A Multi-Aspect Transformer with Explainable AI for Recognizing Implicit Suicidal and Depressive Risk Indicators

Abstract

Early detection of suicidal ideation and depressive risk remains a critical challenge, particularly when individuals express distress implicitly through metaphorical or obfuscated language. Existing approaches primarily rely on explicit linguistic signals, limiting their effectiveness in real-world settings. This paper proposes a unified multi-aspect transformer-based framework that integrates multi-source learning, multi-task optimization, affective feature fusion, and adversarial training to detect implicit psychological risk indicators in textual data. The model jointly learns suicidal ideation detection, depression severity classification, and perceived threat detection, while incorporating emotional representations derived from valence, arousal, and polarity signals. To improve robustness, an adversarial training strategy is employed to simulate obfuscated expressions, enhancing robustness and generalization under linguistic perturbations. Interpretability is ensured through a hybrid explainable AI approach combining attention mechanisms and SHAP-based feature attribution. Extensive experiments conducted on four benchmark datasets demonstrate that the proposed approach achieves state-of-the-art performance (F1-score = 0.91), with statistically significant improvements over strong baselines. Additional analyses, including ablation studies, adversarial evaluation, and calibration assessment, confirm the effectiveness, robustness, and reliability of the proposed framework. These results highlight the potential of the model for deployment in high-stakes applications such as clinical triage and online risk monitoring, where early and interpretable detection of concealed psychological distress is essential.