KMAN Publication Inc. (KMANPUB) Journal of Adolescent and Youth Psychological Studies (JAYPS) 2981-2526 7 Serial Number 43 2026 03 10 Deep Learning Classification of Suicidal Ideation from Electronic Health Records Deep Learning Classification of Suicidal Ideation from Electronic Health Records 1 10 10.61838/kman.jayps.5153 EN Journal Article 2025 11 04 <table> <tbody> <tr> <td> <p><strong>Objective:</strong>  To develop, validate, and evaluate the performance of multimodal deep learning architectures in the automated classification and early detection of suicidal ideation utilizing comprehensively extracted electronic health record data.</p> <p><strong>Methods and Materials:</strong> This retrospective observational cohort study utilized a dataset of unique patient electronic health records from Armenian healthcare facilities spanning to . Data extraction included both structured clinical variables (demographics, diagnosis codes) and unstructured clinical narratives (progress notes). Unstructured text was processed using advanced Natural Language Processing pipelines. Model development involved an /  train-test split, employing the Synthetic Minority Over-sampling Technique (SMOTE) to mitigate class imbalance. We evaluated multiple architectures, including Bidirectional Long Short-Term Memory networks, standalone clinical Transformers, and a multimodal Deep Neural Network integrated with a Transformer via late fusion.</p> <p><strong>Findings:</strong> The overall prevalence of suicidal ideation within the cohort was (  out of records), with significant baseline differences ( ) observed in psychiatric history and demographic distributions between the ideation and non-ideation groups. The multimodal DNN+Transformer model demonstrated superior predictive performance, achieving an Area Under the Curve (AUC) of , an overall accuracy of , and an F1-score of . By comparison, the standalone text-based Transformer achieved an AUC of . Model attention mechanisms revealed that textual tokens such as “hopelessness” and structured features including Major Depressive Disorder and prior suicide attempts were the most heavily weighted predictive variables.</p> <p><strong>Conclusion:</strong> The integration of structured clinical data and unstructured clinical narratives through multimodal deep learning architectures significantly outperforms single-modality approaches in classifying suicidal ideation. This validates the use of advanced computational modeling within existing electronic health record systems as a proactive, highly accurate tool for early clinical decision support.</p> </td> </tr> </tbody> </table> https://journals.kmanpub.com/index.php/jayps/article/download/5153/9348