A multi-stage validation framework for trustworthy large-scale clinical information extraction using large language models

Abstract

Abstract Large language models (LLMs) show promise for extracting clinically meaningful information from unstructured health records, yet real-world deployment is constrained by limited scalable, trustworthy validation approaches. Conventional evaluation relies on annotation-intensive reference standards and incomplete structured data, limiting population-scale use. We propose a multi-stage validation framework for LLM-based clinical information extraction that enables rigorous assessment under weak supervision by integrating complementary validation strategies. The framework combines prompt calibration, rule-based filtering, semantic grounding, judge-LLM evaluation, expert review, and predictive validity analysis to quantify uncertainty without exhaustive manual annotation. We applied this framework to extract substance use disorder (SUD) diagnoses across 11 substance categories from 919,783 clinical notes. Rule-based filtering and semantic grounding removed 14.59% of LLM-positive extractions that were unsupported or structurally implausible. For high-uncertainty cases, the judge LLM showed substantial agreement with subject matter expert. Using judge evaluations as references, the primary LLM achieved an F1 score of 0.80. LLM-extracted SUD diagnoses predicted subsequent engagement in SUD specialty care more accurately than structured-data baselines. By unifying complementary validation strategies, this work presents a generalizable human-in-the-loop framework for establishing the trustworthiness of LLM-based clinical information extraction at scale while minimizing the need for exhaustive manual annotation.