Novel integration of governmental data sources using machine learning to identify super-utilization among U.S....

Area of Science:

• Public Health
• Health Services Research
• Data Science

Background:

• Super-utilizers account for a disproportionate share of resource-intensive healthcare (RIHC).
• Reducing RIHC utilization is a significant challenge for U.S. healthcare systems.
• Identifying super-utilization at a population level, rather than individual, is needed.

Purpose of the Study:

• To predict RIHC utilization among U.S. counties.
• To utilize routinely collected U.S. government data, including consumer spending.
• To offer an alternative method for identifying super-utilization in population units.

Main Methods:

• A machine learning pipeline was developed using 2017 cross-sectional data from five governmental sources.
• Outcome metrics included yearly emergency room visits, inpatient days, and hospital expenditures.
• Predictor features comprised demographic, health, community, and consumer expenditure data.

Main Results:

• The study included 2475 counties with emergency rooms and 2491 with hospitals.
• Model performance (R-squared) ranged from 0.267 to 0.447.
• Demographic and community characteristics were key predictors for all RIHC outcomes.

Conclusions:

• Diverse population characteristics from governmental sources can predict RIHC metrics in U.S. counties.
• The developed model provides a novel and actionable tool for identifying population-level super-utilizers.
• Integrating routinely collected data offers alternative methods for predicting RIHC.