AI Clinical Decision Support
for a 200-Facility Health System
How we cut diagnostic decision time from 22 minutes to under 7 — serving 2M+ patients across 200+ hospitals, with 94% clinician adoption in the first six months.
A regional health system, consolidating across 200 facilities
Our client is a multi-state, not-for-profit health system headquartered in the Southeast US. Following a series of acquisitions, they were operating 200+ hospitals, outpatient clinics, and long-term care facilities — each running different EHR platforms, HL7 configurations, and local clinical protocols. The care network had grown faster than the infrastructure supporting it.
Clinical leadership had a clear mandate: give physicians a consistent, intelligent layer on top of the fragmented data environment. The organization had already piloted three commercial clinical decision support tools over four years. All three had failed to achieve meaningful adoption — not because of technical inadequacy, but because they required clinicians to leave their existing EHR workflow to consult a separate tool.
Diagnostic complexity across fragmented infrastructure
Clinicians were spending an average of 22 minutes per patient encounter navigating fragmented data — toggling between an EHR, a separate PACS system for imaging, a lab results portal, and a medication management tool that didn't talk to any of them. There was no intelligent layer connecting these sources or surfacing patterns across the patient's full longitudinal record.
The technical landscape was a patchwork. HL7 v2 feeds from lab systems used non-standard message segments that didn't cleanly map to any of the three EHR vendors in the network. Imaging data sat in isolated PACS systems with no standard query interface. Medication reconciliation relied on manual pharmacist review because no automated tool could reliably parse the medication history across 200+ facility configurations.
The compliance requirements were non-negotiable: full HIPAA compliance at the application layer, BAA-compliant cloud infrastructure, field-level PHI access controls, and a SOC 2 Type II audit path for the platform within 18 months of go-live. The previous three vendors had all stumbled on the HIPAA/SOC 2 intersection — particularly around model training data governance.
Three things we got wrong in the first eight weeks
HL7 normalization took 3× longer than estimated
We scoped two weeks to build the FHIR adapter. It took six. The variance in HL7 v2 message structure across 200+ facilities — non-standard OBX segments, missing MSH fields, inconsistent PID encoding — was significantly worse than the sample data we were given during scoping. We ended up writing a facility-specific configuration layer with 47 distinct normalization rules before we got clean data out.
Alert fatigue nearly killed adoption before we launched
Our first relevance threshold calibration was too aggressive. In the pre-production pilot at three hospitals, the system surfaced 340% more alerts than clinicians considered actionable. We had to rebuild the signal scoring system with direct input from the clinical advisory board, adding a multi-signal confidence weighting model and a per-specialty suppression layer. This took four additional weeks and pushed our go-live by six weeks.
SOC 2 audit scope was underestimated
We scoped SOC 2 Type II as a 6-week effort. The auditors required us to document 43 control points across the platform, including several around our model training data governance pipeline that we hadn't anticipated. We had to add a formal data lineage logging system and a model training audit trail before the audit could proceed. Lessons learned — compliance scope should always be defined with your auditor before you write a line of code.
Intelligence at the point of care
The architecture was designed as three independent layers. A data ingestion tier normalized HL7 and FHIR feeds from all facility systems, with a facility-specific configuration layer that handled the variance in message structure across vendor platforms. Every clinical event flows through Apache Kafka with exactly-once processing guarantees — including the full audit trail required for SOC 2.
The inference tier runs four specialized PyTorch models: a differential diagnosis ranker trained on 8M+ anonymized clinical encounters, a medication interaction checker, a readmission risk predictor, and a deterioration early-warning model that flags patients likely to enter a sepsis cascade within 48 hours. Models are served from a custom registry with blue-green deployment, enabling zero-downtime retraining cycles.
The front-end was the key architectural decision: instead of building a separate application, we embedded the recommendation interface directly into the existing EHR via SMART on FHIR launch contexts. The React component renders as a side panel inside the EHR's native UI — no context switch, no separate login, no workflow disruption. This is why 94% of clinicians were actively using the system within six months when none of the previous three tools had broken 30%.
Built for compliance, designed to last
ML Inference Pipeline
PyTorch ensemble models served via custom registry with blue-green deployment. Sub-200ms inference latency across diagnostic, medication interaction, and deterioration risk models.
HIPAA-Native Infrastructure
End-to-end encryption at rest and in transit. Field-level PHI access controls, immutable audit logs, and automated control-point scanning. SOC 2 Type II certified on first audit cycle.
FHIR-Native Ingestion Layer
Custom FHIR R4 adapter with facility-specific HL7 v2 normalization rules across 47 configurations. Apache Kafka event streaming with exactly-once semantics and full data lineage tracking.
SMART on FHIR Integration
EHR-embedded recommendation interface via SMART on FHIR launch contexts. Multi-signal alert scoring with per-specialty suppression to eliminate the alert fatigue that killed previous tool adoptions.
Measurable impact on patient outcomes
2M+ Patients Monitored
Real-time clinical insights delivered at the point of care across the full hospital network. Risk stratification runs continuously, not just at admission.
200+ Facilities Deployed
Rolled out across the entire network with zero-downtime deployments and region-specific compliance configurations for state-level regulatory differences.
94% Clinician Adoption
Achieved by embedding recommendations inside existing EHR workflows via SMART on FHIR — clinicians never had to leave their current tool.
HIPAA + SOC 2 Type II
End-to-end encryption, field-level access controls, immutable audit logs, and annual third-party penetration testing. Certified on the first audit cycle.
Within six months of full network deployment, diagnostic decision time dropped from 22 minutes to under 7 minutes per patient encounter. The deterioration early-warning model identified 340+ patients at elevated sepsis risk in its first quarter of operation — interventions the clinical team credited with measurably improved outcomes across the network.
The 94% clinician adoption rate — versus 0–28% for the three prior tools — was entirely attributable to the SMART on FHIR embedding strategy. When you eliminate the context switch, you eliminate the primary reason clinical decision support tools fail. This was the central architectural decision of the engagement, and it was validated by the data.
Key Technical Decisions
- Chose SMART on FHIR embedding over a standalone application — the reason adoption succeeded where three previous tools failed
- Built facility-specific HL7 normalization configs rather than forcing a standard — added 6 weeks but eliminated 90% of data quality issues
- Designed the model training pipeline with a formal data lineage system before audit requirements were confirmed — avoided a costly rework cycle
- Rejected a commercial vector DB in favor of pgvector on RDS — reduced ops complexity and met the client's cloud cost ceiling
The stack
Reference Available Upon Request
This client is referenceable. We can arrange a direct conversation with their CTO or VP of Clinical Informatics for qualified enterprise prospects under mutual NDA. Compliance documentation available separately.
Building in healthcare?
Let's talk.
We build production-grade clinical AI systems that pass compliance audits and earn clinician trust. Tell us about your challenge.