A1330
Title: Design-based provider profiling via constrained double/debiased deep learning
Authors: Wenbo Wu - Johns Hopkins University (United States) [presenting]
Fan Li - Yale University (United States)
Ivan Diaz - NYU Langone Health (United States)
Abstract: Existing profiling models frequently rely on linear risk adjustment approaches that are ill-suited to capture complex relationships between patient characteristics and outcomes. Additionally, many traditional methods prioritize cost containment and overlook the socioeconomic heterogeneity among racial and ethnic groups, limiting their ability to support equity-focused interventions. A novel approach to provider profiling is proposed that incorporates customized comparators. The method harnesses neural networks to relax linearity constraints and employs doubly robust machine learning to reduce bias in estimating provider effects. This approach advances the methodological foundations of provider profiling, ultimately supporting more equitable care decisions by patients and stakeholders and enhancing evidence-based accountability for healthcare providers.
