Elliot Rylands
Lab · Patient State

Let AI read the record before the clinician has to.

Emma Richardson is fictional. The brief is deterministic. AI takes the first pass across blood results, wellbeing scores, sleep notes and an unfinished medication review. It does not diagnose.

How to use it

  1. Press Prepare AI patient brief. Watch the reconstruction.
  2. Select a signal. Read the evidence. That is the whole trust model.
  3. Share with patient to turn the screen around. Same receipts, calmer language.
  4. The figure is a person in context, not an anatomy heatmap. A brain will not glow.

Riverside Practice · Fictional EHR

Emma Richardson

42 · Female · Fictional patient

Fatigue, poor sleep, neck pain, and feeling increasingly overwhelmed

Turns the screen around. Hides clinician-only framing. Keeps the receipts.

Emma Richardson · Fictional patient

Waiting to prepare the brief

  1. Preparing Emma's patient brief
  2. Reviewing longitudinal history
  3. Comparing recent results and assessments
  4. Reviewing sleep, mood and function
  5. Checking medications and follow-ups
  6. Connecting related events
  7. Prioritising recent changes
  8. 5 areas worth reviewing

Signals appear as the brief is prepared

    Demo only. Emma Richardson is fictional. No real patient data, no diagnosis, no causal claims. AI reconstructs. The clinician keeps the judgement.

    Reconstruction, not a chatbot

    The easy version of AI in an EHR is a box. Ask a question, get an answer, still decide what to ask. Patient State is the other version: the record opens, and the system works. Compare. Remember. Notice what never happened. Then stop.

    Mental health is part of the same state, not a red region in the head. Sleep, mood and function change how the whole constellation behaves. A missed review lives in the space around the person. Association is not causation. The copy says so out loud.

    This is a Lab prototype. A real EHR would still belong to a clinician, a record, and a duty of care. The argument is that the clinician should not have to be the first person to read everything.

    How clinicians would use it

    In a real deployment, Patient State would sit beside the EHR, not replace it. The record stays authoritative. The brief is read-only assembly work that runs before the clinician opens the chart.

    Machinery

    Integration would be FHIR R4: Patient, Encounter, Observation, MedicationRequest, DocumentReference. The service pulls a longitudinal view with read-only scopes. Nothing writes back to the chart unless the clinician documents as usual.

    Two triggers. A scheduled batch overnight for active panels, and a pre-visit run when an appointment is approaching so the brief is waiting at desk open. Processing stays inside the organisation boundary — on-prem or a dedicated VPC. Record data does not leave for a public model.

    Every pull and every surfaced signal goes to an audit log: who opened the brief, which evidence was shown, what entered share-with-patient mode. The clinician works on their workstation; share mode turns to a shared screen or iPad. Nothing patient-facing is a diagnosis — only items the clinician approves, in calmer language, with the same receipts underneath.

    AI tools

    The work is narrow and inspectable, not a general assistant pretending to be the chart. Document extraction from scanned notes and PDFs. Trend detection on labs and assessments. NLP on encounter text to surface reported symptoms — neck pain, poor sleep — as entities, not interpretations. Medication and follow-up gap detection against orders and care plans. Temporal grouping when physical, wellbeing and pathway changes overlap.

    Most of that is rules and small models with explicit thresholds. An LLM’s job is summarisation with mandatory citations: one line of clinician copy, every claim linked to a source entry. If the evidence is not there, the signal does not appear. This is not a chatbot drawer. It is a structured brief the clinician can verify in seconds.

    Clinician workflow

    1. Before the visit, the brief is prepared — blood trend, reported neck pain, wellbeing pattern, follow-up gap, chronology.
    2. The clinician opens the patient, scans the signals, clicks each one for the evidence drawer. That click is the trust model.
    3. They consult. Share-with-patient mode for approved items only. Same constellation, same receipts, language the patient can follow.
    4. Decisions and notes go back into the EHR as they do today. AI took the first pass. The clinician kept the judgement.

    Demo only. No patients, no accounts, no model. The brief is a fixture so the interaction can be designed on purpose. The note about why this exists is here.