The advancements in artificial intelligence (AI) have paved the way for innovative approaches in healthcare, particularly in oncology. A multiagent AI system has recently showcased its potential in determining effective immunotherapy options for patients with non-small cell lung cancer (NSCLC) in first-line treatments. Findings presented at the inaugural European Society for Medical Oncology (ESMO) AI & Digital Oncology Congress highlight the significance of this development.1
“Our immune-specialized AI agent system is adept at integrating diverse multimodal patient data to enhance first-line immunotherapy decision-making for NSCLC, demonstrating commendable performance,” remarked Federica Corso, PhD, from the Fondazione IRCCS Istituto Nazionale dei Tumori in Milan.
Background
Currently, the sole biomarker available for predicting immunotherapy response is PD-L1 expression. However, this marker is limited in its ability to identify potential responders due to various unresolved factors. The research team sought to refine predictive tools to optimize first-line immunotherapy decisions for individuals with previously untreated NSCLC.
Federica Corso, PhD
While existing multimodal AI systems have shown predictive capabilities, they typically operate one specialized task at a time, failing to mimic comprehensive decision-making processes. In contrast, a multiagent system can evaluate and perform multiple tasks simultaneously, enabling more thorough treatment recommendations.
Study Design
The study evaluated 58 patients with stage IV NSCLC, who were part of the APOLLO 11 observational study (ClinicalTrials.gov identifier NCT05550961), focused on advanced lung cancer patients receiving innovative therapies. All participants were treated at the Istituto Nazionale dei Tumori in Milan, with 21 receiving immunotherapy alone and 37 undergoing a combination of immunotherapy and chemotherapy.
The research team developed a multiagent system trained on medical knowledge, integrating web searches and accessing a diverse range of patient data, including electronic health records, CT imaging, histology slides, lab work, and molecular reports. The agents’ outputs were assessed by four specialized oncologists, evaluating the accuracy, helpfulness, completeness of the immunotherapy recommendations, and the quality of the AI-generated responses from retrieved information.
[Our] immune-specialized AI agent system is able to integrate the multimodal data of our patients to support the decision-making of first-line immunotherapy for NSCLC with good performance.
— FEDERICA CORSO, PhD
Tweet this quote
The multiagent framework encompassed a React agent that accessed AI tool outputs and a retrieval-augmented generation (RAG) agent for document querying and retrieval. Available AI tools for the agent included the LORIS CLI-Lab model for predicting immunotherapy response and survival rates, the MUSK histology vision-language model for predicting response and histology type, the MedGemma radiology vision-language model for report generation, as well as web search and application programming interfaces (APIs).
Dr. Corso explained that the multiagent system analyzed the clinical content of each patient to determine optimal treatment strategies based on the data and tools at its disposal. It then produced key findings, a rationale for its recommendations, a treatment plan, and insights into any ambiguities or inconsistencies.
Results
The system achieved a correctness rate of 72% in its statements, while the recommendations were deemed helpful 72% of the time and complete 91% of the time. Only 6% of the recommendations were identified as harmful.
Information retrieved during decision-making was meaningful in 98% of instances. The correct usage of tools was noted in 56% of cases, while issues included 25% related to missing or failed data, 11% to incorrect usage, and 8% involved tools that were ultimately ineffective.
Looking ahead, the researchers plan to validate their system by incorporating additional tools and evaluation metrics across a larger cohort of over 700 patients. Dr. Corso added, “We will strengthen the reliability of this system by integrating human-in-the-loop methodologies.”
DISCLOSURE: Dr. Corso has served on advisory boards for EVENTs and Merck.
REFERENCE
1. Corso F, Carminati G, Peppoloni V, et al: Collaborative human-agent for therapeutic decision-making in cancer immunotherapy prediction. ESMO AI & Digital Oncology 2025. Abstract 71MO. Presented November 13, 2025.
EXPERT POINT OF VIEW
Danielle S. Bitterman, MD, Assistant Professor of Radiation Oncology at Harvard Medical School and Clinical Lead for Data Science/AI at Mass General Brigham Digital, shared her insights with The ASCO Post regarding this innovative agentic AI approach in treatment decision-making:
“Agentic AI allows models, such as large language systems, to interact and utilize digital tools within multistep workflows, enabling the execution of more complex clinical tasks. This study, though limited in scale and conducted at a single institution, serves as an exciting proof-of-concept for agentic AI in clinical decision support for immunotherapy. Effective cancer treatment necessitates contextualizing diverse patient data within an expansive and rapidly evolving medical knowledge landscape, and agentic AI presents a compelling solution to these intricate challenges.”
Danielle S. Bitterman, MD
A critical next step towards understanding and realizing the full potential of agentic AI involves establishing standardized evaluation and monitoring frameworks, particularly as these systems become more autonomous and are applied to high-stakes tasks like treatment decision-making. It is crucial to remember that while certain large language models may present reasoning along with a final decision, these outputs do not always accurately reflect the underlying decision-making process.
DISCLOSURE: Dr. Bitterman is part of the Scientific Advisory Board for Mercurial AI and Blue Clay Health LLC, neither of which are related to this submitted work.
