Most clinical risk is invisible
because the signals are scattered.
The next generation of EHR systems will be AI-native, surfacing what matters before clinicians have to ask.
Where we are, where this goes
The same architecture across three increasing scopes
This demo
- ·Static panel JSON: 47 patients, 5 detailed.
- ·Insights are pre-written for the 5 detailed patients with explainability blocks.
- ·Five live AI workflows (callbacks, SBAR, MyChart replies, pharmacy consults, provider notes) drafted by Sonnet 4.6 with claude-3-5-sonnet fallback.
- ·Validators run server-side; Send is gated until they pass.
- ·No EHR connection. No PHI. Demonstration only.
Production at Phelps Health
- ·Read-only Epic feed (FHIR R4 + DocRef + flowsheet streams) into a panel cache scoped to one nurse's outpatient cardiology assignment.
- ·Insight generation runs on a Cloudflare Worker queue triggered by chart events (new lab, new MyChart message, refill gap).
- ·Workflows write back to Epic only via clinician-approved actions: draft to MyChart compose, draft to chart-note compose, draft to InBasket message. Never an autonomous send.
- ·All audit entries logged to a dedicated, immutable audit backend.
- ·Service-account auth + per-clinician identity in every workflow header.
Future ambient layer
- ·Microphone capture in the encounter room produces a live transcript that flows into the same insight engine.
- ·Real-time guideline-gap nudges surface during the visit, not after. "The patient is here, this is the moment."
- ·Predictive panel triage: at the start of the day, the panel is already sorted by risk-weighted need.
- ·Family-proxy MyChart messages auto-classified for prodrome-pattern matching against the patient's personal history.
- ·Continuous learning loop: which insights led to action, which were dismissed, why.
Phelps Health roadmap
Three dated phases. Integration with Epic, not replacement of it.
Read-only pilot, single clinic
- ·Stand up Kairos as a standalone read-only application with its own audit-logging backend, scoped to a single cardiology clinic for an initial cautious rollout.
- ·One-clinic, one-nurse rollout on the existing cardiology nurse workflow surface, scoped to an outpatient cardiology pilot.
- ·Read-only Epic-equivalent JSON feed (initially exported nightly, no live FHIR yet).
- ·No write-back to the chart. Drafts copied to clipboard for manual paste into Epic until clinician trust is established.
Live FHIR + draft-back integration
- ·Switch the panel cache to a live FHIR R4 subscription (Epic Bridges or equivalent) on the Cloudflare edge.
- ·Workflows draft into Epic's native compose surfaces: MyChart message draft, chart-note draft, InBasket message draft.
- ·Clinician sends from inside Epic; Kairos audit log records the handoff timestamp + workflow ID.
- ·Expand from one nurse to the cardiology nursing team (8 nurses, ~400 patients).
Cross-service rollout + ambient encounter
- ·Generalize the rule + validator framework to neurology, nephrology, and primary care. Three additional service lines.
- ·Add the ambient encounter layer: scribe transcript flows into the same engine as chart data; intra-visit guideline gaps surface in the room.
- ·Open a clinician-feedback path for proposing new rule blocks; rules are still code, but the proposal pipeline becomes a clinical product.
- ·Move from "panel review tool" to "clinical perception layer" in framing.
Kairos is positioned as a perception layer over Epic. It reads the chart, surfaces what matters, and drafts into Epic's own compose surfaces. Clinicians stay inside the system they already know. The chart of record stays Epic.
The role of the human
What Kairos preserves on purpose
Decision authority
Every action that touches a patient, a chart, a prescription, or a referral is initiated by a clinician click. Kairos drafts; clinicians decide.
Pattern judgment
The "AI Uncertain" tier exists because some patterns warrant attention without warranting a deterministic claim. The nurse decides whether to escalate; the system surfaces and stays out of the way.
Tone and trust
When Kairos drafts a MyChart reply to a worried family member, the nurse owns the voice. The validators block jargon and unsafe phrasing; the nurse adds the warmth.
Rule curation
New clinical guardrails are a code commit reviewed by clinicians, not a config screen. The const blocks are the policy; the LLM is downstream.
Where the value comes from
Directional framing. Actual numbers will come from the Phelps pilot.
Time the nurse gets back
The minutes per patient per shift currently spent reading labs, scrolling MyChart messages, and reconciling med lists across prescribers. Kairos does this work first, so the nurse arrives at the chart already oriented.
Compounds across panel size and shift length.
Catches that weren't happening
Cross-source patterns no single chart screen exposes: pharmacy refill gaps overlapping vitals drift, family proxy messages matching prior prodrome signatures, drug interactions across prescribers.
One avoided HF readmission ≈ $13-15K direct cost, before counting nurse time, family burden, and patient outcome.
Documentation that writes itself
The drafts already conform to clinical voice, structure, and length norms. The nurse edits for nuance instead of starting from a blank page.
Time-to-callback shortens; quality of the touchpoint goes up.
We will publish actual numbers (callbacks-per-shift delta, time-to-first-touch delta, prevented-readmission count) once the Phase A pilot has eight weeks of paired baseline + Kairos data. Until then we are not going to fabricate them.