Before Dr. Bobby Mukkamala — an ear, nose, and throat specialist in Michigan — prescribed post-surgical opioids recently, he checked state records of his patient’s existing controlled substance prescriptions, as legally required. A score generated by a proprietary algorithm appeared on his screen. Known as NarxCare — and now used by most state prescription monitoring databases, major hospitals and pharmacy chains — the algorithm indicated his patient had an elevated risk of developing an addiction to opioid painkillers.
“I create a lot of pain when I operate,” said Dr. Mukkamala, who leads the American Medical Association’s Substance Use and Pain Task Force. “The nose and the face are very painful places to have procedures done.” Consequently, it is difficult to avoid prescribing opioids to manage pain.
Algorithms like NarxCare and a newly-approved genetic test for opioid use disorder risk known as AvertD, use machine learning techniques to try to help doctors reduce the odds that patients will become addicted to these medications.
Via NarxCare, most Americans now have an opaque equivalent of a controlled substance credit score, which they often don’t even know exists unless a doctor or pharmacist tells them that it’s a problem. (NarxCare’s manufacturer claims that its scores and reports “are intended to aid, not replace, medical decision making.”) And if it ever becomes widely used, AvertD, promoted as a way to use personalized genetics to assess risk, could put yet more difficult-to-challenge red flags on people’s records.
These tools may be well intentioned. But addiction prediction and prevention is a mind-bogglingly difficult task. Only a minority of people who take opioids become addicted, and risk factors vary for biological, psychological, sociological and economic reasons.
Even accurate scores can do harm, since addiction is stigmatized and often criminalized. Some people have been expelled from physicians’ practices for having high NarxCare scores, with no way of appealing the decision. Others were denied post-surgical opioids by nurses or turned away from multiple pharmacies, with little recourse.
These kinds of algorithms could potentially worsen race and class biases in medical decision making. It’s not hard to imagine a dystopian future of unaccountable algorithms that render some people forever ineligible for pain care with controlled substances.
Dr. Mukkamala noted that closer scrutiny of his recent patient’s medical history showed there really wasn’t reason for concern. “What’s inappropriate is for me to look at any number other than zero and say: ‘Boy, this person’s got a problem. I can’t prescribe them anything for their pain,’” Dr. Mukkamala said. Many medical professionals, however, don’t have Dr. Mukkamala’s level of knowledge and confidence. Prejudice against people with addiction is common, as is fear of being charged with overprescribing — and the algorithms’ scores only feed into those concerns. Different, also unaccountable, algorithms monitor physicians’ prescribing patterns and compare them with their colleagues’, so this is not an overblown concern.
When I reported on NarxCare in 2021 for Wired, I heard from patients who were left in agony. One said that she had her opioids stopped in the hospital and was then dismissed from care by her gynecologist during treatment for painful endometriosis, because of a high score. She didn’t have a drug problem; her score seems to have been elevated because prescriptions for her two medically needy rescue dogs were recorded under her name, making it appear she was doctor shopping. Another high-scoring patient had his addiction treatment medication prescription repeatedly rejected by pharmacies — even though such medications are the only treatment proven to reduce overdose risk.
More recent research and reporting confirm that scientists’ concerns about the widespread use of the software remain and that patients are still reporting encountering problems because of potentially incorrect risk assessments and medical staff members’ fears about disregarding NarxCare scores.
To generate risk scores, NarxCare apparently uses variables like the number of doctors someone sees, the pharmacies they visit and the prescriptions they get and compares an individual’s data with information on patterns of behavior associated with doctor shopping and other indicators of possible addiction.
But there is no transparency: The NarxCare algorithm is proprietary, and its information sources, training data and risk variables — and how they are weighted — aren’t public.
Another problem for NarxCare is that opioid addiction is actually quite uncommon — affecting between 2 and 4 percent of the adult and adolescent population, despite the fact that a 2016 study shows some 70 percent of adults have been exposed to medical opioids. “Identifying somebody’s base line risk of opioid use disorder is inherently going to be pretty difficult,” said Angela Kilby, an economist who studied algorithms like NarxCare when she was an assistant professor at Northeastern University. “It’s sort of like trying to find a needle in a haystack.” The rarity of the condition possibly lowers the algorithm’s precision, meaning that most positive tests may be falsely positive simply because the base line rate of the disorder is low.
Research shows that about 20 percent of the time, people who are flagged as doctor shoppers by identifying risk factors similar to those apparently included in NarxCare in fact have cancer: They typically see multiple specialists, often at academic medicine centers where there may be teams of doctors writing prescriptions. The algorithm can’t necessarily distinguish between coordinated care and doctor shopping.
Likewise, someone who is visiting multiple doctors or pharmacies and traveling long distances might be drug-seeking — or they could be chronically ill and unable to find care locally. Some states also put information from criminal records into prescription monitoring databases — and this can lead to bias against Black and Hispanic people simply because racial discrimination means that they are more likely to have been arrested.
There’s also a more fundamental problem. As Dr. Kilby notes, the algorithm is designed to predict elevations in someone’s lifetime risk of opioid addiction — not whether a new prescription will change that trajectory. For example, if someone is already addicted, a new prescription doesn’t change that, and denying one can increase overdose death risk if the person turns to street drugs.
Recently, NarxCare has been joined in the addiction prediction game by AvertD, a genetic test for risk of opioid use disorder for patients who may be prescribed such medications, which the Food and Drug Administration approved last December. Research by the manufacturer, Solvd Health, shows that a patient who will develop opioid addiction is 18 times more likely to receive a positive result than a patient who will not develop it. The test, which looks for specific genes associated with motivational pathways in the brain that are affected by addiction, utilizes an algorithm trained on data from over 7,000 people, including some with opioid use disorder.
But that F.D.A. approval came, surprisingly, after the agency’s advisory committee for the test voted overwhelmingly against it. While the F.D.A. worked with the company behind the test to modify it based on the committee’s feedback, it has continued to raise concerns. And recently a group of 31 experts and scientists wrote to the F.D.A. urging it to reverse course and rescind its approval. Some of the group’s concerns echo the problems with NarxCare and its algorithm.
For a study published in 2021, Dr. Alexander S. Hatoum, a research assistant professor of psychological and brain sciences at Washington University in St. Louis, and his colleagues independently evaluated the algorithm elements used for a tool like AvertD, based on information published by the company. They found that all the iterations they tested were confounded by population stratification — a problem that affects genetic tests because they reflect the history of human ancestry and how it changed over time because of migration patterns.
When AvertD was being considered for F.D.A. approval, Dr. Hatoum and his colleagues wrote a public comment to the agency that said genomic variants used in the test were “highly confounded by genetic ancestry” and did not predict risk any better than chance when population stratification is not taken into account. (At a 2022 meeting, Solvd’s chief executive claimed AvertD adjusted adequately for population stratification; the F.D.A. did not reply directly to a question about this claim.)
Dr. Hatoum’s work also demonstrated that these tests could mislabel people who are descended from two or more groups that were historically isolated from each other as being at risk of addiction. Since most African Americans have such admixed ancestry, this could bias the test into identifying them as high-risk.
“This means that the model can use the genetic markers of African American status to predict opioid use disorder, instead of using any biologically plausible genetic markers,” said D. Marzyeh Ghassemi, a professor at M.I.T. who studies machine learning in health care.
In an email, Solvd said that in its clinical study of AvertD, “no differences in performance were seen by race, ethnicity or gender,” adding that it was undertaking post-marketing tests as required by the F.D.A. to further evaluate the test. The company also critiqued Dr. Hatoum’s methodology, saying that his study “asserts a false premise.”
The F.D.A. said in a statement that it “recognizes that in premarket decision making for devices, there generally exists some uncertainty around benefits and risks,” adding that it had nevertheless “determined that there is a reasonable assurance of AvertD’s safety and effectiveness.”
Still, the agency has placed a black box warning on AvertD, forbidding its use in chronic pain patients and emphasizing that the test cannot be used without patient consent. But this is unlikely to be a genuinely free choice: Patients may fear being stigmatized as potentially addicted if they don’t agree to be tested. And false negatives that incorrectly label someone as low risk may conversely lead to careless prescribing.
Amid the opioid crisis, it is understandable that regulators want to enable technologies that could reduce risk of addiction. But they must ensure that such algorithms and devices are transparent as to their methods and limitations and that they reduce racial and other biases — rather than reinforce them.
Maia Szalavitz (@maiasz) is a contributing Opinion writer and the author, most recently, of “Undoing Drugs: How Harm Reduction Is Changing the Future of Drugs and Addiction.
The Times is committed to publishing a diversity of letters to the editor. We’d like to hear what you think about this or any of our articles. Here are some tips. And here’s our email: letters@nytimes.com.
Follow the New York Times Opinion section on Facebook, Instagram, TikTok, WhatsApp, X and Threads.
