AI agents in oncology: Clinical decision support system

Oncological treatment decisions require the integration of diverse medical information, ranging from radiological images and genetic data to electronic health records and treatment guidelines. AI systems supporting clinical workflows must therefore process multimodal data efficiently and apply reasoning skills similar to those of experienced oncologists. 

To achieve this, the Dresden-based team enhanced the GPT-4 large language model with specialized tools: 

• generation of radiology reports from MRI and CT scans 

• medical image analysis 

• prediction of genetic mutations based on histopathological slides 

• targeted searches in databases such as PubMed, OncoKB, and Google 

• access to more than 6,800 clinical documents and official oncology guidelines 

The system was tested using 20 real-world-inspired, simulated oncology cases. In each scenario, the AI agent selected the appropriate tools, searched for relevant medical information, and proposed a clinical decision. Human experts evaluated the AI’s outputs for accuracy, completeness, and source citation. The AI: 

• delivered correct clinical recommendations in 91 percent of cases 

• correctly cited relevant oncology guidelines in over 75 percent of responses 

• showed reduced hallucination rates due to the use of validated medical sources 

“AI tools are designed to support medical professionals, freeing up valuable time for patient care,” explains Dyke Ferber, first author of the study. “They could help in daily decision-making processes and support doctors to stay updated on the latest treatment recommendations, contributing to the identification of optimal personalized care for cancer patients.” 

The study highlights the potential of AI-based support in oncology but also notes the limitations. The current system was only tested on a limited number of simulated cases and still requires broader clinical validation.  

“To fully leverage the potential of AI agents in the future, it will be crucial to integrate them smoothly into routine clinical practice with minimal disruption,” says Prof. Jakob N. Kather, Professor of Clinical Artificial Intelligence at EKFZ and oncologist at Dresden University Hospital Carl Gustav Carus. 

In the long term, similar AI tools could be adapted for other medical disciplines. The research emphasizes that these agents are intended to support, not replace, clinical professionals. “These agents are designed to support clinicians, but by no means to replace them,” Prof. Kather adds. 

MEDICA-tradefair.com; Source: Technische Universität Dresden