Automated discharge summary drafting

A discharge summary draft is only as trustworthy as the sources you load. Before any model writes a sentence, open the encounter notes that cover this admission and the active problem list that will travel with the patient.

Those two artifacts bound what the draft may claim. Progress notes, consults, procedure notes, and the problem list are fair game. Hallway conversation, nursing memory, and what you usually see with this presentation are not.

Ambient capture from tools in the Abridge, Nuance, Epic, or Oracle Health class can feed the encounter note once a clinician has accepted that capture into the record. Treat it like a typed progress note, not like a second medical opinion. If you are still building the inpatient note with ambient clinical documentation, finish that work first. A summary that cites an unreviewed ambient draft is citing a draft.

Load the full encounter, not only the last 24 hours. Language that reflects only the day of leaving will understate complications, missed consultants, and problems that resolved mid-stay. Every problem you plan to carry forward should already sit on the problem list, with onset and status that match the notes.

If a problem is on the list but never appears in an encounter note, flag it. If a diagnosis appears in a note but never made the problem list, flag that too. The model should not silently pick a winner.

The useful output is not a polished narrative. It is a structured summary in which each clinical sentence points at a note, order, or problem-list row.

Ask the model for a hospital course, procedures, pending results, and follow-up plan. For each sentence, require a pointer: which note, which date, which author or note type. A line that reads "per the cardiology consult on hospital day two" is a cite. A line that reads "course was uncomplicated" with no pointer is not.

A hospital course with no note cite is a failure mode, even when the prose sounds like your service. Fluency is not evidence. If the draft cannot show where a statement came from, delete the statement or replace it with a blank.

Do not let the model invent a diagnosis at discharge to make the course read as complete. New labels belong in your own assessment after you look at the chart. If the notes never named the condition, the summary must not name it either: not demand ischemia that was never documented, not a likely pulmonary embolism that no one imaged, and not heart failure added because a natriuretic peptide was high.

Medication lists in the summary should come from the reconciled discharge list, not from a free-text guess. Run discharge medication reconciliation as its own step. Pasting an unreconciled inpatient list into the summary is how home-dose errors travel.

Empty stays empty. If no encounter note describes the course for a given interval, do not fill the gap with a plausible paragraph. If the problem list has no entry for a suspected condition, do not add one in the summary.

Blanks are a quality signal. They tell you what still needs a human sentence, and they keep the signed note from asserting facts the record cannot support. Common silences include pending studies with no result, specialist advice that was spoken but never written, undocumented code status, follow-up that was "to be scheduled" with no clinic, and dosing details the note never recorded.

Instruct the model to output an explicit placeholder such as "not documented in available notes" rather than smoothing over the hole. Smoothing is how a missing echocardiogram becomes preserved systolic function in a letter the next clinician will trust.

Do not backfill from prior admissions unless those prior notes are part of this encounter's source set and you cite them as such. A last-year catheterization report is not this stay's hospital course.

Take a two-day admission for pneumonia that is already in the chart. The history and physical lists fever, productive cough, and a right lower-lobe infiltrate. The problem list has community-acquired pneumonia and well-controlled type 2 diabetes. Daily progress notes describe improving oxygen need and a plan for oral antibiotics. No echocardiogram was ordered. No new cardiac diagnosis appears in any note.

A defensible draft cites the history and physical for the presenting syndrome and imaging, carries pneumonia and diabetes from the problem list with matching status, cites the daily notes for the oxygen and antibiotic course, leaves cardiac function blank, and leaves specialist follow-up blank if no clinic was booked.

An indefensible draft adds an uncomplicated course with no cite, invents possible demand ischemia because a troponin was once normal, or copies heart failure from a years-old problem that is not on the current list and was not discussed this stay.

If the notes are thinner than this, the draft should be thinner too. Completeness is not a virtue when the chart is silent.

Treating the draft as signed is the fastest way to put unsigned model text into the legal medical record. Until you review, edit, and authenticate, it is a suggestion in a workspace, an inbox, or an unsigned note object in Epic or Oracle Health.

Your sign-off means every retained sentence is true to the encounter notes and problem list you loaded. It does not mean the model usually gets this right. Read the hospital course against the notes, the discharge diagnoses against the problem list, and the pending-results section against outstanding orders.

If a scribe, resident, or documentation specialist prepared the draft, the attending who signs is still responsible for invented diagnoses, missing complications, and blanks that were quietly filled. Do not skip the cite check because someone else already looked at it.

Do not thicken the summary to capture billing complexity. Handle that through automated charge capture from notes against the same source notes. After you sign, generate patient-facing instructions from the signed plan with plain-language discharge instruction generation, not from the unsigned draft.

Keep the chart beside the draft. Confirm you loaded this encounter's notes and the current problem list, not a prior summary reused as a template. Confirm every hospital-course sentence has a note or problem-list anchor. Confirm every silence in the chart is still a silence in the draft. Confirm no diagnosis appeared at discharge that you did not document. Confirm medications and follow-up match reconciliation and actual orders.

If a sentence fails any of those checks, cut it. A short signed summary that cites the chart beats a complete-sounding letter the record cannot defend.

Tools in the Abridge, Nuance, Epic, and Oracle Health class can place the draft next to the note you will sign. Use that adjacency. Split-screen review is the job. Auto-file is not.