Title: Enhancing Self-Harm Monitoring in Emergency Triage Notes via an Evidence-Enhanced Machine Learning Framework

Abstract: Although self-harm represents a significant public health challenge, existing surveillance systems dependent on hospital admission data are often insufficient due to the limited sensitivity of diagnostic coding. Emergency Department (ED) triage notes, captured at the first point of patient contact, offer a concise overview of visits and present a valuable avenue for detecting self-harm incidents. To address this, we designed a three-phase methodology that integrates conventional machine learning with large language model techniques for both screening and evidence extraction. This system was evaluated for its ability to generalize across three distinct Australian hospitals. The model demonstrated high performance, yielding Area Under the Precision-Recall Curve (AUPRC) scores of 0.887 ± 0.016 and 0.884 ± 0.012 during internal and external validation phases, respectively. In prospective testing, the system achieved an AUPRC of 0.881 ± 0.008 at the development location, while maintaining robust performance at two external sites—recording scores of 0.879 ± 0.012 and 0.816 ± 0.015—without the need for site-specific model retraining. A critical benefit of this methodology is its capacity to pinpoint the primary method of self-harm with 95% accuracy, thereby facilitating more detailed surveillance capabilities that extend beyond simple binary classification.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC