A Blood Test Uses Machine Learning to Detect 50+ Types of Cancer

A new test measures the chemical changes in DNA found circulating in the blood, and uses machine learning to identify different forms of cancer.

Methylation is a chemical alteration to DNA that cells use to control which genes are expressed, and is one of the ways cancer cells alter their gene expression. Researchers fed a machine learning algorithm with methylation data from the blood samples of thousands of patients, and had it sort the data based on methylation patterns. They then ‘taught’ the system which patterns reflected which types of cancer.

They then tested the algorithm to see if it could correctly identify the presence and type of cancer based on methylation patterns in another group of patients. The accuracy of the test varied for different cancers at different stages: the system successfully detected 18% stage I cancers and 93% of stage IV cancers. Conversely, 0.7% of patients were falsely identified as having cancer.

In those identified as having cancer, the test correctly distinguished the tissue of origin 93% of the time. The researchers suggest that the test could be useful for early identification of particularly hard to detect cancers such as pancreatic cancer, which was identified with 63% accuracy at stage I.

While these results are generally promising, the team points out that the figures might vary when testing the general population. To test the accuracy of the algorithm, researchers had to use a group that were already known to have cancer, which may mean their cancers were more easily detectable to begin with.

It is also worth noting that, while a false positive rate of 0.7% may not sound like much, this would quickly add up if testing was deployed in the general population. For every 100 000 people tested, 700 would have to undergo unnecessary further investigation, although this still compares favourably to some existing screening programs such as breast cancer screening, which has a false positive rate of 7 to 12 percent. The accuracy of the new test will only improve with time, and this research offers hope for the detection of lethal cancers that currently lack screening programs.