Enhanced prediction of suicidal ideation using dimensional subdomains from DSM-5 Level 1 Cross-Sectional Symptom Scale with neural networks

Identification of suicidal ideation is crucial for clinical intervention and prevention.The study presented below attempts to enhance the predictive validity of neural network models based on suicidal ideation with the help of indirect psychometric indications, but without any direct questioning with regard to self-harm.We analyzed a dataset of 935 individuals aged between 18 and 40 years (mean age 23.5 years) obtained from the Koç University Psychotherapy Center. The initial model included demographic information, lifestyle factors, items of the DSM5 Level 1 Cross-Sectional Scale (except item 11, which directly assesses self-harm) and total scores of the MDQ, SCOFF, PHQ9, GAD7, ASRS, CAGE and WHODAS scales. Suicidal ideation as defined by 9th item of PHQ9. In this item not at all, represented by zero remained as same and other numbers indicating any suicidal ideation (1-3) were considered as 1. Thus, the target variable was dichotomized. The next model utilized the dimensional subdomains of the DSM-5 Level 1 Scale and aimed to reduce the number of input features while maintaining prediction accuracy.The initial model attained a validation AUC of 0.795, demonstrating strong predictive ability. The refined model, using DSM-5 subdomains, slightly improved the validation AUC to 0.798.Our findings indicate that neural networks using indirect predictors from dimensional subdomains of DSM-5 can effectively forecast suicidal ideation, a testament to the potential of machine learning in enhancing mental health assessments.