Today, through the development of more lab and diagnostic tests, our data have multiplied so that we have gigabytes of information that can be parsed in a variety of ways and evaluated using multiple comparisons. The human brain cannot closely examine more than 6 to 8 variables at a time to profile a problem. But the algorithms of predictive modeling can analyze hundreds of data points to make a diagnosis or a prediction of risk. We then arrive at empirical evidence that shapes the way our clinicians treat patients with certain disorders. There is no need for trial and error. We know what will work. We know what treatments will return value for the health care dollars spent.
In the US, the use of predictive modeling is especially important to help control costs. Compared with other industrialized nations, our US health care system consumes twice as many health care resources. We need predictive modeling to help us determine the best use of expensive resources that are becoming financially burdensome not only to the health care system but also to the health of the nation.
For most payers, identifying and managing about 1% of their covered lives can generate significant savings. In a plan with 100,000 covered lives, that’s 1,000 people. This 1% represents the most promising population for potential savings, often because they are people who could benefit from preventive care or specific treatments that they are not currently adhering to or are not currently receiving. These may not be the riskiest patients, because sometimes the riskiest patients do not have the potential for savings. I can think of one patient, for example, who was costing a payer $50,000 per month in prescription drug costs. This patient had hemophilia. He was adherent to therapy. The cost of drug therapy for his condition was relatively stable year after year. A disease management program would not benefit this patient who was already following treatment recommendations. Therefore, despite the high cost of his prescription drug coverage, there was no room to generate savings.
The value of predictive modeling then is that it allows payers to identify the top 1% of covered lives (or another percentage as desired), who are at highest risk and highest cost, and who will most likely benefit from disease management programs. The goal, of course, is to improve health for those patients and to lower costs for the plan in the following year.
Most payers develop plans for patients with the top 6 diseases, including coronary artery disease, congestive heart failure, diabetes, asthma, chronic obstructive pulmonary disease, and depression. The riskiest patients among this group will be selected for disease management programs, which are expected to help the members better manage their health. This is the value of predictive modeling. But are there differences between predictive modeling algorithms? Is one model better than another? The answer is a resounding “yes.”
At Elsevier MEDai, we have built a model that has 10% better predictability (in terms of R2) than others on the market. What does this mean for payers? Let’s compare two predictive models A and B. We have two years of data (medical and pharmacy claims), and we use Year 1 data to predict Year 2 Total Cost. We built two models. Model B has 10% better (higher) R2 than model A. In our comparative study we identify:
- By model A predictions top 1% high-risk\high-cost members with average Year 2 Total Cost of $26,273
- By model B predictions top 1% high risk\high cost members with average Year 2 Total Cost of $27,437
The plan will pay a certain amount of money to manage the top 1% of high-risk, high-cost members irrespective of which model they use to identify these top 1% of members for involvement in disease management programs. Plans can expect to save about 30% of the total cost per member. Therefore, by using better predictive model B, they will save an additional 30% of the difference or $1,164 ($27,437 model B – $26,273 model A). Or they will save about $350 per each one of the top 1% or members or $350,000 per 100,000 members (1% of which is 1,000 members).
This example demonstrates that differences in predictive modeling algorithms can be measured in dollars. The better the model, the greater is the potential for cost savings. Thus, in a nation where soaring health care costs are responsible for a large portion of our national deficit, the benefits of choosing the superior predictive modeling program is very much essential to health care.