05626-8

Image Credits: AI Generated

DOI: 10.1186/s40359-026-05626-8

Keywords: machine learning, depression, men who have sex with men, minority stress model, LightGBM, SHAP, risk stratification, mental health, internalized homophobia, stigma, predictive modeling, sexual minority health