Machine learning models using routinely collected health information, such as blood tests, medical history and lifestyle factors, could help identify people at higher risk of several common cancers, according to an international study reported at the ESMO Congress 2026.
Researchers developed approximately 500 machine learning models to predict risk across 20 common cancers, using data from 463,620 UK Biobank participants who were followed for approximately 18 years. During this period, 57,055 participants developed cancer. The models analysed information including medical history, lifestyle factors, blood tests, genetic and metabolomic data.
To assess how well the models performed in different populations, researchers tested them using data from participants in Scotland and Wales, as well as 42,036 people from the US All of Us research programme and more than three million from Japan’s JMDC database. To reduce the influence of undiagnosed disease, cancers diagnosed within the first year were excluded from the relevant models.
The models outperformed traditional statistical approaches, although performance varied across cancer types. For a combined cancer outcome covering cancers that could be predicted from routine clinical data, they achieved an area under the curve (AUROC) of 0.83 at 15 years, where 0.5 is no better than chance and 1.0 is perfect prediction. Performance was particularly strong for lung, oesophageal and liver cancers, but more limited for ovarian, breast and cervical cancers. For acute myeloid leukaemia and multiple myeloma, performance declined when looking further ahead, suggesting some short-term predictions may reflect early signs of cancer that was already present but undiagnosed.
Much of the predictive information came from routine blood tests and clinical history. Blood tests continued to add predictive value for cancers diagnosed more than five years later, while genetic and metabolomic data added only a small further improvement. The team also developed PRE-Screen, a toolbox enabling artificial intelligence (AI) agents based on large language models (LLMs) to run the risk models consistently. Agents using it performed substantially better than LLMs making predictions unaided.
