Next-gen computers programmed to listen, advise and diagnose


Will computers someday replace physicians as expert diagnosticians? According to some experts, the answer is yes—but that's not necessarily a bad thing for doctors.

“It may seem counterintuitive, but I envision these [computer] tools freeing us to spend more time with our patients,” said Herbert Chase, MD, professor of clinical medicine in biomedical informatics at Columbia University in New York City.

Photo by Thinkstock
Photo by Thinkstock.

Dr. Chase is working with IBM to retrofit the company's “Watson” supercomputer for medical use. “Watson's job is to fully inform the doctor as to what the risks and benefits are, what the guidelines are—having that information right next to me will give me more freedom to interact with my patients,” he said.

Watson, famous for beating top human contestants on the television show Jeopardy!, has the potential to revolutionize the way physicians work by providing a huge store of information at their fingertips, said Dr. Chase. “Watson is a game changer,” he said. “There will be a shift from getting information, which is very time-consuming, to analyzing information once you have it and exerting judgment as to how to proceed.”

Supercomputers like Watson are likely to feature prominently in the hospital once the field of computer-aided diagnosis (CAD) matures. Some hospitals are already using CAD and decision-support software, but the technology has yet to be integrated seamlessly into workflow. Improving integration is a major focus of ongoing research, experts said.

“The work that's being done today to make systems effective is focused on workflow and the cognitive needs of clinicians,” said Edward H. Shortliffe, MACP, PhD, president and CEO of the American Medical Informatics Association and professor in the school of biomedical informatics at the University of Texas Health Science Center at Houston. “The successes you'll see over the next decade will be companies that really understand how to effectively introduce decision support in ways that are acceptable to clinicians.”

“Isabel”: the medical Google

Some hospitals are already using computer programs to assist physicians in making accurate diagnoses. At United Hospital in St. Paul, Minn., for example, hospitalists are participating in a one-year pilot program using Isabel, a Web-based system designed by Isabel Healthcare. Isabel, continuously updated with the latest information from medical textbooks, journals and other sources, quickly prioritizes possible diagnoses based on symptoms entered by the physician combined with the patient's medical history.

“Isabel is like a medical Google,” said Scott Tongen, MD, medical director for quality at United Hospital, which is part of Allina Healthcare. “It doesn't think but it searches and sorts and shows you what it's found in the order it thinks you're most likely to get what you're looking for.”

Isabel is linked to the hospital's electronic health record (EHR) system, allowing it to incorporate the patient's medical history and demographics into its list of possible diagnoses. Physicians like the idea of Isabel, said Dr. Tongen, but it's a challenge to get them to use it consistently.

“One of the difficulties in using software like this is that you have to think about [using] it,” he said. “If you do think about the fact that you might have it wrong, you've already defeated the No. 1 reason that misdiagnoses are made—early closure on what you think the diagnosis is. Once a doctor decides to launch Isabel, they've self-selected themselves into a group of physicians that are thinking more than the average.”

Dr. Tongen encourages hospitalists in the pilot groups to launch it whenever “something just doesn't seem right” and when patients present with multiple problems or medications, he said.

Fitting these programs into workflow has been difficult partly because hospitalists are used to thinking of computer support as static information that pops up during the order-entry process, said Dr. Shortliffe. At some point, though, he believes there will be a shift from thinking of CAD as “not just offering information but guiding you through a process.”

If doctors can start thinking of Isabel in that way, it will become a useful tool, said Dr. Tongen. However, administrators also have to be convinced that it's money well spent (United invested about $100,000 in the pilot). To make a financial case for Isabel, Dr. Tongen has to address a common criticism about diagnostic software—that it increases the cost of medicine by prompting physicians to think of unlikely diagnoses, which triggers them to order a variety of expensive tests without necessarily improving the accuracy of diagnosis.

Over the next year, Dr. Tongen plans to track searches and queries done on Isabel in order to assess the downstream effects and determine if Isabel saved money by preventing misdiagnoses.

“If I can take that data and compare it to those cases of similar diagnoses that did not use Isabel, I can demonstrate whether Isabel is having an effect on cost of stay, cost of care and the end diagnosis,” he said. “If Isabel saved one patient's life (by avoiding missed diagnosis) and prevented a lawsuit, then the hospital's return on investment has been met.”

CAD: targeting specific diagnoses

Concern about misdiagnosis has already spurred development of targeted computer programs designed to help physicians make faster, more accurate diagnosis of specific diseases or conditions. For example:

  • A study published online October 2008 in Brain reported that computers trained to analyze brain scans outperformed radiologists in diagnosing Alzheimer's disease, correctly classifying the disease 95% of the time vs. a median of 89% among a group of radiologists. While well-trained radiologists performed almost as well as computers, the authors noted advantages to automated systems, such as in facilities that lacked trained neuroradiologists or as a screening tool in primary care settings.
  • Face-classification software correctly diagnosed 72% of patients with acromegaly compared with 63% accuracy by specialist physicians. The study was published online April 20, 2011 in The Journal of Clinical Endocrinology and Metabolism.
  • A CAD tool was more accurate than emergency department physicians at diagnosing urinary tract infections, pneumonia and bacteremia in young children, according to an April 2010 study published in BMJ. The tool combined history from patients' electronic records with clinical signs and symptoms (solicited from physicians) related to febrile illness in order to estimate the risk of serious bacterial infection. Since doctors often undertreat bacterial infection, the authors said, a CAD tool could ensure more children get early treatment with antibiotics.

In addition to improving the accuracy of diagnosis, CAD tools that work in tandem with a hospital's EHR system can help ensure that important parts of a patient's clinical profile aren't missed. For example, Dr. Chase at Columbia is working on a chronic kidney disease (CKD) notification tool that reviews patients' charts automatically, identifies those with CKD, and notifies physicians who have not documented the condition in their outpatient notes. Another CKD tool under development gives feedback to physicians on their patient's progress in relation to recommended guidelines.

Clinical decision support also has a huge potential role in tailoring treatments to unique individuals, said Dr. Chase.

“In 25 years, we might know of 100 genes causing hypertension and dozens of potential drugs for treatment of each of those genetic causes. The patients may also have all sorts of comorbidities which will influence which drugs to use,” he said.

If a Watson-like tool could not only compute all that data but also understand non-physicians speaking in natural language, “it could be incredibly powerful,” said James Mazoué, PhD, professor of philosophy and director of online programs at Wayne State University. “The benefits would be tremendous in remote areas such as sub-Saharan Africa where they might have only one doctor for 50,000 people.”

Revolutionizing medical education

Computer tools also have the potential to fundamentally change the way physicians are trained, experts said.

“If we can create computers to diagnose patients, then it will significantly alter our conception of what physicians do,” said Dr. Mazoué. “The question would become, ‘Why do we need to keep training people to develop [diagnostic] expertise?’ If we had a single system as a repository of all information and research that we can tap into and use, that would become the gold standard for diagnosis.”

If computer programs perform the “dirty work” of searching and prioritizing information, medical students no longer will be required to spend most of their time memorizing, said Dr. Chase. “It shifts the educational focus from memorization to analysis.”

Yet the medical education establishment is a long way from making that shift, Dr. Shortliffe said. “It is a rare institution that has made any significant inroads in training medical students about bioinformatics,” he said. “It's not culturally accepted as part of the medical model. And the people that tend to drive the process are from an era where they often can't see the connection between computers and the medicine they know and love.”

Another obstacle is finding enough instructors to lead departments and teach courses in biomedical informatics, he added.

Dr. Mazoué envisions a day when computers replace physicians as diagnosticians, leaving physicians to concentrate on research and patient interaction. But that won't happen if researchers and physicians cling to the idea that diagnosis without human judgment is impossible, he added.

“We need to test the hypothesis that a computer can diagnose without human interpretation,” said Dr. Mazoué. “If we take the approach that we will never be able to construct a computer-based system that will do as well as the typical physician, we are biasing our research.”