A lot of ink has been spilt over the last 10+ years on the potential impact AI will have in healthcare. In fact, it has been over a decade since it was reported that a prominent venture capitalist claimed that “Machines will replace 80% of doctors” because of AI. It turned out what he meant was “80% of what a doctor does can eventually be done by machines.” In either case, the point was AI was going to have an enormous impact on how physicians practice medicine. On the other hand, since 1995 when the Food and Drug Administration (FDA) approved the first AI algorithm to detect abnormal cells in pap-smear tests, the FDA has approved over 500 AI algorithms for clinical use. However, despite all theoretical predictions of the profession of radiology's impending demise because most approved algorithms are in radiology, this has not occurred. I should know because an interventional radiologist saved my life, and there is no way any AI algorithm could have done what he did. Period. I believe we can all agree on one point: AI advances in medicine over the last decade have not had the impact on medical practice that some predicted. There are two primary reasons for this:
1. The change management that healthcare provider organizations must engage in to successfully implement and benefit from AI is poorly understood. It is not as simple as saying, “Let me use this AI algorithm so I can fire physicians and become more efficient.”
2. There is much more to a physician's job than simply examining an image and diagnosing a disease; it requires a great deal of additional skills and the human touch.
That is not to say that AI cannot impact the practice of healthcare. However, to have that impact, we need to take a fundamentally multi-disciplinary approach, that accounts for the perspectives of clinical and non-clinical people in healthcare, businesspeople building companies to solve these problems economically, technologists building modern technologies and solutions, and possibly others. That is what we hope to explore in ThetaRho’s blog as we navigate this multi-dimensional universe. Let’s begin.
1. From 7.2% of the national GDP in 1972 to 18.3% in 2021, healthcare expenditures have steadily increased. However, this increase in healthcare costs has not been accompanied by an improvement in health outcomes; in fact, many population level health indicators have worsened.
2. At an individual level, the experience is worsening as measured by wait times to get appointments and wait times to meet physicians during the appointment.
Healthcare suffers from Baumol's cost disease meaning cost of labor has been rising without a concomitant increase in productivity as measured by economists. Although there is a lot of excitement in using AI for clinical use cases to increase standardization and hence productivity in healthcare. To cure Baumol’s disease, we believe that focusing on non-diagnostic, clerical and administrative use cases can provide a better starting point with greater leverage and generalizability. Even if a fraction of administrative process improvements is applicable across the board, and can be automated with technology, that will result in significant gains in productivity.
At the macro level, although a McKinsey report identified approximately $250 billion in potential administrative expense savings, not all problems are accessible to small businesses. Some problems can be addressed within a single organization (payer or provider), while others require collaboration between organizations (payers and providers, for example), and still others need systemic changes. Brookings Institution's Hamilton Project research found that most administrative costs are associated with healthcare delivery and are borne by the providers. Therefore, we will focus on what can be done for healthcare providers.
At the micro level, looking at individual healthcare provider organizations and their cost structure as reported in their financial statements, the single largest cost is salary for employees. It can easily be 35-40% of revenue for most healthcare providers. Among these employees, the most expensive employees are the clinicians, specifically physicians. Here is where the story gets complicated because clinicians are experiencing burn-out at unprecedented rates, and attrition is high due to clerical and administrative work prior to, during, and after patient encounters. This should terrify healthcare providers because it puts pressure on both the top and bottom lines, squeezing the margin from both sides.
Some studies have shown that physicians spend twice as much time on clerical and administrative tasks as they do on direct patient interactions. If the most expensive resources are spending more time doing clerical and administrative work, the macro measurement of administrative costs in healthcare making up 25-30% of total cost appears to be significant underestimate. One would have expected the opposite, with physicians spending twice as much time with patients as opposed to on electronic health record (EHR) systems doing clerical tasks. AI can help make this happen, by assisting and automating 80% of the tasks currently performed by physicians on EHR systems. Clinicians we have spoken with appreciate this assistance because it allows them to be clinicians rather than data entry clerks. So, if the tasks being automated are clerical or administrative in nature, the venture capitalist mentioned earlier may very well be correct.
1. The technology did not perform as expected. Indeed, there is some evidence for this. Data in even a single EHR system is highly fragmented and difficult to access from the perspective of a physician, let alone data across multiple EHR systems, and physicians frequently need to access multiple systems to put together a comprehensive description of a patient.
2. Rent seeking by incumbents with “market power.” There is some evidence for this as well. According to an article in the New England Journal of Medicine, EHR vendors have made these systems more complex than necessary, hiding behind regulations and spreading fear, uncertainty, and doubt about new solutions to limit innovation.
1. Exactly what processes can be improved or automated by modern AI so that they can be adopted easily? Is it before, during, or after encounters?
2. How can these inherently probabilistic AI tools be enhanced to be repeatable, truthful, reliable, and trustworthy and compliant with all the requirements that businesses in highly regulated industries, such as healthcare, must meet.
We draw inspiration from companies like ServiceNow that achieved success by tackling process improvement issues in enterprise functions like Information Technology (IT) by focusing relentlessly on the user experience. This is the mission we've undertaken at ThetaRho, and you'll learn about our journey across these dimensions in this blog.