From the popular press to the largest health care conferences, promises of artificial intelligence revolutionizing biomedicine are ubiquitous. It often seems as if we are on the cusp of AI systems that can remotely identify a person about to get sick, make a diagnosis (no doctor needed!), select a custom AI-designed pharmaceutical and deliver it to the patient just in time—in an AI-powered self-driving car, of course.
If indeed this is the future, we are far from reaching it. To be sure, the pace of change has been rapid. Deep learning—the fast-growing subfield of AI that enables machines to diagnose pneumonia from chest x-rays or predict health deterioration from medical records—was unfamiliar even to most computer scientists a decade ago. And we do not know what evolutionary or revolutionary advances will drive AI in the coming decades. What we do know is that the success of biomedical AI depends not just on developing the technology but also on developing the people behind it.
Translating algorithmic advances to biomedical breakthroughs requires critically considering both realms of knowledge and endeavor on many levels. What, for example, are the true capabilities of a new technology, and what is simply hype? What problems in biomedicine are most likely to benefit from emerging computational capabilities? And how do we go from an interesting biomedical application of a new technology to the implementation of systems that actually improve human health? These challenging, multifaceted questions will need to be answered by interdisciplinary teams. The teams will require experts in AI, experts in biology and medicine, and, most important, leaders who can motivate and guide individuals with such diverse talents.
Unlike some domains in which AI has been applied, in biomedicine the consequences of failure are weighty. For a social media company, an AI model that is ineffective at increasing ad clicks can be detected and rolled back the same day. When it comes to medicine, however, human lives are at stake. Inadequately informed uses of AI can lead to obvious harm, such as inaccurate diagnostic or therapeutic recommendations, but also to more insidious failures, such as an algorithm that gives racially biased recommendations because it was trained with subtly biased data. Given the complexities of biomedicine and the inscrutable nature of many AI algorithms, it might be years before such a flaw is uncovered. Group leaders—whether in academia, pharmaceutical laboratories or start-ups—must not only understand the technical and scientific issues but also anticipate and articulate the potential risks, benefits and implications of the projects they undertake.
We need men and women who can build AI systems in medicine that improve care. It is relatively easy to generate excitement by solving the technical aspects of a problem, but making those advances useful often involves wrestling with the complex interplay of regulatory, economic and workflow issues in health care systems. Successful leaders benefit from deep knowledge and intuition in both the AI and the biomedical domains. But we face a critical shortage of such versatile individuals.
Tackling this gap is crucial to ensuring the long-term success of biomedical AI. A primary challenge is the length of study required in these disciplines, but a greater one is training students in two realms that could hardly be more different in their approaches to problem-solving. Computer science involves the quantitative rigor of mathematics, statistics and engineering, whereas biology is underpinned by the haphazard products of evolution. Properties of living things are, literally and figuratively, organic. We seek students with the intellectual flexibility and passion to undergo lengthy training in both these contrasting cultures. Are we asking for the impossible?
These individuals do exist, and their numbers are growing. The first approach to their training is to identify individuals who already have a deep background in either biomedical or computational science and then help them become skilled in the other area. Graduate programs (M.S., Ph.D. and M.D./Ph.D.) in biomedical informatics have filled this role since the early 1980s. These programs attract diverse students and have grown to include disciplines that go by various names: computational biology, bioinformatics, clinical informatics, biomedical data science, and so on. All are concerned with different applications of computer science to biomedicine.
But what about training students at the intersection of these disciplines even earlier in their careers—while their intellectual intuitions are still forming? The difference would be like that between learning a second language as an adult and growing up in a bilingual household: fluency is second nature for early starters.
In 2001 we launched an engineering major at Stanford University to enable undergraduates to learn computer science and statistics in the context of biology and medicine. The program creates graduates with a bachelor of science degree who have already wrestled intensively with the challenges of applying computational tools to hard problems in biomedicine. Our students take biology with premedical students and computer science with classmates who will work in Silicon Valley, and each completes a two- or three-quarter-long research project during his or her time at Stanford. They acquire knowledge with breadth across the biomedical and technical fields and depth in a narrower application area. At least one course on the societal and ethical implications of technology is also required.
After almost two decades of training biomedical-computation undergraduates, we can say that the model works. Many of our graduates have gone on to careers in academia, clinical medicine, start-up companies (both in and outside of the biology field), large companies, law firms, venture capital, and elsewhere. And the major has consistently drawn a 50–50 balance of men and women—true for only a minority of quantitatively intensive engineering majors.
For most, the major has shaped their professional identity: they are not “AI people doing bio” or “Bio people doing AI.” Instead both of these intellectual traditions reside comfortably within their minds, each informing their understanding of the other. Whereas it is impossible to learn the entirety of biomedicine and computer science in just four years (or even in 40), these people move freely between the cultures of biology and computer science and have already learned to apply deep technical skills to the hardest societal challenges in biology and human health.
In addition to graduate programs, the development of a robust set of undergraduate programs at the interface of biomedicine and computation could give students who are in a formative period of their education the ability to move fluidly between these very different disciplines. Such programs would accelerate the emergence of the workforce required for appropriate use of AI to advance biology and health care.