The HPI's elements, and which of them the conversation actually contains

The history of present illness is the narrative that turns a complaint into something a clinician can reason about: when it started, where it is, how long it has gone on, what it feels like, what changes it, where it travels, when it is worse, and how bad it is. Clinical teaching packages those as a mnemonic — the StatPearls reference entry on SOAP notes gives OLDCARTS, for onset, location, duration, characterisation, alleviating and aggravating factors, radiation, temporal factor and severity — and every documentation template in use is some arrangement of them.

What the templates do not say is that these elements are not equally available to anyone listening. Some are volunteered by the patient in the first thirty seconds. Some are asked and answered aloud. Some are asked and answered with a nod. And some the clinician never asks at all, because the answer arrives through their hands during the examination. A documentation system that treats all nine as equally transcribable will keep producing an HPI that is thin in the same places every day.

Element by element: what a consultation actually contains

Read the third column as a design specification rather than a criticism of clinicians. Nothing in it is bad practice: an experienced clinician who has already narrowed the differential does not need to work through nine elements out loud, and the consultation is better for it. But the note still has to carry them, and the audio does not.

The nod, the pointed finger, and the examination

Three specific mechanisms account for most missing elements, and each needs a different fix.

• The nonverbal answer. "Does it wake you at night?" followed by a nod produces a transcript containing a question and nothing else. A model has three options and all are wrong: drop the element, record it as not assessed, or infer a yes from the clinician's next sentence. Only the clinician saying "so it does wake you" resolves it, and that is a speaking habit, not a software feature.
• The pointed finger. Patients localise pain by touching. "It's here" is a complete answer in the room and an empty one in a transcript, and no amount of audio quality recovers it. Getting reliable attribution of who was speaking in the first place is a separate hard problem, worked through in diarisation in a consult room.
• The examination finding. Radiation, severity on movement and reproducibility are frequently established by hand rather than by question. They are real findings, and they belong in the note — but they will only reach it if the clinician narrates them or types them, because the microphone has no access to what the fingers learned.

The chief complaint is one line; this is the story around it

The two get conflated constantly, and the conflation causes real damage in a generated note. The chief complaint is the reason for the visit, usually stated in the patient's own words and usually one line: "chest pain for two days". Everything after that first line — the character of the pain, what it does on exertion, what the patient already tried — is the history, not the complaint.

When a template collapses the two into one free-text box, two things follow. Coding and downstream search lose the complaint as a discrete field, so any query that groups encounters by presenting problem stops working. And the model, given one box, fills it with a paragraph — which then reads as though the patient delivered a structured history unprompted. Keeping them as separate slots is the cheapest structural decision available, and it belongs to the template rather than the model, which is why one template for the clinic or one per specialty is a decision worth making deliberately.

Why the fix is at capture rather than in the model

The instinct when an HPI comes back thin is to change the prompt, then the model, then the vendor. None of those can help, for a plain reason: information that was never in the room cannot be recovered from a recording of the room. What changes the result is what happens during the consultation.

1. Make the missing elements visible while the consultation is still happening, not at review. A live panel showing which elements have been captured lets the clinician ask the one question that is missing while the patient is still in the chair.
2. Adopt one speaking habit rather than nine. Reflecting the answer back — "so it started on Tuesday, sharp, worse when you breathe in" — converts several nonverbal answers into audio in a single sentence and doubles as a comprehension check with the patient.
3. Narrate the examination in a short, fixed form. Two sentences at a consistent point in the encounter get the findings into the audio without changing the examination itself.
4. Let the template hold explicit empty slots. An element that was never established should render as an unfilled slot the clinician resolves at signature, rather than being silently omitted so that nobody can tell the difference between not asked and not present.
5. Choose the capture mode to fit the list. A forty-patient outpatient clinic and a twenty-minute consultation reward different arrangements, which is the argument in ambient capture or dictation for a busy list; dictation trades naturalness for completeness, and on a long list that trade sometimes wins.
6. Review the element-level gap monthly, per clinician. The pattern is usually specific and stable — one clinician never asks severity, another never narrates the examination — and a specific pattern is coachable in a way that an overall quality score is not.

The one thing a system must never do is close the gap on its own. An HPI element the model supplies from clinical plausibility rather than from the encounter is an assertion about a patient that nobody made, and it will be signed, because a complete-looking history invites less scrutiny than an obviously incomplete one. The same pressure produces review-of-systems entries nobody asked about, examined separately in the negatives the clinician never asked for.

A model cannot transcribe a question that was never asked. Every element missing from the note is either missing from the consultation or missing from the audio, and those are two different repairs.

Whichever repair applies, the clinician is still the author of record when the note is signed, which is what the attestation on a machine-drafted note has to make explicit. Capture changes of this kind are product decisions rather than configuration, and they usually mean owning the capture surface: whether you can change it at all is the crux of a custom build against a vendor module, and it is why this work tends to arrive as an MVP and product build rather than a settings exercise. The surrounding decisions sit across ambient documentation and the clinical note and our healthcare and MedTech work.

• clinical documentation
• HPI
• ambient scribe
• consultation capture