# Machine learning-based early warning system for hemodynamic deterioration in cardiovascular ICU patients: a bidirectional cross-validation study

## TL;DR

This study developed a machine learning system to predict hemodynamic deterioration in ICU patients, showing strong performance across different hospitals.

## Contribution

A novel bidirectional cross-validation approach was used to ensure model generalizability across diverse healthcare systems.

## Key findings

- The Random Forest model achieved AUROC scores of 0.841 and 0.852 in cross-database validation.
- The model outperformed traditional scores like SOFA and APACHE II in predicting deterioration.
- Key predictors included hemoglobin, history of myocardial infarction, and creatinine levels.

## Abstract

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

We employed a retrospective multi-center cohort design to develop machine learning prediction models using the MIMIC-IV database (46,007 admissions) and the eICU database (50,949 admissions). To rigorously assess model robustness and generalizability, a novel bidirectional cross-validation framework was implemented: models were trained on MIMIC data and validated on eICU data, and conversely, trained on eICU data and validated on MIMIC data. The study defined a strict composite outcome comprising hemodynamic instability, tissue hypoperfusion, and confirmed cardiac etiology. Multiple machine learning algorithms were evaluated to identify the optimal classifier.

The Random Forest model was selected as the optimal classifier. Bidirectional validation demonstrated exceptional cross-database generalizability: the MIMIC-trained model achieved an Area Under the Receiver Operating Characteristic (AUROC) of 0.841 on the eICU cohort, while the eICU-trained model achieved an AUROC of 0.852 on the MIMIC cohort, with performance degradation controlled within a minimal range (<4%). DeLong tests confirmed that the model significantly outperformed traditional clinical scores, including SOFA (AUROC 0.681) and APACHE II (AUROC 0.747). The five-level risk stratification system exhibited a strict monotonic increase in mortality rates, ranging from 0.8% in the very low-risk group to 84.2% in the very high-risk group. SHAP analysis identified hemoglobin, history of acute myocardial infarction, and creatinine as the most significant predictors.

We successfully developed and validated a machine learning-based early warning system for hemodynamic deterioration in cardiovascular ICU patients. The bidirectional cross-validation approach provides robust evidence for model generalizability, while the multi-level risk stratification system and SHAP-based interpretability offer practical clinical decision support. This system demonstrates significant potential to enhance early identification rates, improve patient outcomes, and optimize healthcare resource utilization efficiency.