Artificial intelligence (AI) has become an increasingly common tool in oncology, with one of its more transformative uses—AI-based biomarkers—emerging in recent years. As the number of AI-generated biomarkers continues to grow, experts are working to define standards for evaluating and applying these tools in clinical practice.
One of those experts is Mihaela Aldea, MD, PhD, a medical oncologist at the Gustave Roussy Institute in Villejuif, France, who spoke at the 2026 European Lung Cancer Congress (ELCC) iin March in Copenhagen, Denmark.
During her presentation, titled Artificial Intelligence Requirements for Predictive Biomarkers, Dr. Aldea explored not only how clinicians can trust AI-based biomarkers but also how to determine whether these tools have achieved clinical validation.
Dr. Aldea described a framework called the European Society for Medical Oncology (ESMO) Basic Requirements for AI-based Biomarkers in Oncology (EBAI). “We classify the biomarkers first and then provide minimum requirements to identify which biomarkers are ready for clinical use,” Dr. Aldea said.
Dr. Aldea described how the EBAI framework organizes AI-based biomarkers into three primary classes. Class A focuses on biomarker quantification, Class B on the indirect measurement of existing biomarkers, and Class C on novel biomarkers for outcome prediction, which are further divided into prognostic and predictive categories.
To illustrate each category, she presented a patient case and applied the three classes of AI-based biomarkers: AIM-PD-L1 (Class A), EAGLE (Class B), and TROP2 NMC (Class C).
Dr. Aldea then applied the EBAI criteria to each example, focusing on key questions that determine clinical readiness. These included the need for prospective trials, the selection of an appropriate comparator, the assessment of performance, and the evaluation of generalizability. Additional considerations included explainability, cost, turnaround time, and fairness.
This evaluation often comes down to a central question: can the results be used to guide treatment? In a Class B example, the AI model EAGLE was applied to hematoxylin and eosin (H+E) slides to predict the presence of an EGFR mutation, although key limitations were noted regarding how these results could be used clinically.
“We currently cannot know the mutation subtype from this AI prediction,” she said. “It can predict the probability of the presence or absence of a mutation, but for now, it is not able to—and I am not sure it will be able to—define the subtype.”
Dr. Aldea emphasized that while many studies are applying the EBAI framework as a proof of concept, further work is needed to integrate these tools into routine clinical use. Achieving this requires high-quality validation, including the reporting of multiple performance metrics, calibration measures, and the application of biomarkers exclusively within validated contexts.
“Validation alone is not enough,” Dr. Aldea said. “We need more steps post-validation—which is a bit different compared with standard biomarkers—because AI can have data drift.” She added that continuous monitoring is essential even after validation and emphasized the importance of shared responsibility for errors.
As these requirements evolve, Dr. Aldea suggested that the role of AI in oncology may extend beyond individual biomarkers to more integrated approaches. Rather than relying on single or limited biomarkers to guide treatment decisions, future models may incorporate multi-omic data into unified, patient-specific predictions.
“Our tumor boards include molecular oncologists, pathologists, radiologists, and surgeons. But maybe tomorrow, our tumor boards will include AI engineers and mathematicians to be sure these AI tools are correctly implemented,” Dr. Aldea said. “I would say we need to make [our tumor boards] at least 5% AI engineers in the future.”
