Health IT,Tech AI is Already Substituting for Physicians, Yet Not in the Anticipated Manner

AI is Already Substituting for Physicians, Yet Not in the Anticipated Manner

AI is Already Substituting for Physicians, Yet Not in the Anticipated Manner

When I was a young, charming lad of merely 16, I was fortunate enough to arrange a meeting with Sandy Napel, PhD at Stanford University’s Radiological Sciences Laboratory (RSL). I showcased a prototype algorithm that automatically pinpointed the location and trajectory of arteries on CT scans. This was the initial step in a nearly ten-year-long academic research journey that featured early publications on utilizing neural networks (the forerunners of today’s LLMs). At that time, I was asked, “Will AI take over the role of doctors?” My response would have been a resounding NO. Humans still had to decode the “how” and “teach” the machine. Even with the early neural networks, extensive data engineering was essential before inputting it into the system for classification.

For numerous years, “classification,” the science (and art) of enabling a computer to identify a cat in an image, was standard practice. A considerable amount of human effort was required for each model to function. Successful companies like Hologic and R2 Technologies developed computer-aided diagnosis (CAD) systems for breast imaging to identify cancers. As many radiologists will affirm, these tools were merely “adequate.” As recently as 2015, well-structured studies determined that AI provided no overall benefit to outcomes and that it featured a $400 million debacle for the U.S. healthcare system. Essentially, the decision to purchase a CAD system was driven by the ability to charge for it rather than any tangible real-world benefit.

A decade later, deep learning neural networks and LLMs, along with a significantly larger pool of training data, can perform tasks without human instruction. Even a beginner can utilize and train an AI to suit their requirements. The rate of advancement has surged dramatically, as have AI’s capabilities. Reinforcement learning is reviving with LLMs, but it’s aimed at encouraging better reasoning, which is a progression beyond merely answering a question correctly.

So, we pose the question again. Will AI ultimately usurp the role of doctors? The answer is YES, but maybe not for the reasons you might expect.

To grasp why, we must first recognize that the characteristic that distinguishes humans as a species is our individual uniqueness. If we attempted to forecast the function of every neuron in our brain at any given moment, we would likely still be doing it when the universe comes to an end. It’s not exactly Heisenberg’s uncertainty principle, but it approaches it. Nevertheless, we have been laboring for centuries to suppress uniqueness. Everything from religion to our educational systems imparts a specific set of knowledge and thought processes. In healthcare, we are increasingly conforming to a particular set of guidelines that claim to enhance population health based on the best available data, frequently overlooking subpopulation and individual differences.

You may wonder: Why is this problematic? Shouldn’t we strive to enhance population health overall? Yes, I completely concur. However, when you visit the doctor with a cough, you don’t go there expecting to be diagnosed with a cold. You go because you suspect the cough might signify something more serious or uncommon. You are unaware of what you don’t know. That’s why you seek a doctor’s expertise. AI cannot (yet) envision what it lacks knowledge of because it lacks creativity. A solely AI-driven approach will inevitably lead to the stagnation of knowledge and experience. Humans inject diversity and creativity, propelling advancement.

Here’s the twist. In a sense, the substitution has already initiated.

Hospitals, provider organizations, and now insurance firms have posed the fundamental question: If all we do is adhere to a defined set of guidelines and 90 percent of our patients present with common ailments, who requires experience and knowledge? Moreover, in both the fee-for-service and “value-based” capitation models, minimizing costs and maximizing profit per patient is crucial. Radiologists review 80–100 scans daily without AI assistance, and primary care providers spend