Generative AI-Powered Intelligent Document Processing for Healthcare Operations

myriad_genetics · Healthcare · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (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 document processing

Myriad Genetics needed to lower the cost and latency of healthcare document classification while automating manual key information extraction across complex medical forms for prior authorization workflows.

Constraints:

• documents included Test Request Forms, Lab Results, Clinical Notes, Insurance, and other healthcare document classes
• classification had to preserve or improve an existing 94% accuracy baseline
• the prior system cost about 3 cents per page and $15,000 monthly per business unit
• classification latency was 8.5 minutes per document
• key information extraction required clinical nuance, including negation-sensitive distinctions such as metastatic versus not metastatic
• Women’s Health customer service spent up to 10 full-time employees and 78 hours daily on manual extraction
• the architecture had to fit existing event-driven systems and route outputs into downstream prior authorization workflows

Architecture

Myriad and the AWS GenAI Innovation Center replaced the costly Textract-plus-Comprehend classifier with an AWS GenAI IDP Accelerator implementation using Pattern 2. Incoming order events retrieve documents, concurrency controls coordinate Step Functions jobs and SQS overflow, Textract extracts text/layout/tables/forms, Nova Pro classifies documents with optimized prompts, and Nova Premier performs key information extraction with multimodal document images, few-shot examples, and chain-of-thought style reasoning for ambiguous form marks. Structured results are delivered to Myriad's Authorization System, while the accelerator's evaluation framework, model benchmarks, and SME-refined prompts drive accuracy, latency, and cost optimization.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  order_event_source["Order event source"] --> document_retrieval_and_cache["Document retrieval and cache"]
  source_document_systems["Source document systems"] --> document_retrieval_and_cache
  document_retrieval_and_cache --> concurrency_manager["Concurrency manager"]
  concurrency_manager --> sqs_overflow_queue["SQS overflow queue"]
  concurrency_manager --> step_functions_document_workflow["Step Functions document workflow"]
  sqs_overflow_queue --> step_functions_document_workflow
  step_functions_document_workflow --> amazon_textract_ocr["Amazon Textract OCR"]
  amazon_textract_ocr --> document_classification_service["Document classification service"]
  amazon_textract_ocr --> key_information_extraction_service["Key information extraction service"]
  genai_idp_configuration["GenAI IDP configuration"] --> document_classification_service
  genai_idp_configuration --> key_information_extraction_service
  document_classification_service --> amazon_bedrock_nova_pro["Amazon Bedrock Nova Pro"]
  amazon_bedrock_nova_pro --> document_classification_service
  document_classification_service --> vendor_routing["Vendor routing"]
  vendor_routing --> authorization_system["Authorization System"]
  key_information_extraction_service --> amazon_bedrock_nova_premier["Amazon Bedrock Nova Premier"]
  amazon_bedrock_nova_premier --> key_information_extraction_service
  key_information_extraction_service --> authorization_system
  evaluation_and_sme_loop["Evaluation and SME loop"] --> genai_idp_configuration

Components

Data flow

Incoming order events pull documents from source management systems, reuse cached documents where possible, and enter a concurrency-managed Step Functions pipeline. Textract normalizes visual and textual document evidence, Nova Pro classifies the document from configured prompts, Nova Premier extracts medical fields using text plus visual examples for difficult forms, and structured output is sent through REST operations to Myriad's Authorization System. Evaluation data and SME review feed prompt, model, and configuration updates.

Why this architecture

Pattern 2 gave Myriad the low-cost, configurable path needed for complex healthcare documents: Textract preserved document structure, Bedrock models handled semantic classification and extraction, Nova Pro optimized classification cost and latency, Nova Premier supplied stronger reasoning for ambiguous visual extraction, and the accelerator's built-in evaluation and configuration loops let the team tune prompts and model choices without rebuilding the whole system.

Models used: Amazon Nova Pro · Amazon Nova Premier · Anthropic Claude Sonnet 3.7 · Amazon Bedrock

Scale

• Myriad processes thousands of healthcare documents daily across Women’s Health, Oncology, and Mental Health divisions.
  “Myriad’s Revenue Engineering Department processes thousands of healthcare documents daily across Women’s Health, Oncology, and Mental Health divisions.”article, stated
• The prior classification system cost $15,000 monthly per business unit.
  “Operational costs: 3 cents per page resulting in $15,000 monthly expenses per business unit”article, stated
• The previous classifier took 8.5 minutes per document.
  “Classification latency: 8.5 minutes per document, delaying downstream prior authorization workflows”article, stated
• Manual Women’s Health extraction required up to 10 full-time employees contributing 78 hours daily.
  “Women’s Health customer service requiring up to 10 full-time employees contributing 78 hours daily in the Women’s Health business unit alone.”article, stated
• Model benchmarking used 1,200 healthcare documents across three document classes.
  “We tested four foundation models—Amazon Nova Lite, Amazon Nova Pro, Amazon Nova Premier, and Anthropic Claude Sonnet 3.7—using 1,200 healthcare documents across three document classes”article, stated
• KIE testing used 32 document samples and reached 90% accuracy at about 1.3 minutes per document.
  “Testing across 32 document samples with varying complexity levels via the GenAI IDP Accelerator revealed that Amazon Textract with Layout, TABLES, and FORMS features enabled”article, stated

  “The solution achieved 90% accuracy (same as human evaluator baseline accuracy) while processing documents in approximately 1.3 minutes each.”article, stated
• The planned Women’s Health rollout covers 9,000 prior authorizations and saves 300 hours monthly.
  “This improvement saves 300 hours monthly across 9,000 prior authorizations in Women’s Health alone, equivalent to 50 hours per prior authorization specialist.”article, stated

Alternatives considered (roads not taken)

Keep the Textract plus Amazon Comprehend classifier

Rejected because: It already reached 94% accuracy, but its cost and 8.5-minute latency created operational bottlenecks.

Use the fully managed Bedrock Data Automation pattern

Rejected because: Pattern 1 was recommended for most use cases, but Myriad needed Pattern 2's lower cost and flexibility for custom prompt and model optimization.

Use a specialized SageMaker classification model in Pattern 3

Rejected because: Pattern 3 is positioned for specialized classification, while Myriad's best fit was Pattern 2's Textract plus Bedrock configurability.

Classify documents from text content alone

Rejected because: Similar healthcare documents shared content, so the team added format-based cues and negative prompts to capture layout and class-specific exclusions.

Run key information extraction with text-only OCR and prompts

Rejected because: Checkboxes and ambiguous visual marks required document images, few-shot visual examples, and reasoning over mark positions.

Patterns

tool-use-orchestration (document workflow)

The production pipeline orchestrates document retrieval, concurrency management, queues, Step Functions workflows, Textract OCR, Bedrock model calls, vendor routing, and REST delivery into an authorization system.

structured-output-contract (document extraction)

The system defines document classes, extraction fields, key attributes, and prompts in configuration, then emits structured results for prior authorization workflows.

model-routing (cost and reasoning fit)

Myriad used different Bedrock models for different document-processing subtasks: Nova Pro for low-latency classification, Nova Premier for complex key information extraction, and Claude Sonnet 3.7 reasoning to identify class differentiators.

eval-driven-development (benchmarking and iteration)

The team used the accelerator's evaluation framework, 1,200-document model benchmarks, 32-sample KIE testing, human baselines, and accuracy/cost/latency metrics to choose models and tune prompts.

data-quality-loop (prompt and class refinement)

Class definitions, negative prompts, visual cues, and few-shot examples were refined from sample documents, model reasoning output, and subject matter expert review.

cost-aware-routing (model and prompt cost control)

The architecture routes classification to the cheaper Nova Pro model, reserves Nova Premier for complex extraction, uses Pattern 2 for low-cost flexibility, and applies prompt caching for lengthy few-shot examples.

guardrails-io-filtering (classification constraints)

Negative prompts acted as explicit exclusion rules for confusing document classes, preventing the classifier from treating patient-history forms as lab measurements.

Tradeoffs

Choose the configurable accelerator pattern.

Chose Pattern 2 with Textract and Bedrock models over Pattern 1 Bedrock Data Automation or Pattern 3 SageMaker classification.

Rationale: Pattern 2 balanced low cost, custom prompt/model optimization, and complex-document flexibility without requiring the specialized classification path of Pattern 3.

Split models by document-processing task.

Chose Nova Pro for classification and Nova Premier for extraction over Using one model for both tasks.

Rationale: Classification prioritized low cost and latency, while extraction needed stronger reasoning for ambiguous visual and clinical evidence.

Add visual context and few-shot examples to KIE.

Chose Multimodal document images plus extracted text and examples over Textract text and prompt improvements alone.

Rationale: The extra context was necessary for form layouts, checkbox marks, and ambiguous visual evidence that text-only processing could not handle.

Use explicit negative prompts for classification.

Chose Exclusionary class definitions over Positive-only class descriptions.

Rationale: Negative prompts reduced confusion where forms and results shared patient-history language but represented different document classes.

Roll out classification first.

Chose Phased rollout beginning with document classification in Women’s Health over Immediate broad deployment of classification and KIE across all business units.

Rationale: The article reports a staged business rollout, letting Myriad begin with the validated classification savings before expanding across Oncology and Mental Health.

Gotchas (what bit them)

A high-accuracy classifier was still too slow and expensive for operational workflows.

Root cause: The legacy Textract-plus-Comprehend system reached 94% accuracy but cost 3 cents per page and took 8.5 minutes per document, delaying prior authorization.

Resolution: Benchmark and migrate classification to Nova Pro with optimized prompts, lowering cost and latency while improving accuracy.

Manual key information extraction did not scale across nuanced healthcare forms.

Root cause: Medical forms required contextual interpretation of clinical distinctions, variable formats, and patient or insurance fields that were hard to locate consistently.

Resolution: Automate KIE with Textract, multimodal LLM input, Nova Premier reasoning, and few-shot visual examples.

Similar healthcare documents caused semantic classification ambiguity.

Root cause: Document types shared content and formatting patterns, so simple content matching confused forms, reports, and results.

Resolution: Use model-assisted prompt engineering, document-format cues, and negative prompts that define what each class is not.

Checkbox extraction failed on visual ambiguity that text prompts alone could not resolve.

Root cause: Medical forms contained checkmarks, X marks, handwritten marks, overlapping marks, and content spanning multiple fields.

Resolution: Enable Textract TABLES and FORMS, add visual cue instructions, send images plus text to the model, and use reasoning prompts for ambiguous tick marks.

Few-shot visual examples improved extraction but made prompts long enough to affect cost and latency.

Root cause: Multiple document images and expected outputs had to be included so the model could learn form layouts and marking styles.

Resolution: Use Amazon Bedrock prompt caching to reuse the lengthy examples for similar documents processed within a five-minute window.

Evaluation

Myriad evaluated the system with the GenAI IDP Accelerator's expected-output UI, model benchmarking over 1,200 healthcare documents, classification accuracy/latency/cost comparisons, KIE testing over 32 samples, and comparison against a human evaluator baseline.

Metrics: existing classifier accuracy: 94% · Nova Pro classifier accuracy: 98% · classification cost reduction: 77%, from 3.1 to 0.7 cents per page · classification latency reduction: 80%, from 8.5 to 1.5 minutes per document · model benchmark: 1,200 healthcare documents across three classes · KIE test set: 32 document samples · KIE accuracy: 90%, matching the human evaluator baseline · KIE processing time: approximately 1.3 minutes per document · KIE cost: 9 cents per page

Outcomes

• Classification accuracy improved from 94% to 98%.
• Classification cost dropped by 77%, from 3.1 to 0.7 cents per page.
• Classification time dropped by 80%, from 8.5 to 1.5 minutes per document.
• Automated KIE reached 90% accuracy, matching the human evaluator baseline.
• KIE processed documents in about 1.3 minutes each at 9 cents per page.
• The planned Women’s Health rollout is projected to save up to $132K annually in classification costs.
• Prior authorization submission time is expected to drop from six minutes to four minutes, saving 300 hours monthly across 9,000 prior authorizations.

The solution targets thousands of daily healthcare documents across Myriad's Women’s Health, Oncology, and Mental Health divisions, with a first rollout in Women’s Health prior authorization workflows.