Artificial intelligence is making inroads in the healthcare field, and global demand likely will exceed $5.5 billion by 2022, a recent study by the design and marketing firm Arcluster indicates.
The study found increasing use of AI in healthcare inevitable for everything from medical imaging and diagnostics to emergency room and hospital management to drug research, delivering operational efficiencies that result in better and faster care.
“It’s definitely a good thing,” said Arun Nirman, director of research at Arcluster. “Hospitals will see huge cost benefits. So will pharmaceutical companies. That will all filter back to the patient.”
Unlike industries in which computers and robotics have eliminated jobs, healthcare still will require hands-on personnel to deliver treatment.
“This is not the next job killer — at least not for another 20 years,” Nirman said. “The automation will be for processes, not the delivery of care. AI will impact the process, not the people.”
Unlike those nifty little speakers or robotic floor cleaners popping up in people’s homes, AI for healthcare is not a luxury item nor a curiosity but rather a long-term investment that will pay off in quality of care and scientific breakthroughs, Nirman said.
Big data, big advances
The key is big data, and AI really is the only way to make any sense of it. Cisco estimates machine-to-machine connections in healthcare will grow at a compound annual rate of 30 percent between 2015 and 2020. VentureBeat reported that venture-capital investment in cutting-edge AI medical technologies went from $30 million in 2012 to $892 million in 2016.
More recently, the Food and Drug Administration on Feb. 13 approved an AI algorithm to analyze computed tomography scans for evidence of a stroke. The Viz.AI Contact application sends a text when it spots large blood-vessel blockage. Some 795,000 people in the U.S. suffer a stroke annually, making it the fifth leading cause of death in the U.S. and a major cause of serious disability.
The FDA also has issued guidelines for dealing with artificial intelligence, but it is unclear how the approval process will apply to a device that is constantly changing, said Dr. Paras Lakhani, who specializes in diagnostic radiology and nuclear medicine at Thomas Jefferson University Hospital in Philadelphia.
“Algorithms are trained on gigantic databases to find patterns,” Lakhani said. “AI systems have the ability to continuously learn. Let’s say a computer automatically draws contours around a nodule. If I don’t agree, I can manipulate those contours. Some systems can look at the changes I make and then retrain their own algorithm so it can get better. The FDA has a more traditional approach [and issues approvals] on a particular device. There’s no [regulatory] mechanism that can clear a continuously changing system. Right now, if you make changes, every two or three months, you have to resubmit.”
The World Health Organization projects chronic disease will increase by 57 percent by 2020. AI may be able to curb the costs of treating such conditions by detecting and diagnosing them earlier. A report by Frost & Sullivan found that AI could improve outcomes by as much as 40 percent and cut the cost of treatment in half by reducing the number of human errors and decreasing the number of doctor visits. The savings could be enormous; management-consulting firm Accenture puts the figure at $150 billion annually by 2016 in the United States alone.
Google's DeepMind project aims to develop AI algorithms that will lead to solutions for complex problems. The project currently is working with the U.K.’s National Health Service to develop world-class diagnostic support that will lead to quick initial assessments to help clinicians make treatment decisions.
The AI systems ideally will develop new ways to diagnose conditions by spotting subtle relationships among symptoms. The initial DeepMind research is centered on getting the algorithms to interpret visual information such as head and neck scans and mammograms to identify potential issues and recommend treatment. As the algorithm views more images, it is able to refine its understanding and interpretation.
Gregory Hager, the Mandell Bellmore professor of computer science at Johns Hopkins University, said artificial intelligence shows the most promise for the automatic interpretation, categorization and triaging of scans so that abnormal scans can get faster and more expert attention from radiologists and other specialists.
“There are huge troves of data available and there are annotations to go with that data because you have radiological reports,” Hager said. Coupling the machine findings with people will lead to more efficient and effective treatment, he said.
A second major area is infectious diseases; artificial intelligence could speed diagnosis of these and be used to help determine whether a patient should be hospitalized or sent home, Hager said. Robot surgery companies also are on the cutting edge with systems that can help train doctors.
Quality of care
Jefferson’s Lakhani said he is looking forward to more widespread use of artificial intelligence because of the help it can provide in analyzing images more quickly.
“There’s a lot of fear about our jobs being taken over, but I’m not worried. We have too much work to do,” Lakhani said. “I see [artificial intelligence] as something that can improve our efficiency and performance. If we miss something right now, it’s a medical issue. With AI, we’ll miss fewer problems.”
Johns Hopkins’ Hager noted that artificial intelligence will not be making the final diagnosis. Much patient care “is still fairly nuanced patient-physician interaction,” he said. “You can’t just have an automated system tell a patient he has two weeks to live. Imagine what that would be like for the patient.”
Among the advantages of artificial intelligence is its ability to enhance images with less information, which could lead to lower doses of radiation for X-rays and scans and shorter MRIs.
“An MRI can last an hour. AI potentially could lessen that to 15 minutes. These MRI machines are expensive to buy and maintain. There’s often a backlog. If we can do two or three times as many scans a day without having to buy more scanners, that’s a big win for patients,” Lakhani said.
The trick, however, is getting enough data for the algorithms to learn from, said Dr. Charles Kahn, professor and vice chairman of radiology at the University of Pennsylvania Perelman School of Medicine, who is involved in research for using artificial intelligence to detect cancerous lung nodules.
“Being able to tell which is a lung cancer from a benign nodule [is where the challenge lies],” Kahn said, noting artificial intelligence has been used in some aspects of medicine for a quarter century, but in the past, such systems have been known to spit out too many false positives. “A physician plus AI will be better than either one alone. If this technology helps us take better care of patients, it’s something we want to incorporate into our practice.”
On the business side, artificial intelligence could provide workflow optimization, using analytics to predict no-shows so that doctors could overbook without increasing waiting times, Hager said.
“Right now, AI is on the periphery of medicine. There’s not a lot of interactions [with patients],” Hager said.
Noting that implementation is picking up speed, Hager said the individual nature of every hospital remains a barrier to widespread implementation.
“Systems have to be customized to their settings,” Hager said. “Patient demographics matter a lot. We haven’t really worked through that yet. Those solutions are going to emerge organically. Companies that have a working product will have to develop standardized interfaces. It probably will be another three to five years for those sorts of standards to emerge before you see any rapid scaling.”