Routine Health Data Flags Risk of 20 Cancers

Computer models that read ordinary health records, such as blood tests and medical history, could help spot people at higher risk of 20 common cancers, according to a study presented at the ESMO Congress 2026 in Madrid on 26 October. ESMO is the European Society for Medical Oncology.

The lead author, Dr Paul-Henry Koop of the Else Kröner Fresenius Center for Digital Health at TU Dresden in Germany, said: “For selected cancers, these models could serve as a low-cost pre-screening step using data already available in routine care.”

The findings come with big caveats. The study looked back at existing records, and many people the models flag will never develop cancer.

What the researchers built

According to ESMO’s press release, the team developed “approximately 500 machine learning models” to predict the risk of 20 common cancers. They trained them on records from 463,620 UK Biobank volunteers, who were followed for about 18 years. In that time, 57,055 of them developed cancer.

The models used medical history, lifestyle, blood tests, genetic data and metabolomic data, which measures small molecules in the blood. The team then tested them on volunteers in Scotland and Wales, on 42,036 people in the US All of Us programme and on more than 3 million people in Japan’s JMDC database. Cancers found in the first year were left out, so that tumours already present were less likely to skew the results.

The researchers’ own conference abstract gives different figures: “250 machine-learning models”, trained on “502,309 UK Biobank participants”. ESMO’s release does not explain the gap. The abstract says the work “Has not received any funding”.

How well it worked

Results varied a lot by cancer type. For a combined measure covering cancers that routine data could predict, the models scored 0.83 at 15 years on a standard accuracy scale, where “0.5 is no better than chance and 1.0 is perfect prediction”, the release says.

They did best for lung, oesophageal and liver cancers. They were weaker for ovarian, breast and cervical cancers. That echoes other recent work showing how hard ovarian cancer is to read, such as King’s College London’s AI that looks for a DNA-repair defect in ovarian tumours.

“Cancer types differ substantially in their carcinogenic mechanisms and in how strongly their risk factors and precursor conditions are captured in routine data,” Koop said. Blood tests and clinical history carried most of the useful signal. Adding genetic and metabolomic data helped only a little, which matters because those tests are costly and rare in everyday care.

The team also built PRE-Screen, a toolbox that lets AI chatbots run the risk models in a consistent way. Chatbots using it did “substantially better” than chatbots guessing on their own, ESMO says. Recent tests have shown chatbots struggling with medical judgements when left alone, such as a dermatologist’s test in which a 2015 AI beat two LLMs on skin cancer.

Why a risk flag is not a diagnosis

The promise here is real, and it is mostly about who gets checked. Koop pointed out that “existing screening programmes cover only certain cancers and often use broad eligibility criteria, rather than assessing each person’s individual risk.” In the EU, screening advice has focused mainly on breast, cervical and bowel cancer. A tool that reuses blood results a GP already holds could, in principle, help doctors pick out people for earlier checks on cancers with no screening programme at all.

But the study cannot yet show that. It analysed old records, and ESMO says “further research is needed to establish whether the approach leads to earlier diagnosis or better outcomes.” Koop himself cautioned that many people flagged as higher risk would not develop cancer. “A positive flag should therefore trigger minimally invasive testing first,” he said. Too many false alarms could mean anxious patients and extra scans for health services that are already stretched.

There is also a question of whether models built mostly on one group of UK volunteers will work elsewhere. Dr Marco Lorenzi of Inria in France, who was not involved, said: “What is needed is a comprehensive assessment of the proposed protocol in various kinds of populations and settings, external to the UK Biobank cohort.” Still, he said the method “may represent a valuable instrument for patient prioritisation, early detection and risk assessment,” adding: “The benefit would be alleviating the burden of care on society and the healthcare system.”

The mismatch between the abstract and the release is a further reason for care. Until the work appears in a peer-reviewed journal, the exact figures are hard to pin down.

The wider picture

The study joins a run of AI tools that try to get more out of routine medical data. Sophont’s Holon-S1 aims to predict tumour gene activity from standard tissue slides. And in diabetes care, AI eye exams closed a screening gap for young people, a reminder that the benefit depends on getting tools in front of patients. ESMO notes its release reflects the abstract and “does not necessarily reflect the views or opinions of ESMO”.