Healthcare Hackathon Slate (Abridge event)

Event prompt: pick one clinical or operational workflow; build an agentic system that makes it faster, smarter, or safer; ship something a clinician or patient-facing team could use Monday. Abridge engineers + clinicians pressure-testing on the floor.

Doctrine applied: the prompt's own examples (trial matching, prior auth) are what half the room builds — avoided. Abridge = ambient transcription, so ideas sitting ON TOP of encounter transcripts land hardest with the judges. All demos on synthetic data (generated transcripts + Synthea FHIR); paperwork and logistics, never diagnosis; every irreversible step gated.

---

The slate

1. Loose Ends (first pick — the follow-through engine; merges "Said vs. Signed" + referral-loop closure) — Healthcare's deadliest mundane failure is the dropped thread: the doctor says "we'll get you a cardiology referral and check your thyroid," the note looks complete (increasingly because an AI scribe wrote it), but no referral order exists and no TSH was placed — and even when the referral IS sent, ~half never complete: faxes vanish, patients don't book, reports never return. One agent owns the whole thread lifecycle: (a) catch — diff what was said in the encounter transcript against what was actually ordered and documented, flagging gaps before signature; (b) chase — track every open thread (referrals, orders, results), detect the stall (no state change in N days, urgency- tiered), and nudge through the right channel (fax rail, portal, patient text); (c) close — file the consult note back and mark the loop closed, or escalate. Demo arc: visit transcript promises three things → chart shows two → the missing referral turns red pre-signature; then time-compress three weeks → the urgent derm referral stalls at day 12 → chased → closed. The scribe writes what was said; we catch what was promised and make sure it happens. Spine: eval=catch-rate on transcripts seeded with dropped intentions + stall-detection precision on a synthetic referral cohort · measure=dropped threads per 100 visits, referral closure rate · orch=transcript-differ + thread tracker + channel executors · tools=transcript ingestion, FHIR, fax/portal rails · gate=flags and chases, never places orders; urgency tiers and deadline math are code; sign-off stays human. Why it wins the room: Abridge-native (verification layer on top of their category), a guarantee rather than a drafting wrapper, and a pure appreciation play — the more AI writes the notes, the more the world needs the thing that checks follow-through. Extensions (same engine, more thread classes): tests pending at discharge (open whitespace — see Pending, #4); NOT incidental imaging findings (Rad AI Continuity and inflo Health own that category). ⚠ Floor-test first: Abridge's June 2026 platform keynote includes order generation from conversation — ask their engineers where their roadmap stops; the downstream chase-to-closure half is safely beyond it, the said-vs-signed diff may be adjacent.

2. First 30 Days (the "saving lives" pick — discharge companion, voice-native) — Readmissions cluster in the 30 days post-discharge and the causes are mundane: meds never picked up, instructions not understood, follow-ups never booked, symptoms ignored until day nine. A voice agent runs the transition: plain-language (and native-language) discharge recap, med-pickup confirmation, follow-up booking, short daily check-in calls with memory — and protocol-driven escalation where thresholds come from the discharge orders and live in code, never the model's judgment. Demo: synthetic CHF discharge; the day-3 call catches weight gain plus a missed diuretic pickup and pages the nurse. Spine: eval=escalation precision/recall on seeded red-flag scenarios · measure=readmission-driver interceptions, nurse minutes · orch=recap writer + caller + protocol sentinel · tools=ElevenLabs/telephony, discharge orders, pharmacy status · gate=escalation thresholds are deterministic; it never reassures away a red flag — ambiguity escalates. Risk to prep for: clinician judges will probe escalation safety hard — survivable precisely because the gates are code.

3. Fax Desk (the most on-prompt pick — with an honest caveat) — Clinics receive hundreds of faxes a day (referrals, outside records, results), hand-sorted by burned-out staff; the dangerous failure is the critical result sitting in the pile overnight. Agent: classify, extract, match to patient, route to the right queue, escalate urgent results in seconds — routing accuracy and urgent-catch-rate on screen as the evals. Demo: 50 synthetic faxes → sorted in a minute → one buried critical potassium escalated with a page. Spine: eval=routing accuracy + urgent catch-rate/false-alarm rate · measure=staff hours, time-to- escalation · orch=classifier + extractor + router + sentinel · tools=OCR, fax rail, FHIR · gate=urgent escalations always page a human; nothing files to a chart unreviewed. ⚠ Caveat, stated out loud in the pitch: this category has a funded incumbent (Tennr) — great hackathon demo and Monday-shippable, weak product whitespace. Pick it for execution points, not novelty.

(Cut from the original slate: standalone prior-auth — the example prompt everyone builds, and the differentiated version's best parts [compile-or-fail evidence gate, outcome odds] fold into Loose Ends and Casebook anyway.)

What we're deliberately NOT building (incumbent check, researched 2026-07-06)

Walking in knowing this is ammo — "we checked; that ships already" is a credibility move with clinician judges.

• Ambient scribing / chart summarization / pre-rounding — Abridge (the sponsor) ships pre-round notes as of June 2026; Epic's Art Insights runs 16M uses/month. Oligopoly.
• In-basket drafting — Epic MyChart ART: 150+ orgs, 1M+ drafts/month. Solved at scale.
• Nurse brain sheet / handoff — Google Cloud + HCA Nurse Handoff (scaling to 99K nurses), Abridge for Nurses (250+ systems), Epic end-of-shift notes. Solved, consolidating.
• Radiology incidental-finding tracking — Rad AI Continuity (50+ finding categories) and inflo Health (RSNA 2025) own the category. Deployment gaps exist; the algorithm is claimed.
• PA/benefits phone automation — SuperDial, Prosper AI already do payer IVR calls.

Cognitive-load batch — reduce what lives in a clinician's head

Target: the tracking, chasing, and redundancy clinicians carry mentally between patients. Verdicts and stats from the 2026-07-06 researcher sweep (peer-reviewed where possible; vendor-sourced stats flagged). Each is whitespace or honestly-caveated partial whitespace.

4. Pending (hospitalist — tests pending at discharge; the strongest whitespace in the sweep) — ~40% of hospital discharges leave with a test still pending; 30–40% of those return actionable results; physicians are unaware of 62% of them. Today the coping mechanism is whoever happens to remember. The agent: at discharge, inventories every pending lab/micro/path, binds each to an accountable owner, watches for results, and when an actionable one returns runs an action-forcing loop — not notify-only (the systematic review found notification-only interventions underperform): the result arrives with a drafted next step, requires acknowledgment, escalates up the chain if unacknowledged, and notifies the patient once the owner acts. Demo: synthetic discharge with three pending tests → the day-6 pathology result returns actionable → routed with draft plan → unacked for 48h → escalated → closed, patient notified. Spine: eval=catch + acknowledgment rate on a seeded TPAD cohort vs notify-only baseline · measure=actionable results reaching an accountable owner, days-to-action · orch=inventory + watcher + escalator · tools=FHIR, results feeds, paging/messaging · gate=ownership assignment and escalation ladders are code; it never interprets the result, it delivers it with the chart context. Composes with Loose Ends as a thread class, or stands alone. Evidence base is strong (AHRQ/JHM) and no incumbent owns it.

5. Quiet Read (radiologist — the reading-room phone, and the creative standout) — On-call radiologists field ~72 phone calls per 12-hour shift — ~108 minutes on the phone — and there's a 59% chance of interruption in any 10-minute CT read; interruptions during interpretation are an error risk, not just an annoyance. The agent answers the reading-room phone: identifies the caller and ask ("has the head CT been read?" → answers from PACS status directly), takes structured messages, batches non-urgent queries for delivery at natural between-study breakpoints, and passes true emergencies (stroke, trauma, unstable patient) through instantly by a policy that lives in code. The mental-load frame: protect the read. Demo (voice — ElevenLabs makes this land): a simulated reading session under a call flood — the agent resolves eight inquiries itself, batches seven for the break, and passes the one stroke call through in seconds; interruption count and protected reading minutes on screen. Spine: eval=zero false-holds on an emergency-scenario set (the critical number) + resolution rate on routine inquiries · measure=interruptions per read, phone-minutes returned · orch=answerer + PACS-status resolver + batcher + escalation sentinel · tools=telephony, PACS/RIS status APIs · gate=the pass-through policy is deterministic; when in doubt, interrupt the radiologist — the failure mode must be over-interruption, never under. Whitespace confirmed: triage AI (Aidoc) reduces some call volume upstream, but nobody answers the phone.

6. Page One (on-call/cross-cover — context-wrapped paging) — Cross-covering physicians get blind pages ("pt in 412, BP 180/100") about patients they've never met, then pull the chart cold at 3am; vendor data claims ~70% faster response when context arrives with the call. The agent intercepts the page and wraps it: a three-line snapshot — who the patient is, why they're admitted, code status, the meds/labs/trends relevant to this page, and what the caller wants — assembled in seconds, stamped with data freshness (a stale summary in an emergency is a safety failure, so every snapshot carries its generated-at time and true emergencies bypass entirely). Demo: same page, blind vs wrapped, decision time on a stopwatch. Spine: eval=snapshot factual accuracy vs chart + relevance rating by clinicians

   • zero-delay pass-through on emergencies · measure=time-to-decision, cold chart pulls avoided · orch=page listener + context assembler · tools=paging integration, FHIR · gate=emergency classes bypass unconditionally; staleness threshold refuses to send old data. Partial whitespace: point products exist (healow Genie), no category leader.

7. Form Feeder (cross-role — FMLA/disability/DME/school notes) — Every form is 10–20 minutes of focused physician work — "an appointment slot's worth" — with zero clinical value, and the sweep found no dedicated incumbent. The agent: inbound form → every field mapped to a specific chart element (citation-or-flag — unanswerable fields flag for the human, never improvised, because a hallucinated clinical justification on a federal form is fraud-adjacent, not a typo), drafted for signature, plus the part nobody does: a durable registry of what's pending, expiring, and due for renewal so recurring forms auto-restart. Demo: FMLA cert in → drafted with per-field chart citations + two honest flags in 60 seconds → signature → the renewal appears on the registry. Spine: eval=field accuracy vs labeled forms + zero-fabrication rate · measure=minutes per form, renewals never missed · orch=parser + evidence mapper + registry sentinel · tools=chart/FHIR, form templates, docx/PDF · gate=citation-or-flag; nothing leaves unsigned. RFP Machine's mechanic in clinic clothes.

8. Card Audit (surgeon/OR — preference-card hygiene) — Surgical preference cards drift years out of date, so ORs over-pull instruments and supplies that get opened, unused, and wasted, and circulating nurses guess. The agent reconciles each card against what cases actually used (case logs, supply scans): proposes per-surgeon card diffs with evidence ("laparoscopic tray item X opened in 0 of last 25 cases — remove?"), surgeon one-tap approves. Understudy's learn-from-behavior mechanic in the OR. Demo: 25 synthetic case logs vs a stale card → a diff with dollar figures → approved card v2. Spine: eval=diff precision vs a nurse-manager-labeled card · measure=$ waste per case removed (⚠ be honest that the waste stats are vendor-sourced, not peer-reviewed — clinician judges will know) · orch=usage miner + diff proposer · tools=case logs, supply data · gate=cards change only on surgeon approval; safety-stock items are never proposed for removal. Whitespace: LeanTaaS writes blog posts about this; nobody ships the auto-audit.

9. Day-Of (pre-op — cancellation prevention, honestly caveated) — Day-of-surgery cancellation rates run 2–24% and over 80% are preventable/administrative: the missing cardiology clearance, the lab never drawn, the consent unsigned, the NPO instructions never confirmed. The agent runs the readiness checklist per case: chases each item through the right channel days ahead (assessment ≥24h pre-op cuts postponement risk ~50%), forecasts which cases will cancel ("case 3 cancels Thursday unless cardiology clearance arrives — chased, ETA today 4pm"), and escalates while there's still time to backfill the slot. Spine: eval=cancellation prediction precision on a synthetic cohort · measure=preventable cancellations caught, OR minutes saved · orch=per-case checklist sentinel + channel chasers · tools=scheduling, FHIR, fax/phone rails · gate=checklist completeness is deterministic; clinical fitness calls stay with anesthesia. ⚠ Qventus launched exactly this in Jan 2025 (early-stage, cohort deployments) — the honest angle is the long tail they don't serve: ambulatory surgery centers and small hospitals. Pick this only with that framing.

10. P2P Corner (the peer-to-peer prep, kept deliberately small) — Voice agents now do payer IVR and benefits calls (SuperDial, Prosper), but the physician-to-physician prior-auth call still requires the physician — and they walk in cold. The agent can't take the call (and shouldn't: state laws are moving against autonomous AI determinations) — it schedules it and preps the one-pager: the payer's own policy criteria mapped line-by-line to this patient's chart, the denial's weak points, and what overturned similar denials. The gate IS the pitch: it preps the human, never replaces them. Slice-sized; pairs with Casebook.

Picks

• Overall first pick: Loose Ends (#1) — but pressure-test the said-vs-signed half with Abridge engineers in hour one (their order-generation roadmap may be adjacent); if it overlaps, pivot the same engine to Pending (#4), which is the cleanest confirmed whitespace in the file and shares the thesis: follow-through as infrastructure.
• Creative standout: Quiet Read (#5) — voice-native, visceral demo, protects a safety-critical cognitive task, and nobody in the room will have it. Best "judges retell it later" candidate.
• Fastest to Monday-shippable: Form Feeder (#7) — smallest integration surface, clearest time-savings number, zero-fabrication gate as the trust moment.