A hospital AI screener for opioid use disorder matched usual care, and consulted patients were readmitted less often afterward

The short version

A hospital system put an AI screener into its electronic health records to flag adults who might have opioid use disorder and suggest an addiction medicine consultation. The share of patients who got a consultation stayed about the same as before. Among patients who got a consultation, 30-day readmissions were lower after the screener went in. Readmissions across all hospitalized adults did not change. This was a before-and-after comparison, not a randomized trial.

What they did

The screener ran inside the hospital's electronic health record. It flagged hospitalized adults who might be at risk for opioid use disorder and recommended an addiction medicine consultation.

The researchers compared time periods at one hospital: March to October of 2021 and of 2022 before the screener (16 months), and March to October 2023 with it. Matching the months was meant to limit seasonal differences. The study was designed to test whether the AI screener was non-inferior to usual care. In plain terms, the question was whether it did at least about as well as the existing approach.

The main outcome was the share of patients who completed a consultation with an addiction medicine specialist. A secondary outcome was readmission within 30 days among patients who received a consultation.

The study was published in Nature Medicine and registered as NCT05745480. The authors declared no competing interests. The study included 51,760 adult hospitalizations at UW Health University Hospital in Madison, Wisconsin. It was funded by the National Institutes of Health.

What it found

Consultation rates were 1.35% before the screener and 1.51% with it. That met the study's test for non-inferiority (P < 0.001). The screener did about as well as usual care at getting patients a consultation, and the rate barely moved.

In the secondary analysis, among patients who received a consultation, the screener period was linked to lower odds of 30-day readmission (adjusted odds ratio 0.53, 95% confidence interval 0.30 to 0.91). Across all hospitalized adults, 30-day readmissions did not change (adjusted odds ratio 1.00). The researchers estimated an incremental cost of US$6,801 per readmission avoided.

What it does not show

This was not a randomized trial. Patients were not assigned to screener or no screener. The study compares one period of time with a later one, so other changes at the hospital during those months could explain some or all of the difference.

The readmission result is a secondary outcome. The study was built around consultations, and readmission findings from a secondary analysis are weaker evidence than a primary result.

It took place at one health system. The authors note this may limit how well the results carry over to hospitals with different addiction services, workflows, or patients.

It does not show that the screener improved long-term recovery. The outcomes were consultations and 30-day readmissions.

The cost estimate is uncertain. The confidence interval on the cost per readmission avoided runs from about US$3,900 to about US$50,500. The full paper puts the screener's added cost at US$106,100 for the 8 months it ran and, separately, estimates start-up costs of about US$350,000.

Why it matters

The screener ran in live hospital care rather than being tested only on past records. That makes it a test of the tool in everyday use, not only of the model's accuracy on old records.

Read clinically, the screener was not worse at prompting consultations, and the readmission signal is promising but not proven. A randomized trial or replication at other hospitals would be needed to know whether the readmission drop is real.

If you are a patient, a tool like this works by reading your hospital record. Some substance use treatment records carry extra federal privacy protection under 42 CFR Part 2, depending on where they come from. This study does not address how those rules apply to AI tools.