Large Language Model-Based Clinical Data Extraction for Structured Health Information Systems: Accuracy, Explainability, and Privacy Analysis

Abstract

Clinical documentation contains critical patient information buried within unstructured narrative text that remains largely inaccessible for systematic analysis, quality improvement, and clinical decision support. This research develops and evaluates large language model-based systems for extracting structured clinical data from unstructured medical records while addressing the critical dimensions of accuracy, explainability, and privacy preservation. The study implements fine-tuned versions of GPT-4, Claude, and domain-specific clinical language models to extract key clinical entities including diagnoses, medications, procedures, laboratory values, and clinical relationships from 50,000 de-identified clinical notes. The LLM-based extraction achieved 94.3% accuracy for entity recognition and 89.7% for relationship extraction, substantially exceeding traditional natural language processing methods at 78.4% and 71.2% respectively. Explainability analysis through attention visualization and chain-of-thought prompting demonstrated that models focused appropriately on relevant clinical context in 91% of extraction decisions. Privacy evaluation confirmed that differential privacy techniques maintained extraction accuracy above 92% while providing strong privacy guarantees preventing patient re-identification. The system reduced manual chart review time by 83% while maintaining clinical validity scores of 96.2% when validated by practicing physicians. This research contributes practical frameworks for deploying LLM-based clinical data extraction in healthcare settings, balancing the competing demands of accuracy, transparency, and patient privacy essential for real-world clinical implementation.