Secondary Logo

Journal Logo

Why We Needn’t Fear the Machines

Opportunities for Medicine in a Machine Learning World

Li, David; Kulasegaram, Kulamakan, PhD; Hodges, Brian D., MD, PhD

doi: 10.1097/ACM.0000000000002661
Invited Commentaries
Free
SDC

Recently in medicine, the accuracy of machine learning models in predictive tasks has started to meet or exceed that of board-certified specialists. The ability to automate cognitive tasks using software has raised new questions about the future role of human physicians in health care. Emerging technologies can displace people from their jobs, forcing them to learn new skills, so it is clear that this looming challenge needs to be addressed by the medical education system. While current medical education seeks to prepare the next generation of physicians for a rapidly evolving health care landscape to meet the needs of the communities they serve, strategic decisions about disruptive technologies should be informed by a deeper investigation of how machine learning will function in the context of medicine. Understanding the purpose and strengths of machine learning elucidates its implications for the practice of medicine. An economic lens is used to analyze the interaction between physicians and machine learning. According to economic theory, competencies that are complementary to machine prediction will become more valuable in the future, while competencies that are substitutes for machine prediction will become less valuable. Applications of machine learning to highly specific cognitive tasks will increase the performance and value of health professionals, not replace them. To train physicians who are resilient in the face of potential labor market disruptions caused by emerging technologies, medical education must teach and nurture unique human abilities that give physicians a comparative advantage over computers.

D. Li is research assistant, University of Toronto, Toronto, Ontario, Canada.

K. Kulasegaram is scientist and assistant professor, Department of Family and Community Medicine, Wilson Centre, University Health Network, and Faculty of Medicine, University of Toronto, Toronto, Ontario, Canada.

B.D. Hodges is executive vice president and chief medical officer, University Health Network, and professor, Department of Psychiatry, University of Toronto, Toronto, Ontario, Canada.

Funding/Support: None reported.

Other disclosures: None reported.

Ethical approval: Reported as not applicable.

Correspondence should be addressed to David Li, University of Toronto, 27 King’s College Circle, Toronto, Ontario M5S 1A1 Canada; e-mail: dli294@uwo.ca.

Over the last five years, the accuracy of machine learning models in predictive tasks has surpassed that of human experts in multiple domains. In medicine, deep neural networks can perform as well as or better than board-certified specialists on cognitive tasks such as detecting diabetic retinopathy in retinal fundus photographs and classifying skin cancer based on images of skin lesions.1,2 Machine learning is fundamentally different from technologies like automatic blood pressure cuffs because it can be used to automate complex cognitive tasks that require tacit knowledge and experience. The possibility of significant changes to the role of physicians due to automation has generated substantial discussion and debate.3,4 Whenever there is excessive hype surrounding an emerging technology trend,5 it is important to elucidate the underlying mechanisms of the novel technology to understand its systemic implications.

Almost all of the progress related to machine learning in academia and industry over the last five years has been driven by a type of machine learning called supervised learning. Supervised learning algorithms are trained on massive labeled datasets to infer a function that predicts an output based on input data. In a similar bottom-up approach to learning, physicians develop efficient diagnostic reasoning by experiencing a large number of exemplar clinical cases. However, it is crucial to recognize that current machine learning models can only perform highly specific tasks and lack the generalist and perhaps adaptive competence that a human performing the same task would be expected to possess.6 With potentially disruptive machine learning technologies on the horizon, we must start preparing today by understanding which competencies will be valuable in the future and how we can best educate the next generation of physicians.

Given that disruptive technologies will force health care professionals to learn new skills, medical education must address this inevitable change. Forward-looking medical educators have already recognized the looming challenge.7 Several new curricular initiatives have been launched by Canadian and U.S. medical schools to address the competencies necessary for success in 21st-century medicine.8,9 Although these timely changes are steps in the right direction, strategic decisions about which competencies to focus on should be informed by a deeper analysis of the interactions between physicians and machine learning using an economic lens.

The impact of general purpose technologies throughout history has been studied extensively in economics and serves as a guide to predict how machine learning will change the practice of medicine. Machine learning can be defined as a general purpose technology that allows highly accurate predictions to be made for a very low cost. Economic theory tells us that the value of complements to an emerging technology like machine learning will increase, while the value of substitutes will decrease.10 The complement of machine prediction in clinical decision making is human judgment, the ability to evaluate the benefits and costs of potential treatments based on the patient’s broader context. While structured clinical metrics documented in electronic medical records are typically used to build machine learning models, unique human values and social determinants of health are difficult to model. In the future, physicians will need to combine this unstructured information with machine prediction to augment their human judgment and provide high-quality, patient-centered care. Human capabilities such as empathy and altruism will become more valuable, but the most valuable complement will be the one that no computer can ever replace: the human presence. Offloading highly specific routine tasks to automated technology will not make a physician’s complementary skills unnecessary; it will actually increase their importance and economic value.

Similarly, the value of substitutes for machine prediction will decrease. Societal demand for human predictions will decrease as machine learning models become more accurate and exceed the performance of physicians. Highly accurate machine prediction will augment a physician’s technical skills and allow him or her to spend more time on high-value activities such as forging meaningful relationships with patients. Physicians should specialize in skills and tasks where they maintain a comparative advantage over computers.11 The future value proposition of physicians will be driven primarily by their unique human abilities and not their diagnostic capabilities.

To clearly understand this framework, let’s conduct a thought experiment based on a real application of machine learning in medical practice. The University Health Network in Canada recently deployed a machine learning system that can generate radiation therapy plans based on patient data. It has comparable performance to plans generated by trained health professionals and reduces the planning time from hours to minutes. Currently, a radiation physicist will still review the treatment plan and make further modifications based on his or her judgment.

The complements to machine prediction in this case study are the health professional’s judgment and his or her ability to deliver patient-centered care. The ability to evaluate and modify computer-generated treatment plans to best fit the patient’s broader context will become a valuable skill. This can only be done by taking the time to understand a patient’s unique goals and values. By offloading the time and labor necessary to initially develop treatment plans, health professionals will have more resources to deliver high-quality care that patients want. Guiding patients through cancer treatment requires a high level of empathy and altruism. The substitute to machine prediction in this case—initial treatment plans created by humans—will decrease in value once the software becomes more accurate than the average health professional. Once machine-learning-driven systems are widely implemented, patients will want the technical expertise of artificial intelligence (AI) combined with patient-centered care delivered by empathetic health professionals. Machine learning will augment the performance and value of health professionals instead of replacing them.

From this case study, it is clear that the future role of physicians will likely be redefined as certain tasks become more automated. How does the current training paradigm prepare physicians for this reality? Likely, not very well—or at least, not yet. Discussion about how the workplace will be different is now more necessary than ever. It is also clear that the value proposition of a human physician needs to be defined clearly, and a new training paradigm must be developed in a deliberate manner. Medical schools must teach students cutting-edge medicine without losing sight of the human aspects of medicine. Curricula should focus on two core areas: improving human judgment and delivering patient-centered care. Courses on understanding data-driven decision making and statistics in a medical context should be implemented so that future physicians can apply new technologies appropriately. In addition, changes to the curriculum can help reinforce the role of empathy in the development of professional identity. Role modeling and positive reinforcement of compassionate behavior in clinical environments can help students nurture compassion into sustained practice.

These changes to medical school curricula should be mirrored in residency and continuing medical education/professional development. The former is necessary to reinforce to postgraduates in workplace-based settings the necessity of using AI efficiently and to enhance the care they provide for patients. The latter is necessary partly to ensure that the teachers and supervisors of future students are well prepared to train them for a new practice environment and partly to support the transitioning of the extant workforce. In both settings, new context-specific competencies for all types of AI and a renewed attention to integrating the humanistic aspects of clinical care will be necessary. In some sense, this is not a “new” concern as calls for returning to humanism have been around since the time of Osler and Flexner. But the rapid advance of technology makes the issue especially pressing in medical education today.

The emergence of machine learning presents a tremendous opportunity to transform the way medicine is taught and practiced. According to our analysis, applications of machine learning to highly specific cognitive tasks will increase the performance and value of health professionals. Medical education should teach and nurture unique human abilities that are complementary to emerging technologies. This will ensure that physicians are not displaced by technology. While other industries rush to emphasize machine learning at the expense of everything else, medicine must always emphasize the human element at the center of every patient–physician encounter.

Back to Top | Article Outline

References

1. Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316:2402–2410.
2. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542:115–118.
3. Verghese A, Shah NH, Harrington RA. What this computer needs is a physician: Humanism and artificial intelligence. JAMA. 2018;319:19–20.
4. Cabitza F, Rasoini R, Gensini GF. Unintended consequences of machine learning in medicine. JAMA. 2017;318:517–518.
5. Gartner. Gartner Hype Cycle. https://www.gartner.com/technology/research/methodologies/hype-cycle.jsp. Accessed January 18, 2019.
6. Mylopoulos M, Kulasegaram K, Woods NN. Developing the experts we need: Fostering adaptive expertise through education. J Eval Clin Pract. 2018;24:674–677.
7. Hodges BD. Learning from Dorothy Vaughan: Artificial intelligence and the health professions. Med Educ. 2018;52:11–13.
8. Kulasegaram K, Mylopoulos M, Tonin P, et al. The alignment imperative in curriculum renewal. Med Teach. 2018;40:443–448.
9. Skochelak SE, Stack SJ. Creating the medical schools of the future. Acad Med. 2017;92:16–19.
10. Agrawal A, Gans J, Goldfarb A. The simple economics of machine intelligence. Harv Bus Rev. November 17, 2016. https://hbr.org/2016/11/the-simple-economics-of-machine-intelligence. Accessed January 18, 2019.
11. Ricardo D. On the Principles of Political Economy and Taxation. 1821.3rd ed. London, UK: John Murray.
© 2019 by the Association of American Medical Colleges