Machine learning-based early warning system for hemodynamic deterioration in cardiovascular ICU patients: a...

Area of Science:

• Cardiovascular Medicine
• Artificial Intelligence in Healthcare
• Critical Care Medicine

Background:

• Early identification of hemodynamic deterioration in cardiovascular intensive care unit (ICU) patients is critical for improving clinical outcomes.
• Traditional monitoring and scoring systems often fail to capture dynamic physiological changes.
• Existing machine learning models frequently lack robust external validation across diverse healthcare systems.

Purpose of the Study:

• To develop and validate machine learning prediction models for early detection of hemodynamic deterioration in cardiovascular ICU patients.
• To assess the robustness and generalizability of these models across different healthcare systems using a bidirectional cross-validation framework.
• To compare the performance of machine learning models against traditional clinical scoring systems.

Main Methods:

• Retrospective multi-center cohort design using MIMIC-IV and eICU databases.
• Development of machine learning models with a focus on Random Forest classifier.
• Bidirectional cross-validation (MIMIC-eICU and eICU-MIMIC) to ensure robustness and generalizability.
• Definition of a composite outcome including hemodynamic instability, tissue hypoperfusion, and cardiac etiology.

Main Results:

• The Random Forest model demonstrated strong cross-database generalizability with AUROCs of 0.841 (MIMIC-trained on eICU) and 0.852 (eICU-trained on MIMIC).
• The model significantly outperformed traditional scores like SOFA (AUROC 0.681) and APACHE II (AUROC 0.747).
• A five-level risk stratification system showed a clear correlation between risk level and mortality, and SHAP analysis identified key predictors such as hemoglobin, history of myocardial infarction, and creatinine.

Conclusions:

• A validated machine learning-based early warning system for hemodynamic deterioration in cardiovascular ICU patients was successfully developed.
• The bidirectional cross-validation approach confirms the model's robustness and generalizability.
• The system offers practical clinical decision support through risk stratification and interpretability, potentially improving patient outcomes and healthcare efficiency.