MaternalGuard

The most preventable pattern in US maternal care is silent between visits.

Roughly 800 US maternal deaths per year, sixty percent of which the CDC classifies as preventable. The largest single class of preventable death is preeclampsia signals evolving between visits—rising blood pressure, elevated AST, falling platelets, new proteinuria—because a clinician has to open five different chart views to see them together. MaternalGuard reads BP, AST, platelets, conditions, and SDOH in one MCP tool call and drafts the MFM consult, with the ACOG citation and the Provenance every draft needs for (b)(11) compliance.

The buttons-only huddle Pawan Jindal keeps writing about.

From 1000 Clicks to Copilots: the fewer keystrokes between a clinical intent and a chart write, the more useful the copilot. MaternalGuard's dashboard is designed for the huddle table, not the chart form.

• Approve, Reject, Activate, Dismiss. Zero typing. Server writes the default audit reason so the trail stays populated without asking anyone to type.
• Instant optimistic toast on click: Approving: <task> fires before the FHIR round-trip. The one-to-two-second write completes, then a green Approved banner renders at the top of the auto-refreshed huddle. No scrolling to confirm.
• Ten MCP tools with A2A wiring: AssessMaternalRisk, InterpretLabTrends, ScreenSocialDeterminants, GenerateCarePlan, PredictNeonatalImpact, ProposeMaternalAction, UpdateMaternalAction, OpenMaternalDashboard, ListMaternalActions, MaternalPanelScan.
• Every tool schema is stringify-safe. z.union([native, string]) plus server-side coercion, so upstream agents that occasionally stringify arguments never -32602 the call.
• SHARP-on-MCP: consumes X-FHIR-Server-URL, X-FHIR-Access-Token, and X-Patient-ID straight from Prompt Opinion's FHIR Context Extension.
• Soft-degradation contract: any FHIR read failure returns {status: "no-op", blocksNextStep: false}—never a red banner. The Priority-1 chart-writeback loop always completes.

Every governed write ships an AI Transparency Provenance you can audit.

For a certified body module like Darena Health / MeldRx, the interesting artifact is not the tool call. It is the audit trail behind every chart write. MaternalGuard emits the shape below on every Task and every Flag transition.

• meta.profile[0] = AI-Provenance (HL7 AI Transparency on FHIR IG v1.0.0-ballot, DSTU January 2026).
• meta.security[0].code = AIAST (Artificial Intelligence Asserted, HL7 v3 ObservationValue).
• Contained AI-Device with aiKind = rule-based and the DSI identifier mg-dsi-v0.4.0, linked to a contained AI-ModelCard DocumentReference.
• The DocumentReference points at a resolvable HTTPS URL, not a placeholder. A reviewer clicking the source attribute lands on the live CHAI Applied Model Card JSON or its Markdown companion, not a 404.
• All thirteen 45 CFR 170.315(b)(11)(iv)(A) Evidence-Based source attributes are ALSO carried inline on every Provenance summary extension, for viewers that do not follow references.
• USCDI v3.1 alignment. HEDIS Prenatal & Postpartum Care (NQF #1517) via CPT Category II 0500F/0503F auto-tagged on Task.reasonCode.
• DSI version bumps land in the same commit that touches the JSON, the Markdown, and the MATERNALGUARD_DSI_VERSION constant. That is the "ongoing maintenance" surface the ONC rule expects to see used.

Reproduce the demo in about a minute.

MaternalGuard is deployed as a Prompt Opinion MCP server. The reproduction path assumes you already have a workspace with a FHIR context.

1. Register the MCP server in your workspace.

   Launchpad → Configuration → MCP Servers → New. Endpoint https://maternalguard.jonathanandrei.com/mcp, transport Streamable HTTP, auth API Key on header X-API-Key. Contact for a key. Enable FHIR Context.

2. Load Maria Santos as the patient.

   Import patient-maria-santos-bundle.json via Patient Data → Import. She is a 28-year-old G2P1 at 32 weeks with evolving HELLP: rising BP (118 → 148 systolic), AST 54 U/L, platelets 141 → 100 K/uL, spot proteinuria.

3. Open the Prenatal Visit Prep agent and send this prompt.

   Draft the appropriate follow-up actions for the current patient.

   The agent chains five data tools once each, then ProposeMaternalAction once per RED / YELLOW recommendation, then OpenMaternalDashboard. Three to five seconds. No red banners. The dashboard renders inline with Approve / Reject buttons per task.

4. Click Approve on the top task.

   Instant blue Approving: <task> toast, then a one-to-two-second FHIR round-trip, then a green banner at the top of the refreshed huddle. The task moves from Draft to Recently actioned with the accepted badge.

5. Verify the FHIR Task and Provenance landed.

   The Postman collection in the repo has pre-wired requests. Import it, paste your workspace bearer, run Tasks → GET Task by ID and Provenance → GET Provenance targeting a Task. The Provenance carries the AI-Transparency profile, the AIAST security label, the contained AI-Device and AI-ModelCard, and the thirteen source-attribute inline summary.

Three next steps out of pilot.

Prompt Opinion's own framing is that healthcare AI is stuck in pilots and the issue is assembly, not agents. What follows are the three shortest steps that would move MaternalGuard from prototype to something a covered entity could actually turn on.

1. CDS Hooks bridge alongside MCP. Tools port to CDS Hooks in about a day; tool logic and Provenance shape stay identical. Ships against every SMART-on-FHIR EHR: MeldRx, Epic App Orchard, Cerner code sandbox.
2. Prospective sensitivity/specificity study against an ACOG-benchmarked cohort of 200 to 500 real prenatal charts. Moves the 61-case corpus false-positive claim (under 1%) into a real ROC curve.
3. Design partnership with an OB residency or MFM service for eight to twelve weeks in real morning-huddle use. Turns the "1 in 25 severe preeclampsia" hook into a measured outcome instead of an epidemiology reference.

What ships in every write.

Clinical guideline citations

Regulatory & interoperability