A clinical LLM passed layered checks—not a full human audit
Across 919,783 US veterans' notes, automated filters rejected 14.59% of positive extractions. The remaining accuracy estimate still relied partly on another language model as referee.
The short version
- What happened
- The study processed 919,783 notes from 178,467 Veterans Health Administration patients across 11 substance-use-disorder categories.
- Who it affects
- Clinical-note extraction could surface overlooked needs, but a wrong substance-use label can affect trust, stigma and decisions long after the model run.
- What remains unproven
- One national health system, weak supervision, selective expert review, a judge-model reference for F1 and no prospective test of patient benefit or harm.
Keep in mind
One national health system, weak supervision, selective expert review, a judge-model reference for F1 and no prospective test of patient benefit or harm.
Source analysis · 1 linked source. Primary analysed source date: 16 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
A large deployment problem met a layered test
A US government team tested a way to validate language-model extraction from 919,783 clinical notes without paying for a complete human-labelled reference set. The records covered 178,467 Veterans Health Administration patients with substance-use-related encounters in January 2023. Across 11 diagnostic categories, rule-based plausibility checks and semantic grounding rejected 14.59% of the model's positive extractions as unsupported, irrelevant or structurally implausible. That is useful evidence that a large first pass needs a second line of defence.
The peer-reviewed article combines prompt calibration, lightweight automated filters, a separate higher-capacity judge model, selective subject-matter-expert review and a final outcome check. A calibration set contained 923 expert-annotated notes from 685 patients and was kept out of post-extraction evaluation. The peer-reviewed version is new, but the same work first appeared as a preprint on 7 April 2026; the publication date should not be mistaken for the first public disclosure.[1]
The accuracy number is not an exhaustive human verdict
For uncertain cases—positive extractions without a matching structured diagnosis—the judge model reviewed about 80,000 notes. Its decisions agreed substantially with expert review on a targeted sample, with Gwet's AC1 of 0.80. Using those judge-model decisions as the reference, the primary model reached a relaxed-match F1 score of 0.80. The circularity matters: this shows agreement within a layered weak-supervision system, not accuracy against a complete, independently human-labelled truth set.
The external check asked whether model-derived alcohol-use diagnoses predicted later engagement with specialty care between February and April 2023. The extracted signal reached an area under the curve of 0.80, compared with 0.76 for structured diagnosis codes and 0.84 when both were combined. Of all patients, 51,552—28.89%—used specialty outpatient care in that window. Predictive validity is encouraging, but attendance reflects access, referral and patient choice as well as diagnostic correctness.[1]
The distinction that matters
What each validation layer can establish
Each stage catches a different error, but none supplies a complete independent ground truth on its own.
Filter
Rules and semantic grounding
14.59% of positives rejected
Adjudicate
Judge LLM plus selective expert review
Agreement, not exhaustive audit
Validate
Predict later specialty-care use
AUC 0.80; no care benefit tested
Useful for research does not yet mean safe for care
For veterans, an extraction system could help researchers find needs hidden in free text. It could also attach a sensitive substance-use label to a person incorrectly. The paper did not test clinical decisions, patient communication, stigma, treatment outcomes or appeals against an automated classification. It also states that the models were tailored to the VA and that results should not be generalised outside that system.
The Department of Veterans Affairs sponsored the work and supplied data and biomedical expertise; the VA–Department of Energy collaboration used Oak Ridge National Laboratory resources. The authors declared no competing interests. The initiative was formally classed as non-research by the VA, while Oak Ridge obtained approval for secondary use of patient data; that distinction does not reduce the need for privacy, bias and consent scrutiny. Confidence would rise with a larger blinded human reference set sampled across positive and negative cases, independent health-system replication, subgroup error rates and a prospective workflow study measuring benefit and harm. The result is a scalable validation recipe with transparent limits—not proof that weak supervision can replace human audit.[1]