Structured Prompt-Based NLP for Infection-Related Classification in Clinical Documentation

Abstract

Infectious diseases remain a major contributor to morbidity in hospital settings, and clinically relevant information is often embedded within unstructured Electronic Health Record (EHR) narratives. This study presents a structured prompt-based Large Language Model (LLM) Natural Language Processing (NLP) pipeline for producing infection-related classifications from discharge summaries. The study was designed as a preliminary retrospective methodological feasibility evaluation rather than as a clinical validation study. The pipeline used a structured infection-focused prompt to generate constrained binary classifications and supporting evidence fields. CSV files containing disease-related information were provided as illustrative contextual materials rather than as a verified retrieval, validation, or knowledge-grounding mechanism. Forty randomly sampled MIMIC-III discharge summaries were compared with clinician-annotated reference labels in a single-pass illustrative evaluation. Raw classification counts were reported only to describe workflow behavior in this limited sample. Formal performance metrics were not reported because the small sample size, retrospective design, stochastic inference setting, and absence of repeated-run agreement analysis precluded stable estimation. The contribution of this work is therefore methodological and conceptual, and the findings should not be interpreted as evidence of diagnostic performance, clinical utility, or operational readiness.