OncoAgent: AI system for early cancer detection based on private patient data

Privacy as the Core Principle of Diagnosis

The OncoAgent system represents a revolutionary approach to applying artificial intelligence in oncology clinics. Unlike cloud-based solutions, which require the transfer of confidential patient information to remote servers, OncoAgent operates entirely locally. This means that patient medical history, test results, and personal data remain in the protected clinical network and never leave its boundaries.

Architecture: Two-Level Decision-Making System

The main innovation of OncoAgent lies in its unique architecture with two specialized models. At the first level operates a lightweight Qwen 3.5 model with 9 billion parameters. Its task is rapid initial sorting of incoming patient data: analysis of complaints, basic tests, and medical history. The model determines whether deep analysis is required, or whether the case is routine and can be processed according to established protocols.

For complex cases, the system switches to the second model—Qwen 3.6 with 27 billion parameters. This more powerful version conducts detailed analysis, accessing a database of oncology research and clinical recommendations. The system uses Corrective RAG (Retrieval-Augmented Generation) technique with quality verification of found documents—the algorithm automatically assesses whether the retrieved data is truly relevant to the specific case.

How It Works: From Data to Recommendation

The system is built on the LangGraph framework, which organizes the analysis process as eight sequential logical nodes. Each node is responsible for a specific stage:

• Initial processing and structuring of medical data
• Analysis of clinical symptoms and disease history
• Search for relevant research and recommendations in the knowledge base
• Verification of found information for quality and relevance
• Synthesis of recommendations with indication of algorithm confidence level
• Verification of result using built-in safety validator
• Preparation of recommendation in a format understandable to physicians
• Indication of the need for specialist intervention

On AMD Instinct MI300X equipment, the system achieves processing speeds of complex cases in less than 30 seconds. This allows the physician to receive support for their decision in real time during patient examination.

Security and Standards Compliance

One of the key features is the built-in protected information removal mechanism. Before analysis, the system automatically removes all direct patient identifiers from the text: names, birth dates, document numbers, addresses. The algorithm is trained to recognize and hide even hidden personal data—for example, a rare combination of diseases that could indirectly indicate identity.

The main difference from other decision-support systems is the HITL (Human-In-The-Loop) principle: the system never makes a diagnosis independently. Instead, it provides weighted recommendations, from which the physician makes the final decision. This is critically important in oncology, where a diagnostic error can cost the patient their life.

What This Means

OncoAgent demonstrates a possible path for implementing artificial intelligence in medicine without compromises in security and privacy. The system shows that high analysis accuracy and maximum data protection are not contradictory requirements, but two sides of one architectural task. For medical institutions, this means the possibility of using modern AI methods without concerns about violating patient privacy and complying with international data protection laws.