AI-assisted identification and information extraction of self-harm cases from emergency department reports

The collection and use of medical data have long been integral to healthcare systems. However, the increasing volume of available information makes the identification of relevant cases increasingly challenging [1]. For a study on self-harming behaviour, approximately 160,000 emergency department reports were analysed to identify cases of self-harm and extract clinically relevant information.

As a manual review of all reports would have been impractical, a TF-IDF (Term Frequency–Inverse Document Frequency) vectorizer was used for an initial preselection. Based on around 50 manually classified example reports, the model identified 1,266 likely cases of self-harming behaviour and 428 additional potential cases. These reports were subsequently validated using a Large Language Model (GPT-4.1), which can capture linguistic context and semantic relationships more effectively than TF-IDF. Following validation and additional manual review, 1,107 of the initially identified cases and 108 of the potential cases were confirmed as self-harming behaviour.

In a second phase, the LLM was employed for structured information extraction. Relevant characteristics such as wound number and size, affected body regions, type of self-harm, and treatment measures were automatically extracted from the free-text reports and converted into a structured tabular format.

The findings demonstrate that combining a computationally efficient pre-filtering method with LLM-based validation and information extraction provides a practical approach for analysing large-scale clinical text datasets. Such methods may facilitate future monitoring of self-harming behaviour and support the early identification of emerging trends in mental health crises.

References 1. White, S. (2014). A review of big data in health care: challenges and opportunities. Open Access Bioinformatics, 13. https://doi.org/10.2147/oab.s50519