Ambient clinical documentation

Ambient clinical documentation earns its keep when each sentence in the SOAP draft can point at a span of the visit transcript. If nobody said it, the line stays empty. Fluency is not quality. A clean paragraph that inserts an exam finding you never spoke, and that you did not independently enter, is a failed note even when the history is accurate.

The pattern is the same across this class of tools, including Abridge, Nuance, Nabla, and ambient workflows that file into Epic: speech becomes text, text becomes a SOAP-shaped draft, and a clinician must still sign. The microphone did not examine the patient. The model does not get to invent lung findings, a full ROS, or an assessment you did not reach.

Treat a cite as a pointer, not a decoration. A subjective sentence should map to words the patient or you actually used. An objective exam line should map to words you spoke during the exam, or to data the EHR already holds (vitals, labs) that your workflow is allowed to pull in as facts rather than as speech. If a line has no span and you did not add it yourself from your hands or your screen, delete it. Empty is correct. Guessing is not.

When you are deciding whether to adopt this, the test is whether the draft is safe to sign, not whether it sounds like a note you would have written on a good day.

Start recording the way your clinic already approved: room mic, phone, or EHR-linked device. Confirm consent in the manner your policy requires. Then run the visit. You do not need a special script for the model. You do need the clinical content you want in the chart to exist in speech or in data.

Say aloud the findings you want cited. If you listen to the heart in silence, the transcript has no PE to attach. Either verbalize ("regular rate and rhythm, no murmur") or plan to type the exam yourself after, knowing those lines will not have an ambient span. Do not perform extra ROS questions for the recorder. Ask what the visit needs. Silence is allowed. A later draft that fills every ROS checkbox from a template is a failure mode you prevent here by not feeding the model a fake complete interview.

Capture the whole encounter you intend to document. If you leave to review labs and return with a plan, that return conversation is part of the visit. If a family member answers, the transcript will show a second speaker; on review, do not let the draft attribute those words to the patient unless that is what happened and you accept it.

Stop capture when the visit is over, not when you feel the important part ended. After-visit counseling, dose changes, and follow-up timing are often the assessment and plan. If they never hit the audio, the model has nothing honest to write there either.

Open the draft against the transcript, not against your memory of a normal visit. Walk SOAP in order.

Subjective: keep what was said. If the patient described morning home blood pressures and denied swelling, those belong. If the draft adds "denies chest pain, dyspnea, palpitations, claudication, and syncope" and you only asked about swelling, you are looking at an invented ROS. Strip the extras. Leave unasked systems empty. Do not complete the template because empty looks unfinished.

Objective: vitals and labs may arrive from the EHR. Physical exam may not. A finding with no transcript span is the first failure mode to hunt. "Lungs clear to auscultation" after a visit where nobody mentioned the lungs is not a helpful default. It is an undocumented exam. Remove it unless you actually examined and you add the line as your own entry. Abdominal "soft, nontender, nondistended" is the same test. If you did not examine the abdomen, or you did and never spoke or typed it, the line does not exist.

Assessment and plan: the model will often turn keywords into a diagnosis and a next step. Check that this is your conclusion. Counseling you did not give, referrals you did not discuss, and "continue current meds" when you changed the dose are edits, not nits.

Cites are how you review. Jump from a sentence to its span when the product shows one. If it does not, search the transcript yourself. No supporting words means unsupported. Unsupported does not mean probably true from context. It means do not keep the line unless you author it.

Illustrative walk-through (not a measured case): a 52-year-old returns for hypertension follow-up. She says her home cuff has been "around 150 in the mornings," she is taking lisinopril, and her ankles have not swollen. You take a room blood pressure, say the reading aloud, and discuss increasing the dose. You do not run a full ROS. You do not examine her lungs. A usable draft records the home readings, adherence, the edema denial you heard, the in-room BP you spoke, and the dose change you discussed. A dangerous draft also writes "no chest pain or DOE," "HEENT unremarkable," and "lungs clear," because those fields exist on a SOAP template. The dangerous draft is faster to click through when you are behind. It is also not the visit.

If intake history never entered the room conversation, do not ask ambient documentation to invent a past medical history. Fix that upstream with automated anamnesis collection.

Until you sign, the text is a proposal. Treating the draft as signed is the second failure mode: auto-file, "looks fine," or closing from an inbox without opening the transcript. The legal and clinical document is the signed note. The model's first pass is not.

Review is faster when cites are visible. It is not optional when they are not. Slow down on ROS and exam. Those are where language models complete familiar patterns. History and the plan you just spoke are usually closer to the audio. The templated normals are where fiction appears.

If you add silent-exam findings, write them yourself and know they have no ambient cite. That is honest authorship. Accepting an invented normal because it matches what you usually find is not.

Edit, then sign in the same sitting you would sign any other note. Do not leave a polished unsigned draft where a colleague, coder, or covering clinician could take it as final. If your EHR shows unsigned ambient text in a shared workspace, treat that as still open, not as done.

You own every remaining sentence, including the blanks. Empty PE after a visit that was history and counseling is a true note. A complete PE you did not perform is a false one.

Later automation should read the signed chart, not the raw draft. Coding suggestions belong on what you actually assessed: use diagnosis code suggestion after sign-off, not as a way to complete an unsigned SOAP. Charge capture must follow documented work; automated charge capture from notes is only as clean as the note you accepted. Discharge is a different document with a different audience; automated discharge summary drafting should pull from the signed hospital course, not from an ambient file nobody has stood behind.

Same rule on every downstream step: a line cites speech or authored data, unsaid stays empty, and a clinician has signed.