Heterogeneity of depression and suicidal ideation among college students in Eastern China: a latent profile analysis and machine learning approach

To investigate the heterogeneity of depression and suicidal ideation among Chinese college students and to examine the roles of family and individual psychological factors in differentiating risk profile. A total of 4,368 college students completed measures of depression, suicidal ideation, parental rejection, parent-child relationship, psychological resilience, and self-control. Latent profile analysis (LPA) was applied to identify distinct subgroups based on depression and suicidal ideation. Multinomial logistic regression was conducted to examine factors associated with profile membership, controlling for demographic variables. In addition, eleven machine learning models were compared to evaluate predictive performance and explore the relative importance of risk factors. LPA identified three distinct profiles: a low-risk group (71.09%), a moderate-risk group (23.83%), and a high-risk group (5.08%). Multinomial logistic regression indicated that higher parental rejection significantly increased the likelihood of belonging to higher-risk profiles, whereas stronger parent-child relationships, greater psychological resilience, and higher self-control served as protective factors. Notably, parent-child relationship primarily differentiated low- and moderate-risk groups, while psychological resilience and self-control played a more prominent role in distinguishing moderate-risk and high-risk profiles. Among the machine learning models, CatBoost demonstrated the best overall predictive performance, and feature importance analysis consistently identified parental rejection as the strongest predictor, followed by psychological resilience and self-control. Depression and suicidal ideation among Chinese college students exhibit substantial heterogeneity. Family-related risk, particularly parental rejection, and individual psychological resources jointly shape risk profiles, with their relative influence varying across severity levels. Integrating person-centered analysis with machine learning offers a robust framework for identifying high-risk subgroups and informing targeted prevention and intervention strategies.