A new study suggests that machine-learning tools could help healthcare systems identify people with untreated substance use disorders without requiring every patient to undergo the same intensive screening process.
Researchers from Stanford University, Yale University, the University of Connecticut and the VA New England Mental Illness Research, Education and Clinical Center trained supervised models using electronic-health-record-like data from a nationally representative U.S. survey.
The models were most developed for alcohol use disorder. Exploratory models for opioid and cocaine use disorders also showed promise, although the researchers cautioned that smaller samples limited the strength of those findings.
According to the study, models optimized for precision and recall could reduce the number of people clinicians would need to screen to identify one untreated case—sometimes by more than an order of magnitude compared with universal screening. The systems were also designed to show which features influenced a prediction, an important safeguard when clinical decisions involve complex and sensitive behavior.
The potential use is triage, not diagnosis. A model could flag patients who may benefit from a confidential conversation or validated screening tool, but a clinician would still need to assess the person directly and discuss treatment options. False positives, missing data and unequal performance across populations remain significant concerns.
The researchers also note that implementation would require careful governance. Substance-use information is highly sensitive, and automated risk flags could cause harm if they affect insurance, employment, policing or clinical treatment without informed oversight.
The paper’s immediate contribution is methodological: it demonstrates that routinely available health and demographic information may help health systems direct limited screening resources toward people whose disorders have not yet been recognized. Whether that improves treatment access or outcomes will require prospective studies in real clinical settings.
