Maurie Markman, MD
There is no debate regarding the spectacular role that technology has played and continues to play in the evolution of our understanding of the fundamental molecular biology of cancer and its development, progression, and resistance to currently available therapeutics. Further, the ability to sequence and subsequently interrogate the genome of an individual patient with cancer (somatic or germline) has impressively advanced the paradigm-changing concept of precision cancer medicine, whereby therapy will increasingly be based on documented molecular characteristics within an individual’s cancer or germline profile rather than simply on the site of origin or the histologic appearance of the malignancy.
Understandably, the recent technological advances in medicine and related fields (eg, artificial intelligence [AI]) have encouraged a belief among many that there is little technology will not be able to accomplish in improving cancer-related clinical outcomes. But in this intense rush to move innovative technology into the clinic, it must be acknowledged that clinical medicine, in general, and cancer biology, in particular, are extremely complex arenas.
Success in improving patient welfare will necessitate considerably greater understanding, deliberation, and regulatory oversight than required, for example, in the development of algorithms that successfully increase sales of products available on the internet. Consider for a moment a physician-authored editorial that recently appeared in the New York Times describing the potential problems with efforts to develop AI in healthcare.1
As noted by the author, there remain serious deficiencies in the overall quality of health-related data for substantial population subgroups in our society. Poorer outcomes ascribed to a specific characteristic (eg, being a member of a minority) may in fact be due to an entirely different cause (eg, socioeconomic status), leading to profound implications for any conclusions to be drawn from the data. As poignantly stated in the article, “in a health system riddled with inequity, biases could become automated.”1
An unrelated recently published paper in a high-impact medical journal made a similar point regarding the potential dangers of “machine learning” and provided concrete suggestions for how AI bias can, hopefully, be avoided.2
In my opinion, this article should be required reading for all attempting to develop objectively valid AI algorithms in clinical medicine. In the arena of cancer, we are witness to a concerning blind faith in the power of innovative technology to routinely make a very early diagnosis of the presence of malignant disease. In fact, it is reasonable to state that in general, patients who present with the earliest stages of cancer experience the greatest opportunity for cure and extended survival.
Therefore, why not attempt to detect cancer in the blood before any other signs (eg, abnormalities on imaging studies) or symptoms of malignant disease develop? Today, investigators are actively examining the potential for a “simple blood test” to find cancer at the earliest possible point in its development.3,4
There several concerns with this basic concept, including how clinicians would go about finding the actual anatomic location of the cancer in the absence of abnormalities detected by specific symptoms, physical examination, or, much more likely, imaging studies. Add to this the anxiety a patient would feel upon learning of a diagnosis of cancer whose location is unknown.
How often should further diagnostic testing or even invasive procedures be performed if physicians will not know when or even if a tumor/mass will ultimately appear? But the specific point is to highlight the almost certainly vastly oversimplified assumption that a cancer-specific signature exists at the molecular level such that it can be stated that a patient has a malignancy and that an active search can be initiated and continued until the location of the disease is finally discovered.
Take the following, for example: In an exome sequencing examination of self-declared healthy adults (N = 7974), 1.4% of the population was found to have a molecular variant that has been observed in individuals with genitourinary disorders, a percentage far in excess of the predicted incidence of such disease. This strongly underscores the lack of utility and the actual danger associated with employing such testing in any type of population-based screening.5