Editor's Note: The 2025 European Hematology Association (EHA) Annual Meeting was held in a hybrid format, both online and in-person, gathering top experts and scholars from the global hematology community. The conference focused on the deep integration of cutting-edge technology and clinical practice, with the application of Artificial Intelligence (AI) being one of the core topics. Professor Jakob Nikolas Kather from the Technische Universität Dresden, Germany, was invited to deliver a keynote speech titled "Artificial Intelligence in Hematology: Opportunities, Pitfalls, and How to Bring It into Clinical Practice." He systematically described the revolutionary changes in AI technology from its early explorations to the current era of large models. He particularly highlighted his team's pioneering research on empowering Tumor Board decision-making with AI Agents, charting a course for the future development of AI in hematologic oncology.

In recent years, the wave of artificial intelligence technology has swept across the globe, profoundly transforming various industries. The healthcare sector, especially oncology and hematology, stands at the forefront of this change. At the 2025 European Hematology Association (EHA) Annual Meeting, the renowned scholar Professor Jakob Nikolas Kather from the Technische Universität Dresden, Germany, with his extensive clinical and research background, painted an “evolutionary map” of AI applications in hematologic oncology for us and shared practical pathways and profound insights from current technology implementation to future automated decision-making.

The Wave of AI Development: Evolution from “Watson” to the Transformer Architecture

At the beginning of his report, Professor Kather reviewed the two waves of AI development in oncology. He pointed out that many still remember the first AI boom over a decade ago, represented by IBM’s “Watson for Oncology.” At the time, the industry had high hopes for it, but it eventually faded after failing to deliver on its promises. However, the pace of technology never stopped. With the advent of ChatGPT in 2023, we have entered a new era of AI driven by Large Language Models (LLMs).

The core driving forces of this technological revolution are threefold: first, the emergence of a new neural network architecture called “Transformer” and the maturation of self-supervised learning techniques; second, the application of data on an unprecedented scale; and third, the exponential growth of computing power. Professor Kather emphasized that a key realization is that investing more resources (data and computing power) consistently improves model performance, which has directly spurred massive investments worldwide. Whether it’s the $500 billion pledged by the United States or the over €200 billion planned by the European Union, the primary justification for their societal value points to breakthroughs in healthcare, particularly cancer treatment. “This is not only a tremendous opportunity but also bestows upon us, as clinicians, the responsibility to translate these investments into societal well-being,” said Professor Kather.

From Image Analysis to Clinical Validation: The Solid Implementation and Challenges of AI

On the path toward artificial general intelligence, the clinical application of AI is far from being a castle in the air. Professor Kather noted that before 2023, the vast majority of regulatory-approved medical AI devices were related to medical image analysis. In the field of hematology, AI applications have also been remarkably successful. For instance, AI can perform cytological classification of hematologic malignancies through image analysis and even predict the NPM1 gene mutation status from routine Acute Myeloid Leukemia (AML) images—a feat that human experts cannot achieve by direct observation—thus achieving a leap from image to biomarker.

However, the potential of technology must be validated by rigorous clinical evidence. Professor Kather cited a landmark study published in The Lancet Oncology as an example. This prospective randomized controlled trial, which enrolled one hundred thousand women for breast cancer screening, confirmed that AI-assisted screening was safe and effective in meeting its primary endpoint. “This has set the gold standard for evaluating all AI technologies,” he stated. Based on this, authoritative bodies like the European Society for Medical Oncology (ESMO) are actively developing validation guidelines for AI biomarkers, aiming to provide a standardized path for the healthy development of this emerging field.

Large Language and Vision Models: Reshaping Clinical Decision-Making and Data Utilization

Currently, despite the existence of approved AI medical devices, their penetration into clinical practice is far outpaced by the “shadow use” of general-purpose large models like ChatGPT. Professor Kather cited a study indicating that up to 20% of general practitioners in the UK use such tools in their daily work, and the actual proportion is likely higher. He emphasized that sensitive patient data must never be entered into public commercial AI tools under any circumstances and called for hospitals to urgently deploy secure and controllable in-house large models to mitigate data security risks.

In terms of clinical decision support, large language models have shown astonishing potential. An evaluation of difficult cases published in The New England Journal of Medicine (NEJM) showed that a six-month-old LLM achieved a diagnostic accuracy of 80%, far exceeding the 30% rate of human physicians. Professor Kather pointed out that although LLMs have issues like outdated knowledge bases, their performance can be significantly enhanced through “in-context learning” (i.e., providing the model with the latest guidelines or patient records). Furthermore, LLMs have demonstrated disruptive efficiency in processing unstructured medical texts, such as automatically extracting structured data from tens of thousands of pathology reports or discharge summaries, which will significantly advance real-world studies.

At the same time, Professor Kather introduced the cutting-edge field of Vision-Language Models (VLMs). These models can understand both images and text simultaneously. For example, PathChat, developed by a team at Harvard Medical School, has already shown capabilities surpassing general-purpose models in answering questions related to medical images.

AI Agents: Marching Towards the Future of Automated Tumor Board Decision-Making

The main highlight of the presentation was Professor Kather’s introduction of his team’s latest research on AI Agents. He explained vividly, “A large language model can tell you how to get from Milan to Dresden, but it can’t book the ticket for you. An AI Agent, however, can connect to other software (like a booking website) and perform actions on your behalf.” This marks a leap for AI from passive answering to active execution.

This revolutionary technology is being applied in various fields, including software development and bioinformatics. In the medical domain, Professor Kather’s team is the first to use it to automate the decision-making process of the multidisciplinary tumor board (MDT). In their recently published paper, they constructed an AI agent system based on large language models. The core advantage of this system is that a main model acts as a “central coordinator,” capable of understanding complex case information and intelligently invoking other sub-models specifically trained to analyze different data modalities, such as radiology images, pathology slides, and molecular pathology reports. The system can automatically integrate and iteratively analyze all information to ultimately form an expert-level comprehensive diagnosis and treatment recommendation.

“An automated Tumor Board has been a goal we’ve dreamed of for over a decade, previously unattainable due to technological limitations,” Professor Kather concluded. “We have reason to believe that AI Agent technology, based on large language models, is the right path to achieving this goal.”

Future Outlook and Regulatory Considerations: Navigating Opportunities and Risks

Looking ahead, the role of AI in medicine raises deeper questions. Professor Kather mentioned a legislative draft being discussed in the United States that actually proposes to grant AI the status of a “practitioner” with the authority to prescribe prescription drugs under certain conditions. While this may sound like science fiction, it genuinely reflects the immense impact of this technology.

However, opportunities coexist with risks. As AI agents gain greater autonomy, their potential errors could also have more severe consequences. Professor Kather warned that the medical field must be more conservative and risk-averse than any other, maintaining a cautious approach to the application of new technologies.

At the end of his presentation, Professor Kather called for strengthening discussion and collaboration on AI technology within the hematology community. He also announced that the upcoming ESMO Congress on AI and Digital Oncology in Berlin will delve deeper into these issues. He emphasized that his team is composed of half physicians and half computer scientists, and this deep interdisciplinary integration is key to developing AI tools that genuinely meet clinical needs.

This presentation clearly illustrates the evolutionary path of artificial intelligence in hematologic oncology—from an auxiliary tool to a collaborative partner, and potentially to a future decision-maker. Professor Kather’s research not only brings unprecedented imaginative possibilities to the clinic but also highlights the core value of interdisciplinary collaboration in global health technology innovation, propelling us steadily toward a new era of more precise, efficient, and intelligent cancer diagnosis and treatment.