Overview
While large language models (LLMs) enable scalable extraction of clinical variables from unstructured electronic health records, sparse documentation or complex clinical text can result in extraction errors. When human-labeled data for evaluation are limited, further assessing performance against unlabeled data can strengthen confidence in model performance. This pilot study evaluated an "LLM-as-a-judge" approach to detect inconsistencies in primary LLM predictions across the UK and Germany Flatiron Health Research Databases.
In the pilot evaluation, a primary LLM extracted clinical variables and generated explanations, while a secondary LLM "judge" reviewed predictions against source documents and guidelines to flag potential errors. In a study of LLM-extracted PSMA PET reports, the LLM judge identified the majority of human-reviewed errors (recall: 60% UK, 75% Germany). In a second study across 12 fields spanning initial management, surgery and recurrence events, the judge flagged predictions in 9 fields, and correcting these flagged predictions improved extraction accuracy (F1-score) for 8 of the 9 fields by 0.4 to 6.5 percentage points.
Why this matters
This research indicates that an LLM-as-a-judge method can identify inconsistencies in LLM-extracted clinical variables. This provides a way to gain confidence in performance when labels are limited, complementing the VALID framework in the case of low-prevalence variables. Flagging errors for expert review (or enabling in-place corrections), offers a method to refine automated data curation and enhance confidence in real-world oncology datasets.