To Örebro University

BACKGROUND: Artificial intelligence (AI) is expected to introduce an increasing number of biomarkers in oncology. To bridge the gap between oncology and computer science, it is timely to define recommendations for AI-based biomarkers suitable for routine clinical use. Here, we propose the ESMO Basic Requirements for AI-based Biomarkers In Oncology (EBAI).

METHODS: The EBAI framework was developed using a modified Delphi methodology, involving a multidisciplinary panel of 37 experts who participated in three structured consensus rounds.

RESULTS: AI-based biomarkers were classified as Class A (AI quantification of established biomarkers), Class B (indirect measure of known biomarkers using AI-based alternative methods, to be deployed as pre-screening tests), and Class C (novel AI-derived biomarkers, with C1 for prognosis and C2 for prediction of treatment effect). The EBAI framework addresses AI biomarkers for clinical use. Ground truth, performance, and generalisability were considered essential; fairness was recommended. Minimal validation requirements indicate that Class A requires concordance studies, Class B analytical validation, Class C1 high-quality retrospective real-world or clinical trial data, and Class C2 additionally requires clinical validation in prospective clinical trials for the prediction of response to a new treatment. All biomarker studies should report multiple evaluation and calibration metrics, with a clearly defined primary objective. Generalisability should be demonstrated across all intended use settings, including variability in data acquisition, post-processing, and population characteristics. Biomarkers must not be applied to other cancer types or modalities without supporting evidence.

CONCLUSION: EBAI defines criteria for AI-based biomarker adoption in routine use, providing a common language for physicians, AI developers, and researchers.