Across the U.S., the number of reported events exemplifying the opioid and heroin epidemics continues to skyrocket. U.S. Government Publishing Office data shows that the usage of both prescribed stimulants and prescribed opiates increased by a factor of 19 in just two decades since 1994(1). On Dec. 18, 2015, the U.S. Centers for Disease Control and Prevention (CDC) released a report showing drug overdose deaths reached record highs in 2014, fueled in large part by the abuse of narcotic painkillers and heroin. In 2014, more than 47,000 Americans died from drug overdoses, an increase of more than 14% from 2013. About 61% of those deaths involved the use of opioids. From 2000 to 2014, the report noted that nearly half a million people have died from overdoses in the U.S. In 2014, there were approximately one and a half times more drug overdose deaths than deaths from motor vehicle crashes!(2)
A very worrisome statistic and trend…
For workers’ compensation insurers, opioid use in treating chronic pain has also exploded over the past two decades. Although there appear to be some signs that opioid use is finally cresting, insurers still have a long way to go in helping to ensure that physicians and the injured workers they treat are fully educated on the pros and cons of using opioids with various types of injuries and pain. As the Risk & Insurance article "Paying for Detox – The Opioid Epidemic Is Addressed by Detoxification Programs" notes, some workers’ compensation insurers have been funding tapering and detoxification programs to help dependent or addicted patients wean themselves off the very medications that were designed to ease their pain(3). Unfortunately, recidivism is common, with experts noting that it can take several attempts to wean someone off narcotics.
This article will highlight some of the challenges in front of us and share some innovative ideas on potential ways to help prevent opioid dependency and addiction before the habits requiring tapering and detoxification programs are ever formed.
The Challenge in Front of Us
In January 2011, USA Today shared a powerful story about David Fridovich, a three-star Green Beret general who has become an advocate for warning soldiers about the epidemic of chronic pain and the use of narcotic pain relievers sweeping through the U.S. military(4). Much like others across the country who have suffered a severe back injury, the general began taking narcotics for chronic pain in 2006. Over time, the general became addicted to narcotics. During one 24-hour period the general took five dozen pain pills. After going through a detoxification program, the general has been helping other soldiers avoid the complications he faced because he was unaware of the addictive nature of the pills he was taking.
In a recent book about the opioid and heroin epidemic in the U.S., Dream Land author Sam Quinones shares his research on the history of how we ended up where we are today. From a workers’ compensation perspective, the author shared a story about a prison guard who had injured his back during a fight with an inmate. The doctor, who took the guard off of work for six months, also prescribed opioids to be taken twice a day for 30 days. After becoming severely addicted, the guard said, “It really humbles you. You think you’re doing stuff the way it’s supposed to be done. You’re trusting the doctor. After a while, you realize this isn’t right, but there really isn’t anything you can do about it. You’re stuck. You’re addicted.”
Both stories illustrate how the use of painkillers can lead to dependency and addiction without warning. They also highlight the critical role prescribing physicians play in educating patients about the warning signs and addictive nature of opioid prescriptions. As part of this education process, prescribing guidelines and analytics can play an important role in driving better outcomes.
Opioid Prescribing Guidelines
For workers’ compensation insurers, it is critical to understand the opioid prescribing guidelines that underlie the way physicians are treating injured workers. The more the insurers can help educate physicians on best practices, the better off insurance companies may be in helping to prevent any issues that may arise because of unnecessary or excessive opioid prescribing.
The CDC worked with the National Drug Institute, Substance Abuse and Mental Health Services Administration and the Office of the National Coordinator for Health Information Technology to review existing opioid prescribing guidelines for chronic pain. Their review and analysis of eight prescribing guidelines highlighted a number of important provider actions, such as the review of pain history, medical and family history, pregnancy, prescription drug monitoring programs (PDMP), urine drug screening, evaluations of alternatives to opioids, rational documentation, tapering plans, referrals for medication assisted treatment, evidence review, conflicts of interest and more(5). In January, Kentucky Attorney General Andy Beshear announced his support for national guidelines for prescribing opiates for chronic pain, stating: “In Kentucky, we face a crushing epidemic of addiction. One of my core missions as attorney general is to better address the drug problem faced by our Kentucky families and workforce.”(6) In his speech, the attorney general mentions that he is joining other state attorneys general in voicing support for the CDC guidelines for prescribing opiates for chronic pain.
California’s "Division of Workers’ Compensation Guideline for the Use of Opioids to Treat Work-Related Injuries" documented treatment protocols for three specific pain categories:
The Inspiration for Using Analytics
For more than a decade, Deloitte Consulting’s Advanced Analytics & Modeling practice has been developing claim predictive solutions designed to help insurance companies, self-insureds and third-party administrators better segment and triage predicted high-severity from low-severity claims, enabling business decisions and actions that can help drive loss cost savings of as much as 10% of an organization’s annual claims spending. (See Claims Magazine articles "Analytics on the Cloud: Transforming the Way Claims Leverages Advanced Analytics "(2011)(8), "Enhancing Workers’ Comp Predictive Modeling With Injury Groupings" (2012)(9), "Reaping the Financial Rewards of End-to-End Claims Analytics" (2014)(10) and "The Challenges of Implementing Advanced Analytics "(2014).(11) A large part of the claims modeling success is attributed to gaining actionable insights as early as first notice of loss before adverse chain reactions can set in, and shortly thereafter with the three-point contact investigation where additional information is learned about the patient’s history and co-morbidities.
The authors, having observed the success of predicting claims complexity outcomes early in the claim’s lifecycle, became excited about the application of similar models to help identify early warning signs of future excessive opioid usage by injured workers. With as much as 60% of workers' compensation spending going toward medical costs, one-fifth of that related to prescription drugs(12), we believed the use of predictive models… combined with physician peer-to-peer outreach and proper prescribing guidelines… could help workers' compensation insurers improve the lives of the injured workers while significantly reducing medical expenditures. The following sections explain the analytics journey undertaken to help move the needle on this issue.
Defining the Target Variable: Predicting Future Excess
An important part of any analytics journey is defining the target variable (i.e., what we are trying to understand and predict). Excessive opiates usage is difficult to ascertain, as higher consumption may indeed be necessary for the most severe injuries. Therefore, various tests on the most appropriate target variables were conducted to probe these hypotheses. Many versions of opioid supply days were tested (i.e., ultimate total supply days across all opiates drugs prescribed to, and consumed by, the injured worker). Variations of opiates prescription counts were also considered (i.e., ultimate count of opiates prescriptions through the lifecycle of the claims). Similarly, supply units were analyzed (i.e., ultimate sum of all individual opiates pills prescribed to, and consumed by, the injured worker from the day of the injury until the claim closure). Figure 1 illustrates the calculation of total supply days for three different opiates that were prescribed to, and consumed by, the injured worker over the duration of his workers’ compensation claim:
Figure 1. Supply Day Illustration
Methodology and Data Considered
Using predictive analytics and data science, a number of algorithms were built, tested, iterated and fine-tuned to better understand those like-injury cohorts (i.e., same injury sustained) that consumed more opiates than their corresponding peers who managed to consume a lower amount. Various thresholds of “excess” were analyzed by injury and venues, thus controlling for differences that affect the prescription base.
By testing these algorithms, it was determined that segmentation was similar across the different target variables. However, total supply days seemed to exhibit the most robustness from a modeling perspective and had intuitive interpretability (i.e., number of days an injured worker consumes opioids).
The algorithms used more than eight years of lost time workers' compensation claims to accumulate enough data credibility. Claims were selected for various injury groups where opiates were prescribed and consumed for at least one prescription. The data was organized for a longitudinal study observing a claimant over time and quantifying her consumption of opiates. The comparison to this usage to like-injury counterparts over thousands of cases and using hundreds of attributes is what helped the model shed light on claimants who consumed excessive amounts of opioids relative to the entire population.
Over the years, Deloitte healthcare practitioners and claims professionals used ICD-9 codes that describe a disease or condition, as well as National Council on Compensation (NCCI) nature of injury and body part codes, to create more than 70 proprietary injury groups that are factored into the model to provide enhanced segmentation within like injury claims.(13) For illustration purposes in this article, we presented results for the injury group representing medium- and high-complexity spinal disorders (e.g., ICD-9 codes 722.0 – displacement of cervical intervertebral disc without myelopathy, 722.10 – displacement of lumbar intervertebral disc without myelopathy, 724.9 – other unspecified back disorders, etc.). We selected medium- and high-complexity spinal disorder claims because they are significantly more severe than the average workers’ compensation claim, and, as expected, these claimants typically have more prescriptions filled by their physicians. In addition, the models aren’t run on just any injury group. For example, an injury group containing low-complexity injuries such as finger cuts and minor open wounds would not be part of our analysis. Claimants with these types of low-complexity injuries do not require opioids, given the nature of injury, so it would not make sense to include these injury groups in the model.
Predictive variables
The information attributes used to understand excessive consumption were sourced from similar data sources used in developing our claim-severity models. They are large in number and varied in terms of coverage. They include claimant data (e.g., claimant age, gender, job classification, years of employment, wage, claim filing lag, cause and nature of injury, etc.), prior claims data (e.g., prior frequency and type of claims), employer information (e.g., financial characteristics, years in business, etc.), injury circumstance (e.g. location, type, body part injured), three-point contact information (e.g., co-morbidities, early medical services) as well as other standard external third-party data sources (e.g. lifestyle, behavioral, geo-demographic).
Modeling Results
The lift curves shown in Figure 2 illustrate the segmentation achieved by using multivariate equations to predict total supply days. Each claim below was scored using the model, which generated scores from 1 to 100, with lower scores corresponding to smaller predicted supply days and higher scores corresponding to larger predicted supply days. This score is represented on the x-axis of Figure 2, where each “decile” refers to a group of claims that compose 10% of the data. The actual supply days are tracked and plotted on the y-axis in the appropriate decile.
Figure 2. Lift Curve – GLM model
As one can see from Figure 2, injured workers studied who are predicted to fall in decile 10 have more than 18 times the supply days as workers predicted to fall in decile 1. Injured workers studied who scored in decile 10 consume, on average, more than three and a half years of opioid supply days! This very large and widespread segmentation suggests that individuals sustaining the same injury can still vary significantly in their future consumption of opioids… and this variation ranges from a couple months to more than three and a half years.
In Figure 3, we compare two 24-year-old male claimants with very similar injuries but drastically different predicted outcomes.
Figure 3. Similar Injuries, Drastically Different Outcomes
As one can see from Figure 3, the claimant scoring in decile 10 has a number of variables that correlate with the potential for excessive opioid use. Given the combination of co-morbidities, worker health, reporting lags, employer business conditions and additional attributes collected on the individual from external sources (e.g. lifestyle and behavioral data), it is possible for the insurance company to identify and analyze the early drivers that may lead to future excessive opioid the first few days after receiving notice of the claim.
With more than 60 predictive variables in the model (e.g., co-morbidities, prior claims history, job classes, injury causes, business characteristics, claim characteristics, etc.), the most influential categories and reason codes driving the score represent “eyeglasses” for the insurance company physician. The model helps the insurance company physician weigh together multiple pieces of information but doesn’t replace his judgement. Analogously, many of us wear eyeglasses to read a dinner menu, but those eyeglasses do not order the food for us.
Armed with a plethora of facts and the opioid prescribing guidelines, a physician can open a dialogue with the treating physician to help guide the discussion in a direction that best benefits the injured worker. The physician, using the prediction from the model, can tailor appropriate decisions and actions – from low touch or regular prognosis for the first claimant above, to a much more closely managed case for the second individual.
Figure 4 provides a drill-down into the actual versus predicted supply days achieved in the highest-scoring 30% of medium- to high-complexity spinal disorder claims for the train/test data and validation data. Using the train/test/validation approach, the models were trained and enhanced using approximately 70% of the claims data. The validation results shown below were derived from the remaining 30% of the claims data that was held in “cold storage.” Using this kind of blind-test validation data helps ensure that the model’s estimated “lift” (i.e., segmentation power) is true and unbiased.
Figure 4. Highest Score Drill-Down
Approximately 60% of claims scoring in deciles 8, 9 and 10 exceed one year in supply days. For a quarter of the claims, the injured workers take in excess of four years in supply days of opioids. At the far end of the spectrum, roughly 4% of medium- to high-complexity spinal disorder claims scoring in deciles 8, 9 and 10 will exceed a decade's worth of opioids in supply days.
One Last Check
In addition to the generalized linear models (GLMs) discussed above, focused on predicting the actual supply days, we also ran a logistic regression model focused on predicting which claimants would take more than a year's supply of opioids. Using classical statistical measures of precision (i.e., how many of the positively classified results are relevant), recall (i.e., how accurate the model is at detecting the positives) and specificity (i.e., how good the model is at avoiding false alarms), we achieved the following results: a precision of 59%, a recall of 64% and a specificity of 72%.(14) As one last test of the logistic regression model’s segmentation power, we calculated the receiver operating characteristic (ROC) curve. At almost 80%, it represented a good model from a statistical perspective. Although illustrative, we prefer the GLM model presented above.
Behavioral Economics and Nudges
All across the country, physicians and medical boards are spreading the word about the responsible prescribing of opioids. State and federal agencies are toughening criminal and administrative penalties for doctors and clinics that traffic in prescription drugs. Governors across the country are forming opioid working groups that include senior Health and Human Services professionals, attorneys general, drug courts, hospital professionals, elected officials and more.
Research shows that a number of factors can help insurance companies better understand the severity of claims early on in the life cycle of a claim. Two studies by the National Council on Compensation Insurance, Inc. (NCCI) highlight the effect of obesity on workers’ compensation claims. According to "Reserving in the Age of Obesity," a Nov. 1, 2010, NCCI study by Chris Laws and Frank Schmid, the ratio in the medical costs per claim of obese to nonobese claimants deteriorates over time from a ratio of 2.8 at the end of one year, to 4.5 at the end of three years, to 5.3 at the end of five years.(15) In a following study from May 29, 2012, "Indemnity Benefit Duration and Obesity," authors Frank Schmid, Chris Laws and Mathew Montero found the duration of obese claimants is more than five times the duration of nonobese claimants, after controlling for primary International Classification of Diseases (ICD)-9 code, injury year, state, industry, gender and age for temporary total and permanent total indemnity benefit payments.(16) Deloitte’s claim predictive models have shown that the number of medical conditions at the time of injury plays a significant role in determining the ultimate severity and potential for excess opioid usage (e.g., claims with three or more existing medical conditions are 12 times more costly than claims with no existing medical conditions).
With energy and momentum building around addressing the opioid epidemic, insurance companies can leverage behavioral economics and data-driven nudges to help treating physicians improve outcomes and return to work. Leveraging prescribing guidelines and the model results and reason codes that help explain the top five drivers behind the model prediction, insurance company physicians can be more strategic in shaping the discussions they have with treating physicians. For the highest-scoring claims, the insurance company may want to use a mix of peer-to-peer contact and data-driven nudges (e.g., “did you know that 95% of physicians we work with follow the state prescribing guidelines and only prescribe 30 days of opioids for this type of claim,” ”for injuries of this type, physicians we work with usually prescribe less than x milligrams of strength,” etc.). For lower-scoring claims, the insurance company may touch base with the treating physician but skip any reference to data-driven nudges.
Conclusion
In the end, it is important for workers’ compensation insurers and their medical professionals to clearly understand opioid prescribing guidelines and the internal and external factors that could affect the opioid usage and habits of their injured workers. A Business Insurance white paper titled "Opioid Abuse and Workers' Comp – How to Tackle a Growing Problem," described the challenge well: “Monitoring or managing opioid abuse is another key step for workers' comp managers. It’s not enough to simply dive into the data and look for claimants who appear to be using lots of opioids. Nor is preventing doctors from prescribing opioids a desirable action. The goal is to find claimants who are struggling with a problem they never intended to have, and support those claimants in solving that problem.”(17)
However, our hope is that through the use of predictive analytics (i.e., the ability to identify, in the first few days of receiving a claim, individuals most likely to become high consumers of opioids), prescribing guidelines and physician peer-to-peer outreach, we can help increase insurers’ and treating physicians’ awareness as they work to help prevent injured workers from struggling with dependency and addiction before the behaviors or habits ever form.
As former British Prime Minister Benjamin Disraeli once said, “What we anticipate seldom occurs; what we least expect generally happens.” The science and passion exists today to better anticipate opioid trends and help prevent opioid dependency and addiction before it happens.
As used in this document, "Deloitte" means Deloitte Consulting LLP, a subsidiary of Deloitte LLP. Please see www.deloitte.com/us/about for a detailed description of the legal structure of Deloitte LLP and its subsidiaries. Certain services may not be available to attest clients under the rules and regulations of public accounting.
This communication contains general information only, and none of Deloitte Touche Tohmatsu Limited, its member firms, or their related entities (collectively, the “Deloitte Network”) is, by means of this communication, rendering professional advice or services. Before making any decision or taking any action that may affect your finances or your business, you should consult a qualified professional adviser. No entity in the Deloitte Network shall be responsible for any loss whatsoever sustained by any person who relies on this communication.
Copyright © 2016 Deloitte Development LLC. All rights reserved.
[1] James W. Harris, PhD, CSO Vatex Explorations LLC, www.gpo.gov
[2] http://www.cdc.gov/media/releases/2015/p1218-drug-overdose.html, http://www.cdc.gov/mmwr/preview/mmwrhtml/mm6450a3.htm?s_cid=mm6450a3_w
[3] http://www.riskandinsurance.com/paying-detox/
[4] http://usatoday30.usatoday.com/news/military/2011-01-27-1Adruggeneral27_CV_N.htm
[5] http://www.cdc.gov/drugoverdose/prescribing/common-elements.html
[6] http://harlandaily.com/news/6473/cdc-guidelines-will-help-ky-with-rx-drug-abuse
[7] http://www.dir.ca.gov/dwc/ForumDocs/Opioids/OpioidGuidelinesPartA.pdf
[8] http://www.propertycasualty360.com/2011/02/22/leveraging-analytics-in-workers-comp-claims-handli
[9] http://www.propertycasualty360.com/2012/07/23/enhance-workers-comp-predictive-modeling-with-inju
[10] http://www.propertycasualty360.com/2014/02/03/reaping-the-financial-rewards-of-end-to-end-claims
[11] http://www.propertycasualty360.com/2014/10/01/the-challenges-of-implementing-advanced-analytics
[12] www.ncci.com
[13] http://www.propertycasualty360.com/2012/07/23/enhance-workers-comp-predictive-modeling-with-inju
[14] Precision measures the ratio of true predicted positives to the ratio of true predictive positives plus false predicted positives. Recall, also referred to as sensitivity, measures the ratio of true predicted positives to the ratio of true predicted positives plus false predicted negatives. Specificity measures the ratio of true predicted negatives to the ratio of true predicted negatives plus false predicted positives.
[15] https://www.ncci.com/Articles/Documents//II_research-age-of-obesity.pdf
[16] https://www.ncci.com/Articles/Documents/II_Obesity-2012.pdf
[17] http://www.businessinsurance.com/article/99999999/WP05/120509952
- Opioids for acute pain (pain lasting as much as four weeks from onset)
- Opioids for subacute pain (one to three months)
- Opioids for chronic pain and chronic opioid treatment (three months or more)(7)
The Inspiration for Using Analytics
For more than a decade, Deloitte Consulting’s Advanced Analytics & Modeling practice has been developing claim predictive solutions designed to help insurance companies, self-insureds and third-party administrators better segment and triage predicted high-severity from low-severity claims, enabling business decisions and actions that can help drive loss cost savings of as much as 10% of an organization’s annual claims spending. (See Claims Magazine articles "Analytics on the Cloud: Transforming the Way Claims Leverages Advanced Analytics "(2011)(8), "Enhancing Workers’ Comp Predictive Modeling With Injury Groupings" (2012)(9), "Reaping the Financial Rewards of End-to-End Claims Analytics" (2014)(10) and "The Challenges of Implementing Advanced Analytics "(2014).(11) A large part of the claims modeling success is attributed to gaining actionable insights as early as first notice of loss before adverse chain reactions can set in, and shortly thereafter with the three-point contact investigation where additional information is learned about the patient’s history and co-morbidities.
The authors, having observed the success of predicting claims complexity outcomes early in the claim’s lifecycle, became excited about the application of similar models to help identify early warning signs of future excessive opioid usage by injured workers. With as much as 60% of workers' compensation spending going toward medical costs, one-fifth of that related to prescription drugs(12), we believed the use of predictive models… combined with physician peer-to-peer outreach and proper prescribing guidelines… could help workers' compensation insurers improve the lives of the injured workers while significantly reducing medical expenditures. The following sections explain the analytics journey undertaken to help move the needle on this issue.
Defining the Target Variable: Predicting Future Excess
An important part of any analytics journey is defining the target variable (i.e., what we are trying to understand and predict). Excessive opiates usage is difficult to ascertain, as higher consumption may indeed be necessary for the most severe injuries. Therefore, various tests on the most appropriate target variables were conducted to probe these hypotheses. Many versions of opioid supply days were tested (i.e., ultimate total supply days across all opiates drugs prescribed to, and consumed by, the injured worker). Variations of opiates prescription counts were also considered (i.e., ultimate count of opiates prescriptions through the lifecycle of the claims). Similarly, supply units were analyzed (i.e., ultimate sum of all individual opiates pills prescribed to, and consumed by, the injured worker from the day of the injury until the claim closure). Figure 1 illustrates the calculation of total supply days for three different opiates that were prescribed to, and consumed by, the injured worker over the duration of his workers’ compensation claim:
Figure 1. Supply Day Illustration
Methodology and Data Considered
Using predictive analytics and data science, a number of algorithms were built, tested, iterated and fine-tuned to better understand those like-injury cohorts (i.e., same injury sustained) that consumed more opiates than their corresponding peers who managed to consume a lower amount. Various thresholds of “excess” were analyzed by injury and venues, thus controlling for differences that affect the prescription base.
By testing these algorithms, it was determined that segmentation was similar across the different target variables. However, total supply days seemed to exhibit the most robustness from a modeling perspective and had intuitive interpretability (i.e., number of days an injured worker consumes opioids).
The algorithms used more than eight years of lost time workers' compensation claims to accumulate enough data credibility. Claims were selected for various injury groups where opiates were prescribed and consumed for at least one prescription. The data was organized for a longitudinal study observing a claimant over time and quantifying her consumption of opiates. The comparison to this usage to like-injury counterparts over thousands of cases and using hundreds of attributes is what helped the model shed light on claimants who consumed excessive amounts of opioids relative to the entire population.
Over the years, Deloitte healthcare practitioners and claims professionals used ICD-9 codes that describe a disease or condition, as well as National Council on Compensation (NCCI) nature of injury and body part codes, to create more than 70 proprietary injury groups that are factored into the model to provide enhanced segmentation within like injury claims.(13) For illustration purposes in this article, we presented results for the injury group representing medium- and high-complexity spinal disorders (e.g., ICD-9 codes 722.0 – displacement of cervical intervertebral disc without myelopathy, 722.10 – displacement of lumbar intervertebral disc without myelopathy, 724.9 – other unspecified back disorders, etc.). We selected medium- and high-complexity spinal disorder claims because they are significantly more severe than the average workers’ compensation claim, and, as expected, these claimants typically have more prescriptions filled by their physicians. In addition, the models aren’t run on just any injury group. For example, an injury group containing low-complexity injuries such as finger cuts and minor open wounds would not be part of our analysis. Claimants with these types of low-complexity injuries do not require opioids, given the nature of injury, so it would not make sense to include these injury groups in the model.
Predictive variables
The information attributes used to understand excessive consumption were sourced from similar data sources used in developing our claim-severity models. They are large in number and varied in terms of coverage. They include claimant data (e.g., claimant age, gender, job classification, years of employment, wage, claim filing lag, cause and nature of injury, etc.), prior claims data (e.g., prior frequency and type of claims), employer information (e.g., financial characteristics, years in business, etc.), injury circumstance (e.g. location, type, body part injured), three-point contact information (e.g., co-morbidities, early medical services) as well as other standard external third-party data sources (e.g. lifestyle, behavioral, geo-demographic).
Modeling Results
The lift curves shown in Figure 2 illustrate the segmentation achieved by using multivariate equations to predict total supply days. Each claim below was scored using the model, which generated scores from 1 to 100, with lower scores corresponding to smaller predicted supply days and higher scores corresponding to larger predicted supply days. This score is represented on the x-axis of Figure 2, where each “decile” refers to a group of claims that compose 10% of the data. The actual supply days are tracked and plotted on the y-axis in the appropriate decile.
Figure 2. Lift Curve – GLM model
As one can see from Figure 2, injured workers studied who are predicted to fall in decile 10 have more than 18 times the supply days as workers predicted to fall in decile 1. Injured workers studied who scored in decile 10 consume, on average, more than three and a half years of opioid supply days! This very large and widespread segmentation suggests that individuals sustaining the same injury can still vary significantly in their future consumption of opioids… and this variation ranges from a couple months to more than three and a half years.
In Figure 3, we compare two 24-year-old male claimants with very similar injuries but drastically different predicted outcomes.
Figure 3. Similar Injuries, Drastically Different Outcomes
As one can see from Figure 3, the claimant scoring in decile 10 has a number of variables that correlate with the potential for excessive opioid use. Given the combination of co-morbidities, worker health, reporting lags, employer business conditions and additional attributes collected on the individual from external sources (e.g. lifestyle and behavioral data), it is possible for the insurance company to identify and analyze the early drivers that may lead to future excessive opioid the first few days after receiving notice of the claim.
With more than 60 predictive variables in the model (e.g., co-morbidities, prior claims history, job classes, injury causes, business characteristics, claim characteristics, etc.), the most influential categories and reason codes driving the score represent “eyeglasses” for the insurance company physician. The model helps the insurance company physician weigh together multiple pieces of information but doesn’t replace his judgement. Analogously, many of us wear eyeglasses to read a dinner menu, but those eyeglasses do not order the food for us.
Armed with a plethora of facts and the opioid prescribing guidelines, a physician can open a dialogue with the treating physician to help guide the discussion in a direction that best benefits the injured worker. The physician, using the prediction from the model, can tailor appropriate decisions and actions – from low touch or regular prognosis for the first claimant above, to a much more closely managed case for the second individual.
Figure 4 provides a drill-down into the actual versus predicted supply days achieved in the highest-scoring 30% of medium- to high-complexity spinal disorder claims for the train/test data and validation data. Using the train/test/validation approach, the models were trained and enhanced using approximately 70% of the claims data. The validation results shown below were derived from the remaining 30% of the claims data that was held in “cold storage.” Using this kind of blind-test validation data helps ensure that the model’s estimated “lift” (i.e., segmentation power) is true and unbiased.
Figure 4. Highest Score Drill-Down
Approximately 60% of claims scoring in deciles 8, 9 and 10 exceed one year in supply days. For a quarter of the claims, the injured workers take in excess of four years in supply days of opioids. At the far end of the spectrum, roughly 4% of medium- to high-complexity spinal disorder claims scoring in deciles 8, 9 and 10 will exceed a decade's worth of opioids in supply days.
One Last Check
In addition to the generalized linear models (GLMs) discussed above, focused on predicting the actual supply days, we also ran a logistic regression model focused on predicting which claimants would take more than a year's supply of opioids. Using classical statistical measures of precision (i.e., how many of the positively classified results are relevant), recall (i.e., how accurate the model is at detecting the positives) and specificity (i.e., how good the model is at avoiding false alarms), we achieved the following results: a precision of 59%, a recall of 64% and a specificity of 72%.(14) As one last test of the logistic regression model’s segmentation power, we calculated the receiver operating characteristic (ROC) curve. At almost 80%, it represented a good model from a statistical perspective. Although illustrative, we prefer the GLM model presented above.
Behavioral Economics and Nudges
All across the country, physicians and medical boards are spreading the word about the responsible prescribing of opioids. State and federal agencies are toughening criminal and administrative penalties for doctors and clinics that traffic in prescription drugs. Governors across the country are forming opioid working groups that include senior Health and Human Services professionals, attorneys general, drug courts, hospital professionals, elected officials and more.
Research shows that a number of factors can help insurance companies better understand the severity of claims early on in the life cycle of a claim. Two studies by the National Council on Compensation Insurance, Inc. (NCCI) highlight the effect of obesity on workers’ compensation claims. According to "Reserving in the Age of Obesity," a Nov. 1, 2010, NCCI study by Chris Laws and Frank Schmid, the ratio in the medical costs per claim of obese to nonobese claimants deteriorates over time from a ratio of 2.8 at the end of one year, to 4.5 at the end of three years, to 5.3 at the end of five years.(15) In a following study from May 29, 2012, "Indemnity Benefit Duration and Obesity," authors Frank Schmid, Chris Laws and Mathew Montero found the duration of obese claimants is more than five times the duration of nonobese claimants, after controlling for primary International Classification of Diseases (ICD)-9 code, injury year, state, industry, gender and age for temporary total and permanent total indemnity benefit payments.(16) Deloitte’s claim predictive models have shown that the number of medical conditions at the time of injury plays a significant role in determining the ultimate severity and potential for excess opioid usage (e.g., claims with three or more existing medical conditions are 12 times more costly than claims with no existing medical conditions).
With energy and momentum building around addressing the opioid epidemic, insurance companies can leverage behavioral economics and data-driven nudges to help treating physicians improve outcomes and return to work. Leveraging prescribing guidelines and the model results and reason codes that help explain the top five drivers behind the model prediction, insurance company physicians can be more strategic in shaping the discussions they have with treating physicians. For the highest-scoring claims, the insurance company may want to use a mix of peer-to-peer contact and data-driven nudges (e.g., “did you know that 95% of physicians we work with follow the state prescribing guidelines and only prescribe 30 days of opioids for this type of claim,” ”for injuries of this type, physicians we work with usually prescribe less than x milligrams of strength,” etc.). For lower-scoring claims, the insurance company may touch base with the treating physician but skip any reference to data-driven nudges.
Conclusion
In the end, it is important for workers’ compensation insurers and their medical professionals to clearly understand opioid prescribing guidelines and the internal and external factors that could affect the opioid usage and habits of their injured workers. A Business Insurance white paper titled "Opioid Abuse and Workers' Comp – How to Tackle a Growing Problem," described the challenge well: “Monitoring or managing opioid abuse is another key step for workers' comp managers. It’s not enough to simply dive into the data and look for claimants who appear to be using lots of opioids. Nor is preventing doctors from prescribing opioids a desirable action. The goal is to find claimants who are struggling with a problem they never intended to have, and support those claimants in solving that problem.”(17)
However, our hope is that through the use of predictive analytics (i.e., the ability to identify, in the first few days of receiving a claim, individuals most likely to become high consumers of opioids), prescribing guidelines and physician peer-to-peer outreach, we can help increase insurers’ and treating physicians’ awareness as they work to help prevent injured workers from struggling with dependency and addiction before the behaviors or habits ever form.
As former British Prime Minister Benjamin Disraeli once said, “What we anticipate seldom occurs; what we least expect generally happens.” The science and passion exists today to better anticipate opioid trends and help prevent opioid dependency and addiction before it happens.
As used in this document, "Deloitte" means Deloitte Consulting LLP, a subsidiary of Deloitte LLP. Please see www.deloitte.com/us/about for a detailed description of the legal structure of Deloitte LLP and its subsidiaries. Certain services may not be available to attest clients under the rules and regulations of public accounting.
This communication contains general information only, and none of Deloitte Touche Tohmatsu Limited, its member firms, or their related entities (collectively, the “Deloitte Network”) is, by means of this communication, rendering professional advice or services. Before making any decision or taking any action that may affect your finances or your business, you should consult a qualified professional adviser. No entity in the Deloitte Network shall be responsible for any loss whatsoever sustained by any person who relies on this communication.
Copyright © 2016 Deloitte Development LLC. All rights reserved.
[1] James W. Harris, PhD, CSO Vatex Explorations LLC, www.gpo.gov
[2] http://www.cdc.gov/media/releases/2015/p1218-drug-overdose.html, http://www.cdc.gov/mmwr/preview/mmwrhtml/mm6450a3.htm?s_cid=mm6450a3_w
[3] http://www.riskandinsurance.com/paying-detox/
[4] http://usatoday30.usatoday.com/news/military/2011-01-27-1Adruggeneral27_CV_N.htm
[5] http://www.cdc.gov/drugoverdose/prescribing/common-elements.html
[6] http://harlandaily.com/news/6473/cdc-guidelines-will-help-ky-with-rx-drug-abuse
[7] http://www.dir.ca.gov/dwc/ForumDocs/Opioids/OpioidGuidelinesPartA.pdf
[8] http://www.propertycasualty360.com/2011/02/22/leveraging-analytics-in-workers-comp-claims-handli
[9] http://www.propertycasualty360.com/2012/07/23/enhance-workers-comp-predictive-modeling-with-inju
[10] http://www.propertycasualty360.com/2014/02/03/reaping-the-financial-rewards-of-end-to-end-claims
[11] http://www.propertycasualty360.com/2014/10/01/the-challenges-of-implementing-advanced-analytics
[12] www.ncci.com
[13] http://www.propertycasualty360.com/2012/07/23/enhance-workers-comp-predictive-modeling-with-inju
[14] Precision measures the ratio of true predicted positives to the ratio of true predictive positives plus false predicted positives. Recall, also referred to as sensitivity, measures the ratio of true predicted positives to the ratio of true predicted positives plus false predicted negatives. Specificity measures the ratio of true predicted negatives to the ratio of true predicted negatives plus false predicted positives.
[15] https://www.ncci.com/Articles/Documents//II_research-age-of-obesity.pdf
[16] https://www.ncci.com/Articles/Documents/II_Obesity-2012.pdf
[17] http://www.businessinsurance.com/article/99999999/WP05/120509952
