-
-
Overview
The Last Mile Problem
A 67-year-old man comes in for a routine follow-up. His physician's note from the previous visit reads: "CHF and CKD appear stable on current regimen. Continue metformin, lisinopril, spironolactone, and warfarin. Recheck labs next month."
His labs from that same day:
- eGFR: 18 mL/min/1.73m² (was 38 ten months ago — a 53% decline)
- Potassium: 6.2 mmol/L (critically high)
- INR: 4.8 (significantly supratherapeutic)
- HbA1c: 9.1% (rising despite treatment)
Thirteen days before this visit, urgent care added ibuprofen for knee pain. That provider had no idea his eGFR was 18 or that he was on warfarin with a rising INR. No one connected those dots.
This is not a rare edge case. This is what the EHR looks like when data lives in silos.
The Solution
SafeSignal is a dual MCP + A2A healthcare system built on Prompt Opinion that finds the clinical risks hiding between EHR data silos.
MCP Superpower —The Reusable Weapons: Four FHIR-powered clinical reasoning tools exposed as an MCP server that any agent on Prompt Opinion can invoke:
check_medication_safety— cross-references active medications against current lab values, conditions, and allergies; cites verbatim FDA drug label text alongside FHIR evidencedetect_silent_deterioration— longitudinal trend analysis across 24 months of Observations with explicit trajectory data; detects discrepancies between clinical note language and objective datafind_lost_followups— identifies abnormal findings with no documented follow-up in Encounters, Procedures, or ServiceRequests within clinical timeframes; correctly identifies resolved itemsgenerate_risk_briefing— orchestrates all three into a complete pre-visit risk briefing
A2A Agent — The Superhero: An intelligent agent published to the Prompt Opinion Marketplace that uses SHARP context to receive patient credentials, retrieves FHIR data, and delivers a severity-ordered risk briefing to clinicians before patient visits.
What Makes This Require Generative AI
Finding #7 in Samuel's briefing: "The encounter note from 2026-04-28 states 'CKD appear stable on current regimen.' This directly contradicts an eGFR decline of 38 → 18 over 10 months."
A rule-based CDS system can flag eGFR < 30. It cannot read a free-text clinical note, compare it to a ten-month trend, and write: "the assessment contradicts the data."
Finding #3: "Per FDA Drug Label (OpenFDA) - drug interactions section: '...take a blood thinning (anticoagulant)... are age 60 or older' (severity: documented)" — this is not LLM training knowledge. It is the live regulatory label retrieved from the FDA OpenFDA API, cited against a specific patient's active medications and current INR of 4.8.
Multi-factor temporal reasoning combined with real-time regulatory evidence citation is genuine generative AI territory.
The Knowledge Enrichment Layer (Novel Addition)
Before passing data to the LLM, SafeSignal enriches each medication with evidence from three authoritative public sources:
- RxNorm/RxCUI (rxnav.nlm.nih.gov) — standardised NLM drug identifiers for each medication ingredient, enabling database lookups
- FDA OpenFDA Drug Labels (api.fda.gov) — official boxed warnings, contraindications, and drug interactions text from FDA-approved regulatory labelling
- NLM RxNav Drug Interactions — ONCHigh (rxnav.nlm.nih.gov/REST/interaction/) — per-drug interaction lookup from the National Library of Medicine's high-quality curated clinical subset, filtered to co-prescribed pairs. Where NLM finds no ONCHigh entry, FDA label drug_interactions text is scanned as a fallback.
Each interaction entry carries an accurate source field — "NLM RxNav (ONCHigh)" or "FDA Drug Label (OpenFDA) - drug interactions section" — so the LLM cites the correct regulatory authority.
Demo Results — Samuel Brooks, 67M (CHF, T2DM, AFib, CKD Stage 4)
11 findings detected from a chart whose most recent note said "stable, continue management":
🔴 URGENT — Metformin contraindicated — eGFR 18 (FDA: eGFR < 30 contraindicated; lactic acidosis Black Box Warning)
🔴 URGENT — Spironolactone contraindicated — potassium 6.2 (FDA: contraindicated in hyperkalemia)
🔴 URGENT — Ibuprofen + Warfarin — bleeding risk (FDA drug interactions label cited; INR 4.8, patient age 67)
🔴 URGENT — Ibuprofen with eGFR 18 — nephrotoxic NSAID added by urgent care 13 days prior; prescriber had no FHIR context
🔴 URGENT — Lisinopril (ACE-I) exacerbating hyperkalemia — potassium 6.2, eGFR 18, combined with Spironolactone
🔴 URGENT — Warfarin INR 4.8 — FDA Black Box Warning: major or fatal bleeding risk; immediate review
🟡 WARNING — eGFR 38 → 18 over 10 months (−53%) — encounter note on same day said "CKD stable" — contradiction flagged
🟡 WARNING — Potassium 4.9 → 6.2 over 10 months — progressive hyperkalemia trend
🟡 WARNING — HbA1c 8.4% → 9.1% — worsening glycemic control; encounter note: "continue metformin"
🟡 WARNING — INR 3.4 in February — no documented dose review before April visit
ℹ️ INFO — Systolic BP 144 → 166 over 10 months despite two antihypertensives
✓ RESOLVED — Positive FOBT January 2026 → colonoscopy with polypectomy February 3, 2026 (correctly identified as addressed, not flagged as a gap)
12/12 automated validation checks pass. Three synthetic cases run cleanly (baseline, extreme polypharmacy, sparse-chart edge case).
Why This Architecture Wins
The MCP tools are designed to be reusable: a scheduling agent can call check_medication_safety before confirming a refill; a care coordination agent can call find_lost_followups across a patient panel; a triage agent can call generate_risk_briefing before routing. SafeSignal built infrastructure that makes the entire ecosystem smarter, not just one demo.
Compliance
SafeSignal is designed to support — not replace — clinical judgement. It does not diagnose, prescribe, or make treatment recommendations. Every finding uses language like "warrants clinician review." Every briefing cites specific FHIR resources with dates, values, and IDs. Every output ends with a compliance disclaimer. Follow-up gap findings qualify claims against what FHIR records were actually available (Encounters, Procedures, ServiceRequests) rather than asserting categorically that no action was taken. The system is functionally equivalent to clinical decision support, which the FDA has exempted from device regulation under the 21st Century Cures Act (Section 3060).
Log in or sign up for Devpost to join the conversation.