Chapter 79
Cancer — 6
Mammography for breast cancer screening has been shown to result in false positives in about half of women over the course of a decade even though 88 percent of women will never develop breast cancer. There is also the serious problem of overdiagnosis, unnecessarily detecting and often treating nonthreatening breast cancer, occurring in about one in three women aged seventy to seventy-four years, and nearly half of women aged seventy-five to eighty-four years. Overall, one in seven women aged fifty to seventy-four years has a breast cancer overdiagnosis. A meta-analysis of randomized trials screening tests assessing mammography, colonoscopy, prostate-specific antigen, and lung CT scans, involving over 2.1 million individuals followed for a median of ten years, did not show any improvement in lifespan. Adding to all these challenges is the sharp increases in young people who are developing cancer, often presenting at late stages, who are underage for the cancer screening guidelines.
The premise of screening to detect early cancer, well before it has had a chance to spread, is good. What we have now, with about half of cancers presenting at advanced stages 3 or 4, with metastasis, will not enable a population-wide improvement in health span. Fortunately, we have ample ways to reboot how we screen people who show no symptoms.
The first step is to determine whether they are at high risk (fig. 6.4).
That is the foundation for close surveillance of an individual and promotes the earliest possible diagnosis, when cancer begins to manifest, well before it shows up on a scan. This can now be achieved at an increasingly precise level, using abundant data sources and multimodal AI to integrate the many layers of data (fig. 6.4) that are beyond our conventional clinical risk assessment—beyond merely considering a few factors like smoking or family history and coexisting conditions.
Figure 6.4. Prevention of cancer. A multimodal AI approach of identifying high-risk individuals through comprehensive assessment, with active surveillance of these people, Multicancer Early Detection (MCED), and aggressive preventive strategies.
Risk partitioning capitalizes on AI’s ability to “see” things that human experts can’t. That includes systematically inputting the content from a person’s electronic health records, which now, with generative AI, includes the unstructured text that often provides essential content. A person’s serial lab tests and medical scans may tell an important story when they were previously deemed normal but, when assessed longitudinally, indicate a trend in the wrong direction. Moreover, there are now hundreds of studies that have validated the power of AI to augment the accurate interpretation of scans. We can better avoid missing a lung nodule on a chest X-ray. Machine eyes can pick up abnormalities that expert radiologists wouldn’t even look for, like in the pancreas on a chest CT scan. The genetic data, either in the form of polygenic risk scores or a whole genome sequence (from which these risk scores can be derived), provides key, independent information, as do emerging biomarker blood tests.