Machine learning based identification of suicidal ideation using non-suicidal predictors in a university mental health clinic.
Journal:
Scientific reports
PMID:
40263508
Abstract
Suicide causes over 700,000 deaths annually worldwide. Mental disorders are closely linked to suicidal ideation, but predicting suicide remains complex due to the multifaceted nature of contributing factors. Traditional assessment tools often fail to capture the interactions that drive suicidal thoughts, underscoring the need for more sophisticated predictive approaches. This study aimed to predict suicidal and self-harm ideation among university students using machine learning models without relying on suicidal behavior related predictors. The goal was to uncover less obvious risk factors and provide deeper insights into the complex relationships between psychiatric symptoms and suicidal ideation. Data from 924 university students seeking mental health services were analyzed using seven machine learning algorithms. Suicidal ideation was assessed through the 9th item of the Patient Health Questionnaire-9. Three predictive models were developed, with the final model utilizing only subdomains from the DSM-5 Level 1 Self Rated Cross-Cutting Symptom Measure. Feature importance was assessed using SHAP and Integrated Gradients techniques. To ensure model generalizability, the best-performing model was externally validated on a separate dataset of 361 individuals. Machine learning models achieved strong predictive accuracy, with logistic regression and neural networks reaching AUC values of 0.80. The final model achieved an AUC of 0.80 on the training data and 0.79 on external validation data. Key predictors of suicidal ideation included personality functioning and depressed mood (both increasing the likelihood), while anxiety and repetitive thoughts were associated with a decreased likelihood. The use of non-suicidal predictors across datasets highlighted psychiatric dimensions relevant to early intervention. This study demonstrates the effectiveness of machine learning in predicting suicidal ideation without relying on suicide-specific inputs. The findings emphasize the critical roles of personality functioning, mood, and anxiety in shaping suicidal ideation. These insights can enhance early detection and personalized interventions, especially in individuals reluctant to disclose suicidal thoughts.