Superior machine learning model for post-cardiac arrest mortality prediction: a MIMIC-IV cohort study

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Biomedical Data Science

Background:

• Post-cardiac arrest syndrome (PCAS) remains a significant cause of mortality.
• Despite advances in resuscitation, predicting outcomes in PCAS is challenging.
• Machine learning offers potential for improved mortality prediction in these complex cases.

Purpose of the Study:

• To develop and validate a machine learning model for predicting 28-day mortality in post-cardiac arrest patients.
• To identify key clinical variables that are most predictive of mortality.
• To compare the performance of machine learning algorithms against traditional scoring systems.

Main Methods:

• Analysis of 853 cardiac arrest patients from the MIMIC-IV database with complete 28-day outcome data.
• Extraction and comparison of 99 clinical variables across six domains using five machine learning algorithms.
• Calculation of lactate clearance and evaluation of model performance using AUC-ROC, calibration, and SHAP analysis.

Main Results:

• The XGBoost machine learning model achieved an AUC-ROC of 0.89, outperforming the APACHE III score (AUC: 0.73).
• Lactate clearance rate was the strongest predictor of mortality (SHAP value: 0.24), followed by GCS and SOFA scores.
• Poor lactate clearance (<25%) was associated with significantly higher 28-day mortality (67%) compared to moderate clearance (39%).

Conclusions:

• A machine learning model, particularly XGBoost, demonstrates superior accuracy in predicting 28-day mortality for post-cardiac arrest patients.
• Dynamic lactate clearance and neurological assessment (GCS) are critical, actionable predictors for patient risk stratification.
• These findings support the integration of machine learning tools for enhanced clinical decision-making in critical care settings.