Static Algorithm, Evolving Epidemic: Understanding the Potential of Human-AI Risk Assessment to Support Regional Overdose Prevention

TL;DR

This paper addresses how AI-based regional overdose risk predictions can support local public health decisions in the face of evolving overdose dynamics. It introduces Rx RiskMap, a design probe that visualizes aggregated PDMP-linked risk predictions with SHAP-based explanations, evaluated through focus groups with 11 Pennsylvania public health officials. Key contributions include an early, practice-informed evaluation of a regional risk dashboard, insights into explainable design, model formulation, and adapting centralized data to local contexts, and guidelines for future AI-driven public health decision-support tools. The findings highlight both the potential for improved resource allocation and collaboration, and the challenges of data misalignment, interpretability, and actionability in dynamic, local epidemic contexts.

Abstract

Drug overdose deaths, including those due to prescription opioids, represent a critical public health issue in the United States and worldwide. Artificial intelligence (AI) approaches have been developed and deployed to help prescribers assess a patient's risk for overdose-related death, but it is unknown whether public health experts can leverage similar predictions to make local resource allocation decisions more effectively. In this work, we evaluated how AI-based overdose risk assessment could be used to inform local public health decisions using a working prototype system. Experts from three health departments, of varying locations and sizes with respect to staff and population served, were receptive to the potential benefits of algorithmic risk prediction and of using AI-augmented visualization to connect across data sources. However, they also expressed concerns about whether the risk prediction model's formulation and underlying data would match the state of the overdose epidemic as it evolved in their specific locations. Our findings extend those of other studies on algorithmic systems in the public sector, and they present opportunities for future human-AI collaborative tools to support decision-making in local, time-varying contexts.

Figures (4)

• Figure 1: Validation of the PDMP risk prediction model compared to univariate measures of risk currently used by the Pennsylania PDMP. The rate of overdose-related deaths is shown on the $y$-axis by the percentile of each individual's risk, as computed by the model (blue), the patient's morphine milligram equivalent (MME, brown), or number of days of overlapping opioid and benzodiazepine prescriptions (green). Though the rate of fatal overdose is extremely low overall (less than 0.5%, or 50 deaths per 10,000 people), it shows good correlation with the risk score percentile. Figure adapted from Gellad2023.
• Figure 2: Rx RiskMap allows public health experts to compare AI-based risk scores for geographic regions at the ZCTA (ZIP code) or county level. The visualization tool provides (a) a choropleth map depicting the risk score or other covariates, (b) a slider to show different time periods of data, (c) a time-series plot of variables of interest comparing selected regions against the statewide average, (d) detail panes illustrating the distribution of each variable as well as explanations of the most important features driving the risk score, and (e) descriptions of how the variables have changed over time. Note: the data shown in all screenshots of Rx RiskMap is outdated and provided for illustration purposes only.
• Figure 3: The Important Variables chart shows average feature importance scores for the risk predictions within each selected region and week. Sorting one of the detail cards using the vertical arrow button links the $y$ axis for the charts in the other cards to that region.
• Figure 4: The Weekly Changes pane shows the variables that have changed the most from one week to the next, along with a verbal breakdown of the trend.