Cutting the Bluster on AI in Group Health Underwriting

Generative AI tools like ChatGPT are dominating headlines and causing veteran execs in almost every industry, including insurance, to wonder how such tools will change their business. In the group health space, more and more vendors seem ready to exploit the current fascination with AI by breathlessly invoking it as they peddle their wares.

True to character, our Curv® – Group Health clients are generally skeptical of hype while they also remain understandably curious; generative AI definitely comes up in conversation. And it’s very likely that it will affect almost every industry including ours, but it’s too early to say how it will be used (or regulated) in group health underwriting.

For the time being, predictive models remain the most advanced form of AI in regular use in the insurance industry, and even their workings are still a bit mysterious to many underwriters. Now is good time to clear up some of that mystery, because demand for transparency is growing and privacy regulations are getting stricter and more punitive.

Predictive models are a subset of AI
Curv was the only predictive model for underwriting in the group health space when we introduced it back in 2011. Now, newly minted imitators trumpet their discounted, derivative models as if they’d been downloaded straight from some sci-fi author’s imagination.

Predictive models are a subset of artificial intelligence, which is a broad field that encompasses machine learning and generative AI. Although some AI can seem to have nearly human intelligence, the models we use are “only” statistical tools that use machine learning to find patterns in data. Conceptually, they automate a process very similar to the work underwriters have always done. The key differences are that models are trained on datasets too large for any human to grasp, enabling them to spot patterns too subtle for any human to discern.

Curv makes complex data science look simple
The data that powers Curv consists of five years of prescription histories and medical billing claims for the members of the group you’re quoting. (Currently, aggregate hit rates are about 90%.) The Curv predictive model then compares your group members’ data to the records of millions of other people whose morbidity and healthcare costs have been tracked for years. The model often considers thousands of data points per life.

The complex data science all happens in the training and model-building phases. From our clients’ perspective, Curv is remarkably easy to use. All you provide is a simple group census consisting of basic demographic details. You don’t need employees’ Social Security numbers or HIPAA authorizations, because even though the model uses members’ individual health records, personal health information is deidentified. Behind the scenes, the model predicts future claims risks based on previously observed trends and delivers an easily actionable risk score in moments.

Curv can be used to improve upon some claims histories
Although Curv clients often rely on the model to quote groups for which they have no claims history, the model can also improve your understanding of risk in groups with partially credible data.

In smaller employer groups, even a little employee turnover can dramatically change next year’s outlook. Curv allows you to appropriately rate newly covered employees or dependents based on their available health information. And, because Curv looks back over five years of data, the model is influenced by members whose health status may be changing or evolving; it anticipates the future health status of each member.

Another way that Curv can improve upon prior claims data is that the risk score can be integrated into your own rating formula, allowing you to value the cost of conditions based on experience with your own networks and medical management expertise, rather than just relying on whatever the prior carrier happened to pay.

Training data: Quality matters
Models are only as good as the data that they’re trained on. Data quantity is usually measured in millions of lives in the training data. But data quality is just as important. What data is being analyzed and how is it weighted? If you want to model group morbidity, knowing that a group member has been treated for alcohol dependency is more relevant than knowing what brand of beer he prefers.

Our model also benefits from our deep bench of clinical experts; that’s especially relevant if groups include individuals with rare conditions that don’t often crop up in training data, or individuals who are candidates for new treatments like gene therapies that are effective but breathtakingly expensive.

We protect group members’ privacy to protect you
We did not just build Curv to follow the letter of privacy rules, we built it to respect the underlying intent of the rules and truly protect individual group members’ protected health information. Even we can’t reidentify individuals. That’s reassuring, because new regulations such as the California Privacy Rights Act (which came into effect at the beginning of 2023) or the Colorado Privacy Act (effective July 1, 2023) are changing the landscape when it comes to the use of deidentified data. More states are expected to follow suit and the trend is towards increasing protections for consumer privacy.

We understand why many underwriters would like to see case-level detail in addition to a group risk score. But insurers who insist on seeing too much detail could run afoul of the new regulations.

So instead, we provide as much detail as can within the law, including detail of equal or greater business value, that doesn’t put our clients at risk.

We deliver the business value you need to compete without exposing you to liability
Last year we created a new Curv reporting feature that flags likely high-cost conditions and/or incidence of cancer, end-stage renal disease, or other conditions requiring treatment with expensive specialty drugs. Clients who turn on the single order view feature now find a graphic heat map display that makes it easy to see how a newly scored group compares to other, similar groups when it comes to those commonly encountered key risks. This feature is also broadly customizable; clients can choose to add flags for other condition groupings, or flag different thresholds.

Clients can also benchmark any group against their book of business or against all groups scored from a particular region. By comparing hit rates, clients can reaffirm statistical confidence, and by looking at member-risk distribution they can see whether one (or a few) group members are influencing a larger-than-usual group score. These features all help clients better understand the rationale for a group’s score or when to maximize its rate—without crossing ethical or legal boundaries.

We’ve recently made place-of-service data available for the groups you’re quoting as well. For example, you can see if members are inclined to seek care from their primary care physician or visit a hospital emergency room. Given the cost differential, that’s data with obvious business value.

Curv will pragmatically adopt next-level AI
We’ll never stop developing the Curv features you need to compete, and more enhancements are coming out soon. I expect that someday generative AI will deliver real value to real clients in the group health space. When that day comes, you can count on us to responsibly and pragmatically deliver the next generation of tech—because it’s useful, not because it allows us to market another buzzword.

In the meantime, competitors will continue to imply that their AI is one step away from science fiction. That’s not necessarily harmful (except possibly to our eye-rolling muscles). But you probably should be aware that case-level detail is only nice to have until it turns out that your provider has crossed a line and left you holding data that’s been reidentified—or could be.

Dan Becker is Product Director for Curv® – Group Health at Milliman IntelliScript.