How CBG uses AI to effortlessly process complex medical documents with Amazon Textract, Comprehend Medical, and Bedrock

In this blog post, I’d like to share how CBG is transforming the way we handle complex medical documents using Generative AI (GenAI) and Large Language Models (LLMs) like Amazon Bedrock, Amazon Titan, and Anthropic Claude LLMs. These advanced tools are making a big difference. By automating the processing of lengthy medical documents, including 100+ page faxed Prior Authorizations from providers, we’re making AI conversations more dynamic and boosting productivity with smarter, more intuitive responses.

The Problem

Our service teams and clinicians at CBG often receive large and complex medical documents from providers. These documents come in various formats, sizes, and orientations, often faxed or emailed. Each document needs to be reviewed by our team, and the relevant information has to be entered into our systems both in summarized and detailed formats. This process is time-consuming, prone to errors, and inefficient. We needed a scalable way to automate the extraction of relevant information and its entry into our systems.

The Solution

To tackle this, we use a combination of Amazon Textract, Amazon Comprehend Medical, Amazon Bedrock, Amazon Titan LLMs, and Anthropic Claude LLMs. Amazon Textract automatically extracts text and data from scanned documents. Amazon Comprehend Medical, Amazon Bedrock, and Anthropic Claude LLMs help process this extracted data to generate human-like text. With Textract and Comprehend Medical, we extract text and data from medical documents.

Then, we use Bedrock and Claude LLMs to process this data, making it easier to read and input into our systems. This automation has significantly cut down the time needed to process these documents and improved the accuracy of the extracted information, enhancing the service our members receive.

We specifically use Amazon Bedrock with LLMs in two main ways: to summarize complex documents into readable notes for our team and to extract specific data points for precise entry into our systems.

The Implementation

When we receive inbound medical documents, we start the process with Amazon S3 notifications to EventBridge, which sends document metadata to an Amazon SQS queue. This triggers an Amazon Lambda function that runs an AWS Step Function State Machine. The state machine first uses Amazon Textract to extract text and data from the document.

Next, Amazon Comprehend Medical, Amazon Bedrock, Amazon Titan, and Anthropic Claude LLMs process this data to create human-like responses. These responses help summarize the document and extract key data points. Finally, the summarized document and data points are entered into our systems. This automated process takes only a few minutes.

The Results

This solution has proven to be highly scalable, reliable, and cost-effective. We can process hundreds of documents daily with minimal human intervention. The accuracy of the information extracted has greatly improved, leading to better service for our members.

On the Horizon

We’re excited about the future possibilities of this technology. We’re working on predicting outcomes based on the data extracted from medical documents. We’re also developing smart forms for our team using LLMs and exploring how LLMs can generate helpful responses to user inquiries about our systems. By leveraging these advanced tools, we aim to provide better service to our members and improve the efficiency of our operations.