Patient instruction personalization

Personalized treatment instructions help only when each line traces to something already on the structured care plan. The model may rephrase an order into language a specific patient can follow. It may not add a dose, a restriction, a timing rule, or a "do not" that the plan never stated.

The quality bar is narrow. An instruction line that cites the care-plan order is in scope. If the plan is silent on a topic, that line stays empty. A clinician still reviews before anything reaches the patient.

Nurse educators already do this work by hand: take the after-visit orders, strip jargon, match the patient's preferred language and literacy, and check that nothing extra crept in. The model is a first-pass writer for that same job, not a second source of clinical truth.

EHR and documentation vendors such as Epic, Oracle Health, Microsoft, and Abridge may surface care-plan orders, visit notes, or conversation transcripts. Treat those products as a class of input channels. The rule does not change with the brand: the generator reads the locked plan, not a guess about what the clinician meant to order.

Do not generate until the care plan for this encounter is locked. Locked means the treat-stage orders you will teach from are the signed set: medications, procedures, activity, follow-up, and any explicit teaching points the clinician entered. Unsigned drafts, ambient note fragments, and hallway comments are not the plan.

Once locked, pass only that structured plan into the generator, plus the patient's documented literacy level and preferred language. Do not also dump the entire chart. Extra narrative is how invented restrictions appear. A historical "low salt" note from another visit becomes a new restriction the current plan never ordered.

Draft with cites. Each output line should name the order it restates, using the order identifier, the order text, or both, depending on how your system stores them. If a line cannot point at an order, delete the line. Do not fill the gap with a standard teaching script from a library.

Leave blanks. If the plan lists a medication but no dose, frequency, or route, the instruction for that medication must not invent those details. Show a blank or a "confirm with clinician" placeholder next to the cited order. The educator's review then either pulls the missing field from a signed order that was omitted in the extract, or the line stays out of the patient packet.

Clinician review is the last required step, not a courtesy. The reviewer checks three things: every line cites an order, no line adds a fact the order lacks, and the plain-language rendering still means the same thing as the order. If any check fails, send the line back or remove it. Do not "fix" a missing dose in the teaching sheet without fixing the order first.

Literacy matching and language matching happen after the facts are fixed. First produce a faithful restatement of each cited order. Then simplify vocabulary, shorten sentences, and render the same facts in the patient's preferred language.

If you reverse that order, you get fluent nonsense: a kind-sounding paragraph in another language that includes a dose nobody ordered. Fluency is not quality here. Quality is a cited, empty-if-silent instruction that a reviewer can check against the plan line by line.

When the patient's preferred language is not the language of the order, translate the locked, cited lines only. Do not translate adjacent chart text and then summarize it. That path is how a restriction from an old problem list becomes a new teaching point. For patients who also need wayfinding in their language after the visit, keep this packet separate from a multilingual patient navigation agent. Navigation answers how to get to lab or pharmacy. This page answers what this treat-stage order tells the patient to do.

Discharge packets are a later, broader artifact. Plain-language discharge instruction generation often reuses the same cited lines, but it also covers after-visit logistics the treat-stage teaching sheet may not include. Do not let discharge templates backfill doses or diet rules into the in-clinic teaching sheet.

Literacy on file is a ceiling, not a style preference. If the record says the patient needs simple sentences, do not keep medical abbreviations in the patient-facing line even when the cited order uses them. Expand only terms that the order already specified.

A 14-year-old is seen for an asthma flare. The signed care plan contains three orders: start a named controller inhaler at a stated dose twice daily, continue the rescue inhaler as needed, and return to clinic in two weeks. Literacy on file is "needs simple sentences." Preferred language is English.

The generator should produce three instruction lines, each citing the matching order. One acceptable shape, using placeholders rather than a live order, is: use the named controller at the stated dose in the morning and at night (cite: controller inhaler, twice daily). Keep using the rescue inhaler when breathing is hard (cite: rescue inhaler, as needed). Come back to clinic in two weeks (cite: follow-up, 14 days).

It should not add "rinse your mouth after each use" unless that teaching point is on the locked plan. Mouth rinsing may be good practice. If the plan is silent, the line stays empty. The nurse educator can add it as a clinician-authored teaching point, which then becomes a citable order for the next draft. The model does not get to complete the teaching from memory of typical asthma education.

If the extract of the plan captured the controller drug name but dropped the dose, the draft must show a blank for the dose, still citing the controller order. Filling two puffs because that is a common start is the failure mode this workflow exists to prevent.

After review, the same cited facts can feed later outreach. A medication adherence outreach agent should repeat the reviewed instruction, not generate a parallel version from the chart.

A dose with no order. This is the highest-severity miss. Models complete partial medication lines unless you forbid completion. Any numeric dose, frequency, or route that is not on the cited order is a defect, even if it happens to be clinically typical. Correct by coincidence is still invention. The reviewer should read numbers first, then prose.

Treating the draft as given. Printing or sending the first pass because it looks like something the clinic would say skips the only step that can catch a fluent error. Review against the locked plan, not against tone. Ambient documentation can make this worse. A transcript may mention a dose the clinician discussed and then changed. Ambient clinical documentation is a note-writing input. It is not a substitute for the signed care plan, and it is not a second teaching source.

Inventing a restriction. Diet, activity, avoid ibuprofen, no gym for a week, and similar limits often live in teaching libraries. If the current plan does not order them, they do not belong on this patient's sheet. A restriction that is clinically wise but unordered is still a quality miss. Add it through the clinician, then regenerate so the new line can cite the new order.

Watch two presentation defects after the facts are clean. Literacy overshoot keeps words the documented level cannot support. Language mix drops untranslated medical terms into a non-English sentence. Those matter for comprehension. They are still secondary to factual invention. Fix invention first, then rewrite for literacy and language.

A quieter miss is merging two orders into one sentence so the cite is ambiguous. If the controller and the rescue inhaler share a line, a later edit to one order will not map cleanly. Keep one cite per line.