Meridian Health Network

Mobile application

The challenge

Meridian had grown by acquisition, and each acquisition brought its own record system. A clinician seeing a patient transferred from an affiliated clinic had no reliable way to see prior labs, medications, or imaging without phone calls. Chart review before a routine encounter averaged 11 minutes, and the network's own incident data tied 23 near-miss events in eighteen months to incomplete medication history.

Research

• Sixty hours of clinical shadowing across emergency, primary care, cardiology, oncology, pediatrics, orthopedics, and behavioral health
• Structured interviews with 34 clinicians and 12 health-information-management staff
• Quantitative audit of 400 encounters measuring time-to-relevant-information
• Review of 18 months of safety incident reports for information-availability root causes

The solution

Rather than a rip-and-replace of six EHRs — a multi-year program with a poor success record industry-wide — we built a read-optimized clinical workspace on top of them. A FHIR normalization layer ingests from each source system via HL7v2 feeds and vendor APIs, resolving patient identity through a probabilistic matching service with human review for ambiguous cases. Clinicians work in a single timeline; writes route back to the system of record for the originating site, so no source system loses authority.

UX decisions

Features

• Unified patient timeline spanning encounters, labs, imaging, medications, and notes
• Probabilistic patient identity matching with HIM review queue
• Order entry with real-time drug interaction and allergy checking
• Population health risk stratification across attributed panels
• Mobile companion for rounding with offline-tolerant note capture
• Immutable audit log with per-field access lineage
• CMS and HEDIS regulatory report generation

Technology

Architecture

Source systems publish HL7v2 and FHIR events into Kafka. A Python normalization service resolves identity and writes conformed resources into PostgreSQL with full event history retained, so any resource can be replayed to a point in time — a hard requirement for audit defensibility. The Next.js workspace reads through a FastAPI gateway that enforces role-based scoping at the query layer, not the UI. Everything runs on EKS across three availability zones with per-site traffic isolation, so a failure in one region's integration cannot degrade another's.

Results

Median chart-review time fell from 11 minutes to 5.8 across the measured cohort. Six legacy applications were decommissioned, retiring roughly $2.1M in annual license and support cost. Weekly active adoption reached 94% of licensed providers within two quarters of full rollout — notably without a mandate, which the CMIO had deliberately held in reserve. The platform sustained 99.98% uptime over fourteen months.

Lessons learned

1. 01Clinician trust was gated on provenance, not accuracy. The data was correct months before it was believed; source badges closed that gap faster than accuracy improvements did.
2. 02Rolling out to the two most skeptical clinics first was counterintuitive but decisive — their endorsement carried more weight internally than an easy early win would have.
3. 03Probabilistic identity matching needed a human review queue from day one. Our initial confidence thresholds were tuned too aggressively and produced merges that eroded trust disproportionately to their frequency.