Artificial intelligence: Looking beyond the hype

©metamorworks/stock.adobe.com 

What will the doctor of the future look like?

With new advances in the field of artificial intelligence (AI)-as well as the publication of a handful of high-profile studies demonstrating that AI systems can offer diagnostic capabilities surpassing human doctors in specific cases-many believe the next generation of physicians will have their healthcare capabilities enhanced by these special computer algorithms. Experts believe that these systems will help relieve physicians’ burdens, increase interoperability among different information technology systems, expand patient access to specialty care, and improve treatment plans for chronic conditions.  

Vimla Patel, PhD, director of the New York Academy of Medicine’s Center for Cognitive Studies in Medicine and Public Health, says the idea that AI could help solve many of healthcare’s biggest challenges should not come as a surprise. Modern computer programs are very good at sifting through vast quantities of data. And between electronic health records (EHRs), genetic sequencing, and mobile health data, more and more aspects of healthcare, from diagnosis to clinical decision-making, is becoming data-driven.

“Doctors are human beings. Human beings have limited memory and cognitive resources to apply to what can be very complex medical problems,” she says. “And given that we are overburdening physicians with so many other things, having the extra help from computers to take on that complexity can be a very good thing for patients-and for the doctors themselves.”

But despite growing enthusiasm for AI, can the reality live up to the hype?

Enhancing physician capabilities

With a 2017 Mayo Clinic study suggesting that as many as 88 percent of patients may be misdiagnosed during a first clinical encounter, AI algorithms offer the possibility that doctors could better detect medical conditions in their earliest stages-and recommend the right interventions for treating them.  

Joshua Denny, MD, MS, an internist and medical informatician at Vanderbilt University Medical Center, says that AI programs could help free physicians from medical “busy work” by taking on the rote review of medical images and symptoms, allowing physicians to focus on more complex medical decision-making.  

“I don’t know any physicians who suffer from too little work: so being able to reduce some of the volume of run-of-the-mill stuff so they can focus on more challenging cases-focusing on diversity over volume-I think that would be a good thing,” he says. “And something that would likely result in better patient outcomes.”

In addition, Denny says, AI may offer patients improved access to care. It offers the possibility of performing initial scans for disease more easily and affordably, so instead of having to hunt down referrals or book appointments with a specialist who may be out of network (or even out of town), patients can simply check in with their primary care provider. He says he can imagine a future where, for example, a visual scan of the retina for diabetic retinopathy or a skin cancer check could be done by machines housed in a local chain pharmacy or big box store, much like the blood pressure kiosks we see today.

“AI could reduce some of the health disparities we see, with patients unable to get to or afford a specialist,” he says.

Experts also think that AI systems could help manage one of healthcare IT’s greatest plagues: the lack of  interoperability, or the ability for different medical systems to communicate seamlessly with one another.

Such programs could mine EHRs and health information exchanges for structured data, enabling better communication between physicians at different facilities, decreasing the amount of time doctors have to spend filling in missing patient information in their EHRs, and providing the data required to better inform population health initiatives.

Put those possibilities together, and healthcare organizations can realize new opportunities to streamline  care and better inform treatment selection says Isaac Kohane, MD, PhD, chair of biomedical informatics at Harvard Medical School. Currently, two patients may have the same diagnosis, yet often require different treatment regimens.  As AI programs get better at stratifying patients, especially with more genetic data becoming available, they can also help predict who will respond best to a particular treatment regimen and better inform best practices for managing chronic health conditions. Kohane believes both physicians and patients will benefit as a result.

“Primary care physicians can have all these specialty software programs at their fingertips, providing  them the ability to offer patients specialty care in their own office,” he says. AI applications could cut down on unnecessary referrals to specialists, reducing costs and the risk of repeat tests and procedures. It could also help doctors prescribe the right treatment to each patient the first time-without having to rely on trial and error.

“There are algorithms that can tell you whether a patient is taking their medicine or not,” says Kohane. “So many of these programs can help clarify what is really going on, giving physicians the information they need to make good decisions, avoiding unnecessary changes to more expensive drugs that may have worse side effects.” 

These programs also can offer significant financial benefits to healthcare organizations. Administrative tasks are filled with inefficiencies that could be mitigated with AI algorithms.

Kohane says there are many ways AI could be used outside actual medical care, whether it’s doing automatic eligibility checks, identifying errors in the claims process, or helping to optimize revenue cycle management. If these programs have access to the pertinent data, they can offer significant savings, highlighting ways to boost an organization’s bottom line in areas that hospital administrators might not even think to look.

But is it too good to be true?

While AI offers the possibility of enhancing physicians’ capabilities and streamline business operations,  there are some important caveats. First and foremost, these algorithms are only as good as the data that go into them.  Kohane cautions that AI programs are often trained on very structured data.

Once they go into more widespread use, any data that isn’t as structured may skew the results, whether it’s an AI system used to mine medical information systems for interoperability purposes, a diagnostic tool analyzing a medical image, or a back-office application sifting through claims. While doctors often bristle at being limited to the structured fields in their EHRs, preferring unstructured data to document cases in more natural language or add notes, by doing so they may inadvertently be impeding the interoperability they seek.

“Once you introduce increased diversity of data, data that’s beyond what the system trained on, it becomes more likely that these programs will be fooled,” he says.  

A second limitation is the data or image features AI systems use to predict specific medical conditions. Sometimes these algorithms pick up on unexpected data points that ultimately invalidate the results. Denny cites the example of a machine learning program he was working on to help diagnose colon cancer from magnetic resonance imaging data. He says the system was doing remarkably well at identifying patients who had cancer using a  training data set of patients who had been previously diagnosed and treated for the condition.  But after taking a closer look at what attributes the algorithm was using, he learned it wasn’t relying on the images to predict cases, but rather the unique way vital signs were recorded at the clinic where confirmed cases were sent for treatment.  As such, the algorithm could not successfully be applied to new cases.  It was an “eye opener.”

“It’s always possible that the system will pick up on something you don’t expect,” he says.  

But even when the algorithms work as advertised, they may still bring unexpected administrative annoyances. While AI enthusiasts talk about how these programs can help improve patient care, they can also be used to justify billing and reimbursements. And because medicine is often filled with “noisy” data, key administrative applications will be easier to develop, vet, and deploy. Kohane predicts insurance companies will likely to be the first to adopt AI programs to take a closer look at where, why, and when healthcare dollars are being used-and physicians may have to debate whether their expertise is superior to a computer program’s.

“One of the most obvious uses for AI is to use a program to predict appropriate billing codes for patients with a particular history,” he says.  “If a doctor is billing for things that are different than what the system predicts, you have to imagine that the insurance company will at least want to start a conversation and understand why he or she did so.”

A brave new world for healthcare?

With widespread adoption of EHRs, and larger quantities of health data within physicians’ reach, experts agree that the potential applications for AI systems are numerous.  But while headlines joke that “the robot will see you now,” most observers think that no one has to worry about AI systems putting doctors out of their jobs. Still, these algorithms are likely to change the way doctors do their jobs.  

“These systems will let doctors work in more of their sweet spots, supporting them as they provide the best quality healthcare and, hopefully, also increasing accessibility to care across the population,” says Denny.

Kohane believes that physician input will be critical to ensuring that AI platforms are able to enhance rather than interfere with their workflows.  

“Now, more than ever, I’m convinced that for AI to succeed, it’s going to have to be a joint effort with doctors,” Kohane says. But he says AI programs, while they may not live up to their hype, can add significant value to medicine. He suggests that physicians, instead of waiting to be financially compelled to adopt AI technologies as they were with EHR systems, can benefit by learning about them now-and, if offered the opportunity, providing their IT departments, or even technology developers, feedback on how, where, and when they’d like AI to augment their abilities, now and in the future.  

“Doctors have some time to be strategic, to think about how these technologies can help better align them with their patients-and make sure they are properly vetted before bringing them into their practices,” he says. “But even though it’s early days, it’s important to understand these systems, the good and the bad.  It’s time for physicians to think about how they can best use AI to help reduce their workload, improve their businesses, and help their patients. Because, like it or not, they are on their way.”

What is AI, anyway?

The term “artificial intelligence” was first coined by the English mathematician Alan Turing to describe “the science and engineering of making intelligent machines, especially intelligent computer programs.” Today, AI may be better described as computer programs, trained on massive data sets, that can accomplish tasks that usually require human intelligence, including, problem-solving, and decision-making, says Pedro Domingos, PhD, a computer scientist at the University of Washington and author of The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World.

“You often hear terms like AI, machine learning, and big data used interchangeably-they are related but not the same,” he says. “Machine learning is a type of algorithm that learns from large data sets that are often called ‘big data.’ So AI, in some sense, is the goal-but machine learning is how we get there, with the machine learning algorithms fueled by all that big data.”

Such machine learning algorithms, after being trained by data from medical databases, can take in hundreds of thousands of data elements, analyzing the information over time to identify trends and associations that may not be immediately discernible to the human mind.

And such systems have demonstrated significant diagnostic power. Researchers from Google and Stanford University made headlines in 2017 when their machine learning algorithm, trained on imaging data, was able to diagnose melanomas more successfully than veteran dermatologists.  Other studies have shown that AI can outperform physicians in diagnosing diabetic retinopathy, predicting risk of suicide or heart attack, and managing congenital cataracts.