ICU Mortality Prediction Using XGBoost-based Scoring Systems: A Study from a Developing Country.

BACKGROUND

Accurate mortality prediction in intensive care units (ICUs) is essential for enhancing patient outcomes and optimizing healthcare resource allocation. Traditional scoring systems, such as APACHE, APACHE II, and SAPS, have limitations in handling complex, high- -dimensional ICU data. In this study, multiple machine learning models were compared to establish an efficacious predictive model for mortality tailored explicitly to the Jordanian population and to explicate factors strongly associated with mortality.

METHODS

This study was conducted as a single-center, retrospective cohort investigation, and the XGBoost machine learning algorithm was used to develop a novel ICU mortality prediction model. The model aimed to achieve superior prediction accuracy using a diverse set of readily available clinical data, including demographics, comorbidities, laboratory results, and medication groups. Model performance was evaluated against alternative machine learning algorithms, including logistic regression, conventionally employed in traditional scoring systems.

RESULTS

Comparative analysis revealed that the XGBoost model performed better than other scoring systems, manifesting heightened accuracy (87.91%), sensitivity (92.88%), and Area Under the Receiver-Operating Characteristic Curve (AUC-ROC) Score/Curve (94.29%). Notably, the patient's length of hospital stays, albumin levels, and urea levels emerged as the most substantial predictors for ICU mortality, each exhibiting respective SHAP values of 0.5, 0.41, and 0.37.

CONCLUSION

A locally adapted ICU mortality prediction model was developed, underscoring the pivotal role of predictors such as hospital stay duration, albumin, and urea levels in predicting patient outcomes. The heightened accuracy and sensitivity of the XGBoost model signify its potential as an invaluable tool in the critical task of mortality prediction within the Jordanian ICU context.