Optimization of Tree-Based Machine Learning Models to Predict the Length of Hospital Stay Using Genetic Algorithm

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Length of hospital stay (LOS) is a key metric for patient care quality, hospital efficiency, and operational resilience.
• Accurate LOS prediction is vital for effective hospital resource management and clinical decision support.
• Machine learning (ML) models offer potential for improving LOS prediction accuracy.

Purpose of the Study:

• To enhance the accuracy of length of hospital stay (LOS) prediction through hyperparameter optimization (HPO).
• To identify key variables influencing LOS by reviewing expert physician insights and existing literature.
• To evaluate and improve the performance of various ML models for LOS prediction.

Main Methods:

• Collected electronic medical records from 200 internal medicine patients in an Iranian hospital.
• Applied and evaluated several ML models including k-nearest neighbors (KNN), multivariate regression, decision tree (DT), random forest (RF), artificial neural network (ANN), and XGBoost.
• Utilized the genetic algorithm (GA) for hyperparameter optimization of tree-based models and dummy coding (One-Hot encoding) for categorical features.

Main Results:

• The XGBoost model optimized by the genetic algorithm (XGB_GA) demonstrated superior accuracy and prediction performance compared to other evaluated models.
• The XGB_GA model achieved a mean absolute error of 1.54 days and a median absolute error of 1.14 days on the test dataset.
• This optimized model reduced the mean absolute error by 37%, indicating a significant improvement in prediction reliability.

Conclusions:

• The XGBoost model optimized with the genetic algorithm (XGB_GA) provides a highly accurate and reliable method for predicting length of hospital stay.
• Improved LOS prediction accuracy facilitates better hospital resource management and planning.
• The developed XGB_GA model can be a valuable component in designing robust clinical decision support systems.