Improvement of APACHE II score system for disease severity based on XGBoost algorithm.

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Biomedical Informatics

Background:

• Prognostication is crucial for risk adjustment and decision-making in intensive care units (ICUs).
• Existing models like APACHE II require improvement for more accurate patient severity assessment.
• Developing a superior mortality prediction model for ICU patients is essential to enhance patient outcomes.

Purpose of the Study:

• To develop and validate a novel mortality prediction model for ICU patients.
• To compare the performance of the new model against the established APACHE II scoring system.
• To utilize machine learning for enhanced accuracy in predicting hospital mortality.

Main Methods:

• Utilized the MIMIC-III database for model development and an independent dataset for external validation.
• Compared APACHE II with six machine learning (ML) methods, selecting the best performer (XGBoost).
• Employed TreeSHAP for explaining variable importance within the XGBoost model.

Main Results:

• The XGBoost model demonstrated higher accuracy (0.858) than APACHE II (0.742) using identical variables.
• Expanding the variable set improved XGBoost performance across accuracy, precision, recall, F1, and AUC metrics.
• External validation showed good calibration properties and an AUC of 0.79.

Conclusions:

• The proposed XGBoost model offers superior performance for predicting hospital mortality in ICUs compared to APACHE II.
• TreeSHAP provides valuable insights into feature importance, enhancing model interpretability.
• This model can aid clinicians in prognosis determination and ultimately improve patient care.