Retrieval-Augmented Generation Architectures for Longitudinal Pediatric EHR Analysis

Abstract

This paper evaluates Retrieval-Augmented Generation models for longitudinal electronic health records of pediatrics. Pediatirc electronic health records include rich clinical information about a patient throughout several healthcare visits. The language models tend to overlook historical contexts while conducting clinical analysis processes. However, the RAG integrates information retrieval with Large Language Models and provides grounded clinical responses. In such a manner, the RAG retrieves the pertinent medical records of the patients before generating the evidence-based clinical summaries and recommendations. The present study employed qualitative methodology that involved secondary data from the healthcare literature and datasets. The inductive strategy was used in the systematic interpretation of the published clinical evidence. The results indicated an improvement in the retrieval of longitudinal patient histories during multiple visits. The RAG framework incorporated diagnoses, medications, lab results, immunizations, and clinical notes. The retrieved information helped to provide effective clinical decision support using evidence-based reasoning and output. The analysis demonstrated better continuity of care using integrated patient timeline records. Privacy was still important since the records of children contain confidential developmental and family history details. RAG ensured that unnecessary exposure was minimized by only extracting clinical information needed for the analysis. Age-related interpretations of growth, developmental, milestones, medications, and laboratory value were some of the key findings. Retrieval ensured that time spent on reviewing pediatric clinical documentation was minimized while providing care. The findings have shown that RAG ensures accurate privacy-preserving and efficient analysis of pediatric EHR.