Inferring suicidal ideation without direct questions: A network-augmented machine learning analysis of depression...

Suicidal ideation is assessed through direct screening questions, but depressive-symptom configurations may provide complementary information relevant to suicidal-ideation endorsement. Using the Network-Augmented Machine Learning Utility (NAMU), we examined the incremental predictive value of person-specific symptom-network features derived from PHQ-8 responses beyond symptom models.

Models were trained in the Korea Community Health Survey (KCHS; N = 231,469) and externally validated in the Korea National Health and Nutrition Examination Survey (KNHANES; N = 21,473). We compared a PHQ-8 total-score logistic-regression baseline, symptom-only models, and NAMU-expanded models combining PHQ-8 items with 37 network-derived features. Elastic-net logistic regression, random forest, LightGBM, and XGBoost were evaluated. Incremental PR AUC was tested using paired bootstrap comparisons, and SHAP-guided feature reduction identified a parsimonious specification.

Item-level XGBoost outperformed the PHQ-8 total-score baseline, but network augmentation provided little incremental benefit beyond item-level XGBoost. In KCHS, PR AUC was .5570 for symptom-only XGBoost and .5573 for NAMU-full XGBoost (Δ = .0003, 95% CI [-.0035, .0033], Holm-adjusted p = .958). A six-feature NAMU model preserved discrimination (PR AUC = .5596; ROC AUC = .9234). In KNHANES, NAMU-full achieved PR AUC = .4931 and ROC AUC = .9010, with no significant PR AUC improvement over symptom-only XGBoost (Δ = .0030, 95% CI [-.0036, .0098], Holm-adjusted p = .362).

Most predictive gain arose from modeling PHQ-8 symptoms rather than network augmentation. NAMU serves as a feature-representation framework for person-specific symptom co-activation while maintaining performance comparable to symptom-only modeling. The models predict item-9 endorsement rather than independently verified or concealed suicidal ideation.