Outpatient Synapse — Triage nurse dashboard

Grounded in Mary Kate Banks' email "Triaging Technology Ideas" (Sun 7/19, to Steve and Sam, cc Prachi), written after the Friday 7/17 NYC conversation. Her front-line asks, mapped onto the AI triage engine already built and safety-hardened in the OP sample — with mockups of the combined thing in the outpatient Synapse frame.

Built from her words and a verified repo

The email is the requirements source. Every "exists today" claim below was checked against the OP-sample code on 7/20. Where she offered options — 1–5 levels vs. ER-style colors — this doc recommends one and says why.

▶ Working demo — click through the queue, pathway, and dispositions Open working demo ↗

Section 1

What Mary Kate asked for

Four asks, straight from the email. The teal MK chip marks these as hers, not decisions made in a room.

Portal messages carry no urgency signal. "We have to go through each patient message manually to make sure we are getting to those urgent messages first."

Red-flag lists exist but are applied hit-or-miss by non-clinical staff. "I have developed some neurology specific 'red flags' or panic words… This is all done by non-clinical staff and it's hit or miss if it's done accurately."

Risk-tier every inbound contact from the chart, not just the message. "…categorized into different levels of risk based on their clinical history, neurological conditions, labs, imaging, and current issue they are reaching out about" — as 1–5 levels or ER-style colors.

Turn triage dot phrases into if/then pathways. "Based on the patient answers or nursing assessment either more triaging questions are created OR the nurse is given next steps (send to ER, contact provider, etc)."

Section 2

Most of the dangerous part is already built

Line up her four asks against the OP-sample repo as it stands today, and the gap is much smaller — and much less dangerous — than it looks from the outside.

Her ask	In the OP sample today	The gap
Urgency-tier every inbound contact	7-tier classifier + care-pathway layer (`emergency_now` / `same_day_clinician_review` / expedited / routine / redirect / undetermined) — display colors already in the code.	Triages referral documents only, not portal messages or call-center encounters — new input lanes, same machinery.
Red flags applied consistently, not hit-or-miss	12-syndrome emergency gateway: deterministic anchors with negation/temporality suppression + exact quoted evidence, cross-checked by an independent extractor; disagreement escalates, never silently downgrades.	Her call-center word lists become seed anchors + test cases; a nurse-editable library view doesn't exist yet.
Tier from chart context — history, labs, imaging	The app holds the chart (history, meds, prior visits, imaging reports); the engine accepts only age/sex/referrer today — text-only otherwise.	Wire chart context into the engine's structured input, with provenance shown.
If/then pathways, not static questions	Flat dot phrases exist; a governed pre-approved clarification-question mechanism already exists for the AI historian.	The branching pathway builder itself is new; the governance pattern to build it on already exists.
A nurse queue, worst first	Batch panel sorted by severity exists — for referral batches.	A persistent multi-source queue UI is new.
Patient comms — acknowledgment, follow-up	Portal inbox (patient side), Twilio SMS/voice follow-up agent, and voice historian intake all exist.	None for the demo — staff-side wiring only.

The takeaway

The hard, dangerous part — a safety-gated classifier with an emergency gateway, independent cross-check, adjudication, holds, and a release-gate eval harness — exists and survived adversarial review. What's missing is her front door: the sources, the queue, and the pathways.

Mockup 1

One queue — every inbound contact, worst first

Every source lands in one place, sorted by the engine's care pathway — not by which system it arrived through.

Synapse

Acute Care

Rounding

EEG

Outpatient

Triage

History

Imaging/results

All 7

💬 Portal 4

📞 Call center 2

📄 Referral 1

My queue

Mon 7:42 AM

🚨

EXISTS TODAY Emergency gateway · 7:33 AM — call-center contact matched acute-cerebrovascular red flags. Scheduling locked · 911/ED guidance script shown to rep · covering neurologist paged · awaiting nurse confirmation.

Mirrors the real gateway → lock → EmergencyActionPanel behavior.

Contact now — emergency_now (1)

David Okafor, 67 · post-TIA (2 wks)

📞 Call center 7:31 AM

"face drooping again since breakfast, speech slurred"

EMERGENT — Redirect to ED Immediately 🔎 evidence quoted

11 min Open now

Gateway fired — see strip above.

Same-day — same_day_clinician_review (2)

Margaret Ellison, 41 · MS — ocrelizumab

💬 Portal 6:58 AM

"new numbness in both legs since yesterday, tripping when I walk"

SEMI-URGENT possible relapse; infusion overdue

44 min Start pathway

Robert Chen, 72 · Parkinson's

📞 6:15 AM

"fell twice this morning, no injury — carbidopa-levodopa ran out Friday"

SEMI-URGENT fall risk + abrupt med stop

1 h 27 m Start pathway

Routine — routine_outpatient (3)

Ana Reyes, 31 · migraine

💬 yesterday 9:12 PM

"topiramate refill before travel next month"

ROUTINE

Open

Tom Whitfield, 58 · neuropathy

💬 yesterday 6:40 PM

"question about sleep study results"

ROUTINE

Open

Priya Nasser, 45 · epilepsy f/u

📄 Referral 8:10 AM

Routine neurology referral — seizure-free 14 months, requesting continuity visit.

ROUTINE

Open

Human first — undetermined (1)

Dorothy Klein, 79 · complex hx

💬 7:05 AM

"long message — multiple symptoms over weeks, some conflicting with chart"

INSUFFICIENT DATA — no tier assigned

When the branches disagree or the text isn't enough, the engine already refuses to guess — it holds the case for a human. Same behavior it has for referrals today; nothing silently defaults to "low risk."

Fake demo patients. Tier chips and colors are the engine's own display set, unchanged.

Mockup 2

Margaret's message — what the nurse sees when she opens it

Same engine, two layers shown together: the care pathway that drives urgency, and the pathway questions that drive what happens next.

Synapse

Acute Care

Rounding

EEG

Outpatient

Triage

History

Imaging/results

💬 PORTAL MESSAGE · TODAY 6:58 AM

Since yesterday I've had numbness in both legs and I keep tripping when I walk. No fever, no new medicines. It feels like before my last flare but not as strong. Should I worry or wait for my October appointment?

From the chart — wired in (the extend step)

EXTEND

MS, dx 2019Problem list

Ocrelizumab — last infusion 5 mo ago, overdue Meds

Last visit 3/12/26 — stable, gait normal Note

MRI brain 1/2026 — no new lesions Imaging report

Today the engine only reads submitted text; the chart feed is the wiring this design adds — with provenance on every fact, extracted never invented.

AI assessment

Pathway: SAME-DAY same_day_clinician_review

SEMI-URGENT (seen ≤2 wks)

New bilateral sensory symptoms + gait change in a patient with MS — relapse vs. pseudo-relapse. No emergency red-flag features in the message; the overdue infusion raises suspicion.

"numbness in both legs"Message

"tripping when I walk"Message

Ocrelizumab overdue 5 moMeds

Stable, gait normal 3/12/26Note

Confirm Change

The AI proposes; the nurse disposes. Overrides carry reason codes and land in the calibration dashboard — both already exist in the OP sample.

Pathway — MS: new/worsening sensory symptoms

DRAFT — ILLUSTRATIVE

✓Fever, infection symptoms, or recent illness?

→ No

Any new bladder or bowel changes?

Yes

Yes → escalate to provider now + infection screen

No → "Symptoms >24 h?" → provider callback today

Disposition

911 / ED now Same-day provider Nurse advice + follow-up Routine schedule

Auto-drafted triage note

Source: portal message, 6:58 AM. Symptoms: new bilateral leg numbness + gait unsteadiness since yesterday, no fever, no med changes. Pathway: infection screen negative (patient-reported); bladder/bowel question pending. Working disposition: same-day clinician review — pending nurse confirmation.

🎤 ⚡ ✨

Send to provider — with note Schedule callback · SMS confirm SMS/voice follow-up agent already in the OP sample

Illustrative, not clinical

Pathway content here is illustrative, not clinical guidance. Real pathways are versioned clinical assets — authored by the triage RN + physician, pre-approved the same way the historian's clarification questions already are, and run through the sentinel-style release gate before any patient contact.

Mockup 3

Her red-flag lists become a governed clinical asset

Same anchors she already uses, formalized into the structure the gateway already runs on — versioned, owned, and tested rather than typed from memory.

v3 Owner: triage RN + physician sign-off · Applied to every message and call, identically, every time

Synapse

Acute Care

Rounding

EEG

Outpatient

Triage

History

Imaging/results

Phrase anchors	Syndrome class	Behavior
"worst headache of my life", "thunderclap"	Intracranial hemorrhage / SAH	Gateway: contact now
"face drooping", "slurred speech", "one side weak"	Acute cerebrovascular	Gateway: contact now
"curtain over my eye", "vision went black"	Acute vision threat	Gateway: contact now
"first ever seizure"	Status / recurrent seizure	Gateway: same-day floor

This is not naive string match — the deployed gateway already handles negation ('no facial droop' doesn't fire), temporality ('droop last year, resolved'), and quotes its exact evidence offsets. And a second, independent AI reads the same text; if the two disagree, the case escalates or holds — it never silently downgrades. Her lists make the anchors better and become test cases in the release gate; the model layer catches phrasings no list anticipates.

Section 6

Her two tier schemes

She offered two ways to color-code risk. One means building a taxonomy the engine doesn't have; the other means displaying the one it already does.

OPTION A

1–5 numeric

Granular, sortable, a familiar acuity feel — the ER-triage instinct in its most literal form.

+Granular and sortable; matches how she already thinks about acuity.

−Five levels invite false precision — validation reruns showed borderline cases flip between adjacent tiers on identical inputs, and a 5-way split doubles the borderlands.

−"3 vs 4" tells a nurse nothing about what to DO.

−Duplicates a taxonomy the engine already has.

OPTION BRecommended

Colors = the engine's care-pathway layer

Don't invent a new scale — the engine already outputs two: a care pathway (response time: contact now / same-day / routine / human-first) and a 7-tier scheduling depth (ED-now → 6 months, colors already defined). The nurse queue runs on the pathway layer — every lane IS an action — with the tier chip alongside for scheduling depth.

+Her ER-triage instinct maps 1:1 onto what's built.

+Zero new taxonomy to build, govern, or explain — it's already shipped and display-ready.

+"although in an outpatient setting it is far less dramatic" — her own caveat, acknowledged: muted chips, pulsing reserved for the true emergent case.

Section 7

What combines from where

Sorting every piece of this doc into three buckets: what's shipped, what's shipped but needs a new wire, and what doesn't exist yet.

✓

Exists
Safety-gated classifier (gateway → cross-check → fusion → adjudication → holds) Tier + pathway taxonomies with display colors Scheduling locks + emergency action log Override w/ reason codes + blind-review calibration console Sentinel release-gate harness Patient chart in-app Flat dot phrases Governed pre-approved question mechanism (historian) Portal inbox (patient side) SMS/voice follow-up agent Plans library (127, ICD-10-matched) for next-step content

↗

Extend
Two new input lanes — portal messages + call-center encounters — into the same engine Chart context wired into the engine's structured input (with provenance) Care-pathway lanes as the queue's organizing axis Rerun the sentinel gate against the new input types before live use

New
The nurse queue UI (multi-source, persistent, assignment + timers) The pathway builder (dot phrases → branching questions → dispositions) Red-flag library manager view Callback scheduling + timers on queue rows Portal auto-acknowledgment with standing 911 safety text

Section 8

Open questions

1. Where does this live first? Recommend: prove it in the OP sample — the engine, chart data, and comms plumbing are already there and it's the established demo ring — then port the validated design into Synapse outpatient. Synapse is the destination; the OP sample is the fastest honest prototype.

2. The call-center lane. Near-term the rep types the encounter (as today) and it tiers on entry; the phone-agent R&D work (Kiran's red-flag gate) gives a hands-free path later. Sequencing question, not scope.

3. Red-flag library governance. Owner (triage RN lead + physician sign-off?), review cadence, versioning. Cheap to decide now, painful later.

4. Pilot data for the eval gate. A de-identified sample of real portal messages + call-center encounters, so the sentinel-style gate runs against the new input types before anyone triages live traffic. (The sentinel is a release gate, not clinical validation — same rule here.)

Sources: Mary Kate Banks, "Triaging Technology Ideas" (email, 2026-07-19) · NYC conversation 2026-07-17 (Steve, Sam) · OP-sample repo verified 2026-07-20. Fake demo patients throughout. Series: outpatient-synapse-tabs.html · outpatient-mockups.html · this file · outpatient-triage-nurse-demo.html.