Could AI help primary care clinics spot risky drinking habits?

On any given day in a busy primary care clinic, doctors and others often ask patients about their alcohol use, and try to gauge if it falls into healthy or problematic range.

Patients might even complete an alcohol use questionnaire on a clipboard or their smartphone while they wait for their appointment.

But a new study suggests that artificial intelligence might help increase the chance that people with risky drinking patterns or signs of alcohol use disorder (AUD) will get the outreach and help they need.

That’s important, because only about 10% of people who qualify for the diagnosis of AUD have actually gotten help in the past year. Many may not realize that their drinking habits suggest signs of addiction or that they’re consuming more alcohol than recommended, which can raise the risk of everything from injuries to cancer to sleep problems.

The study, published in the journal Drug and Alcohol Dependence by a University of Michigan Addiction Center team, used a natural language processing tool to analyze the full text of notes, comments and screening tool scores in anonymous electronic health records from more than 133,000 primary care patients at U-M Health.

Going by just the formal parts of the record like diagnosis codes and alcohol use questionnaire scores, 820 of the patients had risky alcohol use or AUD. The NLP tool did almost as well as human reviewers in finding these.

But the natural language processing tool also identified more than 47,500 other patients whose notes showed some sign of risky drinking.

Identifying alcohol issues

Anne Fernandez, Ph.D., the addiction psychologist who helped lead the study, says the findings suggest the majority of people drinking at risky levels are overlooked, and automated tools could help identify them.

This is important because alcohol use causes health problems, interacts with medications, and makes some common medical conditions worse, she said. These are things that doctors want to know, so they can provide the best care possible.

“Doctors can’t read every clinical note from every provider and appointment in a patient’s chart, but automated tools can do this quickly and easily,” she explained. “Our study shows that these notes contain useful information about alcohol use that we hope can improve clinical care in the long term.” Fernandez is an associate professor in the U-M Medical School Department of Psychiatry and clinical psychologist at the U-M Addiction Treatment Service.

She and colleagues, including first author and Psychiatry research assistant Celeste Xintong Ju, applied the same NLP tool that has previously been shown to identify signs of risky drinking in patients scheduled to have surgery.

For the new study, they went a step further, by contacting a sample of those identified by the natural language processing tool as drinking at risky levels, and a matched sample of those identified by standard means.

They asked them to answer questions about their drinking habits and symptoms of alcohol use disorder and used this information to check the accuracy of the electronic health record and natural language processing tool.

In this group of 170 people, the natural language processing tool found 17 more cases of alcohol use disorder that standard data had missed, and 23 more cases of risky alcohol use that standard screening had missed.

The people identified through standard screenings were more likely to have depression or anxiety, and to have sought some form of help for their alcohol use, whether in a self-help group, an outpatient therapy setting or residential care.

Potential implications

Fernandez notes that if the new study is borne out in larger samples using the natural language processing tool, the use of AI could eventually complement what primary care providers already do to identify risking drinking and signs of AUD. 

Despite national recommendations calling for universal screening of adults for risky alcohol use, and insurance coverage under the Affordable Care Act for such screening, many patients aren’t screened.

There’s also the issue of patients not being completely forthcoming or accurate with their health care providers about their alcohol use, especially when they’re filling out standardized questionnaires.

That’s another reason that using natural language processing to scan the text of notes that primary care providers make about a patient’s visit may reveal more than screening tools.

However, natural language processing can’t replace standard alcohol screening. Clinicians still need to do their own screening to check accuracy, because AI can make errors, the electronic health record can be inaccurate, and alcohol use changes over time.

Many patients, and even some providers, may not know that there are prescription medications available to help patients recover from AUD. These are covered by Medicaid, Medicare and most other insurance programs. So is an approach called SBIRT, for screening, brief intervention and referral to treatment.

In addition to Fernandez and Ju, the study’s authors are Jake Solka, Katherine Weber, VG Vinod Vydiswaran, Lewei Allison Lin and Erin E. Bonar.  Fernandez, Lin and Bonar are members of the Michigan Innovations in Addiction Care through Research & Education (MI-ACRE) group, and the U-M Institute for Healthcare Policy and Intervention.

Lin is also a member of the VA Center for Clinical Management Research at the VA Ann Arbor Healthcare System, which is also taking part in the clinical trial of alcohol treatment in primary care.

The research was funded by the U-M Medical School Office of Research Pandemic Research Recovery Program and the National Institute of Alcohol Abuse and Alcoholism of the National Institutes of Health (R01AA029400 and R33AA028315).

The researchers used the U-M Medical School Data Office for Clinical and Translational Research for secure, anonymous access to patient records.

Citation: Unhealthy alcohol use detection in electronic health records: A comparative study using natural language processing, Drug and Alcohol Dependence, DOI: 10.1016/j.drugalcdep.2025.112920, https://www.sciencedirect.com/science/article/pii/S0376871625003734