AI powered patient records analysis for injury surveillance in children and adolescents - a feasibility study from Switzerland.

OBJECTIVE: This study investigates the feasibility of using an automated approach to extract injury-related information from narrative text in a paediatric emergency department to improve data basis on injuries among children and adolescents in Switzerland. METHODS: The dataset comprises paediatric injury cases treated between 2018 and 2022 at the University Children's Hospital Zürich emergency department (N = 30,876). Model development involved (1) adapting EU-IDB, a domain-specific hierarchical coding-system; (2) manual data annotation; and (3) fine tuning and (4) evaluation of a transformer-based model for multi-label text span classification. RESULTS: Interrater reliability improved from moderate (κ = 0.45) to substantial (κ = 0.73) following adaptation of the coding system. Variable level classification performance was encouraging across three training variants (macro F1: ALL = 0.57, IND = 0.64, VAR = 0.63). Performance declined at lower levels, particularly at level 3 (macro F1: ALL = 0.08, IND = 0.23). Automated prevalence estimates correlated strongly with manual annotations (ρ = 0.948). CONCLUSION: Automated text span classification of injury-relevant information from electronic patient records shows promising first results to provide valuable information for injury surveillance and prevention, missing in Switzerland. However, more annotated data and detailed validation are needed to draw robust conclusions.