Machine Learning Based Prediction of Hospital Mortality in Post-Cardiotomy Cardiogenic Shock with Mechanical Support

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence

Background:

• Post-cardiotomy cardiogenic shock (PCCS) patients requiring extracorporeal life support (ECMO) have high mortality rates.
• Accurate prognostication is crucial for optimizing clinical decision-making and resource allocation in these critically ill patients.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting in-hospital mortality in PCCS patients on ECMO.
• To identify key predictors of mortality to improve risk stratification and clinical management.

Main Methods:

• Utilized data from the Extracorporeal Life Support Organization (ELSO) registry (2010-2020) including 5,982 adult patients.
• Trained six ML algorithms (boosting, decision tree, k-NN, random forest, naïve Bayes, neural networks) on integrated and preprocessed data.
• Validated models using a 60% training, 20% validation, and 20% test set split.

Main Results:

• The boosting algorithm achieved the highest predictive performance (AUC = 0.759), followed by random forest (AUC = 0.688).
• Significant predictors of mortality included age, lactate levels during support, arterial pH, and BMI.
• ECMO duration was a less reliable predictor, and transplant outcome prediction was limited by class imbalance.

Conclusions:

• ML models offer moderate predictive capability for in-hospital mortality in PCCS patients on ECMO.
• The random forest model highlights the utility of readily available clinical variables for risk assessment.
• Further external validation and calibration are necessary before clinical implementation; models can serve as dynamic risk assessment tools during ECMO support.