Quality assurance · Production

Flo Health scales medical content review with Amazon Bedrock MACROS solution

The problem

Flo Health publishes thousands of medical articles per year but verifying their accuracy against continuously evolving medical knowledge is a significant challenge — manual review is time-consuming and prone to human error.

Workflow diagram · grounded in source
1
Content and guidelines upload
trigger
“Users begin by gathering and preparing content that must meet medical standards and rules. In the second step, the data is provided as PDF, TXT files or text through the Streamlit UI that is hosted in Amazon Elastic Container Service (ECS).”
2
Optional rule relevance filter
validation
“This smart feature checks whether the set of rules is relevant for the article, potentially saving time and resources on unnecessary processing.”
3
Text chunking into sections
ai_action
“The source text is then split into paragraphs. This crucial step facilitates good quality assessment and helps prevent unintended revisions to unrelated text. Chunking can be conducted using heuristics, such as punctuation or regular exp…”
4
AI review against guidelines
ai_action
“Each paragraph or section undergoes a thorough review against the relevant rules and guidelines.”
5
AI revision of non-adherent content
ai_action
“Only the paragraphs flagged as non-adherent move forward to the revision stage, streamlining the process and maintaining the integrity of adherent content. The AI suggests updates to bring non-adherent paragraphs in line with the latest …”
6
Post-processing and document assembly
output
“the revised paragraphs are seamlessly integrated back into the original text, resulting in an updated, adherent document”
7
Medical expert final validation
human_review
“human-AI collaboration – not full automation – is key to successful implementation. Regular expert feedback and clear performance metrics guided system refinements and incremental improvements. While the system significantly streamlines …”
Reported outcome

The MACROS PoC achieved 80% accuracy and over 90% recall in identifying content requiring updates, exceeded speed targets, and the AI system applied medical guidelines more consistently than manual reviews while significantly reducing the time burden on medical experts.

Reported metrics
Content review accuracy80%
Recall in identifying content requiring updatesover 90%
Processing speed vs targetsexceeded target processing speed improvements
Time burden on medical expertssignificantly reduced
Show all 5 reported metrics
content review accuracy80%
recall in identifying content requiring updatesover 90%
processing speed vs targetsexceeded target processing speed improvements
time burden on medical expertssignificantly reduced
guideline application consistency vs manual reviewsmore consistently than manual reviews
Reported stack
Amazon BedrockAmazon Elastic Container ServiceAmazon API GatewayAmazon S3AWS IAMAWS Step FunctionsAWS LambdaAmazon TextractAmazon CloudWatchStreamlitClaude HaikuClaude Sonnet
Source
https://aws.amazon.com/blogs/machine-learning/scaling-medical-content-review-at-flo-health-using-amazon-bedrock-part-1?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The MACROS PoC achieved 80% accuracy and over 90% recall in identifying content requiring updates, exceeded speed targets, and the AI system applied medical guidelines more consistently than manual reviews while signi…

What tools did this team use?

Amazon Bedrock, Amazon Elastic Container Service, Amazon API Gateway, Amazon S3, AWS IAM, AWS Step Functions, AWS Lambda, Amazon Textract, Amazon CloudWatch, Streamlit.

What results were reported?

Content review accuracy: 80%; Recall in identifying content requiring updates: over 90%; Processing speed vs targets: exceeded target processing speed improvements; Time burden on medical experts: significantly reduced (source-reported, not independently verified).

How is this quality assurance AI workflow structured?

Content and guidelines upload → Optional rule relevance filter → Text chunking into sections → AI review against guidelines → AI revision of non-adherent content → Post-processing and document assembly → Medical expert final validation.

Related quality assurance cases