Using Encounter-Level Data for Risk-Adjustment of Antimicrobial Use Comparisons: Feasibility and Variable Selection.

Area of Science:

• Healthcare Analytics
• Health Informatics
• Antimicrobial Stewardship

Background:

• External comparisons of hospital antimicrobial use (AU) require risk-adjustment using encounter characteristics to inform antimicrobial stewardship program strategy.
• Barriers to encounter-level modeling include data collection feasibility and optimal variable selection for risk adjustment.

Purpose of the Study:

• To measure achievements in sharing validated, encounter-level AU data among a multisystem hospital collaborative.
• To compare variable selection strategies for AU risk-adjustment models using retrospective analyses.

Main Methods:

• Utilized electronic health record data from 50 US hospitals (2020-2021) for model training and testing.
• Compared four input variable strategies: diagnosis-related groups, Elixhauser comorbidities, agnostic Clinical Classification Software Refined (CCSR), and adjudicated CCSR.
• Employed gradient-boosted machine tree-based models to estimate antibacterial days of therapy (DOT), measuring accuracy with mean absolute error (MAE).

Main Results:

• Fifty of 76 hospitals successfully shared validated datasets.
• Modeling strategies with more CCSR inputs yielded the lowest MAE.
• Agnostic and adjudicated strategies showed highly correlated estimates and similar influential variables.

Conclusions:

• Expert adjudication was resource-intensive and did not yield superior results compared to an agnostic approach.
• Risk-adjustment using extensive encounter-level data and machine learning is feasible and valuable for future hospital AU assessments.