From Free-Text Case Notes to Structured Symptoms: AI-Assisted Clinical Documentation

A homeopathic case-taking session does not happen in a structured form. A doctor listens, asks, and writes down what the patient says — often in the patient's own words, because in classical homeopathy the exact phrasing of a symptom can carry diagnostic weight. "Worse before a thunderstorm," "craves company but dislikes consolation," "pain as if a nail were driven in" — these are not administrative details. They are the raw material the entire prescription depends on. Forcing a doctor to stop mid-consultation and slot what they just heard into a dropdown for category, location, modality, and causation breaks that process and risks losing the texture that made the symptom useful in the first place.

Gentle Pills' structure_notes AI task exists to resolve a real tension: the clinical method should stay exactly as classical as it has always been, while the resulting record should still be structured, searchable, and usable for repertory work afterward. The way it does this is specific, and worth describing precisely, because the details are what make it safe to use on a real patient record.

The actual flow

A doctor writes a consultation note the normal way — free text, unstructured, in their own words or the patient's. Nothing about that step changes. After the note exists, the doctor can trigger AI-assisted extraction, and structure_notes reads the free text and proposes a set of candidate symptoms, each broken out along the standard axes used in homeopathic case-taking:

• Category — the general domain the symptom belongs to (for example, mind, generals, or a particular body system)
• Location — where the symptom is experienced
• Modalities — what makes it better or worse (time of day, weather, position, temperature, eating, motion, and so on)
• Causation — any stated or implied trigger

Every candidate keeps the original wording attached alongside the structured fields. The system is not translating the patient's language into clinical shorthand and discarding the source — it is adding structure on top of the original text, so both exist side by side. A doctor reviewing a candidate sees the raw sentence it came from, not just an abstracted category and value pulled out of context.

Candidates, not commitments

This is the part that matters most for trust in the system: nothing produced by structure_notes is written to the patient's record automatically. The output is explicitly a set of candidates — AI-suggested, not AI-confirmed. A doctor reviews each one, and only accepting a candidate converts it into a doctor-confirmed symptom attached to the case. Reject it, edit it, or leave it alone, and it never touches the clinical record.

This mirrors a constraint that runs through every AI-assisted feature in Gentle Pills, not just this one: AI never mutates clinical data silently. The extraction step removes labor — specifically, the tedious, repetitive labor of re-reading a note and manually tagging each symptom along four axes — without removing the judgment call about whether a given extraction is actually correct for this patient. A doctor who disagrees with how the AI categorized a symptom, or who thinks it missed the real point of what the patient said, simply doesn't accept it. The AI has no path to override that decision.

Why the ordering is the whole point

It would be technically simpler to build a structured intake form and skip the free-text step entirely — many clinical software systems do exactly that. The reason Gentle Pills doesn't is a practical argument about what homeopathic case-taking actually requires. A structured form during the consultation forces a doctor to interrupt the patient's narrative to make a categorization decision in real time, before they've heard the whole picture. It also tends to flatten language into whatever options the form offers, which is precisely the kind of flattening that loses the peculiar, characteristic detail a homeopath is often listening for.

Writing first and structuring after inverts that constraint. The doctor captures the full, unfiltered account while it's being given. Structure gets applied once the complete note exists, with the benefit of the AI having the whole context rather than a single utterance in isolation — and with the doctor reviewing the extraction against a note they already trust, rather than trusting a form to have captured things correctly in the moment.

What this looks like as "modern practice"

It's worth being explicit about what is and isn't modern here. The clinical method — individualized case-taking, attention to the patient's own language, repertorization against a classical framework — is unchanged from the classical tradition. What's modern is removing the administrative friction around that method: the manual re-reading, the manual tagging, the manual cross-referencing of a long consultation note against a structured symptom taxonomy. That friction has always existed in paper-based and basic digital practice, and it scales badly with case volume. AI-assisted extraction is a tool for reducing that friction, not a tool for changing the underlying clinical judgment.

That distinction — tooling that reduces friction around a classical method versus tooling that replaces clinical judgment — is the design boundary that separates what structure_notes does from what it deliberately does not do. It will never decide which symptoms matter for a prescription. It only makes it faster to get a doctor's own case-taking into a form the rest of the system — repertory analysis, remedy comparison, follow-up tracking — can actually use.

In practice

For a doctor using Gentle Pills, the day-to-day effect is this: write the consultation note the way you always have. When you're ready, run structured extraction, and you'll see a list of candidate symptoms, each tied back to the sentence it came from. Accept the ones that are right. Edit the ones that are close. Ignore the ones that aren't useful. The record that results is still entirely your clinical judgment — the AI only did the transcription-to-structure work that used to take time away from seeing the next patient.