AI Adoption GuideHealthcareDischarge
Plain-language discharge instruction generation
LLM produces patient-appropriate discharge instructions covering medications, activity, diet, and red flags from structured clinical data.
Healthcare processAccessIntakeAssessDiagnoseTreatDischargeBillFollowup
By Don, DoneThat’s AI coach · updated
Cite the signed plan or leave the field empty
Every medication direction, activity limit, diet instruction, and red-flag warning in the patient-facing draft must point to a line in the signed discharge plan. If the plan does not say it, that field stays empty. The model does not invent a restriction so the packet looks finished. You still review the draft before it reaches the patient.
That is the quality bar. Fluency is not the bar. A kind, easy-to-read sentence that tells someone not to drive, not to bathe, or to return for a fever the attending never specified is a defect. Completeness that is not in the plan is fabrication.
These instructions are what the patient acts on at home. They are not a substitute for the clinical discharge summary. Keep automated discharge summary drafting in the chart workflow. Keep this generator pointed at the signed plan and at language the patient can follow.
Load the signed plan before you generate
Start from the signed discharge plan, not from the admission history, an unsigned order set, or a nursing note from yesterday. The source of truth is the attending's signed activity, diet, follow-up, wound, and monitoring instructions, together with the reconciled home medication list.
If a resident drafted the plan and the attending has not attested, treat it as unsigned. Do not generate patient instructions from a plan that is still in flight.
Whatever product sits on the discharge workspace (Epic, Microsoft, Abridge, Oracle Health, or another vendor in this class), bind the draft to that signed plan. Do not bind it to a worklist that is still in progress. Product names and buttons differ. Your check stays the same: which document is this draft reading?
Take medications from completed discharge medication reconciliation, not from the inpatient MAR. An inpatient hold does not become a home instruction unless the signed plan says so.
If the plan is unsigned or required fields are empty, stop. Do not fill the gap with typical post-op wording. Send the incompleteness back to the discharging clinician. Generating from an incomplete plan is how invented restrictions get a head start.
Write each instruction with a cite
Draft the blocks the patient actually uses: medications, activity, diet, and red flags. Add follow-up only if the plan names a visit, a call, or a window. Every sentence that tells the patient to do something, stop something, or call for something should carry a cite to the plan line it came from. The cite can be a field name, order text, or note anchor, depending on what your system stores. You need to walk from the sentence to the signed source during review.
Plain language is required. It does not license a new fact. Patient instruction personalization is how you match reading level, language, and how the patient names their drugs. This page is about fidelity. The easier sentence still has to be true.
Keep lines short. One action per line when you can. Put the drug name, dose, timing, and purpose in words the patient uses, but only when those facts sit on the reconciled list.
Here is a single worked example, not a measured case. A patient is going home after laparoscopic cholecystectomy. The signed plan states: regular diet; walk as tolerated; no lifting more than 10 pounds for two weeks; oxycodone 5 mg by mouth every 6 hours as needed for pain; acetaminophen 650 mg every 6 hours; surgeon follow-up in 10 to 14 days; call for fever, vomiting that will not stop, or worsening abdominal pain. The plan is silent on driving, bathing, and return to work.
A sound draft restates regular diet, walking as tolerated, the lifting limit (cited to the activity line), both home pain medicines (cited to the reconciled list), the follow-up window, and the three call-back symptoms (cited to the red-flag line). Driving, bathing, and work stay blank. A defective draft adds "Do not drive while taking oxycodone" or "No tub baths for two weeks" because those rules are familiar. Familiar is not signed. Delete the extra lines or turn them into a question for the attending. Do not leave them in patient text.
Leave blanks when the plan is silent
A blank field is an honest status. It means the discharging clinician has not decided yet, or the topic was never addressed. Filling driving, work, or bathing with a cautious default hides the gap from you in review and from the patient at home.
Do not write household help into the medical instructions unless the plan documents it. "Your daughter will help with meals" is not an instruction if no one documented who is at home. That belongs in social support gap detection and then in a conversation, not in generated prose.
Keep unused sections visible on the review screen. Collapsing them makes the packet look done when it is not.
If the generator cannot find a source, it should emit a blank or an explicit "not specified in the signed plan" marker. It should not emit a plausible sentence.
Catch the three defects that look like care
A restriction with no plan cite. "No driving for two weeks" appears because opioid and lifting language sat nearby. If you cannot point to a signed activity or safety line, it is not an instruction. It is a guess. Remove it or ask.
Treating the draft as already handed to the patient. Model output is a working copy. Printing it, sending it to the portal, or tucking it into the after-visit packet before you read it is how a hallucinated diet or a missing anticoagulant direction leaves the unit. If your after-visit summary pulls this draft, review still happens before print. Census pressure at discharge does not change that.
Inventing a red flag. Adding "return for calf pain or chest pain" after an uncomplicated laparoscopic case because those warnings are taught in orientation is the same error as inventing a lifting limit. Precautions you believe in still need a signed plan line, or an edit you make and the clinician signs.
Watch for tone that over-explains. Extra "why" sentences often smuggle in unsigned rules ("your bile ducts need time to heal, so no fatty food"). If the plan said regular diet, do not add a fat restriction in the explanation.
Review, then release the patient copy
Read the draft against the signed plan, line by line. Confirm every restriction has a cite. Confirm every blank is truly silent in the plan, not a field you skipped. Confirm medication names and directions match the reconciled list. Edit. Attest or sign per your unit's policy. Only then produce the patient copy.
Review is required because the prose sounds right. A wrong instruction in plain language is easier to follow, which is the harm.
After you sign, do not regenerate from an older plan version. If the attending changes activity or medicines, reload the new signed plan and draft again. Stale cites fail the same quality bar as missing cites.
Keep citations in the clinician view. The patient should see instructions, not a source appendix. What they carry out the door is the plain-language lines you approved, nothing the model added after your review, and nothing the plan never said.
Is this worth automating for you?
Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.
DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.
Measure the baseline first