Establishment of ICU Mortality Risk Prediction Models with Machine Learning Algorithm Using MIMIC-IV Database.

Area of Science:

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

Background:

• High mortality rates in Intensive Care Units (ICUs) necessitate accurate prognosis assessment.
• Existing scoring systems like APACHE III and LODS aid clinicians but can be improved.
• Machine learning offers potential for enhanced prediction of ICU patient mortality.

Purpose of the Study:

• To develop and compare machine learning models for ICU mortality prediction.
• To evaluate the efficacy of Acute Physiology Score III (APS III) and Logistic Organ Dysfunction Score (LODS) in these models.
• To identify the optimal model for predicting mortality risk in critically ill patients.

Main Methods:

• Utilized a dataset of 14,110 patients from the MIMIC-IV database, balanced for deceased and surviving individuals.
• Employed XGBoost, logistic regression, support vector machine, and decision tree algorithms.
• Input variables included LODS and APS III subscores; model performance was assessed using AUC, specificity, sensitivity, and calibration curves.

Main Results:

• XGBoost achieved the highest Area Under the Curve (AUC) of 0.918, outperforming logistic regression (0.872), SVM (0.872), and decision tree (0.852).
• XGBoost demonstrated superior performance across various metrics, including ROC curve, sensitivity, and specificity.
• Calibration analysis showed XGBoost excelled in the 40-70% prediction range, while logistic regression and SVM performed better at the extremes.

Conclusions:

• XGBoost, utilizing APS III and LODS data, provides a highly accurate method for predicting ICU patient mortality.
• This model can significantly assist clinicians in evaluating in-hospital outcomes, particularly for patients with uncertain prognoses.
• The findings support the integration of advanced machine learning techniques into critical care decision-making.