From medical documents to a reliable patient history

A medical document is more than a block of text. Its meaning depends on who produced it, when it was written, what examination it describes and whether a recommendation was later changed. A discharge letter, a referral and a patient’s note can contain similar terms while serving different purposes. Document processing must preserve these relationships before it can produce a useful clinical summary.

The first methodological distinction is between transcription and interpretation. Text recognition attempts to reproduce what appears in the source. Interpretation determines what that content means in the current case. When an assistant silently resolves an unclear number or expands an abbreviation, it crosses that boundary. The resulting text may sound more complete while becoming less faithful.

A structured evaluation should include clean digital documents, scanned pages, tables, rotated images and ambiguous fragments. The reference should identify not only the expected words but also information that cannot be read with confidence. Successful processing includes recognising uncertainty. Guessing a missing dose is not a better outcome than flagging it for review.

Temporal relationships deserve their own checks. “Recommended”, “started”, “stopped” and “currently taking” describe different states. A system that retrieves an old prescription correctly but presents it as current has preserved the words and lost the clinical meaning. A timeline should therefore retain the source date and distinguish the date of an event from the date it was documented.

For evaluation, we propose comparing extracted facts with a reviewed reference and examining medication names, doses, units, dates and negations separately. We also ask whether each summary statement can be traced back to the source and whether corrections remain visible. Average text accuracy alone can conceal errors concentrated in the details that matter most.

The intended workflow is a draft history that a specialist can inspect efficiently. The clinician confirms, corrects or leaves an item unresolved. The patient does not need to repeat every detail merely because it was filed in a different document. This is the practical direction of our document research; improvements in clinical outcomes would require separate prospective evaluation.