Feed

Suicide is the second-leading cause of US childhood mortality after 9 years of age. The accurate measurement of pediatric emergency service use for self-injurious thoughts and behaviors (SITB) remains challenging, as diagnostic codes undercount children. This measurement gap impedes public health and prevention efforts. Current research has not established which combination of electronic health record data elements achieves both high detection accuracy and consistent performance across youth populations.

This study aims to (1) compare the detection accuracy of electronic health record-based methods for identifying SITB-related pediatric emergency department (ED) visits: basic structured data (International Classification of Diseases Version 10, Clinical Modification codes, chief concern), comprehensive structured data, clinical note text with natural language processing, and hybrid approaches combining structured data with notes; and (2) for each method, measure variability in detection by youth demographics and underlying mental health diagnosis.

Multiple human experts reviewed clinical records of 3828 pediatric mental health emergency visits (28,861 clinical notes) to a large health system with 2 EDs (June 2022-October 2024). The reviewers used the Columbia Classification Algorithm for Suicide Assessment to label the presence of SITB at the visit. Random forest classifiers were developed using 3 data modalities: (1) structured data (low-dimensional [International Classification of Diseases codes and chief concerns], medium-dimensional [adding Columbia Suicide Severity Rating Scale screening or mental health diagnoses], and high-dimensional [all structured data or augmented case surveillance, aCS]); (2) text data (general-purpose natural language processing, medical text-specific trained natural language processing, and Large Language Model Meta AI-derived scores), and (3) hybrid data (combining aCS with each text approach). Model performance was evaluated using area under the receiver operating characteristic curve (AUROC).

Of the 3828 visits, 1760 (n=1760, 46.0%) were SITB-related. Detection performance improved with dimensionality: low-dimensional (AUROC=0.865), medium-dimensional (AUROC=0.934-0.935), and high-dimensional (AUROC=0.965). Low-dimensional structured (International Classification of Diseases codes and chief concerns) showed high variability in detection, with lower accuracy among preadolescents (AUROC=0.821 vs 0.880 for adolescents); male participants (AUROC=0.817 vs 0.902 for females); and patients with neurodevelopmental (AUROC=0.568-0.809), psychotic (AUROC=0.718), and disruptive disorders (AUROC=0.703). Hybrid modality (aCS+Large Language Model Meta AI) achieved optimal performance (AUROC=0.977), with AUROC ≥0.90 for all 20 demographic and 12/15 diagnostic subgroups.

This cross-sectional retrospective study identified that, relative to diagnostic codes and chief concern alone, hybrid structured-text detection methods improved accuracy and mitigated unwanted detection variability. The findings offer a scaffold for future clinical deployment of improved information retrieval of pediatric suicide and self-harm-related emergencies.