Vinod Khosla was right: Machine learning/algorithms can replace plenty of what physicians do

In Medicine and Machine Learning by H. Jack West, MD0 Comments

Speaking in September, 2012, technology expert and Silicon Valley entrepreneur/venture capitalist Vinod Khosla made the bold prediction that computer algorithms could replace 80% of physicians.  Coming on the heels of the excitement of IBM’s Watson, a supercomputer of sorts that decisively vanquished the storied human champions of Jeopardy! in 2011 and was promptly turning to applications in medicine, there was a mixture of equal parts optimism and uncertainty in the air. Watson was honing its skills in a partnership between IBM and Memorial Sloan Kettering Cancer Center, and I took the view that no algorithm or even the most sophisticated collection of algorithms could mimic the work I do.  I believe that my hubris was a mistake, and Vinod Khosla was right: even though automation can’t, shouldn’t, and won’t replace everything that physicians in general and oncologists in particular do, they can do a great deal.  More importantly, they can do it better, faster, without risk of burnout or time off for family vacations, and no demand for health care benefits.  It is not only possible but inevitable that automation will consolidate much of health care.

artificial-intelligenceIn fact, there are many contributing factors toward my mistaken thought process. First, though I work as a subspecialist and can identify many times throughout the day in which a patient’s care requires integration of many complex challenges in order to develop an optimal plan or strong list of management options, this actually amounts to a regrettably small proportion of my day. Instead, much of the rest of a busy clinic involves copious paperwork, writing and signing routine orders, clicking through 20 boxes to order one simple medication, and overseeing the routine pass-throughs of many patients whose visits are relatively perfunctory but are expected and contribute to the gaping maw of our system’s requirements for well productivity, the engine of payment.  We have cultivated an inefficient volume-based health care system that is wasteful of training and skills because it burdens the most expensive and highly trained physicians with inexhaustible paperwork and electronic busywork, while also financially rewarding the blessing of a revolving door of patients who require low complexity of management input.  Even the best-trained subspecialists spend far more time clicking boxes on the EMR and waving patients through ongoing treatment based on threshold lab values than performing any higher cortical processing that might pose a challenge to even a basic collection of algorithms, let alone a Jeopardy! champion level supercomputer.

Another critical point I needed to recognize is that even mundane medicine as practiced by your friendly neighborhood human doctor is best directed by guidelines and algorithms. This takes several forms, starting with the basic “Checklist Manifesto”, in which the Rhodes Scholarship and Emmy Award-winning surgeon Atul Gawande enumerates the evidence supporting better care by following basic checklists in the operating room and ICU.  Increasingly, widely respected national and international guidelines define national and even global standards of care for everything from vaccinations to cancer screening and treatment.  The National Comprehensive Cancer Network provides an emerging consensus view of the most appropriate steps of cancer care from initial workup and molecular marker testing to how best to integrate local and systemic therapy options.  These guidelines define expectations for what insurers should be expected to cover and how our practitioners should measure the quality of care we deliver.  These and sometimes other guidelines from widely respected sources such as professional societies and UpToDate, as well as institutional or network-based “pathways”, have converged to shape our sense of a gold standard for treatment. And they are are enumerated in the form of clearly defined algorithms, ideally suited for navigation by “Dr. A”, as Vinod Khosla refers to algorithm-based medicine.

But even beyond automated algorithms working to justify a place alongside experienced humans who can follow ever-proliferating guidelines, artificial intelligence-based oversight will soon become a highly desirable if not required safety net as we contend with the uncomfortable realization that even the brightest and most motivated humans cannot keep up with the torrent of new clinically relevant information that is shaping cancer care.  The volume of new data being published in the over 200 oncology journals today (a list that continues to grow every few months) can only be processed by the voracious capacity of computers that don’t max out with the memory or attention capacity of a human brain. And with the complexity and data richness of “precision medicine” ushering in exponentially greater genomic data, molecularly targeted therapies, and clinical trial options, we are on the cusp of an era in which cognitive processing unencumbered by the limitations of a brain that operates at the limits of what the biological world has developed.

Finally, there is the mundane but critical argument in favor of the economic attractiveness of automation that can work in tandem with health care personnel from medical assistants to nurses, physician assistants, as well as primary care and specialists physicians.  Incorporating technology in our clinical work flow can provide both the backstop of essentially infinite knowledge base with the indefatigable ability to perform the rote tasks that represent a wasteful diversion of time better spent by humans on tasks that actually justify the talents that remain the province of human brains. A model of effectively implemented automation integrated with and complementing human care is also one in which it is true that fewer humans are required, but those who are can be fully engaged and operating at a level in which they spend far more of their time appropriately challenged and performing activities that resonate with their talents rather than pursuing drudgery far below their skill set and pay grade.

I believe that the biggest issue limiting use of algorithm-based medicine in directing healthcare won’t be its abilities, but rather the regulatory environment and human patient attitudes about how critical it is to have a human present with their hands on the wheel, even if artificial intelligence is a better driver.  There should always be a role for humans in the personal connection and discussion about complex, emotionally charged issues that pervade medicine, but I believe that if we actually seek and evaluate the evidence fairly, we’ll find that algorithms can practice medicine better than humans in far more settings than physicians might want to recognize.

Leave a Reply