By Joe Anderson, Director of Analytics and Andrew Spaner, Senior Statistical Analyst
Workers’ compensation payers are continuously seeking ways to mitigate claim costs and improve outcomes. Pharmacy-related claim expense is often an area of emphasis, not only because use of medications to treat workplace injury and illness is prevalent, but because of the influence medication utilization has on overall claim cost and duration. Long-term pharmacy use correlates with high overall claim cost and has been found to delay recovery and return to work. Thus, the pharmacy benefit manager (PBM) can be a valuable resource. Not only does a PBM offer network discounts on prescriptions, but they can help payers with understanding and optimizing the medication therapy regimen to improve overall claim outcomes. In recent years, there has been increasing interest in the use of early intervention programs to mitigate potential high-risk, misuse, abuse situations. However, before intervention can occur, the claim must first be identified.
Defining Triggers and Predictive Modeling
There are a variety of methods PBMs can use for the early detection and identification of high-risk claims, two of which are clinical triggers and predictive modeling.
Clinical triggers flag particular behaviors associated with a claim, such as using two short-acting opioid analgesics or going to multiple prescribers. The presence of a trigger typically results in an alert to the adjustor notifying of the potential for the claim to become high-risk. In addition, triggers act as a benchmark to compare one set of claims to another. For example, consider morphine equivalent dose (MED). As claims age, the trigger can be used to determine if a particular claim requires additional review because an industry-accepted benchmark, in this case 90 mg MED, is being surpassed.
Predictive modeling is a process that paints a holistic picture of a claim by analyzing several aspects of the claim at one time. Utilizing a statistical model, predictive analytics looks for a specific result, such as pharmacy cost, and combines all the known data elements to provide the most accurate prediction of that result. Modeling is not limited to a single aspect of transactional data about the medications. Rather, modeling also takes into account more descriptive data, such as the types of doctors a patient is visiting, demographic data, details about the injury, and an aggregation of spend and prescriptions over the analysis period. As an example, at Helios, the level of risk associated with the claim is determined by our predictive model’s risk score output, which is the culmination of all the variables to predict what the claimant’s future pharmacy spend will be.
Once high-risk claims are identified, either using triggers or a predictive model, a PBM’s clinical team can analyze the claim further to determine whether intervention is necessary. Understanding which claims are more likely to become high risk and which ones are not helps determine where to apply clinical resources and intervention tools based on the severity of the concern.
Demonstrating the Difference
In order to demonstrate the difference between triggers and predictive modeling, we compared how accurately the two methods identified claimants with long-term pharmacy use. To achieve this comparison, we analyzed a set of claimants injured in the first quarter of 2010. For this group, we predicted their pharmacy costs using predictive modeling and we determined which claims would have been flagged by an industry-standard clinical trigger. For the testing portion we looked at the current long-term spend for this claim population to understand:
- How many claimants actually had long-term spend
- How much long-term pharmacy cost is captured by looking at the high-risk claims from the predictive model
- How much long-term pharmacy cost is captured by looking at the claims that meet the criteria of individual triggers
In each case the predictive modeling approach allows us to target claims with actual long-term spend more efficiently. In fact, the trigger that flagged the most long-term spend was for claims with a morphine equivalent dose (MED) of at least 90 mg. While 34% of the claims in the analysis met the criteria of this trigger, the claim population accounted for only 47% of the long-term spend. However, using the predictive model, not only did different cases comprise 34% of the highest-risk claims, but the identified claim population accounted for 80% of the actual long-term cost, a 33 percentage point difference. Following is a sample of the results of our analysis.
Claimant-A: High Spend Trigger
Claimant-A ruptured her disc at work. She met the criteria for the clinical trigger of high spend because she had an annualized pharmacy spend of at least $3,000 in the first three months of the claim life. However, the predictive model assigned a low risk score to this claimant. In this example, contemplating data in addition to cost allowed the model to accurately assign a low likelihood of long-term pharmacy utilization:
- There were not any opioid analgesic pain medications used during the first 4 months of the claim
- The claimant did not use other pharmaceuticals that are indicative of long-term, high-cost claims, such as anticonvulsants, anti-anxiety, and cardiovascular medications
It is important to review all of the information at your disposal. Using predictive modeling facilitates this. While the claimant was using anti-rheumatic medications, which happen to be expensive, in our experience, these medications are not correlated with long-term claims.
Claimant-B: More than 90 mg MED
Claimant-B suffered from a shoulder sprain. He met the criteria for the clinical trigger of more than 90 mg Morphine Equivalent Dose (MED); however, the predictive model’s risk score assigned to this claim was low. In addition to the claimant’s use of opioid analgesics, the model also considered the following, assigning a lower score as a result:
- The claimant had very low spend in their first four months of service at just 25% of the average
- The claimant was not taking high-risk medications
- Though he met the high daily MED threshold, his overall MED for a four month period was low at just 50% of the average claim in this analysis
- The claimant went to a family medicine physician, which in our experience is considered a low-risk prescriber
While a high level of MED can be a concern, the predictive model considers broader factors. In this case, by also considering additional factors, the predictive model accurately predicted that this would not become a high-cost claim.
Claimant-C: Multiple Prescribers
Claimant-C experienced a strain to his lower back. He met the criteria for the clinical trigger of going to multiple prescribers in the same month. Tracking whether an injured worker receives prescriptions from multiple prescribers can be important because if there is not a knowledge of a patient’s entire medication therapy regimen, there is the potential for drug-drug interactions, therapeutic duplication, side effects, or other clinical concerns. Additionally, the use of multiple prescribers could be indicative of drug seeking behavior. However, while it is important to understand the prescribers involved with the claimant’s care, the use of this data point as an indicator of potential high-cost is not always the best predictor. In this case, the claimant was low-risk because:
- The claim had average pharmacy spend through four months of activity
- The claimant was not using high-risk medications
- The claimant had a very low-risk injury
In this example, reliance on the trigger alone may have also resulted in the unnecessary use of resources, such as clinical intervention tools.
Claimant-D: No Triggers – High Risk
Claimant-D also ruptured his disc, but unlike Claimant A discussed above, did not meet the criteria for any clinical triggers. However, the claim was assigned a high risk score using the predictive model because of the following:
- The claimant had 10 times the average spend of an average four-month-old claim in the analysis, but not enough spend to meet the “annualized spend of $3,000” trigger
- The claimant was using a large amount of a high-risk medication
- The claimant had a very high-risk injury
Utilizing clinical triggers, this claim would not have been flagged as high risk. As such, it is likely there would not have been clinical resources applied to the claim until much later on, if at all.
Claimant-E: No Triggers – Low Risk
Finally, Claimant-E was diagnosed with a hernia. She did not meet the criteria for any clinical triggers. The predictive model also found her claim to be low risk due to the following:
- Very low spend through four months of service at less than 5% of the average
- The claimant was not taking any high-risk medications
- The claimant had a low MED at 30% of the average
- The claimant had a very low-risk injury
- The claimant was seeing a physician with a very low-risk specialty
A predictive model is able to identify low-risk claims as well as high-risk, whereas triggers cannot. The function of a trigger is to predict long-term pharmacy spend, therefore, a trigger can only cause the utilization of clinical resources. Triggers cannot tell us when not to allocate resources to a claim. An advantage of predictive modeling is not only to produce predictions of long-term spend, but to ensure the right clinical resources are being utilized at the right time, helping ensure claimants receive safe and efficacious therapy.
While triggers can be useful in raising awareness for potential behaviors or the presence of a potential clinical concern, there are limitations. Furthermore, a trigger’s narrow focus can be misleading, resulting in the misapplication of resources, inefficiency, and lost opportunity. Payers are therefore encouraged to consider the value of predictive modeling to identify potential high-cost claims. As highlighted by the examples above, the ability to examine multiple factors can positively influence the outcome of a claim, both clinically and financially.
About Joe Anderson
As the Director of Analytics, Joe Anderson researches ways to utilize data to develop new services for our clients. These innovative programs utilize historical pharmacy data to predict which injured workers will have long-term pharmacy spend, or long-term complications with prescription drug misuse or abuse issues. Joe leads a team of statisticians to develop and implement these programs.
Joe is an expert in the predictors of pharmacy spend and addiction issues. Prior to joining Helios in 2010, Joe worked with Stax, a strategy consulting firm, where he performed corporate strategy analytics and private equity due diligence.
Joe Anderson has a bachelor’s degree from the University of Chicago and a Masters of Business Administration from Northwestern University’s Kellogg School of Management.
About Andrew Spaner
As a senior statistical analyst with more than five years of experience, Andrew Spaner analyzes industry trends and all available information from the workers’ compensation industry’s largest set of pharmacy data, identifying potential high-risk claims and addiction issues in order for our Clinical Services team to intervene.
Andrew Spaner has a bachelor’s degree in Mathematics with a concentration in Actuarial Science from Kent State University.
Helios brings the focus of workers’ compensation and auto no-fault Pharmacy Benefit Management, Ancillary, and Settlement Solutions back to where it belongs – the injured person. This comes with a passion and intensity on delivering value beyond just the transactional savings for which we excel. To learn how our creative and innovative tools, expertise, and industry leadership can help your business shine, visit www.HeliosComp.com.
Disclosure: Helios is a WorkCompWire Ad Partner.
This is not a paid placement.