Beyond Clinical Symptoms: A Psychosocial and Machine Learning Approach to Understanding Suicidal Ideation (Preprint)

BACKGROUND Suicide is a major public mental health challenge, and many individuals experiencing suicidal ideation do not meet diagnostic criteria for psychiatric disorders or disclose symptoms during clinical encounters. This limits the effectiveness of symptom-based screening commonly used in mental health services. OBJECTIVE From an epidemiological and social psychiatry perspective, identifying psychosocial risk and protective factors beyond psychopathology is essential for population-level prevention and early intervention. This study examined whether suicidal ideation can be predicted using non-psychopathological psychological measures and identified socially and developmentally meaningful risk and resilience markers through explainable machine-learning models with external validation. METHODS Data were obtained from a mixed clinical and community-based sample of 190 adults. Suicidal ideation was defined using self-reported endorsement of suicide-related thoughts. Predictors consisted exclusively of validated non-pathological psychological measures relevant to social psychiatry and life-course epidemiology, including chronic and perceived stress, childhood adversity, personality traits, attachment styles, coping strategies, and emotion regulation capacities. Five supervised machine-learning algorithms were implemented using stratified five-fold cross-validation. To address demographic confounding and class imbalance, analyses were conducted on both the full dataset and a demographically matched subsample (n=110) balanced for age, sex, education, and income. Hyperparameter optimization targeted both area under the curve (AUC) and F1-score to ensure relevance for low-prevalence outcomes. Model generalizability was assessed using an independent external validation sample of university students (n=84). Model interpretability was examined using SHapley Additive exPlanations (SHAP). RESULTS Random Forest models showed the most robust and generalizable performance. In the matched sample, Bayesian-optimized models achieved an AUC of 0.878 and an F1-score of 0.819. External validation yielded an AUC of 0.791 and an F1-score of 0.549, representing a 2.4-fold improvement over chance-level prediction. Explainable analyses consistently identified perceived loneliness as the strongest predictor of suicidal ideation, followed by childhood emotional abuse, limited access to emotion regulation strategies, and low perceived self-efficacy. Higher conscientiousness and extraversion emerged as stable protective factors. CONCLUSIONS Suicidal ideation can be predicted using psychosocial and developmental indicators independent of psychiatric symptom severity. These findings highlight loneliness and emotion regulation as key modifiable targets for population-level suicide prevention and identify personality-related resilience factors relevant for stratified risk assessment. Integrating such markers into mental health services and surveillance frameworks may enhance early identification beyond diagnostic approaches. CLINICALTRIAL -