An AI-generated clinical note can be factually supported and still leave out something relevant. That is the omission problem. It is serious, but it is not well described by saying omissions are always more dangerous or impossible to detect.
A 2025 npj Digital Medicine study gives us a better frame. The researchers evaluated 450 consultation transcript-note pairs produced across 18 experimental configurations. Clinicians reviewed 12,999 generated-note sentences for hallucinations and 49,590 transcript sentences for omissions. They also marked an error as major when it could affect diagnosis or management if left uncorrected.
The useful lesson is methodological: accuracy and coverage require two different audits.
The study checks in two directions
To find hallucinations, reviewers started with the generated note and asked whether each sentence was supported by the consultation transcript. Unsupported content was labelled a hallucination.
To find omissions, they reversed the direction. Reviewers started with clinically relevant information in the transcript and checked whether it appeared in the note. Relevant information that was absent was labelled an omission.
This exposes the limit of a one-way grounding check. Checking every sentence that was written can test support for those sentences. It cannot, by itself, test whether all relevant source information was covered. Omissions do not escape every possible evaluation; they escape an evaluation that only starts from the output.
The two percentages do not share a denominator
Across the study, 191 of 12,999 generated-note sentences contained hallucinations, a rate of 1.47%. Of those hallucinated sentences, 84 were classified as major, or 44%.
For omissions, the denominator was the source transcript: 1,712 of 49,590 transcript sentences were omitted, a rate of 3.45%. Of the omitted sentences, 286 were major, or 16.7%.
Those numbers should not be turned into a simple ranking. Hallucination rate was calculated over note sentences; omission rate was calculated over transcript sentences. The study evaluated generated notes from consultation transcripts, not longitudinal hospital charts, and it did not measure patient outcomes. It shows that both error classes can carry clinical significance. It does not establish that one is universally more dangerous.
Where major omissions appeared in this experiment
The paper grouped major omissions into three areas. Current issues accounted for 55%, past medical, medication, family and social history for 35%, and information and plan for 10%. In the study taxonomy, medication history included allergies.
That finding supports targeted coverage checks for those sections in similar note-generation workflows. It does not justify a universal list of the most dangerous fields in every EHR. A longitudinal chart summary has different source material, and its own evaluation should define what counts as required coverage before the summary is generated.
What Meddies currently does
The current Meddies patient-summary instructions begin by compiling FHIR data into a structured patient artifact. The artifact includes conditions, medications, allergies, recent tests, vital signs and missing-data categories, with provenance recorded for each field. The instructions explicitly distinguish “no allergy data” from “no allergies” and require a final check across active conditions, medication doses, allergy status and trends.
That is a workflow contract in the current implementation. It is not a published omission rate, a guarantee that every relevant fact will appear, or evidence that the summary improves clinical outcomes. We have not run a study comparable to the npj evaluation on Meddies summaries.
The next useful test is clear. Build a coverage specification for the source chart, review the source and summary in both directions, classify missed information by clinical consequence, and report the errors that remain. A source link helps a clinician inspect a statement. Only a coverage audit can tell us whether a required statement never appeared.