Large language models for electronic health records in pediatric and surgical care: A systematic review

Large language models (LLMs) are promising tools in healthcare, particularly for accessing unstructured, text-based electronic health record (EHR) data. This systematic review evaluates the applications of LLMs in the EHR for pediatric and surgical care, model performance compared to traditional methods, and proposed clinical potential to improve healthcare processes, patient outcomes, and overall quality of care.

A systematic search of ten databases from inception to November 2024 was conducted according to PRISMA guidelines, focusing on LLM use in EHRs. Two reviewers independently screened studies for inclusion, with a third reviewer resolving conflicts. Risk of bias was assessed using PROBAST.

Among 4326 identified studies, 44 met the inclusion criteria - 30 (68.2 %) in all surgical specialties, 3 (6.8 %) in pediatric surgical subspecialties, and 11 (25 %) in pediatrics. Most studies (59.1 %) were published in 2024. LLM types included Bidirectional Encoder Representations from Transformers (BERT) and their variants (52.3 %), ChatGPT (29.5 %), and other models (18.2 %). Most studies (90.9 %) relied solely on retrospective unstructured data, and 40.9 % focused on classification tasks. LLMs demonstrated performance improvements in 78.1 % of studies with a traditional comparator. Clinical documentation assistance (54.5 %) and diagnostic and clinical decision support (36.4 %) were the most commonly proposed applications for LLMs.

While LLMs offer opportunities for EHR analysis in pediatric and surgical care, most studies remain early-stage, with notable limitations including limited external validation and lack of evaluation in actual clinical workflows. Future research should prioritize rigorous validation and real-world testing to support their safe and effective use in practice.