The use of AI tools in healthcare and by insurance providers has raised the question of whether high denial rates are a glitch — or a feature.
In the wake of the murder of its CEO this week, UnitedHealthcare has come under greater scrutiny for its use of an allegedly flawed AI algorithm that overrides doctors to deny elderly patients critical heathcare coverage.
UnitedHealthcare CEO Brian Thompson was fatally shot in a targeted attack outside a New York City hotel on Dec 4. The shooter fled on an e-bike, leaving shell casings with possible motive-related messages, though the actual intent remains unclear. (The words “deny,” “defend” and “depose” were written on the shell casings.)
One motive floated by many is that the murder might be connected to high treatment rejection rates or UnitedHealthcare’s (UHC) outright refusal to pay for some care. Healthcare providers and insurers have been automating responses to care requests using generative AI (genAI) tools, which have been accused of producing high denial of care rates, in some cases, 16 times higher than is typical.
UHC uses a genAI tool called nH Predict, which has been accused in a lawsuit of prematurely discharging patients from care facilities and forcing them to exhaust their savings for essential treatment. The lawsuit, filed last year in federal court in Minnesota, alleges UHC illegally denied Medicare Advantage care to elderly patients by using an AI model with a 90% error rate, overriding doctors’ judgments on the medical necessity of expenses.
Some have argued that the genAI algorithm’s high rejection rate is a feature, not a flaw. An investigation by STAT News cited in the lawsuit, claims UHC pressured employees to use the algorithm to deny Medicare Advantage payments, aiming to keep patient rehab stays within 1% of the length predicted by nH Predict.
According to the lawsuit, UnitedHealth started using nH Predict in November 2019. nH Predict, developed by US-based health tech company NaviHealth (now part of UnitedHealth Group), is a proprietary assessment tool that designs personalized treatment plans and recommends care settings, including hospital discharge timing.
“Despite the high error rate, defendants continue to systemically deny claims using their flawed AI model because they know that only a tiny minority of policyholders (roughly 0.2%) will appeal denied claims, and the vast majority will either pay out-of pocket costs or forgo the remainder of their prescribed post-acute care,” the lawsuit argued. “Defendants bank on the patients’ impaired conditions, lack of knowledge, and lack of resources to appeal the erroneous AI-powered decisions.”
Last year, UnitedHealth Group and its pharmacy services subsidiary Optum rebranded NaviHealth following congressional criticism over the algorithms it used to deny patient care payments. More recently, in an October report, the US Senate Permanent Subcommittee on Investigations criticized UHC, Humana, and CVS for prioritizing profits over patient care.
“The data obtained so far is troubling regardless of whether the decisions reflected in the data were the result of predictive technology or human discretion,” according to the report. “It suggests Medicare Advantage insurers are intentionally targeting a costly but critical area of medicine — substituting judgment about medical necessity with a calculation about financial gain.”
Using millions of medical records, nH Predict analyzes patient data such as age, diagnoses, and preexisting conditions to predict the type and duration of care each patient will require. nH Predict has faced criticism for its high error rate, premature termination of patient treatment payments (especially for the elderly and disabled), lack of transparency in decision-making, and potential to worsen health inequalities.
UHC declined to comment on its use of genAI tools, opting instead to release a statement on how its dealing with the loss of its CEO.
The healthcare industry and insurers have long embraced AI and generative AI, with providers now leveraging it to streamline tasks like note-taking and summarizing patient records. The tech has also been used to assess radiology and electrocardiogram results and predict a patient’s risk of developing and worsening disease.
Insurers use AI to automate processes such as prior authorization, where providers or patients must get insurer approval before receiving specific medical services, procedures, or medications. The high denial rates from AI-driven automation have frustrated physicians, leading them to counter by using AI tools themselves to draft appeals against the denials.
Asthma drugs, new weight loss drugs and biologics — a class of drugs that can be life-saving for people with autoimmune disease or even cancer — are routinely denied coverage by insurance companies. Data shows that clinicians rarely appeal denials more than once, and a recent American Medical Association survey showed that 93% of physicians report care delays or disruptions associated with prior authorizations.
“Usually, any expensive drug requires a prior authorization, but denials tend to be focused on places where the insurance company thinks that a cheaper alternative is available, even if it is not as good,” Dr. Ashish Kumar Jha, dean of the School of Public Health at Brown University, explained in an earlier interview with Computerworld.
Jha, who is also a professor of Health Services, Policy and Practices at Brown and served as the White House COVID-19 response coordinator in 2022 and 2023, said that while prior authorization has been a major issue for decades, only recently has AI been used to “turbocharge it” and create batch denials. The denials force physicians to spend hours each week challenging them on behalf of their patients.
GenAI technology is based on large language models, which are fed massive amounts of data. People then train the model on how to answer queries, a technique known as prompt engineering.
“So, all of the [insurance company] practices over the last 10 to 15 years of denying more and more buckets of services — they’ve now put that into databases, trained up their AI systems and that has made their processes a lot faster and more efficient for insurance companies,” Jha said. “That has gotten a lot of attention over the last couple of years.”
The suspect in the Wednesday shooting of Thompson has not yet been captured, nor has there been any claims of motive.