AI Models Show Promise in Predicting Risk of 20 Common Cancers

INEN

2 sources · 6 quotes

Limited corroboration · 2 of 2

Started 2D · Updated 1D

Researchers have developed approximately 500 machine learning models that can predict an individual's risk of developing 20 common types of cancer using routinely collected health data, according to a study presented at the European Society for Medical Oncology (ESMO) congress in Madrid.

The models were trained on data from 463,620 participants in the UK Biobank, tracked for about 18 years, during which 57,055 developed cancer. The algorithms incorporated variables such as medical history, lifestyle factors, blood test results, and genetic and metabolic data.

Performance and Accuracy

The models outperformed traditional statistical methods, achieving an AUROC of 0.83 at 15 years for cancers whose risk could be estimated from routine clinical data. (AUROC ranges from 0.5, equivalent to random chance, to 1.0, perfect discrimination.) The predictive power varied by cancer type: results were especially promising for lung, esophageal, and liver cancers, but more limited for ovarian, breast, and cervical cancers.

Key Findings

Most predictive power came from routine blood tests and clinical history; blood tests remained informative for cancers diagnosed more than five years later.
Adding genetic and metabolomic data only slightly improved model performance.
For acute myeloid leukemia and multiple myeloma, model performance declined over longer prediction horizons, suggesting short-term predictions may detect early signals of already-present tumors.

The team also developed a tool called 'PRE-Screen' to help AI agents run the risk prediction models systematically, yielding significantly better results than unsupported predictions.

This storyTap to open