Automated anamnesis collection

Quality for automated anamnesis is a history field that cites the patient's answer. If the patient did not answer, the field stays empty. The agent must not invent a medication. A clinician still reviews before those fields are treated as charted.

Transcript length and a complete-looking medication widget do not count. The test is whether a reviewer can point to the patient turn that supports the written value, and can see which items were never answered in this conversation.

Conversational intake products such as Nabla and Klara, and the intake and charting surfaces in systems such as Epic and athenahealth, belong to one class. They can run a structured dialogue and emit EHR-shaped data before the visit. If a value has no cite from this dialogue, it is not a patient-reported history for this encounter.

Prior charts are a different source. Pulling last year's list is prior record ingestion and summarization, not anamnesis. Keep that label. Do not paste a prior list into current medications as if the patient just confirmed it.

Open the session with a scripted sequence that maps to fields, not with "tell us everything." Ask chief complaint first. Then duration and onset. Then associated symptoms you actually need. Then current medications. Then allergies. Then the slice of past history this visit type requires. Each question has a destination field. If the patient answers sideways, ask one clarifying question that still maps to that field. Do not diagnose, reassure, or suggest treatment in the intake thread.

Capture the patient's wording. The cite is the utterance, a transcript snippet, or the form-item identifier your vendor stores. "Left knee swollen since Saturday, twisted it on the stairs" is a cite. A model paraphrase with extra laterality, mechanism, or "no red flags" is not, unless those extras were asked and answered.

Drive the script to a real close: answered, declined, or timed out. A decline is not "none." Write no structured value for a declined item. Write "none" only when the question was any-versus-none and the patient said they take nothing or have no allergies. "I will bring the bottles" is incomplete, not a list.

If someone other than the patient is typing, flag that when the tool allows it. A spouse-completed history can still be used, but the clinician needs to know it is second-hand.

Every write is a pair: the EHR value and the cite from this dialogue. A coded chief complaint without the sentence is incomplete. A transcript dump without a field is a note, not anamnesis. Unanswered items are omitted from the payload. Do not send a placeholder that downstream systems will display as a confirmed negative.

Here is one walkthrough, not a measured case.

A 54-year-old patient opens the pre-visit dialogue for a same-week orthopedic slot. The agent asks for the main problem. The patient types, "Left knee swollen and painful since Saturday, twisted it on the stairs." The agent writes chief complaint as left knee pain and swelling, and stores that sentence as the cite. It asks about current medications. The patient replies, "I take ibuprofen when it hurts. I don't remember the blood pressure one." The agent writes ibuprofen as a patient-reported as-needed medication, cited to that turn. It does not write a blood-pressure agent. It does not copy a tablet name and dose from last year's list into this encounter's current-meds field. That blank, or an incomplete-meds flag, is the input to the medication reconciliation agent at rooming or at the start of the visit. The agent asks about drug allergies. The patient skips. The allergy field stays empty. The agent does not write NKDA.

EHR-ready means types match what Epic, athenahealth, or your intake layer expect, and that silent items are absent. It does not mean the note is signed, the med list is reconciled, or allergies are attested. If dose, route, or frequency were not stated, do not invent them to satisfy a structured med model. Name plus as-needed, with the cite, is enough. Missing details stay missing.

The dangerous outputs look finished.

A medication with no answer. The patient skipped the question, said they were unsure, or promised to bring bottles. The agent still writes a list, usually from an old reconciliation, a fill feed, or a guess that patients like this are on an ACE inhibitor. A populated widget in the EHR reads as confirmed today. If the only source is historical, leave this encounter's patient-reported list empty and keep the historical list in its own lane until a person reconciles it.

Treating the form as signed. Staff see intake complete and skip rooming questions, auto-attest allergies, or lock the HPI. A conversational pre-visit packet is not a signed note, not an order, and not a legal attestation. The patient did not co-sign the fields. Do not auto-attest. Do not skip "what allergies do you have?" when the allergy field is empty. Do not treat a completed dialogue as consent for the plan of care.

Inventing a history. The template had slots for duration, laterality, mechanism, and a row of symptom negatives. The model fills them because the complaint was knee pain. A template slot is not an answer. If the agent never asked about fever, calf swelling, or inability to bear weight, it cannot write those negatives. The same rule applies to family history, smoking, and "no past surgery." Silence is not a negative finding. If the patient said "maybe since Saturday," do not write onset as a precise date. Keep the hedge in the cited text or leave duration empty.

Send the packet to a human before the visit when there is time, and always before copying fields into the signed note. The reviewer checks that every written field has a cite from this dialogue, that empty fields are empty on purpose, that nothing was promoted from prior records without a new confirmation, and that the packet does not contain a diagnosis or a plan the agent was never allowed to make.

Rooming can clear obvious gaps such as incomplete meds or skipped allergies. The treating clinician still owns the chart. If ibuprofen is listed as-needed and the blood-pressure line is blank, the visit starts with that gap, not with a false complete list.

If you also score intake for urgency, keep that score off the history fields. A flag from acuity risk stratification at intake can change who reviews first or how soon the patient is seen. It must not backfill symptoms the patient never reported.

Once the visit starts, in-room capture is a second source. Do not merge it silently into pre-visit cites. Keep pre-visit answers labeled as pre-visit. Ambient clinical documentation can update the note from what is said in the room. It should not overwrite a blank pre-visit allergy with a guessed NKDA, and it should not treat the intake form as already signed.

Automated anamnesis should arrive as a cited, partial, reviewable packet. Stated chief complaint and history can shorten rooming. Named medications can seed reconciliation. Blanks tell the team what to ask first.

Do not wait for a perfect script. Ship a short tree, require a cite on every write, block auto-sign, and train staff that empty means unasked or unanswered. Judge the workflow by whether clinicians still chase the blanks, not by whether the form looks full. A full form with an invented med is a quality failure. A sparse form with honest gaps is the intended output.

When prior records, medication reconciliation, acuity flags, and in-room documentation all touch the same chart, source labels are the safety mechanism. This conversation is one source. Everything else is another. The history field that can show the patient's sentence is usable. The empty field that stayed empty because they never answered is also usable. The field that cannot show either is not.