Using Machine Learning Approaches for Emergency Room Visit Prediction Based on Electronic Health Record Data

Area of Science:

• Healthcare Informatics
• Machine Learning in Medicine
• Clinical Data Analysis

Background:

• Emergency room (ER) visit prediction is critical for hospital resource allocation and patient health forecasting.
• Existing machine learning approaches, primarily general linear models, struggle with complex correlations inherent in clinical data.
• The limitations of linear models hinder accurate prediction of ER visit likelihood and frequency.

Purpose of the Study:

• To investigate the efficacy of non-linear machine learning models for ER visit prediction.
• To compare the performance of XGBoost and Recurrent Neural Networks against traditional methods.
• To improve the accuracy of predicting both the occurrence and count of ER visits.

Main Methods:

• Utilized XGBoost, a gradient boosting algorithm, for non-linear predictive modeling.
• Employed Recurrent Neural Networks (RNNs), a class of deep learning models, for sequence-aware prediction.
• Evaluated model performance on clinical datasets to predict ER visit status and count.

Main Results:

• Both XGBoost and Recurrent Neural Networks demonstrated superior performance compared to existing linear models.
• Non-linear models effectively captured complex correlations between clinical features and ER visit outcomes.
• Improved accuracy in predicting both the probability and the number of emergency room visits was observed.

Conclusions:

• Non-linear models, specifically XGBoost and RNNs, offer significant advantages for ER visit prediction.
• These advanced models provide more accurate forecasting, enabling better healthcare resource management.
• The findings suggest a shift towards utilizing non-linear approaches for complex clinical prediction tasks.