AI-Powered Fax Processing Automation for Healthcare Referrals

providence · Healthcare · 2025 · source Agentic: no · promotional bias: high · summary-only (no fetched source) · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Healthcare referral fax processing

Providence needed to automate high-volume referral fax processing that created backlogs, delayed care, and consumed clinical staff time.

Constraints:

The organization processes more than 40 million faxes and 160 million pages annually
Healthcare referral formats vary across clinics, individuals, and document types
Extracted data must integrate securely and correctly with Epic EHR workflows

Architecture

Providence's summary-only extraction describes a Databricks/MLflow fax-referral pipeline that combines OCR, GPT-4.0 extraction, experiment tracking, parameterized OCR/prompt sweeps, historical simulation, validation, CI/CD-generated YAML jobs, and Epic EHR integration. The system targets referral triage from highly variable fax documents across a large health system, with MLflow used as the lifecycle and observability backbone.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  fax-ingestion-stream["fax-ingestion-stream"]
  ocr-extraction-layer["ocr-extraction-layer"]
  gpt4-referral-extractor["gpt4-referral-extractor"]
  mlflow-experiment-tracker["mlflow-experiment-tracker"]
  parameterized-job-runner["parameterized-job-runner"]
  yaml-cicd-config["yaml-cicd-config"]
  validation-and-simulation-layer["validation-and-simulation-layer"]
  epic-ehr-handoff["epic-ehr-handoff"]
  fax-ingestion-stream --> ocr-extraction-layer
  ocr-extraction-layer --> gpt4-referral-extractor
  ocr-extraction-layer --> mlflow-experiment-tracker
  gpt4-referral-extractor --> validation-and-simulation-layer
  gpt4-referral-extractor --> mlflow-experiment-tracker
  mlflow-experiment-tracker --> parameterized-job-runner
  parameterized-job-runner --> yaml-cicd-config
  validation-and-simulation-layer --> epic-ehr-handoff

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`fax-ingestion-stream`	Referral fax input stream covering diverse document formats and pages.	`ocr-extraction-layer`
`ocr-extraction-layer`	OCR model layer, including Azure AI Document Intelligence, for extracting text from fax images.	`gpt4-referral-extractor`, `mlflow-experiment-tracker`
`gpt4-referral-extractor`	GPT-4.0 extraction workflow for referral fields from OCR text.	`validation-and-simulation-layer`, `mlflow-experiment-tracker`
`mlflow-experiment-tracker`	Experiment-management and lifecycle layer for model, prompt, OCR, and hyperparameter comparisons.	`parameterized-job-runner`
`parameterized-job-runner`	Reusable runtime job framework for sweeping OCR models, prompt templates, and configuration inputs.	`yaml-cicd-config`
`yaml-cicd-config`	CI/CD-generated configuration for large-scale experiment jobs.	—
`validation-and-simulation-layer`	Historical fax simulation and validation layer used to reduce production risk before Epic handoff.	`epic-ehr-handoff`
`epic-ehr-handoff`	Integration path into Epic EHR referral workflows after validation.	—

Data flow

Fax documents enter the Databricks pipeline, where OCR and GPT-4.0 extraction models convert unstructured referral pages into structured fields. MLflow tracks experiments across OCR models, prompts, and hyperparameters; parameterized jobs and YAML configuration support large sweeps; historical fax data simulates downstream behavior; validated results are prepared for Epic EHR handoff and referral triage.

3 supporting quotes

“The system combines OCR with GPT-4.0 models to extract referral data from diverse document formats and integrates seamlessly with Epic EHR systems”summary, stated

“Providence uses parameterized jobs to systematically sweep across combinations of OCR models, prompt templates, and other hyperparameters.”summary, stated

“The parameterized job framework enables dynamic input configuration at runtime”summary, stated

Why this architecture

Providence needed systematic experimentation and simulation because referral faxes are variable, healthcare compliance is strict, and Epic integration is operationally sensitive. MLflow gives a shared record of OCR/prompt/model behavior, parameter sweeps avoid manual tuning, and historical simulation reduces the risk of dropping extraction models directly into live referral workflows.

3 supporting quotes

“the automated processing of massive volumes of referral faxes that previously required manual intervention and created significant care delays.”summary, stated

“This provides the team with clear visibility into model performance across different document types and referral scenarios”summary, stated

“Providence's use of historical data for simulation represents another sophisticated LLMOps practice.”summary, stated

Models used: GPT-4.0 · OCR models · Azure AI Document Intelligence

Scale

Providence processes more than 40 million faxes annually.
“The organization processes more than 40 million faxes annually”summary, stated
Those faxes total more than 160 million pages.
“totaling over 160 million pages”summary, stated
Providence operates 51 hospitals and more than 1,000 outpatient clinics.
“serving vulnerable communities through 51 hospitals, over 1,000 outpatient clinics”summary, stated
The health system has more than 130,000 caregivers across seven states.
“more than 130,000 caregivers across seven states.”summary, stated
The old manual workflow created multi-month backlogs.
“This manual process created multi-month backlogs, delayed patient care”summary, stated

Alternatives considered (roads not taken)

Manual referral-fax transcription into Epic

Rejected because: Manual review created multi-month backlogs and delayed patient care.

“a significant portion requiring manual review and transcription into their Epic electronic health record (EHR) system.”summary, stated

Manual fragmented prompt/OCR optimization

Rejected because: MLflow and parameterized jobs systematically compare OCR models, prompts, and hyperparameters.

“the real engineering value comes from the MLflow-orchestrated pipeline that automates what would otherwise be manual and fragmented development processes.”summary, stated

Direct live deployment after ad hoc tests

Rejected because: Historical fax simulation is used to reduce production risk before model outputs affect referrals.

“By leveraging existing fax data to simulate downstream outcomes, they can refine their models before production deployment, significantly reducing risk and accelerating the deployment cycle.”summary, stated

Patterns

`eval-driven-development` (document-extraction-experimentation)

Providence uses MLflow and parameterized jobs to compare OCR models, prompts, and hyperparameters before production deployment.

“Providence uses parameterized jobs to systematically sweep across combinations of OCR models, prompt templates, and other hyperparameters.”summary, stated

`observability-tracing` (experiment-tracking)

MLflow provides centralized experiment tracking and logging to compare model behavior across referral documents and scenarios.

“Central to their LLMOps strategy is comprehensive experiment tracking and logging through MLflow.”summary, stated

`structured-output-contract` (ehr-handoff)

Extracted referral data is formatted, validated, normalized, and securely delivered into existing Epic infrastructure.

“All extracted referral data must be seamlessly formatted, validated, and securely delivered to the existing healthcare infrastructure.”summary, stated

`rag-retrieval` (planned-healthcare-knowledge-support)

The broader architecture includes Azure Search for RAG workflows and future Mosaic AI exploration to improve accuracy and responsiveness.

“Providence's broader technical infrastructure includes Azure Kubernetes Service (AKS) for containerized deployment, Azure Search to support retrieval-augmented generation (RAG) workflows, and Postgres for structured storage.”summary, stated

`batch-offline-processing` (historical-simulation)

Historical fax data is used to simulate downstream outcomes and refine extraction models before live deployment.

Tradeoffs

How to reduce production risk in healthcare document automation.

Chose simulate downstream outcomes using historical fax data over deploying extraction models directly into live referral workflows after ad hoc tests.

Rationale: Simulation allows refinement before production and reduces patient-care risk.

“This is particularly important in healthcare settings where errors can have significant patient care implications and where integration with established systems like Epic requires rigorous validation.”summary, implied

How to manage prompt and OCR optimization.

Chose MLflow-orchestrated parameter sweeps over manual and fragmented development processes.

Rationale: A unified experimentation framework makes combinations reusable, comparable, and CI/CD-friendly.

“the real engineering value comes from the MLflow-orchestrated pipeline that automates what would otherwise be manual and fragmented development processes.”summary, stated

Gotchas (what bit them)

Healthcare fax inputs are highly variable.

Root cause: The documents range from handwritten notes to typed PDFs and differ across clinics, roles, and workflows.

Resolution: Providence uses systematic experimentation across OCR, prompts, and document scenarios rather than one-off prompt tuning.

“The diversity of input documents - from handwritten notes to typed PDFs - creates a wide range of processing challenges that require sophisticated prompt engineering and model tuning.”summary, stated

Universal automation pipelines are hard in healthcare operations.

Root cause: Workflow differences between clinics, roles, and individuals make representative testing difficult.

Resolution: The experimentation framework supports rapid iteration and validation against diverse document types and workflow patterns.

“The lack of standardization between clinics, roles, and individuals creates significant challenges for defining universal automation pipelines or creating test scenarios that reflect real-world complexity.”summary, stated

Epic integration adds regulated enterprise complexity.

Root cause: Referral data must be normalized and delivered into established EHR systems with compliance constraints.

Resolution: Databricks pre-processes and normalizes information before handoff to Epic.

“Databricks plays a critical role in pre-processing and normalizing this information before handoff to the EHR system, requiring careful attention to data quality and format consistency.”summary, stated

Evaluation

Providence evaluates fax automation through MLflow experiment tracking, historical simulations, document-type comparisons, and downstream Epic workflow validation.

Metrics: backlog reduction · document extraction accuracy · OCR/prompt configuration performance · deployment cycle speed · manual review reduction

“This provides the team with clear visibility into model performance across different document types and referral scenarios, enabling efficient comparison of results without duplicating experimental effort.”summary, stated

“The elimination of two to three-month backlogs in some regions directly impacts patient care timelines”summary, stated

Outcomes

Providence automated referral fax processing for a workflow that handles over 40 million annual faxes.
The system eliminated months-long backlogs in some regions.
Automation frees clinical staff to focus on patient care across 1,000+ clinics.

This is a summary-only extraction because the source fetch failed; claims are grounded in meta.full_summary.

“Providence Health System automated the processing of over 40 million annual faxes using GenAI and MLflow on Databricks”summary, stated

“eliminating months-long backlogs and freeing clinical staff to focus on patient care across 1,000+ clinics.”summary, stated