Turn healthcare documents into structured data healthcare teams can trust. Nutrient Data Extraction API extracts and validates information from forms, records, and scanned documents, giving teams traceable, production-ready data while directing human attention to exceptions — not every field. Encrypted transport and access controls help support the security requirements of HIPAA-sensitive workflows.
Trusted by enterprises, governments, and teams building document workflows at scale
USE CASES
Extract patient demographics, insurance details, consent fields, checkboxes, signatures, and supporting information from intake documents.
Structure provider, payer, procedure, diagnosis, date, authorization, and supporting information for review and routing. See prior authorization automation for the full workflow.
Automate extraction of payer details, dates, amounts, codes, patient responsibility, payment details, and structured tables from EOBs and claims documents. See a coding agent build an underpaid-claims checker.
Extract test names, values, reference ranges, dates, prescription details, provider information, and structured fields from lab and prescription documents.
Prepare mixed scanned and digital medical records for review, search, AI, and downstream workflow use. See release of information automation for records-request workflows.
WALKTHROUGH
GOVERNED EXTRACTION
LLMs can reason over documents, but medical workflows need deterministic, auditable output grounded in the source file — not generated answers that vary between runs.
Typed output remains tied to the source document — not generated answers that change between runs.
Flag uncertain values before they move downstream into patient records or administrative systems.
Every extracted value is anchored to its source location for traceability and human review.
Lab tables, intake forms, checkboxes, and key-value regions preserved — not flattened into plain text.
Support validation steps before structured data enters patient records, claims systems, or administrative workflows.
Source context and page detail support audit trails across charts, claims, and lab results.
WHAT YOU CAN EXTRACT
Names, DOBs, NPI numbers, addresses, and identifiers from intake and referral documents.
ICD-10, CPT, and related codes extracted with confidence scores for validation before downstream use.
Payer names, policy numbers, group IDs, and coverage details from insurance and claims documents.
Test names, result values, units, and reference ranges preserved from lab reports.
Structured key-value regions, handwritten fields, and checkbox states captured with ICR — with confidence scores for review.
Tables from EOBs, prior authorization forms, and claims documents, preserved with row and column context.
HOW IT WORKS
Text
For fast extraction from born-digital records and forms with a reliable text layer.
Structure
For OCR-backed extraction from scanned records with simple layouts.
Agentic
For documents that require deeper visual interpretation, review, or recovery.
OUTPUT FORMATS
Confidence scores, coordinates, and page context included.
Choose output: "json" or Markdown per request.
Spatial JSON
For extraction · validation · review
{ "type": "keyValueRegion", "id": "kv_d4e5f6", "bounds": { "x": 82, "y": 128, "width": 260, "height": 92 }, "confidence": 0.98, "page": { "pageIndex": 0, "pageNumber": 1 }, "pairs": [ { "key": { "entityType": "QUESTION", "value": "Patient name", "bounds": { "x": 82, "y": 128, "width": 100, "height": 24 }, "confidence": 0.99 }, "value": { "entityType": "ANSWER", "value": "Sarah Chen", "bounds": { "x": 190, "y": 128, "width": 100, "height": 24 }, "confidence": 0.99 }, "relationshipConfidence": 0.98 }, { "key": { "entityType": "QUESTION", "value": "Diagnosis code", "bounds": { "x": 82, "y": 164, "width": 100, "height": 24 }, "confidence": 0.97 }, "value": { "entityType": "ANSWER", "value": "J45.40", "bounds": { "x": 190, "y": 164, "width": 60, "height": 24 }, "confidence": 0.97 }, "relationshipConfidence": 0.95 } ]}Markdown
For RAG · search · knowledge bases
# Patient Intake Form
**Patient** Sarah Chen**DOB** 1985-03-12**Insurance ID** UHC-8821047
## Diagnosis
| Code | Description || --- | --- || J45.40 | Moderate persistent asthma |
**Provider** Dr. Marcus Webb**Date** 2024-11-08Reviewed extraction example
This is one reviewed run on a selected public form populated with privacy-safe demo values. It isn’t an accuracy, production, or performance benchmark.
A selected public Centers for Medicare and Medicaid Services (CMS) prior authorization form populated with privacy-safe demo values and no protected health information (PHI).
Grounding makes mistakes reviewable. It doesn’t make them correct.
The videos use an extract → source evidence → decision gate schematic. It’s a workflow concept, not a shipped product UI.
“Reviewed” means the evidence audit was recorded. It doesn’t mean a reviewer approved the result.
0.93 accuracy (vision engine), independently benchmarked
200 real-world documents. Three metrics. Results published with every release.
SOC 2 Type 2 audited
Backed by Nutrient’s SOC 2 Type 2 security practices and built for use in business-critical and compliance-sensitive workflows.
Trust and compliance
TLS encryption by default
All API communication is TLS-encrypted. Documents in transit are protected end to end.
HIPAA-compatible architecture
No document retention on paid plans, encrypted transport, and access controls designed to support HIPAA-sensitive workflows.
Yes. For hand-completed fields, checkboxes, and handwritten values, the API applies intelligent character recognition (ICR) to capture what standard OCR misses. Every extracted value includes a confidence score so uncertain fields can be flagged for human review before moving downstream.
Data Extraction API processes PDFs; images (including scans and photos); and Word, Excel, and PowerPoint files. It handles scanned PDFs, fillable forms, and mixed digital/image-based documents without requiring a separate OCR pipeline.
Every extracted element includes confidence scores, page references, and coordinates so you can compare outputs, flag low-confidence fields for human review, and trace values back to the source document before they move into patient records, claims systems, or administrative workflows.
Nutrient Data Extraction API is an extraction layer — not an EHR, claims platform, or workflow management system. It extracts structured data from medical documents so that data can be reviewed, validated, and routed into the systems your team already uses. See the Data Extraction API overview for the full capability set, or the comparison hub for how it stacks up against other extraction platforms.
Yes, with the Data Extraction SDK. The Data Extraction API runs in Nutrient’s cloud, but the SDK runs the same engine in Python or Java on servers you control. Text and layout recognition run on those servers, and if you also host the optional AI model, patient records never leave your infrastructure. Talk to Sales about a self-hosted setup.
To try it on a set of sample records first, use Nutrient Extract Studio, a desktop app for Apple silicon Macs that runs the same engine. With an AI model running on that Mac, the records stay on the Mac. You sign in with a free Nutrient account, runs use Data Extraction credits, and Nutrient receives only page counts. Learn about Extract Studio for macOS.
Use the schema generator in Studio — upload example forms, records, or claims documents and describe the document type. It then generates a JSON Schema you can use directly with /extract. You can also write the schema manually for fields like patient details, diagnosis codes, dates, and payer information. Refer to the documentation for supported field types and limits.
Use agentic mode for medical documents that require the deepest visual understanding — degraded scans, faxed records, low-quality images, handwriting, and documents with embedded images or diagrams where understand mode produces insufficient results. Agentic mode adds a vision language model (VLM) on top of the understand pipeline for complex visual content. It costs 18 credits per page.