Predicting Readmission Charges Billed by Hospitals: Machine Learning Approach.

Area of Science:

• Health informatics
• Machine learning in healthcare
• Predictive analytics

Background:

• Healthcare costs are rising, with significant contributions from patient readmissions.
• Machine learning applications are increasingly benefiting various healthcare sectors.
• Predicting readmission costs is crucial for managing escalating healthcare expenditures.

Purpose of the Study:

• To identify effective machine learning models for predicting hospital readmission charges.
• To address the underexplored area of machine learning in readmission cost prediction.
• To compare the performance of various predictive models, from interpretable to deep learning approaches.

Main Methods:

• Utilized the Nationwide Readmission Database (2013) for all-cause readmission category (RADC) and same-day readmission category (RSDC).
• Compared six machine learning algorithms: linear regression, lasso, elastic net, ridge, eXtreme Gradient Boosting (XGBoost), and Multilayer Perceptron (MLP).
• Employed 10-fold cross-validation to evaluate model performance on predicting readmission charges.

Main Results:

• XGBoost and MLP models demonstrated superior performance in predicting readmission charges.
• Both models achieved low error metrics, with XGBoost showing a Mean Absolute Percentage Error (MAPE) of 3.121% for RADC and 3.171% for RSDC.
• MLP models achieved MAPE of 3.103% for RADC and 3.202% for RSDC, with statistically significant lower RMSE compared to other models (P<.001).

Conclusions:

• XGBoost and MLP models are highly suitable for accurately predicting hospital readmission charges.
• Major diagnostic categories (MDCs) enhance the accuracy of these predictive models.
• The findings support the use of advanced machine learning for better financial forecasting in hospitals.