Automated Surgery Delay Detection from Synthetic Pediatric Clinical Notes Using Large Language Models

Abstract

Prompt identification of surgery delays is imperative for enhancing the quality of pediatric surgery and optimizing hospital operations. Nevertheless, delay information is contained in unstructured medical notes, which makes manual extraction of such information cumbersome and inconsistent. The current study focuses on using Large Language Models (LLMs) to detect delays in surgery through processing synthetic pediatric clinical notes. In this study, a qualitative research design was used, together with secondary synthetic clinical data to maintain patient confidentiality and at the same time retain realistic pediatric surgery contexts. Inductive reasoning was employed in the analysis process, and thematic analysis was performed to reveal common patterns and categories of delays in clinical stories. It was found out that LLMs could effectively identify clinically relevant delay markers from unstructured text. The LLMs were also able to interpret temporal data and establish a timeline of clinical events, thereby determining the stage when delays happened. Analysis of themes showed that there were similar delays happening at each step of the pediatric surgical pathway, thus proving the usefulness of language context understanding in clinical document analysis. It is concluded that LLMs serve as an effective tool for automated detection of surgeries delay in the clinical domain and for clinical text processing. Synthetic clinical documents give a safe ground for creation and evaluation of decision support systems based on LLMs without raising any patient privacy concerns.