Fairness Gaps in AI Tools for Opioid Treatment

Doctors struggle to keep people in opioid treatment programs. Many patients stop early, which hurts recovery.

Researchers used a big U.S. database of treatment records from 2015 to 2019. They built four computer models to guess if a patient would quit early or stay in care for 180 days. The data covered people who got outpatient medication for opioid use disorder.

When they checked the results, the models were off more for Black patients and for older adults. They also saw that race and age influenced the predictions differently for each group. This shows fairness problems even when overall accuracy looks okay.

Trying to fix the bias with pre‑processing, in‑processing, or post‑processing steps lowered some gaps. But these fixes often raised errors for other groups or made the model less accurate overall. So the tools need careful balancing.