Development and Validation of a Multivariable Machine Learning Model for Mortality Prediction among Intensive Care...

Area of Science:

• Critical care medicine
• Biomedical informatics
• Machine learning applications

Background:

• Accurate prediction of mortality in intensive care units (ICUs) is crucial for patient management and resource allocation.
• Traditional scoring systems like APACHE II and SOFA may not fully capture the complexity of critical illness.
• Electronic health record (EHR) data offers a rich source for developing advanced predictive models.

Purpose of the Study:

• To develop and internally validate machine learning models for predicting ICU mortality.
• To compare the performance of machine learning models against traditional scoring systems.
• To leverage routinely collected EHR data for enhanced predictive accuracy.

Main Methods:

• A retrospective cohort study of 5,553 adult ICU admissions.
• Development of Logistic Regression, Random Forest, and XGBoost models using demographic data, APACHE II, SOFA scores, comorbidities, and ventilatory support.
• Model performance evaluated using AUROC, precision, recall, and F1 score, with data split into 80% development and 20% testing cohorts.

Main Results:

• ICU mortality rate was 33.6% among the 5,553 patients.
• Random Forest (AUROC 0.842) and XGBoost (AUROC 0.835) outperformed Logistic Regression (AUROC 0.833).
• APACHE II and SOFA scores were significant predictors across all models; ensemble models captured non-linear relationships.

Conclusions:

• Machine learning models, particularly Random Forest and XGBoost, demonstrate high efficacy in predicting ICU mortality.
• These models offer potential improvements over traditional scoring systems for risk stratification.
• Integrating machine learning into ICU care can enhance clinical decision-making and patient management.