Blog Post

How Cognitive Computing Can Help Predict Diabetes Risk

Share this post:

Peter Bouman is a lead scientist at Truven Health Analytics, part of the IBM Watson Health business.

According to the American Diabetes Association, thirty million people currently are living with diabetes in the United States. Another 84 million have pre-diabetes, and every year, the cost of treating these patients totals $322 billion. That’s a potentially crippling financial commitment to a healthcare system already burdened with increasingly unaffordable spending.

As it stands now, physicians diagnose adult-onset diabetes at a later stage, leaving providers, health plans, employers and patients to find a way to deal with the soaring costs. But what if there were a way for analytics to predict which patients are at higher risk, and help physicians possibly intervene to prevent the development of diabetes? That’s what predictive modeling aims to help physicians do.

Big Data is too often a buzz phrase, and it’s easy to lose sight of the power of its practical applications. But for data scientists, its potential impact has never been more exciting. By providing clear and actionable information to expand the view of patients, Big Data with predictive analytics has the power to help more quickly isolate important patterns in large populations. This is information that healthcare professionals can use to intervene sooner and more effectively, which should ultimately help to streamline costs and drive better outcomes.

To prove the power of predictive analytics, we conducted a study tracking the real-world experiences of 27,528 individuals to see how accurately we could predict the occurrence of diabetes among those who had not been diagnosed as either diabetic or pre-diabetic. Using the Truven Health MarketScan® Research Databases, which capture the real-world annual medical and drug claims data for individuals with employer sponsored insurance, along with data from health risk assessment surveys completed by the individuals themselves, we were able to create a model that let us track hundreds of thousands of discrete data points over a five-year period.

This model, known to statisticians as a “decision tree,” captured risk factors such as body mass index, cholesterol, age, history of gestational diabetes, hypertension, smoking, number of work absences due to poor health, etc.  These factors were then weighed against one another to determine the best predictors of diabetes diagnosis within five years for each individual within the full study group.  Ultimately, we were able to define four distinct risk levels, or strata, based on each individual’s probability of being diagnosed with diabetes.

This analysis found that patients who were identified to have the highest risk were indeed very likely to be diagnosed with diabetes when our model was compared with real-world outcomes. Overall, the mean incidence of diabetes for our highest risk population in our model was nearly eight times higher than those who fell in the lowest-risk strata.

In the right hands, this is crucial clinical intelligence. Particularly for diseases like diabetes, for which the incidence rate can be reduced through early interventions that motivate lifestyle changes, the ability to target the right communications at the right time may have a huge difference in the health experience of the most at-risk population.  It can also potentially have a tremendous impact on cost.

The cost of modeling data and developing interventions such as wellness programs pales in comparison to the cost of treating a chronic disease. Modeling does have its limitations, however, and experienced oversight is needed to know when and how to intervene.

For example, in our study, the highest risk population had a predicted incidence rate over 50%. Likely, any care manager, clinician, or population health professional looking at that data would say they need to focus their preventive efforts on that group. However, the group represented just 187 people out of the 12,528 in the study. While this group is certainly in need, some level of interpretation is required to determine how far down the risk curve we should go when targeting preventive care.

An objective of most long-term population health initiative is to optimize resources and impact.  Just as it would not be cost effective to target the lion’s share of preventive care at a tiny population, it also wouldn’t make sense to invest heavily in preventive support for a large population that has a very low probability of developing the disease.

Though interpreting the data and developing interventions to act on it is still very much a balancing act, the predictive capability enabled by Big Data has indeed become a science. Armed with these models, large employers, health plans, clinicians, and other population health stakeholders have the power to develop the types of targeted interventions that were once the stuff of dreams. In a world where, for many populations, access to affordable care and favorable outcomes remain an elusive proposition, predictive models can be a game changer.



More Blog Post stories

IBM Watson Health taps hospital clinical data to improve sepsis care

Written by Watson Health | Blog Post, Value-Based Care

DeKalb Medical and AHMC Healthcare leveraged data and clinical care team capabilities to make improvements in sepsis care. more

Schneck Medical Center: Improving population health management with near real-time data

Written by Watson Health | Blog Post, Value-Based Care

This blog post contains an interview with members of the Schneck Medical Center team on how they use data to develop programs and processes that improve population health management efforts. more

Watson Health: Setting the Record Straight

Written by Dr. John E. Kelly III | AI, Blog Post, Oncology & Genomics

We at IBM have a lot to be proud of, including our pioneering work with Watson Health. Unfortunately, some media reports, including an August 11th story published by The Wall Street Journal, distort and ignore facts when suggesting IBM has not made “enough” progress on bringing the benefits of AI to healthcare. more