Locally Deployable AI Agent Matches Hematology Tumor Board Decisions in Clinical Evaluations

A locally deployable artificial intelligence agent designed to assist in diagnosing and treating blood cancers has demonstrated performance comparable to human hematology tumor boards in multiple evaluations, according to a study published in Nature Medicine on June 30, 2026. Researchers found the AI system achieved high concordance—meaning its recommendations closely matched those of expert panels—across retrospective, external, and prospective assessments of hematological malignancies.

The AI agent, developed as a case-grounded large language model, was tested against real-world tumor board decisions for patients with blood cancers, including leukemias, lymphomas, and myelomas. In retrospective evaluations using historical patient data, the AI’s recommendations aligned with those of human experts in a substantial proportion of cases, according to the study. When applied to external datasets—data from institutions not involved in its training—the concordance rate remained strong, suggesting robustness beyond the specific environments where it was initially validated.

Prospective testing, where the AI evaluated new, real-time cases before treatment decisions were finalized, showed the system’s recommendations were adopted in a significant number of cases. “This level of agreement is unprecedented for an AI tool in oncology,” said a co-author of the study. “It suggests that such systems could serve as a reliable second opinion or even a primary decision-support tool in resource-limited settings.”

The study highlights several key advantages of the AI agent over traditional decision-making models. Unlike generic large language models, this system was trained on anonymized, de-identified patient records linked to specific tumor board discussions, allowing it to learn not just from clinical guidelines but from the nuanced reasoning of expert panels. This “case-grounded” approach may explain its higher accuracy compared to earlier AI tools in oncology, which often relied on broader medical literature or synthetic data.

However, the research also underscores limitations. The AI’s performance varied by cancer subtype, with lower concordance rates observed in rare or highly aggressive malignancies where even expert opinions diverge. Additionally, the study notes that the system does not yet replace human judgment entirely—it is intended as an adjunct tool to flag inconsistencies, suggest alternative treatments, or provide explanations for its recommendations. “The goal isn’t to automate decision-making but to augment it,” said the co-author. “Physicians still need to interpret the AI’s suggestions in the context of a patient’s overall health, preferences, and clinical trajectory.”

This development comes as hospitals and clinics increasingly adopt AI tools to address physician shortages and reduce diagnostic delays, particularly in oncology, where treatment decisions often require synthesizing vast amounts of data.

The Nature Medicine study builds on earlier work demonstrating AI’s potential in medical imaging and genomic analysis but represents one of the first rigorous validations of an AI agent’s ability to replicate the complex, iterative discussions of a tumor board. Previous attempts to use AI for treatment recommendations in hematology often struggled with overfitting—performing well on training data but failing to generalize to new cases. The authors attribute the success of this model to its focus on real-world clinical scenarios rather than abstract patterns.

What happens next will depend on regulatory pathways and clinical uptake. The U.S. Food and Drug Administration (FDA) has begun evaluating AI tools for oncology under its Software as a Medical Device framework, but approval requires demonstrating not just technical performance but also real-world impact on patient outcomes. Meanwhile, European regulators are exploring similar frameworks, with the European Medicines Agency (EMA) issuing draft guidelines in 2025 for AI-driven diagnostic aids.

For now, the study’s findings suggest that AI agents could play a growing role in hematology, particularly in regions where access to specialized tumor boards is limited. A computational biologist called the results “a significant step forward” but cautioned that further validation is needed. “We’re still in the early stages of understanding how these tools integrate into clinical workflows,” the biologist said. “The next challenge is ensuring they don’t introduce new biases or reduce the human element of care.”

As research progresses, experts anticipate that such AI systems could evolve to handle more complex scenarios, including predicting treatment responses or identifying novel biomarkers. However, the focus remains on collaboration between AI and clinicians rather than replacement. “The most promising applications will be those where the AI highlights what humans might miss,” said the co-author. “That’s where the real value lies—not in replacing judgment, but in sharpening it.”