Cancer Care AI Assistant

Inspiration

Cancer treatment decisions often require doctors to review complex pathology reports, molecular findings, patient histories, and clinical guidelines. This process can be time-consuming, especially when important information is scattered across different healthcare systems.

We wanted to make this process easier by building Cancer Care AI Assistant, an agentic AI-powered solution that helps healthcare providers turn pathology reports into evidence-supported clinical insights while keeping doctors in control of treatment decisions.

Our goal is simple: Upload a pathology report, ask a clinical question, and receive understandable insights with supporting citations.

What it Does

Cancer Care AI Assistant allows a healthcare provider to upload a patient's pathology report and ask questions such as: "Based on this pathology report, what treatment options should be considered?" The system is designed to:

• Extract important information from the uploaded pathology report.
• Identify cancer-related findings and available biomarker information.
• Retrieve relevant clinical evidence using Retrieval-Augmented Generation (RAG).
• Generate understandable clinical insights with source citations.
• Highlight missing patient information that may affect treatment decisions.
• Present potential treatment considerations for physician review rather than making autonomous treatment decisions.

How We Built It

Our architecture combines agentic AI, RAG, large language models, and clinical safety controls. The workflow begins when a healthcare provider uploads a pathology report through the application. After validating the document and patient context, a coordinator agent routes it to the appropriate clinical processing workflow. The system extracts the document text, divides it into clinically meaningful chunks, generates embeddings, and stores the information in a searchable vector index. When the provider submits a clinical question, the coordinator routes the request to the Clinical Intelligence RAG workflow. The retrieval system identifies relevant information from the pathology report and available clinical evidence. A context-engineering layer combines the question, retrieved evidence, patient context, source references, and safety instructions before passing them to the language model. The language model generates a response containing clinical insights, citations, and information requiring further verification. The architecture also incorporates a proposed HAARF-aligned governance layer to support authorized access, tool-use restrictions, auditability, and human oversight.

Challenges We Faced

Designing AI for cancer care requires more than generating convincing answers. One major challenge is maintaining clinical accuracy when pathology reports contain complex terminology, incomplete findings, or inconsistent formats. Another challenge is preserving source traceability so providers can verify where each AI-generated statement originated. We also need to maintain patient-specific context across multiple AI components without mixing information between patients. Finally, we must ensure that AI-generated treatment considerations are presented responsibly, with appropriate safeguards and physician review.

What We Learned

This project reinforced that healthcare AI must combine intelligent information retrieval with strong governance. We learned the importance of separating document ingestion, evidence retrieval, context engineering, and response generation into clearly defined stages. We also recognized that AI should support clinical judgment, not replace it.

What's Next

Our next steps are to expand support for structured clinical data through FHIR integrations, improve retrieval quality, strengthen multilingual capabilities, and evaluate outputs against clinician-reviewed test cases. We also plan to measure response time, citation accuracy, retrieval completeness, and the effectiveness of safety controls. Our vision is to make evidence-supported cancer care information more accessible, understandable, and actionable for healthcare professionals.

Built With

• agentic-ai
• amazon-bedrock
• amazon-web-services
• context
• fastapi
• fhir
• healthcare-ai
• llm
• mcp
• natural-language-processing
• opensearch
• python
• rag
• streamlit
• vector-database