Structured extraction from oncology clinical notes

This is an OncoWiz educational overview of a research area, written for clinicians. It summarises the shape of the evidence and the questions worth asking of it. It is not a summary of any single study, and it reports no individual trial’s results.

Most of the oncology record is unstructured text. Stage, histology, biomarker status and treatment intent are stated in prose in radiology, pathology and clinic notes, which is why extracting them computationally is a precondition for almost every downstream analytic ambition a cancer service has.

What makes clinical text hard

Clinical language is compressed, abbreviated, locally idiosyncratic and heavily negated. The failure modes of extraction systems are therefore systematic rather than random, and they cluster in predictable places:

• Negation and uncertainty. “No evidence of metastatic disease” and “cannot exclude metastatic disease” both contain the target term and mean different things.
• Temporality. A note describing prior disease status, family history or a planned intervention is easily read as current fact.
• Attribution. Findings belonging to a differential, a quoted outside report, or a relative must not be attributed to the patient.
• Local abbreviation. Site-specific shorthand does not transfer between institutions, and neither does a model that learned it.

Extraction accuracy is per-field, not global

A single headline accuracy figure for a document is close to meaningless. Fields differ enormously in difficulty: an explicitly stated biomarker result is far more reliably extracted than a stage that must be inferred across several documents, or a treatment intent that is never stated in so many words. Evaluation should be reported field by field, with the error direction stated.

Large language models change the shape of the problem

General-purpose language models reduce the engineering cost of extraction substantially and handle phrasing variety better than rule-based pipelines. They also introduce a failure mode rule-based systems do not have: fluent, confident output that is wrong, with no signal distinguishing it from output that is right. For extraction feeding a registry, a trial-matching pipeline or an audit, the output has to be verifiable against the source span rather than trusted on presentation.

Before relying on an extraction pipeline

• Is accuracy reported per field, with error direction, rather than as one document-level number?
• Was it validated on notes from your institution, written by your clinicians?
• How are negation, uncertainty and temporality handled, and how were they tested?
• Can each extracted value be traced back to the span of text it came from?
• What is the plan for monitoring accuracy as documentation practice changes?