How to Use AI to Reduce Diabetes Cost | Geneia

How to Use AI to Reduce Diabetes Cost

August 21, 2019
Jasmine McCammon, Ph.D., Principal Data Scientist


Preliminary results show a health plan with 1 million members using this model may realize an annual savings of approximately $1.5 million

The American Diabetes Association estimates that one of every seven healthcare dollars are spent to treat diabetes and its complications. Estimated costs in 2017 were $327 billion, up from $245 billion in 2012. According to the International Diabetes Federation, “In almost all high-income countries, diabetes is a leading cause of cardiovascular disease, blindness, kidney failure and lower limb amputation.” Other known diabetic complications are stroke, high blood pressure, neuropathy, eye complications, ketoacidosis and ketones, skin complications, gastroparesis and hyperosmolar hyperglycemic nonketotic syndrome (HHNS).

Diabetic complications are far too common, and touch most Americans, including me. My grandmother had type 2 diabetes and suffered from many of the associated complications. The National Diabetes Statistic Report, 2017 reported a total of 7.2 million adult hospital discharges in 2014 were with diabetes as any listed diagnosis:

  • 1.5 million for major cardiovascular diseases
  • 400,000 for ischemic heart disease
  • 251,000 for stroke
  • 108,000 for a lower-extremity amputation, and
  • 168,000 for diabetic ketoacidosis

There also were 14.2 million diabetic-related emergency department visits in 2014.

That’s why I’m so encouraged by the work we’re doing in the Geneia Data Intelligence Lab

In case you missed it, we invest in research, models and data science talent to address major cost drivers and help our health plan, hospital, physician and employer clients lower healthcare costs. Our models better predict risk, cost of care, likelihood of adverse events and outcomes. They also help our clients to intervene earlier with patients whose risk is expected to rise and/or who are at risk for major, expensive conditions, allowing the mitigation of future costs. Diabetes is a particular focus for Geneia’s clinical team, allowing them to leverage their expertise in creating and validating related models.

Diabetes Complication Model

Take the Diabetes Complications Model, for example. It’s one of which I’m especially proud. After all, I spent more than six months creating and improving the model. Through careful research of diabetic complication literature as well as extended conversations with Geneia’s clinical team, I worked on decoding an algorithm that predicts which type 2 diabetics will experience a diabetes-related complication:

  • Who: People with type 2 diabetes (T2D)
  • What: Likelihood of experiencing a T2D-related complication 
  • When: Within the next year 
  • How: Based on variables from the previous year
  • Why: Improved health outcomes and cost savings

For the sake of this blog, let me condense the project into nine primary steps:

  1. Identify the cohort of people likely to have type 2 diabetes during intake year.
  2. Identify cohort members that do experience a diabetes complication the following year.
  3. Gather variables for people in this cohort with possible relationship to experiencing diabetes complications. Among the variables we considered are age, gender, medical costs, prescription counts and more.
  4. Compare variables between those that did and did not experience a complication to build a predictive algorithm.
  5. Use 75 percent of the data to develop a model that sees the relationship between the variables and the complication outcome.
  6. Validate the model on the remaining 25 percent of data where algorithm only sees the variables and has to predict the outcome. Compare predictions to whether or not each person actually experienced a complication.
  7. Based on validation performance accuracy, evaluate potential performance improvements and iterate through the entire process by calculating new variables and selecting the best ones, adding other data sources and trying different algorithm parameters.
  8. Test the best model on production data. Confirm that it performs well in the setting where it will be applied.
  9. Stratify final output predictions based on prediction probability. This critical step supports clients customizing the number of people to be targeted for an intervention based on the client’s resources and desire for return on investment.

Model development and utilization process

For type 2 diabetics at risk of a complication, this model identifies the likelihood of complications. Healthcare organizations can then direct them towards high-value preventive and self-care options. Preliminary results show that a health plan with one million members using this model to predict and intervene with those diabetics determined to be at risk for a diabetes-related complication may potentially realize an annual savings of approximately $1.5 million. Next up on the list of enhancements to this model is looking beyond the probability of a complication for each member to provide the end user with information about why the model predicts what it does, e.g. which variables and their values contributed to the prediction. This increases model transparency and interoperability, and ultimately client and consumer buy-in.

To learn more about the models the Geneia Data Intelligence Lab is creating to drive lower healthcare costs, download our white paper, Using AI to Drive Lower Healthcare Costs.

Download the white paper

 

Related Blogs