Artificial intelligence seems poised to revolutionize healthcare. Whether it’s the promise of supercomputers to find cures for disease, identify vulnerable patients before they become ill, or better personalize testing and treatments, research is moving fast to apply computing power to healthcare problems large and small. At the same time that clinical laboratorians are exploring analytics to mine the wealth of data available to labs, clinicians are seeking to streamline processes and boost their diagnostic and predictive abilities with advanced algorithms and supercharged software.

“Right now, 15% of medical decision-making in care delivery is automated,” said Richard Zane, MD, chief innovation officer at UCHealth and professor of emergency medicine at the University of Colorado School of Medicine in Aurora, Colorado. “We would like it to be 80 to 85% that is automated.” That is not to say computers will replace doctors, Zane added. “There’s no substitute for a clinician. It’s really the inverse of what we need the clinician’s brain to do,” he said.

Computers are good at gathering information to support and inform a clinician’s decisions. For example, looking at all the antibiotics that could be prescribed and which would work best considering a patient’s history of drug reactions, resistance, and travel, as well as the regional and hospital antibiogram, is a task well suited to computers.

“All of that is simply information gathering,” Zane explained. “The real art of it is diagnosing the patient with pneumonia. We want the clinician to be at the interface with the patient, making important decisions and not spending time gathering information.”

Targeting Chronic Diseases

While banks, casinos, and billion-dollar companies like Netflix and Amazon speed ahead with artificial intelligence projects, the healthcare industry is just waking up to the possibilities, according to Zane. “People talk all the time about, ‘What’s the Uber of healthcare?’ and we’re still in the Polaroid era,” he said.

Yet artificial intelligence is slowly finding its way into patient care. For example, Vanderbilt University Medical Center in Nashville has created a predictive algorithm for genetic testing that recommends who should be tested for gene variants that affect their response to prescription drugs such as warfarin or clopidogrel. “I think there will be a day in the future in which we will perform broad spectrum [genetic] testing of people, maybe when they’re born, but the cost model isn’t there yet,” said Joshua C. Denny, MD, MS, a professor of biomedical informatics and of medicine at Vanderbilt.

Vanderbilt’s algorithm assigns a score representing a patient’s likelihood of someday needing a drug that requires gene-driven prescribing. When the algorithm was in development, Denny ran it using his own patients’ data and found some surprises. For example, an elderly patient who he thought was high risk received a low score. Upon reflection, he said, this patient hadn’t developed cardiovascular disease by age 90 so she probably wouldn’t in the future, and if she developed an arrhythmia, she probably wouldn’t be prescribed a blood thinner because she was at risk for falls.

“I sort of just worked through the contraindications and the indications for the different meds and realized that she really was low risk,” Denny said. “The algorithm was right and my gut reaction was wrong.” In the future, artificial intelligence tools like this will extend the abilities of clinicians, he said, augmenting their skills rather than replacing them.

Another example of artificial intelligence in clinical practice is a computer model developed at Mayo Clinic by Rozalina G. McCoy, MD, an assistant professor of medicine and scholar at Mayo Clinic Rochester, Minnesota. She and her colleagues built the model to identify patients who need proactive diabetes management.

Typically, clinicians consider patients with diabetes low risk if they have normal blood glucose readings at doctor visits, McCoy said, “but we know that there are patients who have been doing ok but still develop problems—the key is to figure out who those people are so we can monitor them closely and treat them pre-emptively.”

McCoy and her colleagues built a machine learning algorithm based on billing data from patients who appeared to be low risk for diabetes complications yet went on to experience such sequelae. The model’s machine learning core enables it to predict risk at both the population and individual levels.

For example, the model may identify heart failure as a risk factor for diabetes within a population, perhaps at a particular clinic or hospital. Yet when the algorithm considers an individual patient who has heart failure, it may determine he or she is not actually at risk because of other potentially protective factors. And vice versa, a patient may seem to be low risk based on population-level statistics, but without intervention is in fact likely to have an adverse outcome.

“You don’t apply the same formula to every patient,” McCoy noted. “The method allows for individualized risk prediction and population risk prediction at the same time.” The risk model is not used clinically yet, but her team is working toward testing it among patients with type 2 diabetes (Med Care 2017; doi: 10.1097/MLR.0000000000000807).

McCoy agrees that healthcare lags behind other industries in taking advantage of machine learning and artificial intelligence. Catching up will require healthcare professionals to take the lead and be more engaged, she said.

For laboratorians and data scientists exploring this nascent field, it is critical to partner with clinicians from the beginning. “I think that part of the reason why we don’t have more useful tools is that clinicians aren’t as involved as they should be,” McCoy said. “We know where the need is the greatest. We know the areas where there are gaps in our knowledge and where we really need to improve patient care and health outcomes. We also know what kind of information will be actionable and useful in the real world.”

Zane agreed. “Clinicians are central to this,” he said. His team at UCHealth is about to deploy continuous glucose monitors for patients in a hospital inpatient unit, with the goal of using artificial intelligence to find patterns in data. “We’re going to find signals of disease that we otherwise would not have recognized,” Zane said. “What we expect to find is that there are patterns that are going to predict ensuing crisis that will allow us to avoid crisis.”

Imagine knowing that, say, a change in pulse combined with a particular trend in glucose levels is a warning that a patient will have a stroke in 3 days, Zane said. Imagine finding patterns that predict sepsis or myocardial infarction or surgical complications. “Think about what that means, how many lives that could be saved,” he commented.

Taking Small Steps Toward Big Promises

Meeting the loftiest goals of artificial intelligence may take time. A report by STAT News in September found that IBM’s supercomputer Watson for Oncology may have failed to live up to its marketing promise to discover new insights in cancer care. The report revealed that, rather than relying on artificial intelligence to diagnose cancer and suggest treatments, Watson for Oncology relies mainly on human experts at Memorial Sloan Kettering Cancer Center to laboriously input recommendations.

MD Anderson Cancer Center ended its partnership with IBM Watson in February after 4 years of work and an investment of more than $62 million, without ever guiding patient treatment. IBM declined to comment for this story.

Despite some bumps along the way, smaller scale artificial intelligence solutions are easily within reach, McCoy said. Machine learning is being used successfully in many areas of risk prediction, and also to interpret images, such as retinal eye images, to predict who might develop diabetic eye disease.

The possibilities for machine learning tools are endless. In terms of laboratory tests, McCoy would like to see, for example, a model that could improve utilization of hemoglobin A1c testing and identify which patients with diabetes are likely to develop complications. “Once you sit down and understand what the methods are trying to do, it makes sense,” McCoy says. “I have no formal computer science, informatics, or math training. I just learned by doing and I am constantly learning from my data science colleagues. They are experts on the how of machine learning, but my job is to make sure that everything that we do ultimately helps improve patients’ lives.”

For the generation of patients and healthcare professionals who grew up with Google, relying on algorithms to offer suggestions will become second nature, McCoy noted. “I may not want a robot treating me,” she said, “but I certainly want a computer model using information from thousands of other patients to give recommendations to my doctor on how to give me the best, most personalized care possible.”

Julie Kirkwood is a freelance journalist who lives in Rochester, New York. +EMAIL: julkirkwood@gmail.com