Learning from the past, structuring the future: using large language models to unlock a century of paediatric research in <i>Archives of Disease in Childhood</i>.

Abstract

The centenary of <i>Archives of Disease in Childhood</i> (<i>ADC</i>) presents an opportunity to reflect on a century of paediatric research and consider how best to leverage this ever-growing repository for future use. While content is indexed via PubMed and medical subject headings terms, this provides a superficial representation of complex journal content, leading to limited accessibility. We discuss the potential utility of large language models (LLMs)-advanced artificial intelligence systems that can understand, summarise and generate human-like language-and demonstrate their feasibility for structuring historical <i>ADC</i> articles, proposing a future pipeline to enhance indexing, retrieval and discoverability. For demonstrative purposes, five articles from <i>ADC</i> December 1999 issue were locally downloaded and processed using a closed deployment of an LLM, Mistral (V.0.3, 7B). A structured prompt was used to extract key metadata. Outputs were manually compared with source texts and scored for accuracy. Hallucinations, fabricated or incorrect outputs, were recorded. The LLM achieved a mean accuracy of 86.9%, aligning with previous benchmarks for medical research assistance. No hallucinations were identified. Some repetition and verbosity were noted, likely due to chunk-based processing, but key fields were accurately extracted when explicitly present. <i>ADC</i> holds a vast but underutilised body of research. This article shows that lightweight, locally hosted LLMs could structure <i>ADC</i> content without compromising intellectual property. Such methods could enable improved access, support automation of systematic reviews and enhance discoverability through biomedical ontologies, laying the foundation for a searchable, semantically enriched <i>Archive</i>s that bridges historical insight with modern research needs.