This study utilized patient data from two databases: the Medical Information Mart for Intensive Care IV (MIMIC-IV) database and the eICU Collaborative Research Database dataset (eICU-CRD). MIMIC-IV (version 2.2) is an extensive critical care database, encompassing detailed records of over 190,000 ICU patients from 2008 to 2019. This dataset aggregates a wealth of clinical information, including patients’ demographic profiles, laboratory test results, medication histories, and additional comprehensive data sets (22). The eICU Collaborative Research Database (eICU-CRD) pools detailed data from more than 200,000 patients in various U.S. intensive care units, collected between 2014 and 2015, making it an essential tool for advancing research in critical care medicine (23).

This study focuses on patients with AMI-CS, and obtains the relevant patient data from two databases using ICD-9 and ICD-10 codes. The exclusion criteria for the study population are as follows: (1) under the age of 18, and (2) no ICU experience or an ICU stay less than 24 h. For patients with multiple admissions or a history of ICU stays, only the first ICU experience during their first admission is included. In our research, cardiogenic shock (CS) was defined based on several clinical parameters, primarily including the following: ① Systolic blood pressure (SBP) < 90 mmHg or the need for vasopressor support to maintain SBP ≥ 90 mmHg; ② Signs of reduced cardiac output leading to poor organ perfusion (e.g., low urine output, cold and clammy skin, altered mental status); ③ Exclusion of other causes of shock, such as hypovolemic or septic shock.

In this study, we included ICU patients diagnosed with AMI-CS, from whom we extracted the following data: (1) Demographics: age, gender, height, and weight; (2) Vital Signs: temperature (T), heart rate (HR), respiratory rate (R), systolic blood pressure (SBP), diastolic blood pressure (DBP), mean blood pressure (MBP), and peripheral oxygen saturation (SpO2); (3) Laboratory Indicators: complete blood count, liver and kidney function tests, electrolytes, lipid profile, blood gases, coagulation profile, and cardiac enzymes; (4) Comorbidities: hypertension, diabetes mellitus, hyperlipidemia, chronic obstructive pulmonary disease (COPD), pneumonia, chronic kidney disease (CKD), and atrial fibrillation (AF); (5) Surgical Indicators: coronary angiography (CAG), percutaneous coronary intervention (PCI), percutaneous transluminal coronary angioplasty (PTCA), and intra-aortic balloon pump (IABP); (6) Medication data: angiotensin-converting enzyme inhibitors/angiotensin II receptor blockers (aceiorarb), beta Blockers, furosemide, spironolactone, dobutamine, dopamine, epinephrine, milrinone, norepinephrine, and phenylephrine; and (7) Other Indicators such as Mechanical Ventilation and APACHE II score. The primary outcome was 28-day all-cause mortality, defined as death from any cause within 28 days starting from admission to the ICU.

The raw data is refined through advanced calculations. This involves computing Body Mass Index (BMI) from height and weight, and assessing in-hospital mortality within 28 days using hospitalization duration and survival status. In both the MIMIC-IV and eICU-CRD datasets, we rigorously addressed missing data by eliminating entries with missing rates exceeding 30%. Subsequently, we employed the K-Nearest Neighbors (KNN) imputation method to handle remaining missing values. KNN imputation leverages sample similarity, utilizing observed values from the K nearest samples to predict and fill missing values effectively. Exploration of variable relationships was conducted using Spearman correlation coefficients, visually represented through heatmaps. Multicollinearity among variables was meticulously assessed using Variance Inflation Factor (VIF) values. To optimize predictive performance and prevent overfitting, we selectively pre-screened predictive variables with high correlations or VIF exceeding 5.

After preprocessing, we initially included 38 predictive factors to construct a model for predicting in-hospital mortality within 28 days for patients with AMI-CS. To improve the stability of the predictive model, all continuous variables were standardized, scaled to have a mean distribution of 0 and a standard deviation of 1. The Boruta algorithm was used for feature selection. Its principle is to determine the most relevant features in the dataset based on the Random Forest by comparing with randomly generated “shadow” features. This method has the advantage of not requiring assumptions, which enhances the robustness of the model and simplifies the feature selection process (24). The XGBoost method was employed to rank the selected features based on their importance in a professional manner. The MIMIC dataset was used to construct the model, employing a 10-fold cross-validation method to generate training and validation sets. Four machine learning models were established and validated, including Logistic Regression (LR), eXtreme Gradient Boosting (XGBoost), Adaptive Boosting (AdaBoost), and Gaussian Naive Bayes (GNB). Model comparisons are conducted on the validation set. The Area Under the Receiver Operating Characteristic (ROC) curve (AUC) is utilized to assess the model's discriminative ability, while calibration curves and Brier scores are employed to evaluate model accuracy. Additionally, Decision Curve Analysis (DCA) curves are utilized to assess clinical utility. Additionally, confusion matrix metrics were included for evaluation, such as accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. We also compared the predictive performance of our model with commonly used severity scoring systems such as SOFA, APSIII, SAPSII, and OASIS.

The SHAP (SHapley Additive exPlanations) method is an approach for explaining machine learning model predictions (25). It leverages Shapley values to decompose the impact of each feature on the model output, providing insights into the prediction process. By considering interactions between features, SHAP offers intuitive explanations for individual predictions, aiding in the identification of key features and prediction sources. In SHAP plots, red and blue points represent the SHAP values of each sample, with red indicating higher feature values and blue indicating lower feature values. Observing the distribution of red and blue points helps understand the contribution and direction of each feature towards the model output.

Patients were categorized into two groups based on their 28-day mortality status. Continuous variables were summarized using means and standard deviations and compared using t-tests (or Wilcoxon rank-sum tests). Categorical variables were presented as percentages of the total and compared using chi-square tests (or Fisher's exact tests). A P-value < 0.05 was considered statistically significant. Statistical analyses were conducted using R version 4.2.3 and Python version 3.11.4.

According to the inclusion and exclusion criteria, a total of 961 patients with AMI-CS were enrolled in this study, including 570 in the MIMIC-IV database and 391 in the eICU-CRD database. The screening process is illustrated in Figure 1. In the MIMIC-IV database, 215 cases of AMI-CS patients died within 28 days (mortality rate: 37.7%), compared to 102 cases in the eICU-CRD (mortality rate: 26.1%). Differences in baseline characteristics are summarized in Table 1. In both the MIMIC-IV and eICU-CRD databases, patients who died had higher levels of age, blood urea nitrogen (BUN), prothrombin time (PT), creatinine (Cr), international normalized ratio (INR), respiratory rate (R), and APACHE III score, as well as lower levels of serum albumin (ALB) and pH (p < 0.05). Furthermore, differences were observed in the use of angiotensin-converting enzyme inhibitors/angiotensin receptor blockers (aceiorarb), beta-blockers, spironolactone, epinephrine, and ventilation (p < 0.05) between the survival and mortality groups. However, there were no significant differences in comorbidities such as hypertension, diabetes, and chronic obstructive pulmonary disease (COPD) between the two groups.

We removed features with a missing rate exceeding 30%, as shown in the Supplementary Figure S1. Additionally, we illustrated the correlation heat map for features with correlations exceeding 0.5, along with the results of features with VIF exceeding 5, as depicted in the Supplementary Figure S2 and Supplementary Figure S3. For features exhibiting high correlation and VIF, we conducted screening before model construction and excluded the following features: red blood Cell count (RBC), coronary angiography (CAG), aspartate aminotransferase (AST), mean blood pressure (MBP), diastolic blood pressure (DBP), international normalized ratio (INR).

We employed the Boruta algorithm for feature selection and generated a plot, shown in Figure 2. Boruta assesses feature importance by creating shadow features (derived from shuffling original feature values) and training them alongside original features in a random forest. In the plot, green denotes important features, enhancing model prediction and thus included. Red represents unimportant features excluded from the model, while yellow signifies uncertain importance, necessitating further investigation. Blue indicates shadow features used for comparison but not in model training. Through Boruta selection, we identified 13 features for inclusion: angiotensin-converting enzyme inhibitors or angiotensin II receptor blockers (aceiorarb), beta blockers (betablockers), furosemide, dobutamine, norepinephrine, age, albumin (ALB), glucose (GLU), alanine aminotransferase (ALT), blood urea nitrogen (BUN), creatinine (Cr), prothrombin time (PT), and temperature (T), along with two potential features: hemoglobin (Hb) and sodium (Na).

We utilized the XGBoost method to rank the importance of features, as depicted in Supplementary Figure S4. The top 10 variables, ranked from highest to lowest importance, are: “aceiorarb”, “betablockers”, “PT”, “age”, “BUN”, “GLU”, “T”, “Na”, “dobutamine”, and “Hb”.

This study employed four binary classification machine learning algorithms: Logistic Regression (LR), eXtreme Gradient Boosting (XGBoost), Adaptive Boosting (AdaBoost), and Gaussian Naive Bayes (GNB). Utilizing the MIMIC database, we employed a 10-fold cross-validation technique to divide the dataset into training and validation subsets, and evaluated the model performance on the validation subsets. Figure 3A and Table 2 describe the performance of these predictive models, with results indicating that the LR model exhibits better discriminative ability, achieving an AUC of 0.841 in the test queue, compared to other ML models (AUC: XGBoost = 0.835; AdaBoost = 0.839, GNB = 0.826). Furthermore, LR's calibration curve closely approximated the ideal line (Figure 3B), exhibiting the lowest Brier score, indicative of superior calibration. Decision Curve Analysis (DCA) illustrated in Figure 3C revealed LR's highest net benefit within the 0%–80% threshold range. As shown in the PR curve (Figure 3D), the LR model maintains high precision while capturing more positive samples, demonstrating superior performance. Hence, LR was selected for model development, incorporating five predictive variables: betablockers, aceiorarb, PT, age, and BUN. Model parameter optimization through hyperparameter tuning and grid search yielded the following settings: tol (convergence criterion) = 0.0001, penalty (regularization type) = l2, max_iter (maximum number of iterations) = 100, and C (regularization factor) = 10.0.

We utilize the SHAP method to interpret the model results. For global interpretation, the absolute Shapley values are averaged across all instances in the data, yielding importance values for each feature. The results are visualized as SHAP summary plots, where the Y-axis represents the features, and the X-axis indicates the magnitude of the feature's impact on the outcome. Each point represents a sample, with red points indicating high-risk values and blue points indicating low-risk values. As shown in SHAP summary plot (Figure 4A), the feature importance from top to bottom is: aceiorarb, betablockers, age, BUN, PT. Patients using aceiorarb (red points) have a reduced risk of death, while patients not using betablockers (blue points) have an increased risk of death. Similarly, higher values of age, BUN, and PT are associated with higher risk of death.

For local interpretation, each observation has its own set of Shapley values, which can be utilized to explain the contribution of each sample and feature to the prediction. The results are visualized as SHAP force plots, where each Shapley value is represented as an arrow that can either push the prediction up (positive value) or down (negative value). Due to the standardization of numerical variables, the age, BUN, and PT values in the plot do not represent the actual values of individuals. As depicted in Figure 4B for patient 1, as age increases, the predicted mortality risk rises, while a decrease in BUN value corresponds to a decrease in predicted mortality risk. As depicted in Figure 4C for patient 2, a decrease in age and the use of ACEI or ARB drugs are associated with a reduction in predicted mortality risk.

Based on the LR model, which exhibited the best performance, we established an online prediction model. As shown in Figure 5, we present the survival prediction for actual patients using the online website. The patient is 64 years old, with PT of 17s, BUN of 22.08 mg/dl, has used ACEI/ARB, and has used beta-blockers. The probability of the occurrence of the disease is: 12.0% (the threshold of the occurrence of the disease is: 45.7%). To facilitate understanding, we randomly selected four survivors and four deceased patients from the MIMIC-IV and eICU databases in the past. We had used the online prediction model to estimate their mortality risk. According to the LR model, when the predicted mortality risk was less than 50%, the patient was inferred to have survived beyond 28 days, and when it was greater than 50%, the patient was inferred to have died within 28 days. The final results, which were presented in Supplementary Table S3, demonstrated a 100% accuracy in predicting the outcomes for these eight patients. We extracted 391 cases of AMI-CS patients from the eICU-CRD database as an external validation dataset. The data characteristics are outlined in Table 2. The 28-day in-hospital mortality rate in the eICU-CRD database was 26.1%, lower than that in the MIMIC database. Prior to external validation, we processed the eICU-CRD data using the same methods as the MIMIC data. Due to the substantial imbalance between survival and death cases, we applied Synthetic Minority Over-sampling Technique (SMOTE) to balance the dataset. The results of external validation revealed an AUC of 0.755 and a Brier score of 0.229. Other metrics for external validation, such as accuracy, sensitivity, etc., are available in Table 2. Additionally, an analysis of commonly used severity scores using the MIMIC-IV dataset revealed model AUC values, including SOFA (AUC = 0.620), APSIII (AUC = 0.710), SAPSII (AUC = 0.660), and OASIS (AUC = 0.640).