To identify oral and throat cancer, lesions have to be big enough for doctors to see. A saliva test could enable much earlier diagnosis
4 August 2022
A new diagnostic tool uses artificial intelligence to detect oral and throat cancers from saliva samples with more than 90 per cent accuracy.
Estimates suggest there will be 54,000 new cases of oral cancer and 20,640 new cases of oesophageal cancer in the US alone this year. The respective 5-year survival rates for these cancers are 68 and 20.6 per cent, but when detected early, those numbers jump to more than 86 and 47 per cent.
The issue is that most oral and throat cancers aren’t detected early. Current screening methods rely on visual examinations by a healthcare provider. This means lesions must grow large enough until they can be detected by the naked eye.
Previous research has shown oral bacteria in people with these cancers differ from those in people who are cancer-free. This inspired Guruduth Banavar at the biotechnology company Viome in New York City to see if he and his colleagues could create a more accurate diagnostic tool by looking at changes to the microbiome.
To do this, they collected saliva samples from 1175 people who were 50 years and older or had a history of tobacco use – both risk factors for these cancers. From each sample, they extracted genetic material from bacteria, fungi and skin cells.
The team used genetic data from 945 samples, 80 of which came from people with oral cancer and 12 came from people with throat cancer, to train a machine learning model. The model identified 88 distinct changes to gene expression in people with oral and throat cancer, as well as 182 genetic features unique to the bacteria found in samples from those people.
They then tested their model with the remaining 230 samples, 82 of which came from people with cancer. It accurately identified 90 per cent of samples from people with cancer and 95 per cent of samples from people without, meaning there were very low rates of false negatives or false positives.
The test they developed with this work is called CancerDetect and, based on preliminary data, it was given breakthrough device designation by the US Food and Drug Administration in April 2021. The provision grants expedited review to products that can improve treatment or diagnosis of life-threatening illnesses.
Under rules from the US Centers for Medicare & Medicaid Services, the test is now available for purchase in the US: people at high risk for oral or throat cancer can fill out a questionnaire with Viome, buy the test online and get results in around two weeks. The company will continue to pursue FDA approval, which, if granted, would mean the test would be covered by health insurance providers and more widely available.
Yet just how useful this test will be in the near term is unclear. Oral and throat cancer specialist Brett Miles at Northwell Health Cancer Institute in New York says he welcomes the idea in principle. “That you are going to visually inspect somebody and wait until [a lesion] is large enough to look suspicious is really archaic when we have all these technologies,” says Miles.
But he points out that the diagnostic tool was tested in quite a small number of people and that it doesn’t actually find cancer, only changes associated with cancer. “There is kind of a cause-and-effect question,” says Miles. “Are these bacteria changing because the cancer is there? Or are there changes in bacteria in certain people that then develop cancer later?”
There is another practical issue too, Miles says. Even if the test were 100 per cent accurate, what are doctors meant to do with that result? If a cancer is too small to detect visually, it cannot be biopsied and without a biopsy to confirm it is cancer, insurance companies won’t authorise treatments, he says.
However, Banavar says the technology will improve over time. “A major advantage of machine learning is that the more data we collect, the more accurate the tool will become,” he says.
More on these topics: