AGA Tech Summit

Predictive analytics with large data sets are being pursued to individualize IBD therapy



SAN FRANCISCO – The potential power of machine learning to improve therapeutic choices in inflammatory bowel disease (IBD) is at the threshold of clinical applications, according to data presented at the 2019 AGA Tech Summit, sponsored by the AGA Center for GI Innovation and Technology.

Joel Austell/MDedge News

Dr. Akbar K. Waljee

This type of work is relevant to many fields of medicine, but studies conducted in IBD have provided particularly compelling evidence that predictive analytics will improve outcomes and lead to more cost-effective delivery of care, according to Akbar K. Waljee, MD, MSc, an associate professor in the division of gastroenterology, University of Michigan, Ann Arbor, and a staff physician and researcher at the VA Ann Arbor Healthcare System.

Having now published several papers on the role of precision analytics and big data to improve treatment choices in IBD as well as other diseases, Dr. Waljee said, “We collect large amounts of clinical data every day in the delivery of health care but we are now only just beginning to leverage [these] data to guide treatment.”

Based on the experience in IBD, these analyses are relevant for selecting who to treat and not to treat with a given drug.

In one study of how this technology can be applied, data from 1,080 IBD patients taking thiopurines were used to develop a machine learning algorithm that analyzed multiple readily available variables, such as a complete blood count with differential and a chemistry panel, to predict if the patient was in remission. The area under the receiver operating characteristics curve (AuROC) was 0.79 or much higher than previous prediction using drug metabolites, according to Dr. Waljee.

The mean yearly rate of clinical events (new steroid prescriptions, hospitalizations, and abdominal surgeries) was then compared between those who did and those who did not have an algorithm-predicted remission. The lower mean rate in those predicted to be in remission (1.08 vs. 3.95 events) provided support for the conclusions that the algorithm is clinically viable.

“The heterogeneity of response to therapies for IBD is well established. If machine learning predicts effective choices, there will be an opportunity to accelerate the time to disease control as well as save costs by avoiding therapies not likely to be effective,” Dr. Waljee explained.

In another example, the focus was on predicting response to vedolizumab, a monoclonal antibody targeted at a gut-specific mediator of inflammation. In this case, machine learning was applied to predicting corticosteroid-free remission at 1 year in patients with Crohn’s disease patients evaluated 6 weeks after initiating therapy. The machine-learning algorithm was based on an analysis of numerous variables, including laboratory data, sex, and race. Based on the model drawn from the analysis of 472 patients, 35.8% of the patients predicted to be in corticosteroid-free biologic remission at 1 year achieved this endpoint, whereas only 6.7% of the patients predicted to fail achieved the endpoint.

“This suggests that we can use an algorithm relatively early in the course of this biologic to predict who is going to respond,” reported Dr. Waljee. Again, patients with a low likelihood of response at 6 weeks can be started on an alternative therapy, potentially accelerating the time to disease control and avoiding the costs of ineffective and expensive treatments.

IBD is a particularly attractive focus of precision analytics with big data. IBD has a relatively unpredictable relapsing-remitting course and a variable heterogeneous response to available therapies. Algorithms predictive of response circumvent the inherent delays from evaluating disease control over an extended period.

“With ever increasing concern about costs of care and access to care, these treatment algorithms promise to use resources more efficiently,” Dr. Waljee said.

Next Article: