AI in the chart. PHI never in the cloud.
eSpiral deployed a SMART on FHIR chart assistant at Infirmary Health, giving residents AI-powered clinical support during live patient care, with zero PHI exposure to external servers.
Staging environment — all patient data is synthetic. Deployed identically in live hospital settings.
Deployment
Live Hospital
Infirmary Health system
Program
Residency
Physician training support
Integration
SMART on FHIR
Native EHR chart access
Background
AI during residency. The stakes are real.
eSpiral is a clinical intelligence platform built for hospital systems. At Infirmary Health, it runs natively inside the EHR as a SMART on FHIR application, meaning physicians never leave the chart to access AI support.
Residency programs present a specific challenge: trainees move quickly between patients, need rapid clinical context, and operate under attending supervision. AI assistance has clear value. But traditional AI tools require data to leave the building, a non-starter in any HIPAA-regulated environment.
The constraints they had to solve
AI value is real
Residents benefit from on-demand clinical guidance during active patient care: differential diagnosis, drug interactions, evidence-based protocols.
PHI can't leave the building
Hospital compliance and HIPAA prohibit sending identifiable patient data to external AI APIs. Consumer LLMs are off the table without de-identification.
The workflow has to be seamless
A solution requiring manual data export or a separate login would fail adoption. It had to live inside the chart.
The product
Three panels. One clinical picture.
Temporal condition chart
ICD-10 conditions mapped by quarter across years — complex patient histories rendered as a radial timeline, readable at a glance.
Problem list, organized
49+ active conditions surfaced, searchable, and grouped by year of onset. The longitudinal view that gets buried in tabular EHR views.
Chart Assistant
Physicians and residents ask clinical questions. The assistant answers grounded in the actual chart data — differential diagnoses, drug interactions, screening gaps.
In production
The chart assistant, live in the workflow.
The radial visualization maps a patient's full condition history by quarter and year. The Chart Assistant answers clinical questions grounded in that data, all without PHI reaching external servers.
Staging environment shown. All patient data is synthetic. Deployed identically in live hospital settings.
“Residents are eager to use the latest tools. That’s never been the challenge. The challenge is giving them access in a way the institution can actually stand behind. MedScrub solved that. We know exactly where the data goes and where it doesn’t.”
Dr. David Clarkson
Program Director, Infirmary Health Residency
The solution
How MedScrub makes it work.
01
Physician opens a patient chart
The eSpiral SMART on FHIR app launches inside the EHR. The patient's full problem list, labs, and visit history are loaded — all within the hospital network.
02
PHI is stripped before the AI sees anything
Patient identifiers (names, dates, MRNs) are replaced with opaque tokens by MedScrub's self-hosted proxy. The de-identified payload travels to the AI model. PHI stays on-premise.
03
The Chart Assistant responds in context
Physicians and residents ask clinical questions and receive evidence-based answers grounded in the actual chart data. The proxy re-applies token mappings on the way back.
Clinical data
Patient: John Doe
DOB: 03/14/1978
MRN: 4829301
CC: knee pain, bilateral
Provider: Dr. Sarah Park
Inside hospital network
MedScrub Proxy
De-identification
18 HIPAA Safe Harbor identifiers stripped and tokenized
Self-hosted, on-premise
AI Model
Patient: [NAME_1]
DOB: [DATE_2]
MRN: [ID_3]
CC: knee pain, bilateral
Provider: [NAME_4]
PHI never transmitted
Non-PHI text passes through unchanged. Only identifiers are replaced with opaque tokens.
Compliance & Security
Built for the requirements hospitals actually have.
MedScrub's self-hosted proxy is Docker-deployed — it runs inside the hospital's own infrastructure. PHI is de-identified using a three-layer pipeline covering all 18 HIPAA Safe Harbor identifiers before any data reaches an AI model.
PHI never transmitted to external AI servers
HIPAA Safe Harbor compliant — all 18 identifiers covered
Self-hosted proxy runs inside the hospital infrastructure
SMART on FHIR launch — no separate login or data export
Temporal condition chart — ICD-10 history across years
Resident-facing AI assistant with clinical guardrails