The Food and Drug Administration ended 2018 by awarding Milliman, a health insurance consulting company, with extra resources and expedited review for a new technology. The system uses artificial intelligence and machine learning to predict the likelihood of future of opioid use disorder (OUD) diagnoses in individual patients.
Predicting “the likelihood of receiving an OUD diagnosis in the next 90 days,” Milliman’s algorithm assigns a patient a “risk score” from zero to one. In addition to Milliman, the companies HBI Solutions and MIT-IBM Watson AI Lab are among the key players predicting patients’ OUD risks.
Intended to assist in OUD prevention, the technologies could increase profit margins for health insurers—but possibly at the expense of patient agency and privacy.
Doctors could use the risk scores generated to help determine if they will continue to prescribe opioids to a patient, or even start them on the medication in the first place.
Critics are concerned that a higher risk score may “pigeonhole people without their knowledge and give doctors an excuse to keep them from ‘getting the drugs they need,’” Lorraine Possanza of the ECRI Institute told Politico.
Asked for comment on such criticism, Milliman’s spokesperson told Filter that the company had “no interest in engaging with critics.”
Product Manager Joseph Boschert, who worked on Milliman’s predictive analytics services, has suggested that risk scores will not supplant doctors’ own judgments about what their patients need. “Pain is one of those things that’s really hard to objectively measure,” he said, “and it really comes down to the prescriber and the patient relationship and figuring out what the best path forward is.”
But Politico’s reporting also raises concerns about how the scores are calculated in the first place.
In Milliman’s product, for example, a patient’s risk score is calculated by “considering all of their prior medical history, all of the diagnoses they’ve had and the procedures and drugs they’re taking.” Factors include “history of behavioral or mental health issues, whether that’s mood disorders or anxiety disorders, bipolar,” “chronic pain conditions,” and “certain geographic areas where it’s more socially acceptable—so a lot of the Appalachian areas, where it’s just more prevalent.”
HBI Solutions’ tech considers even more factors, using “non-medical conditions and situations like housing status, income and even zip code.”
The health claims data used by Milliman and other risk management companies to calculate risk are housed by databases such as MarketScan—and these are contentious crossroads for privacy violations. The confidentiality of a patient’s substance use disorder information was threatened by the Substance Abuse and Mental Health Services Administration in 2018. A bill sponsored and introduced in 2017 by Senator Edward J. Markey would require data brokers to provide greater transparency about the data they collect.
Politico reported that many companies using predictive analytics for opioids refused to comment on their data sourcing. Instead, they prefer to focus on what the service will provide—a backstop against prescribing opioids to patients deemed at high risk for misuse.
While identifying the technology’s utility for doctors, Milliman has been less clear about how insurers could use opioid risk scores.
Milliman’s director of Media Relations and Public Affairs, Jeremy Engdahl-Johnson, denied in a phone conversation with Filter that the company’s opioid risk scoring will be used by insurance companies.
Yet in contrast to this denial, Boschert said on a Milliman podcast last year that predictive analytics “can be used in the context of payers [insurers], when they have a medical management team that helps manage the members on a plan.”
What’s more, OUD risk scores could be used by insurers to boost their profits. The Affordable Care Act (ACA) prohibits insurers from refusing to cover patients because of pre-existing conditions. Instead, Obamacare’s risk adjustment model requires insurers “found to have healthier customers (according to a federal formula) [to] pay money into the program, and then insurers who have sicker customers receive money out of it.”
This model includes a substance use disorder diagnosis as a factor that increases risk for insurers. And sometimes, the riskier the patient, the more profit the insurer sees. This seems counterintuitive—especially given the history of insurers maximizing profits by minimizing the number of high-risk members. But insurers often end up receiving more money from risk adjustment payments for insuring members with high-risk conditions than the actual cost associated with insuring them.
It’s not necessarily that the more risky conditions, the higher the profit—rather that there can be an optimal amount of risk in a patient to maximize an insurer’s profits. In 2014, for example, insured members with 54 risk-adjusted conditions brought in market-average profit margins of 20-100 percent; those with seven risk-adjusted conditions reaped profit margins north of 1,000 percent.
Scott A. Weltz, a consulting actuary for Milliman, accordingly urges health insurers to “consider risk adjustment impact in every key decision they make.”
So it is hard to believe that Milliman’s predictive analytics will not be used by insurers. After all, the company’s development of OUD risk scoring happens to coincide with a movement towards specifying OUD as a condition included in Medicare and Medicaid’s risk adjustment program.
To be clear, the Affordable Care Act and its required risk adjustment program have improved access to treatment for people diagnosed with substance use disorders (SUD). Around 1.6 million people with SUD have new access to insurance coverage in Medicaid expansion states, and insurance providers can no longer deny coverage to people with OUD. At the same time, some states that have accepted Medicaid and Medicare expansion still limit access to OUD medications.
Photo by Goh Rhy Yan on Unsplash