AI-Driven Multiagent System for Guiding First-Line Immunotherapy for NSCLC

The Shift Toward Agentic AI in Oncology

For years, artificial intelligence in cancer care has largely functioned as a series of specialized tools—one model to read a scan, another to analyze a pathology slide. However, the horizon is shifting toward “Agentic AI.” Unlike traditional models that perform a single task, agentic systems utilize multiagent architectures to replicate the broader decision-making process used by human clinicians.

A recent proof-of-concept presented at the European Society for Medical Oncology (ESMO) AI & Digital Oncology Congress highlights this evolution. Researchers developed a multiagent system capable of integrating multimodal data to guide first-line immunotherapy for non-small cell lung cancer (NSCLC). This system doesn’t just provide a “yes” or “no” answer; it uses a combination of a React agent and a Retrieval-Augmented Generation (RAG) agent to query documents and analyze patient data.

By utilizing specialized tools—such as the LORIS CLI-Lab for survival prediction, the MUSK histology vision-language model and the MedGemma radiology model—the AI can output key findings, a detailed rationale, and a treatment plan, whereas as well identifying potential contradictions in the data.

Beyond PD-L1: The Power of Multimodal Data

The current gold standard for predicting immunotherapy response in NSCLC is PD-L1 expression. While valuable, this single biomarker offers limited guidance due to various unresolved issues, leaving clinicians searching for more precise tools to identify potential responders.

The future of precision medicine lies in multimodal integration. Instead of relying on one marker, emerging AI platforms are synthesizing diverse data streams, including:

• Electronic Health Records (EHR): Patient history and clinical context.
• Radiomics: CT imaging and radiology reports.
• Pathology: Histology slides and molecular reports.
• Omics: Genomic and proteomic data to predict overall survival (OS) and progression-free survival (PFS).

Pathology-Driven Predictions with Path-IO

One significant leap in this direction is Path-IO, a machine learning platform developed at The University of Texas MD Anderson Cancer Center. Unlike molecular-only approaches, Path-IO utilizes pathological data that is already routinely gathered from patients. This platform has demonstrated the ability to outperform the current standard-of-care biomarker in predicting how metastatic NSCLC patients will respond to immunotherapy.

Because Path-IO was designed for clinical translation, it focuses on making explainable decisions based on known factors, ensuring the results are consistent across different data sets. Learn more about the Path-IO platform here.

Bridging the Gap to Clinical Reality

While the potential of AI is vast, the transition from a research setting to the clinic requires rigorous validation. For instance, while the multiagent system mentioned above showed high completeness, 6% of its statements were found to be harmful, and tool usage was correct in 56% of cases.

To solve this, the industry is moving toward two critical trends: larger validation cohorts and “human-in-the-loop” (HITL) approaches. Researchers are currently working to validate these agentic systems in cohorts of more than 700 patients to ensure reliability.

As noted by Dr. Danielle S. Bitterman of Harvard Medical School, the development of standardized evaluation and post-deployment monitoring frameworks is essential. Because AI reasoning is not always a reliable indicator of how a system arrived at a decision, maintaining human oversight is the only way to ensure safety in high-risk, high-reward tasks like treatment decision support. [Internal Link: The Role of Human Oversight in Medical AI]

Frequently Asked Questions

What is NSCLC?
Non-small cell lung cancer (NSCLC) is the most common histological subtype of lung cancer and a leading cause of cancer-related mortality worldwide.

How does AI improve immunotherapy decisions?
AI can analyze multimodal data—such as pathology slides, CT scans, and genomic data—to predict treatment efficacy more accurately than relying on a single biomarker like PD-L1.

What is “Agentic AI” in medicine?
Agentic AI refers to systems where multiple AI agents can interact, use digital tools, and follow multi-step workflows to perform complex clinical tasks, rather than performing a single specialized function.

Is AI replacing oncologists?
No. Current trends emphasize “human-in-the-loop” approaches, where AI acts as a decision-support tool to provide data-driven insights that are then verified and implemented by specialized physicians.