AI-powered patient intake automation
Meridian Health Network needed to eliminate a manual patient intake process that was slow, error-prone, and consuming hundreds of staff hours every month. We built an AI document processing pipeline that extracts, validates, and routes intake data automatically.
73%
faster processing
12 hrs
saved per day
99.1%
data accuracy
The Challenge
What Meridian Health Network was facing
Intake staff spent an average of 45 minutes per patient case manually entering data from insurance cards, referral letters, and demographic forms into the EMR system.
Error rates hovered around 8%, causing rejected claims, delayed authorizations, and rework across billing and clinical teams.
The bottleneck was worst during Monday morning surges and post-holiday periods, when backlogs of 200+ cases were common and patients waited days for appointments to be confirmed.
Previous attempts to solve this with basic OCR tools failed because they could not handle the variety of document formats, handwriting, and multi-page fax submissions.
Our Approach
How we solved it
We designed a multi-stage pipeline: document classification (insurance card, referral, demographics), data extraction using fine-tuned vision models, cross-validation against payer databases, and automated EMR population via FHIR R4 APIs.
For ambiguous or low-confidence extractions, the system routes cases to a human review queue with the extracted data pre-filled, so reviewers correct rather than re-enter from scratch.
Every action is logged with an immutable audit trail for HIPAA compliance, and the system encrypts data at rest and in transit with role-based access controls.
We deployed incrementally: one department first, validated accuracy against manual entry for two weeks, then expanded network-wide over the following month.
Key Features
What we delivered
Intelligent Document Classification
Automatically identifies insurance cards, referral letters, lab results, and demographic forms from scanned documents, faxes, and uploaded images.
AI Data Extraction
Fine-tuned vision and NLP models extract structured data from unstructured documents with 99.1% accuracy, handling handwriting and varied layouts.
FHIR-Based EMR Integration
Validated data flows directly into the EMR via FHIR R4 APIs, eliminating manual entry and ensuring interoperability with Epic and downstream systems.
Human-in-the-Loop Review
Low-confidence extractions are routed to a review queue with pre-filled data, reducing reviewer effort by 80% compared to manual entry.
HIPAA-Compliant Audit Trail
Every document, extraction, and routing decision is logged with timestamps, user IDs, and version history for compliance audits.
Real-Time Processing Dashboard
Operations managers monitor intake volume, processing times, accuracy rates, and queue depth through a live dashboard.
Our Process
How we got there
Discovery and EMR Audit
We spent two weeks on-site mapping the existing intake workflow, documenting EMR integration points, and quantifying error rates and processing bottlenecks.
Model Training and Validation
We collected 3,000 anonymized intake documents, trained classification and extraction models, and validated accuracy against manually entered ground truth data.
Pipeline Development
We built the end-to-end processing pipeline with document ingestion, AI extraction, cross-validation, FHIR integration, and the human review interface.
Pilot Deployment
Deployed to one department for a two-week parallel run, processing cases through both the AI pipeline and manual entry to compare accuracy and speed.
Network-Wide Rollout
After pilot validation confirmed 99.1% accuracy, we rolled the system out across all departments over four weeks with on-site training and support.
Results
Measurable outcomes
73% Faster Processing
Average intake processing time dropped from 45 minutes to 12 minutes per case, with most of that time spent on human review of flagged cases.
12 Hours Saved Per Day
Across the network, the equivalent of 12 staff-hours per day was freed from data entry and redirected to patient-facing work.
99.1% Data Accuracy
Extraction accuracy exceeded the 96% manual entry baseline, reducing rejected claims and downstream billing corrections.
Technology Stack
What we used
AI / ML
Backend
Healthcare Integration
Infrastructure
Compliance
Business Impact
The bigger picture
Within three months of full deployment, Meridian reduced intake-related claim rejections by 41% and shortened the average time from patient referral to first appointment by two days. The intake team was reassigned from data entry to patient communication and care coordination, improving both staff satisfaction and patient experience scores.
“CareOnix delivered a working system in six weeks that our team adopted immediately. The accuracy exceeded what we expected.”
Have a similar challenge?
Tell us what you're working on. We'll come back with a clear plan and honest technical guidance.