Chapter 83
Cancer — 10
Now let’s consider how AI is making a difference for breast cancer. Recall that only 12 percent of women in the United States will ever manifest breast cancer in their lifetime. If we could determine who is at risk, it could save the vast majority frequent screening, false positives and the follow-up scans and biopsies they engender, no less the angst and cost that arises from them, and overdiagnosis.
One way that has been pursued is partitioning based on the mammogram, for which AI can provide important prognostic information for five-year risk. That brings up another key point. Just knowing a person is at risk during a lifetime isn’t enough—we also want to know more about the likelihood of when a cancer will first manifest. A mammographic texture model assessed in over 280,000 screenings in the Netherlands and Denmark found the highest 10 percent risk of women developed about 40 percent of the breast cancer cases during follow-up. A mammography AI model assessed in a four-country case-control study in Europe identified women at high risk who were told their mammogram was normal. In a study conducted at four centers in three countries, the United States (Mass General and Emory), Sweden (Karolinksa), and Taiwan (Chang Gung), a mammography AI-based model for detecting future risk was far more efficient—as defined by reduction in number of mammograms and earlier detection—than routine, age-based screening.
But just pinning risk stratification on the mammogram or, if performed, ultrasound or MRI, is only one dimension. Other layers of data would include lifestyle+ (obesity, alcohol consumption, exercise), polygenic risk score, genome sequencing, reproductive factors such as age at menarche, age at menopause, gravidity, parity, and age at first birth, and breastfeeding. Such a multimodal model has not been prospectively assessed, but it would be expected to help partition low, intermediate, and high risk and supersede age-based mass screening in the future. Many prospective clinical trials of tailored screening are in progress.
TESTS OF THE TINIEST
While the multiple layers of data can certainly help to pick up people at high risk for cancer and develop smarter screening strategies, tests that directly measure microscopic evidence of cancer in an individual have drawn intense interest. The imprecise term liquid biopsy has frequently been used but isn’t quite right. Various tests seek evidence of cancer such as cell-free tumor DNA in the blood plasma, fragmentation of tumor DNA, methylation of DNA, circulating tumor cells in the blood, circulating exosomes, proteins, metabolites, cell-free tumor DNA in the urine or stool, cell-free RNA tests, repeat DNA elements, and alterations of T cell receptors. Such Multicancer Early Detection (MCED) assays are constantly getting refined. For example, the amount of cell-free tumor DNA in a blood sample is very limited, but priming agents can be added to reduce its clearance and improve accuracy of testing. Deep-learning AI has been shown to overcome the low-quantity problem with a radical signal-to-noise enrichment, making these assays ultrasensitive. One of the most difficult cancers to diagnose early is ovarian, but AI of DNA methylation markers has been shown to improve on our current assessment.