OncoAgent: An Open-Source Privacy-Preserving Clinical Decision Framework
- •OncoAgent is a new open-source system designed for oncology clinical decision support.
- •The framework uses a dual-tier agent architecture to ensure privacy by enabling local, on-premises deployment.
- •Built with LangGraph, the system achieved full-dataset fine-tuning in under one hour using AMD hardware.
OncoAgent is a new, open-source AI framework designed to provide clinical decision support for oncology. The system is engineered to solve three major limitations in current clinical AI: high rates of hallucination, reliance on external cloud APIs that risk patient data, and architecture saturation during complex medical cases.
The framework utilizes a dual-tier approach, routing clinical queries to either a 9B parameter model for speed or a 27B parameter model for deep reasoning. To maintain privacy, OncoAgent is built for on-premises deployment, ensuring patient data sovereignty by eliminating the need to send information to third-party servers. It includes a dedicated "Zero-PHI" redaction node that identifies and replaces personal health information before any text is processed by the AI.
System reliability is maintained through a four-stage retrieval pipeline using Corrective RAG (a technique that checks retrieved documents for relevance before using them) and a safety validation loop. The team fine-tuned the models using QLoRA (a method to reduce memory usage during training) on AMD MI300X hardware. By using sequence packing, they completed full-dataset fine-tuning in approximately 50 minutes, significantly faster than typical API-based training.