Editor’s Note: The following article has been contributed by Scott Palmer, president & CEO of Injury Sciences LLC.
Predictive analytics is a valuable tool and is being applied to a growing number of areas in auto insurance operations. In recent months, I have fielded questions about the benefit of applying predictive analytics to a combination of auto physical damage data and medical data to identify questionable injuries.
To be sure, predictive analytics has shown benefits in claims operations by improving fraud referrals, identifying subrogation opportunities and “right tracking” claims assignments. Consequently, on the surface, the approach sounds appealing. However, a closer look into what predictive analytics can offer and the constitution of the data employed reveals a problematic landscape.
In the book “Competing on Analytics,” authored by Tom Davenport and Jeanne Harris, analytics is described as “the extensive use of data, statistical and quantitative analysis, explanatory and predictive models and fact-based management to drive decisions and actions.”
Further, in the table below, Davenport and Harris outline a succinct progression of business analytics and intelligence. Note the table presented in “Competing on Analytics” was adapted from a graphic produced by SAS and used with permission. Each method presented requires increasing sophistication—that is, optimization is a more sophisticated level than predictive modeling.
In the deployment of analytics, most will concur that the usefulness of results will depend greatly on the quality of the data, the appropriateness of the data analysis, and the quality of assumptions employed. It is also important to note that when models are properly deployed, they do not provide answers. Rather, they yield information about a tighter distribution on possible outcomes.
‘Smoke Alarms’ for Claims Organizations
Some have suggested the opportunity to use predictive analytics to identify problematic or questionable injury claims is analogous to providing a claims organization a “smoke alarm.” The analytics alert one to a problem early before it becomes a bigger problem. This analogy, while inaccurate, is actually instructive regarding the proper use of predictive analytics.
A smoke alarm detects smoke, which is an outcome that is caused by an actual fire. Hopefully it alerts one of an actual fire before it becomes a bigger fire. Alternatively, effective predictive analytics applications are more consistent with alerting one to conditions that are favorable for a fire taking place. One would want to investigate the actual existence of the fire before taking action—calling 911, activating a fire suppression system, and so on—because there can be considerable costs to such actions or decisions when there are false alarms. Alternatively, there are also instances when a fire does not matter because it may be a better outcome to just let it take its course if it in fact develops.
So, while there still may be value in knowing favorable conditions for a fire exist early, an investigative step or process is usually recommended to confirm the actual existence or potential impact of the problem before decisions are made to act on the information. This is why in fraud analytics applications, fraud investigations are usually conducted after potentially fraudulent claims are flagged and before final decisions about fraud are made. In subrogation applications, subrogation opportunities are normally further evaluated before being acted upon.
Investigation Beyond Analytics
Let’s examine why further investigative steps might be required when basing analytics, in part, on medical data. Medical data usually available to the auto insurance industry is accumulated from injury claims presented by claimants, their attorneys or their medical care providers. This information is often audited for reasonableness and appropriateness, in both first and third-party injury claims. The entire process of collecting the data assumes that the basis for the claim, specifically the auto accident, actually caused the claimed injury and the need for treatment. Next, we will explore how this assumption creates the opportunity for the data to include outcomes not caused by the accident (no causality) and outcomes that include over-treatment of an injury potentially caused by the accident (one form of an abuse in the system).
The assumption that causation exists is actually perpetuated by medical providers. Physicians, for example, are classically trained to take a history from the patient, conduct a physical examination, and then incorporate the findings of the previous two steps into his/her diagnosis and treatment recommendations. Therefore, if someone claims that a condition is directly related to an auto accident, then the physician factors this information into his or her treatment plan and cites the auto accident as the cause.
This problem is further compounded when the treatment plan is tailored to the severity of the injury as perceived by the medical provider, not by an analysis of the severity of the collision and the imparted stresses and strains. Subsequently, treatment guidelines are often used, regardless of the severity of the impact—in terms of collision energy—to identify when treatment is reasonable and appropriate. However, this practice still helps identify other forms of abuse in the system.
Flaws In Medical Data
These real-world practices begin to illuminate the flaws in the medical data maintained by auto insurance companies and their service providers: Medical treatments are often not prescribed or subsequently evaluated in light of the physical events required to produce the need for them. This flaw is inherent in the data maintained by utilization review or medical audit providers. From a scientific perspective, injuries are actually caused by a very specific physical stress or strain, or specific combinations of stresses and strains unique to the injury. These requisite stresses and strains can be found to exist (or not exist) based on an analysis of the physics (vehicle accelerations or decelerations) of the accident. An investigation would also take into account the position of the occupant in the vehicle, the use of restraint systems, and various other factors. Additionally, when requisite stresses and strains exist, they still must exceed an individual’s tolerance to same, before the injury can be caused.
Reliable Models To Assess Injury Claims
Without such a scientific analysis, how can a professional determine which medical data within a data set is questionable? Wouldn’t this knowledge be required to build a sound model to identify questionable claims so that auto physical damage relationships can be developed for both groups?
From a holistic view, it would also seem there would be much learned about questionable injury claims by understanding data in instances when injuries are not claimed as a result of an auto collision. The good news is that accidents that do not cause injuries occur very frequently. This outcome is corroborated by human subject testing in low-speed crash tests. More than 75 percent of the approximately 4,000 scientific human subject test exposures known to the author produced no injury. The bad news is that little claim data is typically collected by an insurer on an uninjured passenger when an injury claim is not made and no injury feature is created.
Sometimes no data is available because no claim is made as a result of an accident. Wouldn’t some information about the condition and physical attributes of the individual not making an injury claim, his or her seating position in the vehicle, use of available restraint systems, and so on be relevant to a reliable model that predicts questionable injury claims? Other times, injuries occur but data collected is incomplete because there is a determination that there is either no coverage or liability. Again, wouldn’t more complete information under these circumstances be important to a reliable model?
Combining Auto Physical Damage and Medical Data
Structurally, one can begin to see how trying to apply predictive analytics to a combination of auto physical damage data and medical data to identify questionable injuries can be problematic. To illustrate, let’s examine a hypothetical temporomandibular disorder (TMJ) injury claim from a low-energy frontal collision from a statistical and scientific perspective. Statistically, data will exist that involve accidents with varying degrees of physical damage to the claimant’s automobile accompanied by a TMJ injury claim.
Most likely, these injuries will be observed at a low to very low incident rate. Data will also likely show reasonable and customary costs for treatment of the TMJ injury and likely reflect deviations from these standards. So, does the statistical model suggest the injury should be questioned? If so why, then why? Was this because a treatment period was too long or not properly coded? Or, perhaps the medical provider was not a physician? What about the location of the clinic? What gives the claims adjuster the basis to make a decision or take an action and defend it?
Scientifically, the answer is straightforward. To traumatically injure TMJ, there must be contact between the mandible and an object with sufficient force to create the stresses and strains to cause injury. Simply, if there is no mandible contact with an object, then there is no opportunity for a TMJ injury (none of the previously referenced human subject test exposures experienced a mandible strike or a TMJ injury). Can an occupant in a vehicle involved in a frontal collision strike his or her mandible on an object? It is certainly possible. The answer becomes clearer once we know where the occupant was seated in the vehicle and whether he or she was restrained. Incidentally, both of these critical facts are typically not found in the medical data.
‘Old School’ Attributes
The scientific analysis as described above can be applied to a variety of injuries, including neck, back, shoulder and knee, to actually determine when questionable injuries were (or were not) caused from a collision. To use a popular example illustrated in the book Moneyball by Michael Lewis, there was an important difference in the predictive outcome in baseball games between the “old school” use of batting average to evaluate players and newer analytics using on-base percentage. While admittedly not a perfect analogy, using medical data in the proposed approach in lieu of scientific analysis certainly has many “old school” attributes.
So what can be done? First, understanding the limitations of the data and the resulting implications are critical. Use of statistical methods for dealing with unknowns and data limitations is common in predictive analytics. However, when these methods are used in claims applications, they should be followed by investigative processes that resolve the related unknowns and validate the assumptions employed. In the instance of questionable injury claims, an assessment of causality and, when appropriate, an analysis of potential abuses in the system can help the decision maker more consistently reach accurate outcomes.
Secondly, there is an opportunity to improve the data collected when there is no injury claim made in an accident. Based on the TMJ injury example previously provided, collection of scientifically relevant facts during a claim investigation can, over the long term, help offset many of the significant limitations found in the medical data. Third, consider using analytics that, while not considered as only predictive analytics, actually define and describe the event in an accurate and defensible way. For example, referring back to the book “Competing on Analytics,” one form of analytics described was employed by VisViva Golf Inc. The company uses nanotechnology embedded in golf clubs that is connected to Bluetooth radio technology to calculate and measure technical aspects of a golfer’s swing—the swing speed, acceleration, deceleration, and so on—as well as use the data with predictive analytics to provide guidance to improve the golfer’s swing.
Today, automated analytics are available for claims organizations which scientifically predict or actually determine acceleration and deceleration of vehicles in accidents and the resulting implications to injury potential—or, in other words, scientifically identify questionable injuries. More specifically, the scientific analytics described in the previous TMJ example can be systematically applied to claims data such as repair estimate information and injury claim information.
In conclusion, using medical data in predictive analytics applications to identify questionable claims could be, well, questionable. While all the numbers and formulas associated with today’s analytics suggest objectivity, experienced managers understand that the “garbage in, garbage out” phenomenon has never been truer. Realizing the power of analytics requires being realistic about what models can and cannot do, improving the quality of the data feeding models, and creating the appropriate managerial processes around them.
In the context of injury claims, statistical models based on medical data should be accompanied by an investigative process that will provide an adjuster with actionable information and the basis for defensible decisions. Otherwise a claims organization may find itself in a position of systematically creating more fires than they are trying to avoid. The use of predictive analytics in this approach begs the following: Why not use defensible analytics in the identification of questionable claims? Would this not eliminate an unnecessary step?