New article: AI ethics is not a panacea. McLennan S, Lee MM, Fiske A, Celi LA
From machine learning (ML) and computer vision to robotics and natural language processing, the application of data science and artificial intelligence (AI) is expected to transform health care (Celi et al. 2019). While the rapid development of technological capabilities offers paths toward new discoveries and large-scale analysis, numerous critical ethical issues have been identified, spanning privacy, data protection, transparency and explainability, responsibility, and bias.
Last year, for instance, a commercial prediction algorithm affecting millions of patients was shown to exhibit significant racial bias, dramatically underestimating the health needs of Black patients (Obermeyer et al. 2019). Trained using health care cost as the proxy for the need for more comprehensive care, the algorithm had been designed specifically to exclude race as a feature, in an attempt to avoid bias—but cost was clearly not a race-neutral measure of health care need. Studies have repeatedly illuminated racial disparities in the provision of primary care services: Black patients incur approximately US$1800 less in medical costs per year compared to white patients with the same number of chronic conditions, and are less likely to be identified as high-risk for complex care in the future. But even if another proxy, such as probability of death, had been used to train the algorithm, would it have led to a “better” algorithm and fair patient outcomes?
At present, a key evaluation metric for machine learning in health care applications is accuracy. To inspect an algorithm for bias, an additional step is often undertaken to measure performance across different subpopulations, aiming for consistent accuracy across race, gender, country, and other categories where disparities exist. But just because an algorithm is deemed accurate does not mean it will support fairness in health care applications. In an ideal world, only individual patient health and disease factors would determine—and guide prediction of—clinical outcomes. However, studies have repeatedly demonstrated that this is far from the case. For example, mortality from critical illness has been shown to be higher in disproportionately minority-serving hospitals after adjustment for illness severity and other biological factors that pertain to the patient and to the disease (Rush et al. 2020; Danziger et al. 2020).
Data routinely collected in the process of care are heavily influenced by long-standing social, cultural, and institutional biases. Unless the underlying inequities in our communities are addressed, algorithms will perpetuate, if not magnify, existing health disparities.
The bioethics community has an essential role in providing thoughtful, ethical consideration of machine learning health care applications (ML-HCAs), and the model pipeline framework proposed by Char and colleagues (2020) is an important step forward in promoting the systematic identification of relevant ethical concerns. However, it would be misguided to think that an insufficient consideration of ethics is the sole factor leading to poorly designed systems that harm users (Mittelstadt 2019). Indeed, despite a wide range of AI ethics frameworks and principles being published in recent years, the challenges of building ethical AI are as acute and pervasive as ever. It remains unclear whether the influx of guidelines has actually made any impact on improving the ethical development and implementation of AI.
As the bioethics community increasingly turns its attention to ML-HCAs, it is therefore crucial for individuals and organizations to think beyond narrow definitions of machine learning or specialized disciplines, and to engage in addressing disparities with a holistic perspective. ML-HCAs exist in complex societal contexts, rife with biases, disparities, and ethical issues—requiring deeper commitment than a general adoption of ethical frameworks and principles. Efforts that do not actively address the disparities they are informed by, and within which they operate, will always fall short (500 Women Scientists Leadership 2020). As Timnit Gebru has noted, it is not just a matter of biased inputs leading to biased outputs. Ethical considerations of fairness or bias need to be considered beyond any given ML or AI application (Gebru and Denton 2020). In a health care setting for the benefit of patients, this means stakeholders including bioethicists must intentionally extend the typical range of concerns considered when thinking about what constitutes ethical or fair AI. As Chen and colleagues note:
Researchers often frame theoretical problems of disparity around achieving algorithmic fairness. Looking forward, these conversations should be expanded to acknowledge the systematic dimension of health disparity, taking into account that data is collected in the context of a flawed and unjust system. The research community itself should continue to promote and drive diversity within the field of AI, as more diverse perspectives will ensure that the right questions are asked. (Chen et al. 2020, 16)
A model pipeline framework such as the one proposed by Char and colleagues can certainly help to design better ML-HCAs, but only if the scope of ethical deliberations is attuned to matters of structural inequity and how technologies can be used and co-opted to marginalize subpopulations.
With the scale of patient impacts at stake, the AI community must recognize the call to action, for instance, in hiring, building, and supporting more diverse teams of developers, engineers, and research scientists. For the AI community and the field of bioethics, conversations around how to address persistent inequities in relation to both AI development and clinical consequences are deeply needed—and so too are scholarship and discussion focused on topics including social determinants of health, racial and social justice, LGBTQ+ ethics, disability ethics, and factors that contribute to vulnerable populations. As Keisha Ray recently noted in an opinion piece on #BlackBioethics, the field’s bias in this respect represents a failure to address critical needs and scholarship (Ray 2020). For the worlds of both AI and bioethics to enable the change needed in modern health care, it will be essential to co-design approaches with those in marginalized communities who have experienced harmful effects of technologies (Technology can’t fix this 2020).
The model pipeline framework is a solid step forward for AI ethics in health care, but ultimately frameworks are never enough. If we want to make sure that AI applications in health care contribute to a society we want to live in, we each have to ensure we are building that society—both in AI ethics and beyond.