A model picked Rhode Island neighborhoods at highest overdose risk. The trial's posted numbers do not show that it helped

The short version

Researchers built a machine-learning model to predict which Rhode Island neighborhoods would have the most overdose deaths in the next six months. The state health department wanted a list short enough to act on: 20% of neighborhoods holding at least 40% of the deaths. The model met that target and beat simply picking the places with the most past deaths.

Whether acting on the list saves lives is the question the model's randomized trial was built to answer. The trial's posted main result points the wrong way: towns that used the model had a higher overdose rate than towns that did not. Those are raw numbers with no statistical test, and no peer-reviewed analysis has been published, so they do not yet show whether the model helped, hurt, or made no difference.

What they did

The team, from New York University, Brown, the University of California, Berkeley, and the Rhode Island Department of Health, split the state into 809 populated census block groups. Each holds roughly 600 to 3,000 residents, and the paper treats them as neighborhoods.

The model used data most state health departments already hold: overdose deaths, ambulance runs for nonfatal opioid overdoses, the prescription monitoring program (including buprenorphine prescriptions), census survey data, and the locations of treatment programs, health care, and social services. It combined two machine-learning methods and updated every six months.

The rules were set with the health department before testing. Pick no more than 20% of neighborhoods, the share the department said outreach groups could realistically cover. Include at least one neighborhood in each of the state's 39 cities and towns. Capture at least 40% of statewide overdose deaths. The comparison, standing in for usual practice, was the 20% of neighborhoods with the most past overdose deaths under the same one-per-town rule.

What it found

In two test periods, from July 2019 to June 2020, the model's picks held 40.2% of the state's overdose deaths on average. The past-deaths approach held 33.5%. In the following year, July 2020 to June 2021, used to validate the model before the randomized comparison began, the model's picks held 44.1% on average, against 36.2%.

Overdose deaths were spread thin. The typical neighborhood had one over the full four and a half years of study data. The state's overdose death rate rose from 29.4 per 100,000 residents in 2016 to 33.9 in 2020.

The model also chose different places. Among urban neighborhoods, it picked a larger share of majority non-white neighborhoods (54.4% against 40.2% in the test periods) and a smaller share of majority white ones (4.5% against 11.7%).

What the trial registry shows so far

The trial randomly assigned Rhode Island's 39 cities and towns to two groups. In 20, the health department and community groups got the model's picks to guide where naloxone, outreach, and other prevention resources went. The other 19 received prevention resources as usual. Every town kept getting the state's regular overdose surveillance reports.

The trial ran from November 2021 to August 2024, and results were posted on ClinicalTrials.gov in April 2026. The main measure was fatal and nonfatal overdoses per 10,000 residents over the trial. The registry lists 43.5 for the towns that used the model and 37.1 for the towns that did not.

Those two numbers need care. The registry posts no statistical test of the difference and no pre-trial overdose rates for each group, so it does not show whether the groups started out alike. The trial's posted analysis plan explains why that matters. It calls for comparing the groups after accounting for each town's pre-trial overdose trend, notes some imbalance between the groups in racial and ethnic makeup, and plans extra checks for a June 2023 Providence police policy that sent all overdose-related 911 calls to EMS, which changes how many nonfatal overdoses get counted. Without the investigators' full analysis, the numbers cannot show whether the model helped, hurt, or made no difference. A PubMed search on September 29, 2026 found no published paper with the trial's main result.

What it does not show

The paper tests a prediction, not a prevention strategy. It shows where risk was concentrated, not whether sending naloxone or outreach there changed anything.

The authors list their own limits. The model had only seven six-month periods to learn from, so fast shifts in where overdoses happen could throw it off. It was trained on data from before COVID-19 and tested mostly on data from after. Some inputs, such as buprenorphine prescriptions, only see people already connected to services, and methadone treatment data were not available. The model favored accuracy over being easy to explain. Rhode Island has unusually rich data, so the model may not carry over to other states. The one-neighborhood-per-town rule, chosen for fairness and the trial design, may send some resources to lower-risk places.

The data are not public because of health department restrictions. The team posted demonstration code, but outsiders cannot rerun the analysis on the real data.

The work was funded by the National Institute on Drug Abuse, and the authors declared no conflicts of interest. The health department helped set the rules for success, and two of its staff are authors.

Why it matters

Most machine-learning work on overdose, the authors note, has scored individual people. PROVIDENT scores places, which fits decisions health departments already make, such as where to distribute naloxone and send street outreach.

For health departments, the paper shows that data they already collect can point to next season's hardest-hit neighborhoods somewhat better than last season's death counts. Whether acting on the forecast lowers the count waits on the trial's full analysis.