"OncoAgent: A Dual-Tier Multi-Agent Framework for Privacy-Preserving Oncology Clinical Decision Support"

OncoAgent is an innovative, open-source AI platform developed to provide privacy-preserving clinical decision support in oncology. It uniquely integrates a dual-tier, fine-tuned Large Language Model (LLM) architecture with a sophisticated multi-agent LangGraph topology. This system is further enhanced by a four-stage Corrective RAG pipeline, drawing upon over 70 physician-grade guidelines, and incorporates a three-layer reflexion safety validator to ensure strict adherence to a Zero-PHI (Protected Health Information) policy.

The system intelligently routes clinical queries based on complexity. Simpler cases are handled by a 9-billion parameter, speed-optimized model (Tier 1), while more intricate scenarios are directed to a 27-billion parameter deep-reasoning model (Tier 2). Both tiers are fine-tuned using QLoRA on a vast corpus of over 266,000 real and synthetic oncological cases, leveraging the Unsloth framework on AMD Instinct MI300X hardware.

A key feature of OncoAgent is its ability to perform full-dataset fine-tuning rapidly, achieving approximately 56 times throughput acceleration compared to API-based generation. The Corrective RAG (Retrieval-Augmented Generation) pipeline boasts a 100% success rate in document grading, effectively preventing hallucinations by ensuring the clinical relevance of retrieved information.

OncoAgent's architecture includes a robust Reflexion Safety Loop, featuring a three-layer validation cascade that critically reviews output before it reaches a Human-in-the-Loop (HITL) gate. This ensures that safety enforcement is deterministic and cannot be bypassed. The entire system is 100% open-source and deployable on-premises, thereby eliminating dependencies on proprietary cloud APIs and upholding patient data sovereignty. It represents a significant step towards closing the knowledge gap between evolving medical evidence and clinical practice in oncology.