AWS re:Invent 2025 - Architecting Compliant GenAI Systems for Healthcare Workflows (PEX319)

Architecting Compliant GenAI Systems for Healthcare Workflows

Business Context and Problem Statement

• In the US, 4.6 million insurance claims are processed daily, with 10% denied due to medical coding errors
• Medical coding is done using the ICD-10 (International Classification of Diseases, 10th Revision) system, which has over 70,000 standardized codes
• Current medical coding process is error-prone, relying on rule-based systems and manual validation by coders
• Inaccurate coding leads to denied insurance claims and large medical bills for patients, as well as complexity and inefficiency for providers

Challenges with Traditional Approaches

• Rule-based systems using regular expressions struggle with variations in medical terminology and typographical errors
• Manual coding by human coders is time-consuming and prone to mistakes
• Complexity increases when patients have multiple conditions that require mapping to multiple ICD-10 codes

Proposed AI-based Solution Approaches

1. Foundation Model Prompting: Using a pre-trained language model (e.g., Anthropic's Claude, Nvidia's NovaPro) to generate ICD-10 codes directly from medical impressions

   • Pros: Straightforward to implement
   • Cons: Prone to hallucination and not specialized for the task

2. Retrieval-Augmented Generation (RAG):

   • Leverage a vector database (e.g., Amazon OpenSearch) to retrieve relevant ICD-10 code examples based on the input medical impressions
   • Use the retrieved examples to guide and ground the language model's response
   • Pros: Improved accuracy by leveraging relevant context
   • Cons: Requires additional steps for retrieval and integration

3. Fine-Tuning:

   • Start with a pre-trained language model (e.g., Amazon NovaPro)
   • Fine-tune the model on a dataset of labeled medical impressions and ICD-10 codes
   • Pros: Highest accuracy as the model's knowledge is baked into the weights
   • Cons: Resource-intensive fine-tuning process, may require significant training data

Technical Implementation Details

1. RAG-based Approach:

   • Create an Amazon OpenSearch index with ICD-10 code descriptions and sample medical impressions
   • Use Amazon Titan embeddings to enable semantic search within the index
   • Implement a Lambda function that:
     • Applies Bedrock guardrails to sanitize and protect sensitive data
     • Performs a semantic search to retrieve top matching ICD-10 code examples
     • Passes the examples and user input to a language model (e.g., Anthropic Claude) to generate the final ICD-10 codes
   • Deploy the solution as a Streamlit app or API for end-user interaction

2. Fine-Tuning Approach:

   • Prepare a dataset of labeled medical impressions and ICD-10 codes
   • Fine-tune an Amazon NovaPro model using the Bedrock Model Customization API
   • Deploy the fine-tuned model for on-demand inference using Bedrock's serverless deployment options
   • Implement a Lambda function to invoke the fine-tuned model and return the ICD-10 codes

Comparison and Results

• The fine-tuned model achieved 92% precision on real-world radiologic data, 70% faster than the manual coding process
• The fine-tuned model was 8% faster in latency compared to the RAG-based approach
• The solutions enabled:
  • Flexible integration options (Streamlit UI, API)
  • Reduced billing cycles and claim rejections
  • Freed up clinicians' time to focus on more strategic work

Best Practices and Lessons Learned

1. Security and Responsible AI:

   • Implement comprehensive Bedrock guardrails for both prompt engineering and model output
   • Ensure data privacy and protection, especially for sensitive PHI (Personal Health Information)
   • Leverage contextual grounding to mitigate hallucination and improve accuracy

2. Cost and Performance Optimization:

   • Use model evaluation to compare and select the most appropriate model
   • Leverage batch inference and prompt caching to improve performance
   • Design for robustness, including cross-region deployment and failure handling

3. Operational Excellence:

   • Utilize CloudWatch, CloudTrail, and other monitoring tools for observability
   • Implement a feedback loop to continuously improve the system based on user input

Conclusion

The presented AI-based solutions, leveraging techniques like retrieval-augmented generation and fine-tuning, demonstrate the potential to transform medical coding workflows. By automating the ICD-10 coding process, these solutions can reduce errors, streamline insurance claim processing, and free up clinicians' time to focus on patient care. The key is to balance the trade-offs between accuracy, speed, and resource requirements, while ensuring robust security and responsible AI practices.