Enhancing Korean adolescent suicide risk prediction with TabR: A generative AI and explainable retrieval-based deep learning approach.

Adolescent suicide is a growing public health crisis, particularly in South Korea, which has one of the highest youth suicide rates among Organization for Economic and Development countries. Suicide is the leading cause of death among South Koreans aged 10 to 39 years, driven by academic stress, mental health conditions, and social pressures. The early identification of at risk individuals is crucial for timely intervention. This study presents a novel application of TabR, a retrieval-augmented residual neural architecture, for predicting suicide attempts among adolescents, using the Korea Youth Risk Behavior Survey dataset from 2020 to 2022, comprising 1,61,646 participants. To address the extreme class imbalance and limited positive cases, we employed a tabular variational autoencoder for synthetic data augmentation, improving minority class representation while preserving feature distribution fidelity. TabR was compared against 8 baseline machines learning and deep learning models: AdaBoost, CatBoost, LightGBM, XGBoost, Random Forest, SAINT, TabNet, and FT-Transformer. TabR trained on TVAE-augmented data achieved the highest accuracy of 0.9726, AUC of 0.8963, and F1-score of 0.4754. While our findings demonstrate notable improvements in performance, limitations remain, particularly regarding the dataset size, class imbalance, and lack of external validation. Future research should prioritize external validation in independent cohorts, prospective or longitudinal study designs, and explicit evaluation of clinical impact and potential harms (including consequences of false positives and false negatives) before any consideration of deployment.