Extract Data from Health Documents

Learning Objectives

After completing this unit, you’ll be able to:

• Describe the document-processing workflow from uploading documents to saving records.
  
• Identify how the feature extracts entities from a referral and auto-populates records.
  
• Explain how human review, confidence scores, and record matching validate extracted data before it’s saved.
  

Process a Referral from Upload to Record

In the previous unit, you learned how Document AI for Health extracts information from unstructured documents and maps the reviewed results to Salesforce records. Now take a closer look at the user workflow that moves a document from upload through extraction, review, matching, and saving.

The workflow reduces the need to read and reenter every detail manually, while preserving the review steps required to confirm extracted values and connect them to the correct records.

At Cumulus Health, referral coordinator Dana Okafor uses this process for patient referrals from external providers. When a new referral arrives, she uploads the document and reviews the proposed records before making the information available to the care team.

In this unit, you learn how Document AI for Health processes a health document from upload to connected Salesforce records. You also see how confidence scores, human review, and record matching help ensure that extracted information reaches the correct records.

Upload the Document

Patient referrals can arrive as emailed PDFs, scanned forms, faxed letters, or photographs of paper documents. Their layouts and image quality vary by source, but the intake team still needs to identify the same core information and enter it into the appropriate records.

From the App Launcher, Dana finds and selects Document Extraction Request, then selects Extract Document.

She can choose a file already stored in Salesforce, upload a new one, or select a file directly from a connected Amazon S3 bucket. For this referral, she uploads a file and selects Next.

At this stage, Dana doesn’t need to create a separate intake process for each provider’s layout or delivery channel. Supported PDFs and image files enter the same extraction workflow, where the selected template determines what information the system looks for.

Extract the Data

A patient referral can contain information that belongs across several related records. It may identify the patient and referring provider, describe the clinical reason for the referral, name the requested service, and include coverage details. Intake staff must interpret those relationships correctly before the referral can move forward.

After uploading the file, Dana defines the scope of the run on the Review Extraction Configuration page:

• Select Pages determines whether the feature processes the full document or a specified page range.
  
• Review Template Type identifies the extraction template for the document. You can select a template manually or have the feature auto-identify it.
  

The selected template defines the entities and attributes the feature attempts to extract. The Records to Extract list shows the target record types configured for that template, including the patient and the clinical service request that connects the referral information.

Dana keeps All Pages, selects the Patient Referral template, reviews the configured record types, and selects Extract.

When the request is processed, the request status changes from Extraction In Progress to Review Not Started. The extraction can identify configured values in labeled fields, tables, or narrative text without depending on one fixed page position. Each result remains proposed data until a user reviews and saves it. When a run fails, the request status indicates the error and provides details for troubleshooting.

Review and Match the Results

Before extracted information becomes part of the patient record, Dana reviews the proposed values and confirms where they belong. An incorrect patient match, diagnosis, referring provider, or requested service can delay the referral or create duplicate and conflicting records.

From the new extraction request, Dana opens the action menu and selects View.

The Extraction Result page brings the source document and the proposed record structure into one review workspace.

The left panel lists the extracted entities defined by the template. When Dana selects an entity, the middle panel shows its proposed field values, confidence scores, and record options. The original referral remains visible in the right panel so she can compare each value with the source.

Dana reviews the entities one at a time. Confidence scores help her identify values that may require closer inspection, but she verifies the proposed information against the document before saving. She can correct any extracted value and verify where each entity belongs. By default, the feature automatically matches a patient to an existing account when one already exists, so Dana confirms the proposed match rather than searching for it. When there's no match, she creates a new record with the confirmed fields already populated.

This review step confirms both the content of the extraction and its place in the Health Cloud data model. By the time Dana finishes the review, the referral information is ready to be saved without creating an unnecessary patient record or losing the relationships among the referral, condition, coverage, and requested service.

Save to Records

After confirming the extracted values and resolving the record matches, Dana saves the record. Salesforce creates or updates the connected records defined by the Patient Referral template, then shows a Review Extracted Records summary of the saved results.

One of those newly created records is the Clinical Service Request, which serves as the anchor for the referral.

It connects the patient with the requested service and provides the record the care team can use to review, accept, and schedule the referral.

The remaining extracted information is saved in related records according to the Health Cloud data model.

These relationships preserve the structure of the referral rather than placing every extracted value on a single record. They also allow each part of the referral to participate in the workflows and reporting that use that Health Cloud object.

After the records are processed, the status of the extraction request changes to Save Completed. The original document remains available for reference, while its reviewed information is now represented in connected records that the care team can act on through the referral lifecycle.

What’s Next?

Document AI for Health has now turned a patient referral into reviewed, matched, and connected Salesforce records that the care team can use.

That workflow depends on configuration completed before the document arrives. An extraction template defines what information to identify, how to structure it, and where the confirmed values belong. Batch settings can also automate how high volumes of documents enter the process.

In the next unit, you learn how admins configure templates and batch processing, and how Salesforce tracks each extraction from the source document through the transform, review, and save stages.

Resources

• Salesforce Help: Use Document AI for Health