Building a crystal ball with pharmacy claims (predictive modeling)

Medicare members have higher opioid prescription rates; and while those rates are declining, they are declining slower than other populations. 

Across the U.S., one in three Medicare Part D members had an opioid claim in 2018.¹ The rate of opioid claims for Prime’s Medicare members is comparable to that. When you compare opioid prescription rates across Prime’s books of business, Medicare members have a higher prescription rate than Medicaid or commercial members. Prescription rates for all three groups have gone down since 2015. But the rate of decline for the Medicare group is slower than for the other two groups. 

Comparison of opioid claims and rates of decline by line of business, 2015- 2018²

High doses lead to increased risk

The Centers for Disease Control and Prevention (CDC) associates high dose opioid therapy with increased harm.³ High dose is defined as > 90 morphine milligram equivalent (MME) daily dose.

Given that our role is to help people safely get the medicine they need, we asked ourselves some important questions about this problem: 

• Can we improve care for Medicare members with opioid prescriptions? Yes.
• Could we find a reliable way to determine which members were more at risk of becoming a high dose utilizer before it happened? Yes. 
• Can we use claims data to target interventions to members most at risk to escalating daily dose increases? Yes.
• Can we prove this connection to make it compelling to health plans and providers? Yes.

Building a tool to intervene upstream – a way to red flag potential issues

Across the board, our benefit design and clinical intervention programs work most effectively when applied preventively. Nowhere is that truer than in prescriptions for high dose opioids. No surprise, research shows high dose opioids lead to more harm³, we want to help the member before he or she has a prescription for high dose opioids.

We built a predictive model to identify these members upstream. It wasn’t long before we realized we needed two models:

1. For members who were new to opioid prescriptions
2. For members who had a previous opioid prescription

Artificial intelligence at work: we started with hundreds of predictors and millions of permutations 

Our health outcomes team set out to develop a high dose opioid predictive modeling process for Medicare members. Our data scientists are specialized in health care claims data and the latest predictive modeling techniques. Using artificial intelligence and predictive modeling software, predictive models were built and statistically validated. We used a mapping process of more than 170 potential predictors including:

• Pharmacy claims
• Demographic information
• Hospice 
• Low-income status
• Pharmacy travel distance
• Prescriber information

Two separate data sets were constructed.

1. New opioid users (no opioid in past six months) 
2. Current opioid users (1+ opioid claims in past months) 

Over 10 predictive modeling methods were used to assess both data sets. Here are the results: 

The best fitting predictive model for new opioid users was the decision tree model. The important independent predictors included:

• MME (i.e., the amount of morphine equivalent amount in each dose) of first opioid claim 
• Whether the first claim is a long-acting opioid 
• Number of other medications the individual has in their claims history

The best fitting predictive model for current opioid users was the logistic regression model. Against all models tested and against all factors, this model proved to be the most accurate. The most important independent predictors were:

• Average MME in most recent month 
• Type of opioid 

For both models, the prescriber specialty (e.g., family medicine) was an independent predictor. 

Will other models in the marketplace let you see “behind the curtain?” We do.

There are a few other predictive models in the marketplace. But they show only one model – for members with a history of opioid use. In addition, those other models in the marketplace don’t provide clear information on the data used, how the model was developed, top predictors, or the model’s performance. As a customer, that wouldn’t satisfy me. 

Our models have proven highly accurate

Our opioid prediction models were sound with 95 to more than 98 percent accuracy and specificity. We can identify the data sets we used, and provide details on our methodology.⁴

Prescription drugs are an important part of today’s opioid epidemic. And while managed care can’t influence certain aspects of the opioid epidemic, like heroin or illegally manufactured fentanyl, we can have an impact within our own sphere of influence. From all its opioid prescriptions, we can help a health plan identify the specific members who are more likely to progress to higher doses of opioids. This gives the health plan options. It can:

• Contact that member’s doctor
• Contact that member’s pharmacist
• Contact that member
• Track that member’s next prescription  

Our highly accurate predictive models score and rank a Medicare member on their future likelihood to escalate to high dose opioids. The score helps the health plan select a targeted outreach and intervention to prevent that future high-risk opioid use. 

References

1. HHS OIG Data Brief. Opioid Use in Medicare Part D Remains Concerning. June 2018. https://oig.hhs.gov/oei/reports/oei-02-18-00220.asp
2. Prime book of business.
3. Dowell D, Haegerich TM, Chou R. CDC Guideline for Prescribing Opioids for Chronic Pain — United States, 2016. MMWR 2016;65(1):1-49. doi:10.15585/mmwr.rr6501e1.
4. Yang Q, Gleason PP. Predicting High Dose Opioid Utilizers: Different Models are Required for New and Current Opioid Users. Academy of Managed Care Pharmacy; October 2018: Orlando, FL. J Manag Care Spec Pharm 2018;24(10-a Suppl):S94.