OncoAgent: An AI System for Early Cancer Detection Built on Private Patient Data

Privacy as the core principle of diagnosis

OncoAgent takes a radically different approach to applying artificial intelligence in the oncology clinic. Unlike cloud solutions, which require sending confidential patient information to remote servers, OncoAgent runs entirely locally. Medical histories, test results and personal data stay inside the protected clinical network and never leave it.

Architecture: a two-tier decision-making system

OncoAgent's main innovation is its architecture, built around two specialized models. The first tier runs a lightweight Qwen 3.5 model with 9 billion parameters. Its job is fast initial triage of incoming patient data: reviewing complaints, basic test results and medical history. The model decides whether a deeper analysis is needed or whether the case is routine and can be handled under established protocols.

For complex cases, the system switches to a second model, Qwen 3.6, with 27 billion parameters. This more powerful model performs a detailed analysis, drawing on a knowledge base of oncology research and clinical guidelines. The system uses Corrective RAG (Retrieval-Augmented Generation), which checks the quality of retrieved documents: the algorithm automatically assesses whether the extracted data is actually relevant to the specific case.

How it works: from data to recommendation

The system is built on the LangGraph framework, which organizes the analysis into eight sequential logical nodes. Each node handles a specific stage:

• Initial processing and structuring of medical data
• Analysis of clinical symptoms and disease history
• Search for relevant studies and guidelines in the knowledge base
• Quality and currency check of the retrieved information
• Synthesis of recommendations, with the algorithm's confidence level
• Verification of the result by a built-in safety validator
• Preparation of the recommendation in a format clear to the doctor
• Flagging cases that require a specialist's intervention

On AMD Instinct MI300X hardware, the system processes a complex case in under 30 seconds. That lets a doctor get support for a decision in real time during a patient visit.

Security and standards compliance

One of the key features is a built-in mechanism for removing protected information. Before analysis, the system automatically strips all direct patient identifiers from the text: full names, dates of birth, document numbers and addresses. The algorithm is trained to recognize and hide even indirect personal data, such as a rare combination of diseases that could point to a person's identity.

What sets it apart from other decision-support systems is the HITL (Human-In-The-Loop) principle: the system never issues a diagnosis on its own. Instead, it provides weighted recommendations, and the doctor makes the final decision. This is critical in oncology, where a diagnostic error can cost a patient's life.

What this means

OncoAgent demonstrates one possible path for deploying artificial intelligence in medicine without compromising on security and confidentiality. The system shows that high analytical accuracy and maximum data protection are not conflicting requirements but two sides of the same architectural problem.

For medical institutions, this means being able to use modern AI methods without worrying about violating patient privacy or international data protection laws.