What's Happening?
An international study presented at the ESMO Congress 2026 by ASCO AI has developed and validated approximately 500 machine learning models designed to stratify risk for the 20 most common cancer types. These models utilize routinely collected clinical
data, including demographic information, lifestyle factors, ICD-10 coding history, blood biomarkers, genetic data, and metabolomic data. The research, led by Dr. Paul-Henry Koop of the Else Kröner Fresenius Center for Digital Health, TU Dresden, Germany, aims to move cancer screening eligibility from broad criteria to individual risk assessment. The models were trained on a cohort of 463,620 individuals from the U.K. Biobank, with follow-up extending up to 18 years, and included 57,055 patients with incident cancer. External validation was conducted using data from 42,036 individuals in the U.S. All of Us program and over 3 million individuals from the Japan Medical Data Center database, alongside data from Scotland and Wales.
Why It's Important?
This development is significant for U.S. healthcare as it offers a scalable foundation for structured pre-screening and clinical decision support, potentially improving cancer outcomes through targeted prevention and surveillance strategies. Current screening programs often rely on broad eligibility criteria, which may not be optimal for individual risk assessment. By using readily available routine clinical data, these machine learning models could serve as a low-cost pre-screening step, making cancer risk stratification more accessible and efficient. The ability to identify individuals at higher risk could lead to earlier interventions and more personalized care pathways. While the models showed strong performance for cancers with clear prediagnostic signatures like lung, esophageal, and liver cancers, their predictive ability for sporadic cancers such as ovarian, breast, and cervical cancers was limited, indicating areas for further refinement. The inclusion of U.S. data in the external validation highlights the direct applicability and potential impact on American patients and healthcare systems.
What's Next?
The researchers have developed a toolbox called PRE-Screen to support the implementation of these risk stratification models, enabling large language model-based agents to run them consistently. However, further research is needed to definitively establish whether this AI approach will lead to earlier diagnoses or improved patient outcomes. ESMO discussant Marco Lorenzi, PhD, Research Director at Inria Sophia Antipolis, Université Côte d’Azur, France, emphasized the need for a comprehensive assessment of the proposed protocol in diverse populations and settings beyond the U.K. Biobank cohort. If successful, this approach could represent a valuable instrument for patient prioritization, early detection, and risk assessment, potentially alleviating the burden on healthcare systems and society. The ongoing evaluation will determine the practical utility and widespread adoption of these AI-driven risk stratification tools in clinical practice.
Beyond the Headlines
The integration of machine learning into cancer risk stratification raises important ethical and practical considerations. While the models aim to personalize care, the caution from Dr. Koop that many individuals stratified as high-risk may never develop cancer underscores the need for minimally invasive testing first to avoid unnecessary anxiety or procedures. The reliance on routine clinical data also brings to light questions about data privacy, security, and equitable access to such advanced screening methods across different socioeconomic groups. Furthermore, the varying performance across cancer types suggests that a 'one-size-fits-all' AI solution may not be feasible, necessitating continuous development and validation for specific cancer contexts. The potential for AI to transform preventive medicine is immense, but its responsible deployment will require careful navigation of these complex issues to ensure patient benefit and trust.













