News Research Prognostic & Predictive Models Decision-Making Support Lung Cancer

Multicancer Risk Stratification Machine Learning Models Enable Prediction Using Routine Data

November 16, 2026 ASCO AI Staff 4 min read
Share Share via Email Share on Facebook Share on LinkedIn Share on Twitter

Using routinely collected clinical data, population-scale machine learning models stratified risk for patients across multiple common cancers. Findings from an international study were presented at the ESMO Congress 2026 (Abstract 280RO).

“Existing screening programmes cover only certain cancers and often use broad eligibility criteria, rather than assessing each person’s individual risk. For selected cancers, these models could serve as a low-cost prescreening step using data already available in routine care,” said lead author Dr. Paul-Henry Koop, Else Kröner Fresenius Center for Digital Health, TU Dresden, Germany.

“We provide a scalable foundation for structured pre-screening and clinical decision support, supporting targeted prevention and surveillance strategies with potential to improve cancer outcomes,” the study authors noted.

Study Methods

Researchers developed and validated about 500 machine learning models for stratifying risk for the 20 most common cancer types. The models used routine clinical data, including demographic, lifestyle, ICD-10 coding history, blood biomarker, genetic, and metabolomic data, to risk stratify patients. These models were trained on a cohort of 463,620 individuals from the U.K. Biobank who were followed for up to 18 years. The cohort included 57,055 patients with incident cancer.

Performance of the models was assessed with grouped cross-validation, geographic hold-out testing, and external validation. The external validation cohorts included 42,036 individuals from the U.S. All of Us program and more than 3 million individuals from the Japan Medical Data Center database, plus data from participants in Scotland and Wales for the geographic hold-out testing.

Cancers diagnosed within the first year were excluded from models to reduce the impact of undiagnosed diseases.

Key Findings

The machine learning models outperformed traditional linear approaches to cancer risk stratification, but performance did vary by cancer type. Among cancers with strong prediagnostic signatures—such as lung, esophageal, and liver cancers—the models' area under the receiver operating characteristic curve was 0.83 at 15 years, whereas the models had limited predictive ability for sporadic cancers—such as ovarian, breast, and cervical cancers.

For select cancers, including acute myeloid leukemia and multiple myeloma, performance declined with long-term prediction. The study authors suggested that such short-term–only predictions may have caught early signs of cancer that were already present but undiagnosed.

“Cancer types differ substantially in their carcinogenic mechanisms and in how strongly their risk factors and precursor conditions are captured in routine data,” Dr. Koop explained. He cautioned that many individuals stratified as having a higher risk of cancer would never develop cancer, so they should undergo minimally invasive testing first.

Feature weighting revealed that blood biomarkers and clinical history, among other routine features, impacted much of the models' predictability. Blood tests were especially valuable for cancers diagnosed after more than 5 years, but genetic and metabolomic data did not add significant benefit to the models' predictive ability. Shapley additive explanations (SHAP) analysis pointed to shared and cancer-specific risk patterns of cancer prevention strategies.

To better support implementation of the risk stratification models, the researchers developed a toolbox, called PRE-Screen, for large language model–based agents to autonomously run the risk stratification models consistently. The agents outperformed large language models conducting zero-shot prediction of incident cancer.

The study authors noted that further research is needed to establish whether the AI approach results in earlier diagnoses or better patient outcomes.

Insights

“What is needed is a comprehensive assessment of the proposed protocol in various kinds of populations and settings, external to the U.K. Biobank cohort,” said ESMO discussant Marco Lorenzi, PhD, Research Director at Inria Sophia Antipolis, Université Côte d’Azur, France.

“The automated analysis protocol proposed in this work may represent a valuable instrument for patient prioritization, early detection and risk assessment,” Dr. Lorenzi suggested. “The benefit would be alleviating the burden of care on society and the health-care system.”

DISCLOSURES: The study authors reported no conflicts of interest.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.

KOL Commentary
Watch

Related Content