Philosophical Issues in Cancer Diagnosis and Treatment

Justin Biddle (Georgia Institute of Technology) - Cancer screening—or testing for cancer in the absence of symptoms—is the subject of much debate. Screening has the potential to save lives by identifying and treating cancers in early stages. However, not all cancers cause symptoms, and the diagnosis of these cancers can lead to unnecessary treatments and subsequent side-effects and complications. The debate over cancer screening is a part of a larger discussion about overdiagnosis of disease, and at the very least, the debate has highlighted the difficulties involved in balancing the risks of failing to treat against the risks of overdiagnosis (e.g., Welch et al. 2011). This paper will focus on the debate over prostate cancer screening. In 2016, the United States Preventative Services Task Force (USPSTF) published draft recommendations that men ages 55 to 69 “make an individualized decision about prostate cancer screening with their clinician” (USPSTF 2016). This recommendation is updated from 2012 guidelines recommending that no man of any age undergo screening. This shift in recommendations, which opens the door to more men getting screened, is based on a shift in values. The 2012 guidelines are paternalistic and based on the norm of beneficence; the USPSTF determined on the basis of a risk-benefit calculation that it was not in the best interests of men to undergo screening, and as such, it recommended against screening. The 2016 recommendations, by contrast, are based on the norm of respect for patient autonomy; they attempt to ensure that “each man has an opportunity to understand the potential benefits and harms of screening and to incorporate his values and preferences into his decision” (USPSTF 2016). Respect for patient autonomy requires, at a minimum, that doctors communicate clearly to patients the risks and benefits of treatment options. Drawing upon recent work on inductive and epistemic risk, I examine the processes of risk assessment in prostate cancer screening, and I argue that prostate cancer diagnosis is pervaded by epistemic risks that presuppose value judgments (Biddle and Kukla 2017). These risks include inductive risks and data formation risks—in particular, risks involved in the assignment of Gleason scores to biopsied cells. The pervasiveness of these epistemic risks creates significant and under-explored difficulties for physician-patient communication and, more generally, the achievement of autonomous patient decision making. I will examine the obstacles that must be overcome if genuine respect for patient autonomy is to be reached and argue that the obstacles are sufficiently high as to call into question the feasibility of the updated guidelines.

Bennett Holman (Underwood International College, Yonsei University) - Last year may be remembered as the pivotal point for artificial “deep learning” and medicine. A large number of different labs have used Artificial intelligence (AI) to augment some portion of medical practice, most notably in diagnosis and prognosis. I will first review the recent accomplishments of deep-learning AI in the medical field, including: the landmark work of Esteva et al. (2017) which showed that AI could learn to diagnose skin cancer better than a dermatologist; extensions of similar projects into detecting breast cancer (Liu et al., 2017); Oakden-Rayner et al.’s (2017) work showing AI could create its own ontological categories for patient risk; and through analyzing tumor DNA identify more possible sites for intervention (Wrzeszczynski et al., 2017). I will next argue that a forseeable progression of this technology is to begin automating treatment decisions. Whether this development is positive or negative depends on the specific details of who develops this technology and how it is used. I will not attempt to predict the future, but I will run out some emerging trends to their logical conclusions and identify some possible pitfalls of the gradual elimination of human judgment from medical practice. In particular some problems could become significantly worse. It is the essence of deep learning AI that reasons for its outcomes are opaque. Many researchers have shown that industry has been adept at causing confusion by advancing alternative narratives (e.g. Oreskes and Conway, 2010), but at the very least with traditional research there were assumptions that could, in principle, be assessed. With this deep learning AI there are no such luxuries. On the other hand, I will argue that properly implemented deep learning solves a number of pernicious problems with both the technical and the social hindrances to reliable medical judgments (e.g. the end to a necessary reliance on industry data). Given the multiple possible routes that such technology can take, I argue that consideration of how medical AI should develop is an issue that will not wait and thus demands immediate critical attention of philosophy of science in practice.

Sarah Wieten (Stanford University) - The popularity of direct-to-consumer genetic testing is on the rise. The largest and most visible of these companies, 23andMe, had been providing not only “ancestry” targeted testing, but “health” related testing with FDA approval since March 2018. In this talk I argue that there are linked epistemological and ethical issues with the method and content of the 23andMe-style direct-to-consumer genetic testing. I will focus on the case study of the development of 23andMe testing for BRCA1 and BRCA2 genes as indicating risk of breast cancer. These linked epistemic and ethical issues are pervasive and, in some cases, are beyond the scope of issues anticipated by the Bioethics Commission in 2013. For example, 23andMe is careful to say that their testing isn’t substantial medical screening, but consumers seem to treat it as if it is, making health decisions on the basis of incomplete information. Companies like 23and Me and their customers both seem to be operating under the impression that more knowledge is always better, while examples from pervious screening efforts suggest this need not be the case. Lastly, the high rate of false positives reported in 23andMe testing suggests an undefended asymmetry in the risk associated with Type I and Type II errors, with all the ethical difficulties that asymmetry has brought in other settings. While other “direct to consumer” companies have attempted to resolve this entanglement of epistemic and normative issues by suggesting that consumers “talk to their doctor,” individual clinician explanation is unlikely to solve this issue in the case of direct-to-consumer-genetic testing, given that many general practitioners are not trained to interpret the information provided by these tests.