Title: Challenges in Generalizing Self-Harm Prediction Models: The Impact of Lexical and Semantic Shifts in Emergency Department Triage Records

Abstract:

Presentations to emergency departments (EDs) involving self-harm are closely linked to an increased risk of suicide. While natural language processing (NLP) systems have demonstrated strong capabilities in identifying self-harm behaviors from triage notes within individual hospitals, their accuracy frequently diminishes when applied across different institutions. This study investigates the underlying reasons for this performance gap by comparing triage documentation from two distinct hospitals, focusing on lexical traits, key predictive features, and prominent topics. The analysis uncovers significant disparities in how self-harm is lexically expressed and which features hold predictive weight across sites, even though fundamental themes like self-injury and self-poisoning remain consistent. These variations in clinical documentation contribute to the decline in cross-institutional model performance. Our results offer valuable insights into how institutional differences influence the detection of self-harm in clinical narratives and suggest avenues for enhancing the generalizability of predictive models.

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