Humana also using AI tool with 90% error rate to deny care, lawsuit claims

Humana, one the nation’s largest health insurance providers, is allegedly using an artificial intelligence model with a 90 percent error rate to override doctors’ medical judgment and wrongfully deny care to elderly people on the company’s Medicare Advantage plans. According to a lawsuit filed Tuesday, Humana’s use of the AI model constitutes a “fraudulent scheme” that leaves elderly beneficiaries with either overwhelming medical debt or without needed care that is covered by their plans. Meanwhile, the insurance behemoth reaps a “financial windfall.” The lawsuit, filed in the US District Court in western Kentucky, is led by two people who had a Humana Medicare Advantage Plan policy and said they were wrongfully denied needed and covered care, harming their health and finances. The suit seeks class-action status for an unknown number of other beneficiaries nationwide who may be in similar situations. Humana provides Medicare Advantage plans for 5.1 million people in the US. It is the second lawsuit aimed at an insurer’s use of the AI tool nH Predict, which was developed by NaviHealth to forecast how long patients will need care after a medical injury, illness, or event. In November, the estates of two deceased individuals brought a suit against UnitedHealth—the largest health insurance company in the US—for also allegedly using nH Predict to wrongfully deny care. Humana did not respond to Ars’ request for comment for this story. United Health previously said that “the lawsuit has no merit, and we will defend ourselves vigorously.”

AI model

In both cases, the plaintiffs claim that the insurers use the flawed model to pinpoint the exact date to blindly and illegally cut off payments for post-acute care that is covered under Medicare plans—such as stays in skilled nursing facilities and inpatient rehabilitation centers. The AI-powered model comes up with those dates by comparing a patient’s diagnosis, age, living situation, and physical function to similar patients in a database of 6 million patients. In turn, the model spits out a prediction for the patient’s medical needs, length of stay, and discharge date. But, the plaintiffs argue that the model fails to account for the entirety of each patient’s circumstances, their doctors’ recommendations, and the patient’s actual conditions. And they claim the predictions are draconian and inflexible. For example, under Medicare Advantage plans, patients who have a three-day hospital stay are typically entitled to up to 100 days of covered care in a nursing home. But with nH Predict in use, patients rarely stay in a nursing home for more than 14 days before claim denials begin. Though few people appeal coverage denials generally, of those who have appealed the AI-based denials, over 90 percent have gotten the denial reversed, the lawsuits say. Still, the insurers continue to use the model and NaviHealth employees are instructed to hew closely to the AI-based predictions, keeping lengths of post-acute care to within 1 percent of the days estimated by nH Predict. NaviHealth employees who fail to do so face discipline and firing. ” Humana banks on the patients’ impaired conditions, lack of knowledge, and lack of resources to appeal the wrongful AI-powered decisions,” the lawsuit filed Tuesday claims.

Plaintiff’s cases

One of the plaintiffs in Tuesday’s suit is JoAnne Barrows of Minnesota. On November 23, 2021, Barrows, then 86, was admitted to a hospital after falling at home and fracturing her leg. Doctors put her leg in a cast and issued an order not to put any weight on it for six weeks. On November 26, she was moved to a rehabilitation center for her six-week recovery. But, after just two weeks, Humana’s coverage denials began. Barrows and her family appealed the denials, but Humana denied the appeals, declaring that Barrows was fit to return to her home despite being bedridden and using a catheter. Her family had no choice but to pay out-of-pocket. They tried moving her to a less expensive facility, but she received substandard care there, and her health declined further. Due to the poor quality of care, the family decided to move her home on December 22, even though she was still unable to use her injured leg, go the bathroom on her own, and still had a catheter. The other plaintiff is Susan Hagood of North Carolina. On September 10, 2022, Hagood was admitted to a hospital with a urinary tract infection, sepsis, and a spinal infection. She stayed in the hospital until October 26, when she was transferred to a skilled nursing facility. Upon her transfer, she had eleven discharging diagnoses, including sepsis, acute kidney failure, kidney stones, nausea and vomiting, a urinary tract infection, swelling in her spine, and a spinal abscess. In the nursing facility, she was in extreme pain and on the maximum allowable dose of the painkiller oxycodone. She also developed pneumonia. On November 28, she returned to the hospital for an appointment, at which point her blood pressure spiked, and she was sent to the emergency room. There, doctors found that her condition had considerably worsened. Meanwhile, a day earlier, on November 27, Humana determined that it would deny coverage of part of her stay at the skilled nursing facility, refusing to pay from November 14 to November 28. Humana said Hagood no longer needed the level of care the facility provided and that she should be discharged home. The family paid $24,000 out-of-pocket for her care, and to date, Hagood remains in a skilled nursing facility. Overall, the patients claim that Humana and UnitedHealth are aware that nH Predict is “highly inaccurate” but use it anyway to avoid paying for covered care and make more profit. The denials are “systematic, illegal, malicious, and oppressive.” The lawsuit against Humana alleges breach of contract, unfair dealing, unjust enrichment, and bad faith insurance violations in many states. It seeks damages for financial losses and emotional distress, disgorgement and/or restitution, and to have Humana barred from using the AI-based model to deny claims.