Artificial intelligence is helping providers match patients to clinical trials and offering treatment recommendations to specialists. What does the next round hold?
Physicians know that when patients come to the emergency department complaining of “pressure” in their chest, it could be a sign of chest pain caused by myocardial infarction. But getting an artificial intelligence (AI) platform that can “read” notes in an electronic health record to understand that “pressure” in this case means chest pain rather than hypertension (i.e., “blood pressure”) requires collaboration between data scientists, computer scientists, and clinicians.
Paul Tang, MD, vice president and chief health transformation officer at IBM Watson Health, believes that with proper training from humans, AI platforms can provide clinicians with better access to evidence-based treatment recommendations tailored for a specific patient.
“Only 15 percent of the decisions that clinicians make are supported by evidence from randomized controlled trials,” Tang says. Even when such evidence exists, it is often based on data from fairly homogeneous populations. “But AI applications can help provide personalized health care that takes into account a more holistic view of an individual beyond their demographics to include socioeconomic status, education, and preferences, which can affect treatment compliance and outcomes,” he says.
Specifically, AI applications can create precision cohorts that allow clinicians to review treatment outcomes in specific population subsets that share the same characteristics as their patients.
The latest AI technology offers clinical options in what can be described as a type of knowledge matchmaking, in which the platform pairs the most relevant evidence with individual patients based on their clinical and life context.
“The ability for a human to do that kind of matchmaking is fairly limited,” Tang says. “So what we are asking the machine to do is to narrow the search space and try to find the best match given our knowledge of this particular individual, and present those options to the human professional, who makes the judgment and comes up with a relevant treatment plan.
“This is not about replacing an individual but rather offloading the data preparation and assimilation to a machine and, in a sense, offering complementary or augmented intelligence.”
Today’s AI platform can organize vast amounts of data from medical literature and other sources into “knowledge graphs,” which connect related concepts (for example, a disease would be connected to a symptom, which would be connected to a treatment).
This process requires a unique kind of training using machine-learning approaches. Scientists might interview an expert such as an oncologist, who can describe ideal treatments for specific conditions or review and annotate the platform’s treatment recommendations with correct answers. They also may incorporate unsupervised, continuous-learning approaches that capture when users accept or ignore the recommendations, which helps the AI platform make better suggestions in the future.
“Underpinning all of this is a constant review and assessment of the quality of the feedback,” says Eric W. Brown, PhD, director of foundational innovation at IBM Watson Health. “You want experts to be correcting the system, and you want to validate that the training data you are using is accurate, comprehensive, and effective.”
Applications For Cancer
One of the latest AI applications for oncology was trained by subspecialists at New York City’s Memorial Sloan Kettering (MSK) Cancer Center who worked alongside computer scientists and project managers.
The team used National Comprehensive Cancer Network guidelines to teach the system the universe of possible treatment options. MSK cancer subspecialists then prioritized the options into recommendations.
For each cancer type, the subspecialists identified hundreds of cases from MSK that included every clinical attribute they believed was important to help a clinician make the right decision. These cases also included what the subspecialists determined was the right “answer” to a clinician’s question. For example, subspecialists described how they would treat a patient with metastatic lung cancer who is a certain age, has a specific previous treatment history and specific lab results, and is taking certain medications. From there, the system “learned” that those attributes pointed to a given answer, or treatment recommendation.
This process was repeated hundreds of times to ensure the system can ultimately learn what weight to give each patient attribute when recommending a therapy. Over time, the system becomes better and better at mimicking the thought process of the human expert.
In a blog post, Mark Kris, MD, a medical oncologist who leads the training project at MSK, discussed the potential healthcare applications of AI. “Doctors treating these illnesses now know how different they are from person to person,” he wrote. “We need better ways to help us understand the complexity and variation of these diseases to improve care and research. Textbook and guideline-based treatments are a good place to start, but they can’t address the many biological and other factors affecting the course and aggressiveness of cancers.”
AI systems also can be trained to search their body of knowledge for evidence such as American Society of Clinical Oncology (ASCO) guidelines, textbook information, and peer-reviewed literature to support their recommendations.
AI platforms with natural-language processing abilities also have applications for clinical trial matching. A feasibility study presented at an ASCO conference found that an AI platform cut the time required to screen patients for clinical trial eligibility by 78 percent. During the pilot, the platform “read” the clinical trial protocols at one large Arkansas medical practice and evaluated data from patient records and clinical notes to automatically exclude ineligible patients from the pool.
Such clinical-trial matching technology has been used the longest at Mayo Clinic, notes Andrew Norden, MD, MPH, MBA, deputy chief health officer for oncology and genomics at IBM Watson Health. Preliminary data from Mayo demonstrate improvement of more than 50 percent in the breast cancer trial accrual rate from 2015 to present day.
Laura Ramos Hegwer is a freelance writer and editor based in Lake Bluff, Ill.
Interviewed for this article: Eric W. Brown, PhD, director of foundational innovation, IBM Watson Health, Yorktown Heights, N.Y.; Andrew Norden, MD, MPH, MBA, deputy chief health officer for oncology and genomics, IBM Watson Health, Cambridge, Mass.; Paul Tang, MD, vice president and chief health transformation officer, IBM Watson Health, Cambridge, Mass.
This article is based in part on a presentation at the American College of Healthcare Executives 2017 Congress on Healthcare Leadership in Chicago.