A comparison of models for predicting early hospital readmissions

Area of Science:

• Health services research
• Computational epidemiology
• Machine learning in healthcare

Background:

• Financial penalties from the Centers for Medicare and Medicaid (CMS) and risk-sharing agreements incentivize hospitals to reduce early readmissions.
• Existing 30-day readmission risk models often lack the predictive accuracy required for clinical application.

Purpose of the Study:

• To compare the performance of various predictive models, including novel approaches and deep learning methods, for forecasting 30-day hospital readmissions.
• To identify the most cost-effective conditions for intervention among those penalized by CMS.

Main Methods:

• Evaluation of several predictive models, including regression techniques and methods from deep learning.
• Application of models to predict readmissions for the five conditions targeted by CMS penalties.
• Development of a framework to assess the cost-effectiveness of interventions.

Main Results:

• Novel predictive models demonstrated superior performance compared to traditional regression methods.
• Deep learning approaches were applied to analyze readmission risks for CMS-targeted conditions.
• A framework was proposed to guide the selection of high-impact intervention targets.

Conclusions:

• Advanced predictive modeling, particularly deep learning, offers improved accuracy for identifying patients at high risk of 30-day readmission.
• Targeting specific high-cost conditions identified through this framework can optimize resource allocation for readmission reduction efforts.