.png)
A regional health network embedded a HIPAA-compliant AI layer into an EHR serving 340 clinicians without replacing the underlying platform.
.png)
A regional health network operating across 12 clinical sites — two acute care hospitals, four specialist outpatient clinics, and six primary care facilities — had deployed an EHR serving 340 clinicians and 280,000 active patients. Their problems were not with the EHR itself. They were with what it could not do. Clinicians were averaging 2.1 hours of documentation per shift, and the pharmacy team had identified that clinically significant drug interactions were being missed at a rate that had appeared in two consecutive patient safety committee reviews.
Two previous vendors had concluded that AI integration required replacing the EHR platform — an 18-month programme the board had declined to fund. Verttx found a path through the existing architecture. A HIPAA-compliant clinical decision support layer went live across all 340 clinicians and all 12 sites 12 weeks after the initial discovery call.
The documentation burden was measurable and serious. Emergency department clinicians were averaging 2.8 hours of documentation per shift. Primary care physicians completing complex wellness visits regularly exceeded 3 hours. The EHR's note templates had been designed for billing optimisation, not clinical efficiency — clinicians moved between structured fields, free-text sections, and external reference tools with no intelligence layer connecting the patient's existing record to the documentation being written.
The drug interaction problem was more urgent. The EHR's prescribing module had not been updated since 2021 and was configured to suppress Class C and Class D interaction warnings by default to reduce alert fatigue. A pharmacy team audit of 6,400 prescriptions written over 18 months found that 312 — 4.9% of the total — contained drug combinations that would have triggered a suppressed flag, and that 47 of those combinations had resulted in documented adverse drug events requiring clinical intervention.
Verttx conducted a four-week technical assessment of the EHR environment before agreeing any build scope. We identified three integration points within the existing architecture — including an unused SMART on FHIR interface that the EHR supported but the network had never activated — that could support AI capabilities without changes to the validated clinical platform.
All AI processing runs within the network's existing HIPAA-compliant AWS GovCloud tenancy under a Business Associate Agreement. PHI is never transmitted to third-party model APIs. Audit logs for all AI-assisted clinical interactions are retained for six years in OCR-ready format under 45 CFR §164.312.
A documentation layer embedded within the EHR's existing note editor — no new application, no workflow change — operates across three modes: structured data extraction that populates EHR fields as clinicians type or dictate, note completion suggestions surfacing relevant content from the patient's longitudinal record, and discharge summary generation that drafts from structured encounter data in 23 seconds for a standard acute admission. The model was fine-tuned on 840,000 de-identified encounter notes from the network's own EHR to reflect the network's documentation conventions and patient population.
Rather than reinstating suppressed alerts at full volume — which would have generated 340 additional alerts per day and been rejected by clinicians immediately — the module uses a patient-specific risk scoring model to surface only interactions that are clinically significant for the specific patient being prescribed to. A Class C interaction is only flagged if the patient's renal function, age, weight, and concomitant medication burden place them in a risk tier where that interaction carries a meaningful probability of adverse outcome. The result: more clinically relevant alerts at a lower total alert volume — alerts that clinicians actually act on.
Documentation time per shift fell from 2.1 hours to 1.3 hours — 48 minutes saved per clinician per day. Emergency department physicians saved an average of 61 minutes per shift. The freed time translated directly into patient capacity: available appointment slots increased by 12% within 90 days as clinicians completed documentation within consultation time rather than carrying it into after-hours work.
The drug interaction module surfaced 2,840 clinically significant alerts in its first 90 days. Of those, 2,190 resulted in the prescribing clinician modifying or cancelling the original prescription. Missed clinically significant interactions fell by 67% against the 18-month pre-implementation baseline. Clinician EHR satisfaction scores improved from 3.2 to 5.1 out of 7. Burnout index scores improved by 18 percentile points against the regional peer group, moving the network from the bottom quartile to the median. The board approved a phase two scope within 60 days of the phase one results presentation.
Both AI modules, the fine-tuned clinical language model, the RAG knowledge base, the FHIR integration layer, and all compliance infrastructure were transferred to the network's engineering team at handover — the network owns the system entirely and operates it independently within its own HIPAA-compliant environment.
Two vendors told us we needed to replace the EHR. We had already decided we couldn't afford that. Verttx found a way through the existing architecture and had the system live in twelve weeks. The drug interaction work alone has changed how safe our prescribing environment is. — Chief Medical Officer, Regional Health Network
.png)
Documentation time per clinician fell from 2.1 hours to 1.3 hours per shift — 48 minutes saved per clinician per day. Missed clinically significant drug interactions dropped by 67%. Clinician EHR satisfaction scores improved from 3.2 to 5.1 out of 7. The system went live across all 340 clinicians and 12 sites in 12 weeks, with no patient safety incidents during rollout, no EHR platform changes, and no disruption to live clinical operations.
.png)
.png)
