Machine learning for predicting acute myocardial infarction in patients with sepsis

Abstract

Acute myocardial infarction (AMI) and sepsis are the leading causes of high mortality rates in intensive care units. While sepsis frequently affects the cardiovascular system, distinguishing between sepsis-induced cardiomyopathy and AMI remains challenging due to overlapping biomarkers. Misdiagnosis can hinder timely treatment and increase risk of complications. This study used multidimensional clinical data and machine learning techniques to develop and validate a novel predictive model for identifying AMI in critically ill patients with sepsis. Data from patients with sepsis were extracted from the Medical Information Mart for Intensive Care-IV database. Six machine learning algorithms were employed for model construction. Additionally, the machine learning-based models were compared with traditional scoring systems. Model performance was evaluated in terms of discrimination, calibration, and clinical applicability. In total, 2,103 critically ill patients with sepsis were included, 459 (21.8%) of whom experienced AMI during hospitalization. A total of 26 variables were selected for model construction. Among all models, the Gradient Boosting Classifier model demonstrated the best predictive performance in terms of discrimination, calibration, and clinical applicability. Machine learning models have the potential to serve as tools for predicting AMI in patients with sepsis. The Gradient Boosting Classifier model developed herein demonstrated promising predictive performance, supporting clinicians in identifying patients at high-risk of sepsis and implementing early interventions to reduce mortality rates.

Keywords: Acute myocardial infarction; MIMIC-IV database; Machine learning; Prediction model; Sepsis.

Fig. 1

Flowchart of patient selection. MIMIC-IV medical information mort for intensive care, ICU intensive care unit.

Fig. 2

Feature selection based on the Boruta algorithm. The horizontal axis represents all variables, and the vertical axis represents the Z value of each variable. The box plot shows the Z values of each variable during model calculation. The green boxes represent the first 26 important variables, the yellow boxes represent tentative attributes, and the red boxes represent unimportant variables.

Fig. 3

Receiver operating characteristic curves of the eleven models. Figure (a) Includes 6 machine models; Figure (b) includes five traditional scoring systems. LR logistic regression, KNN k-nearest neighbors, RF random forest, GBC gradient boosting classifier, SVC support vector, DT decision tree, APS III acute physiology score III, LODS logistic organ dysfunction system, CCI Charlson comorbidity index, SOFA sequential organ failure assessment, OASIS Oxford acute severity of illness scale, AUC area under the curve.

Fig. 4

Recall vs. decision boundary T curves for sepsis patients with acute myocardial infarction (AMI) and non-AMI patients for the top performing model, the gradient boosting classifier. At the default classification of T = 0.5 (middle dotted line), the recall rate was 0.94 for non-AMI patients and 0.25 for AMI patients. For T = 0.77 (right dotted line), the recall rates for non-AMI and AMI patients were both 0.75, representing a large improvement in identifying AMI patients, with only a small decrease in the recall rate (19%) for non-AMI patients.

Fig. 5

Recall vs. decision boundary curves for AMI patients and non-AMI patients according to all 11 ML models. Figure (a) includes 6 machine models; Fig. (b) includes five traditional scoring systems. The gradient boosting classifier model performed the best. The metrics are shown in Table 2.

Fig. 6

ROC curve for the gradient boosting classifier. The 2-standard deviation spread representing the testing performance for models trained on N_boot bootstrapped samples is also plotted.

Fig. 7

Decision curve analyses of the six models.

Fig. 8

Feature importance derived from the gradient boosting classifier model. SHAP revealed the top 10 features that had the greatest impact on the GBC model predictions.