Balsam Semantica Computable clinical meaning

The Clinical Meaning Is Shaped
Before the AI Writes a Word

A recent study on AI-generated discharge summaries raises a question that matters more than the performance of the language model itself: what happens to clinical meaning before it reaches the model?

What Osborne et al. Found

Osborne et al., publishing in npj Digital Medicine, compared GPT-5.2-generated discharge summaries with physician-authored summaries across 60 internal medicine hospitalizations lasting 7–21 days. Physicians preferred the AI summaries in 95% of encounters, rating their overall quality at 4.71 out of 5 against 3.16 for the physician-authored ones. The AI summaries also carried fewer annotated errors, mainly because they omitted less.

95%Encounters preferring AI
4.71AI quality out of 5
3.16Physician quality out of 5

One methodological detail
deserves closer attention.

Before GPT-5.2 generated a single summary, the team had to decide what clinical information would enter the model. The records contained 654 distinct note types. A physician informaticist reviewed them, and 370 were retained as inputs, cutting EHR text volume by an average of 55% per encounter.

654Distinct note types
370Retained as inputs
55%EHR text volume cut

One methodological detail
deserves closer attention.

That step is itself a clinical decision about what information matters. The model worked from a curated representation of the record.

And the evaluation was built around that curation. The reviewers judged completeness against the notes that had been fed to the model, so any omission they found was an omission from the curated input.

One methodological detail
deserves closer attention.

Whatever the selection step removed sat largely outside what the evaluation was designed to see.

So the study measured the quality of the summaries well. It was not designed to measure what was lost on the way in.

That gap matters because clinical information does not keep its meaning when it is lifted out of context.

The study's own error analysis shows this. Among the omissions in the AI summaries were relevant comorbidities that predated the admission. Whether a historical condition still bears on the current care plan is a question of temporality and clinical relevance, and it is exactly the kind of distinction a summary can drop while still reading as coherent and complete.

That gap matters because clinical information does not keep its meaning when it is lifted out of context.


Negation, uncertainty and attribution belong to the same category: a summary can be fluent and still lose whether something was present, suspected or ruled out, or whether it belonged to the patient at all.

Completeness depends on
who will use the information

Primary care physicians also found more omissions than hospitalists. Part of that reflects how each group was asked to review. Part of it reflects something often overlooked in AI documentation evaluation: completeness depends on who will use the information. What is sufficient to close an inpatient admission may be insufficient for the clinician who takes over care after discharge.

Completeness depends on
who will use the information

The task, then, is less about generating better clinical text and more about ensuring that information entering an AI system keeps its clinical identity, context and relationships through every transformation. Terminology standards such as SNOMED CT can represent negation, temporality and uncertainty, but only when those distinctions are captured at the source and carried forward.

Completeness depends on
who will use the information

Provenance, who said something and where it came from, sits largely outside the terminology altogether.

Where clinical
meaning can be lost

Clinical AI evaluation needs to look at the whole pathway:

A more eloquent summary is not evidence that the clinical meaning underneath it survived.

In healthcare, that difference is foundational.