Predicting suicidal ideation in academic communities using machine learning methods: a cross-sectional study

Background: Research consistently shows that depression and suicidal ideation (SI) often cooccur. However, SI can arise without elevated depressive symptoms, suggesting that additional factors may also contribute. This study investigated the protective and vulnerability factors associated with SI beyond depressive symptomatology in the academic community. Methods: We employed multiple kernel learning (MKL) to distinguish participants with SI from those without SI. MKL incorporates the contribution of each psychometric instrument and specific items to the predictive model. Data were drawn from a large-scale online survey of the Brazilian academic community (N = 3828; 67.6% women; 33.2% black; mean age 38.28 years; SD 12.89 95% CI: 31.6-33.1). The models incorporated measures of depressive symptoms, optimism, loneliness, childhood maltreatment, and demographic characteristics. Findings: The MKL model accurately distinguished individuals with and without SI, achieving a mean balanced accuracy of 77.61% (95% CI: 77.40-77.82) and an area under the curve (AUC) of 0.862 (95% CI: 0.860-0.864). While depressive symptoms were strong predictors, other variables, such as optimism, loneliness, childhood emotional maltreatment, and demographic characteristics together accounted for half the total weight in the classification model. Interpretation: These findings underscore the need for suicidal ideation screening protocols that consider a broader range of emotional and behavioral factors beyond depressive symptoms, particularly within academic communities. These insights may inform the design of targeted interventions to promote mental well-being in academic settings. Funding: Carlos Chagas Filho Foundation of Research Support in Rio de Janeiro (FAPERJ: E-26.201.678/2022; E-26.201.118/2021).