Chapter 1: Document Intelligence

Healthcare runs on paper. Faxes remain the dominant interoperability mechanism for prior authorizations, referrals, and records requests. Every integration eventually hits a document extraction problem, whether it's scanned claims forms, handwritten physician notes, or multi-page attachments. This chapter covers extracting structured, actionable data from the unstructured documents that flow through payer operations daily. PHI is present in virtually every document, confidence scoring is non-negotiable for human-in-the-loop workflows, and multi-page mixed-format documents (tables, checkboxes, free text, handwriting) are the norm rather than the exception.

AWS Services Featured: Amazon Textract, Amazon Comprehend Medical, AWS Lambda, Amazon S3, Amazon DynamoDB, Amazon A2I (Augmented AI)

What You'll Learn

By the end of this chapter, you'll know how to:

  • Extract structured fields from images — turning a phone photo of an insurance card into a JSON payload your eligibility system can consume (Recipe 1.1)
  • Digitize multi-section paper forms — handling tables, checkboxes, and mixed layouts from patient intake forms (Recipe 1.2)
  • Combine OCR with medical NLP — layering Amazon Comprehend Medical on top of Textract output to pull ICD-10 codes and clinical entities from lab requisitions (Recipe 1.3)
  • Process multi-page clinical documents — tackling the prior authorization problem: extracting structured data from 5-20 page faxed submissions where every page is different (Recipe 1.4)
  • Build document classification and routing pipelines — classifying incoming attachments by type, routing each to the right extractor, and merging results into a unified claims record (Recipe 1.5)
  • Handle handwritten text with confidence scoring and human review — the hardest extraction problem in healthcare, solved with a tiered confidence pipeline and Amazon A2I (Recipe 1.6)
  • Extract medication data from prescription labels — mapping pharmacy label fields to RxNorm concept IDs for downstream medication reconciliation (Recipe 1.7)
  • Parse Explanation of Benefits documents — normalizing payer-specific table layouts into a canonical financial schema with math-based validation (Recipe 1.8)
  • Process medical records request forms — extracting routing identifiers and validating HIPAA authorization elements before records are released (Recipe 1.9)
  • Migrate historical paper charts at scale — bulk OCR, document segmentation, FHIR R4 mapping, and HealthLake import for legacy chart digitization programs (Recipe 1.10)

Chapter Prerequisites

Before diving into these recipes, make sure you have:

AWS Account & Services:

  • An AWS account with a signed AWS Business Associate Addendum (BAA) — non-negotiable for any workload touching PHI
  • Amazon Textract, Amazon Comprehend Medical, Amazon S3, AWS Lambda, and Amazon DynamoDB enabled in your target region
  • For Recipe 1.6: Amazon A2I (Augmented AI) and an A2I workforce configured

IAM & Security:

  • An IAM role for Lambda with least-privilege access to Textract, Comprehend Medical, S3, and DynamoDB
  • S3 bucket policies enforcing aws:SecureTransport (TLS-only) and default SSE-KMS encryption
  • VPC endpoints for Textract and S3 if operating in a private subnet (recommended for production)
  • CloudTrail enabled for audit logging of all API calls touching PHI

Sample Data:

  • Each recipe includes notes on synthetic test data. For real-world testing, use de-identified documents or CMS sample forms — never use actual PHI in development environments
  • The CMS Forms Library provides real-world form layouts for testing

Development Environment:

  • Python 3.9+ with boto3
  • AWS CLI configured with appropriate credentials
  • Cost estimates in each recipe assume us-east-1 pricing as of early 2026

Chapter Architecture

How the 10 recipes relate to each other and where PHI flows between them:

                            Chapter 1: Document Intelligence
                            ================================

    SIMPLE                     MODERATE                        COMPLEX
    (Textract core)            (+ LLM reasoning)               (multi-stage pipelines)

    ┌─────────────┐
    │  1.1 Ins.   │─ confidence gating ─┐
    │  Card Scan  │                     │
    └─────────────┘                     │
    ┌─────────────┐                     │    ┌──────────────────┐
    │  1.2 Patient│─ confidence gating ─┼───>│  1.6 Handwritten │
    │  Intake     │  (async Textract)   │    │  Notes (A2I      │
    └─────────────┘                     │    │  human review)   │
    ┌─────────────┐                     │    └──────────────────┘
    │  1.3 Lab    │─ Comprehend Medical ┘              │
    │  Requisition│  (ICD-10 codes)          ┌─────────┘
    └─────────────┘                          │  (shared patterns)
           │                                 v
           │ (code validation        ┌──────────────────┐
           │  pattern reused)        │  1.4 Prior Auth  │──> Ch.2 Clinical
           │                         │  (LLM + tiering) │    Criteria
           └────────────────────────>└──────────────────┘
                                             │
    ┌─────────────┐                          │ (fan-out pattern)
    │  1.7 Rx     │                          v
    │  Label OCR  │              ┌──────────────────────┐
    │  (RxNorm)   │              │  1.5 Claims          │──> Ch.3 Claims
    └─────────────┘              │  Attachment (boundary │    Adjudication
           │                     │  detect + matching)   │
           │                     └──────────────────────┘
           │
           │ (medication         ┌──────────────────┐
           │  reconciliation)    │  1.8 EOB         │──> Ch.4 Payment
           └────────────────────>│  Processing      │    Integrity
                                 │  (payer profiles)│
                                 └──────────────────┘
                                 ┌──────────────────┐
                                 │  1.9 Med Records │──> Ch.5 Member
                                 │  Request (HIPAA  │    Access
                                 │  auth validation)│
                                 └──────────────────┘
                                 ┌──────────────────────┐
                                 │  1.10 Chart Migration │
                                 │  (batch inference,    │──> HealthLake
                                 │  FHIR R4, capstone)   │
                                 └──────────────────────┘

    Shared infrastructure across all recipes:
    S3 (SSE-KMS) + DynamoDB + CloudTrail + VPC endpoints + Lambda

    Technology progression:
    Textract only ──> + Comprehend Medical ──> + Bedrock LLM ──> + Vision models
       (1.1-1.2)         (1.3, 1.7)              (1.4-1.5)        (1.6, 1.10)

Each recipe's "Related Recipes" section identifies specific dependencies. Recipes 1.1 and 1.2 establish foundational patterns (confidence gating, async processing, HIPAA infrastructure) that every subsequent recipe builds on.

Recipes

# Recipe Complexity Phase
1.1 Insurance Card Scanning Simple ⭐ MVP
1.2 Patient Intake Form Digitization Simple ⭐ MVP
1.3 Lab Requisition Form Extraction Moderate 🔶 Phase 2
1.4 Prior Authorization Document Processing Moderate ⭐ MVP
1.5 Claims Attachment Processing Complex 🔶 Phase 2
1.6 Handwritten Clinical Note Digitization Complex 🔷 Phase 3
1.7 Prescription Label OCR Simple 🔶 Phase 2
1.8 EOB Processing Moderate 🔶 Phase 2
1.9 Medical Records Request Extraction Moderate 🔶 Phase 2
1.10 Historical Chart Migration Complex 🔷 Phase 3

Reading order: Recipes build on each other. Start with 1.1 — each successive recipe introduces new concepts while referencing patterns established earlier. If you're only here for one thing, Recipe 1.4 (Prior Auth) is the most common real-world ask and can be read after 1.1-1.2 for context.

Next: Recipe 1.1 — Insurance Card Scanning →