Background: Traditional suicide risk assessment relies primarily on quantitative measures, yet predictive accuracy remains limited. This study examined whether systematic analysis of qualitative clinical interview content using advanced Natural Language Processing (NLP) enhances suicide risk prediction beyond conventional assessment approaches.


Methods: Data from 1,443 patients recruited from psychiatric emergency departments across ten Spanish university hospitals following suicide attempts were analyzed. Participants completed the Columbia-Suicide Severity Rating Scale (C-SSRS), Brief Symptom Inventory, Barratt Impulsiveness Scale, Reflective Functioning Questionnaire, and Acquired Capability for Suicide Scale. Qualitative C-SSRS responses were classified using BERT-based models into categories including desire to die, method-specific ideation, and preparatory acts. Three hierarchical models were compared: sociodemographic variables only, sociodemographic plus quantitative psychological measures, and a complete model incorporating NLP-derived qualitative features.


Results: BERT-based classification achieved high accuracy across suicide-related categories (77.3%-94.44%), with exceptional performance for active ideation (F1-score: 0.98) and preparatory acts (F1-score: 0.97). The complete multimodal model demonstrated superior predictive performance across all outcomes. Most notably, for Potential Lethality, incorporating qualitative variables increased performance from below-chance (pseudo-R² = -0.002) to excellent prediction (pseudo-R² = 0.895). For Ideation Intensity, explained variance increased from 17.2% to 44.2%.


Conclusions: Systematic integration of qualitative clinical interview content through NLP significantly enhances suicide risk prediction. These findings suggest that incorporating automated analysis of patient verbal expressions into clinical protocols could substantially improve the identification of high-risk individuals.