Related Experiment Videos
Development and external validation of a machine learning model for predicting 28-day mortality in patients with
Jiechun Yao1, Chun Lin2, Qingbo Xu1
1Cardiac and Circulatory Medicine Section, Maoming People's Hospital, Maoming, Guangdong, China.
Background:
Acute myocardial infarction (AMI) complicated by malignant ventricular arrhythmia (MVA) carries high 28-day mortality. Existing risk scores inadequately capture this population. We aimed to develop and externally validate an interpretable machine learning model for predicting 28-day mortality in AMI-MVA patients.
Methods:
This retrospective study included 952 AMI-MVA patients from MIMIC-IV (training n = 668, internal validation n = 284) and 100 patients from Maoming People's Hospital, China (external validation). Feature selection integrated multivariable logistic regression, LASSO, and Boruta, yielding eight predictors: age, lactate, fasting blood glucose, RDW, antiplatelet agents, beta-blockers, ACEI/ARB, and norepinephrine. Seven machine learning algorithms were compared by AUC, calibration, and decision curve analysis. Treatment-free sensitivity models, propensity-score matching (PSM), and inverse-probability-of-treatment-weighting (IPTW) were performed to assess confounding by indication.
Results:
28-day mortality was 32.46% (309/952). The SVM model (linear kernel, C = 0.1) achieved AUC 0.865 (95% CI 0.824-0.909) in internal validation and 0.898 (95% CI 0.835-0.949) externally, with excellent calibration (Brier score 0.128). SHAP analysis identified beta-blockers, lactate, and ACEI/ARB as the three most influential predictors. A treatment-free sensitivity model (biomarkers only) achieved AUC 0.801 (95% CI 0.748-0.851) internally and 0.771 (95% CI 0.675-0.853) externally. PSM and IPTW attenuated but did not eliminate treatment-outcome associations, consistent with a mixture of genuine prognostic signal and residual confounding.
Conclusion:
This interpretable SVM model supports bedside risk stratification for AMI-MVA patients within the first 24 h of ICU admission. The model should be understood as a prognostic snapshot rather than an early prediction tool. Prospective, multicentre validation with standardised severity scores and coronary-anatomy variables is required before routine clinical use.
Insights
A new machine learning model accurately predicts 28-day mortality in acute myocardial infarction (AMI) patients with malignant ventricular arrhythmia (MVA). This interpretable SVM model aids bedside risk stratification for improved patient outcomes.
Area of Science:
- Cardiology
- Machine Learning
- Medical Informatics
Background:
- Acute myocardial infarction (AMI) with malignant ventricular arrhythmia (MVA) has high mortality.
- Existing risk scores are insufficient for this patient group.
- Need for improved risk stratification in AMI-MVA patients.
Purpose of the Study:
- Develop and externally validate an interpretable machine learning model.
- Predict 28-day mortality in AMI-MVA patients.
- Enhance clinical decision-making for high-risk cardiac patients.
Main Methods:
- Retrospective analysis of 952 AMI-MVA patients (MIMIC-IV) and 100 external validation patients.
- Feature selection using logistic regression, LASSO, and Boruta identified eight key predictors.
- Support Vector Machine (SVM) model developed and validated using AUC, calibration, and decision curve analysis.
Main Results:
- The SVM model achieved high predictive accuracy: AUC 0.865 (internal) and 0.898 (external).
- Excellent calibration (Brier score 0.128) and interpretability via SHAP analysis.
- Key predictors included beta-blockers, lactate, and ACEI/ARB, highlighting treatment and biomarker importance.
Conclusions:
- An interpretable SVM model provides effective bedside risk stratification for AMI-MVA patients within 24 hours of ICU admission.
- The model serves as a prognostic snapshot, not an early prediction tool.
- Further prospective, multicenter validation is recommended before routine clinical implementation.