The baseline characteristics of patients across the four centers included in the cohort are detailed in Table 1. The data demonstrates significant inter-centre variability in several key clinical parameters, which merit further attention.
The prevalence of cardiac arrest occurring before or at admission was markedly higher in Centers 3 and 4 (~3.0%) compared to Centers 1 and 2 (around 1.0%). Center 3 also exhibited significantly lower systolic blood pressure than the other centers. A history of diabetes mellitus was observed in approximately 17.0% of patients in Centers 1 and 3, whereas this proportion increased substantially to around 35.0% in Centers 2 and 4.
The duration from symptom onset to hospital admission was notably shorter at Center 4, with a median of 4 hours, compared to a significantly longer median of 7–8 hours at the other three centers. This disparity highlights the unique distribution of PCI-capable centres in developing countries, such as Vietnam, which are predominantly located in central urban areas. Consequently, most patients in this cohort experience prolonged travel times from their first medical contact to a PCI-capable center.
Anterior STEMI was the predominant type in Centers 2 and 4, comprising approximately 55.0%–60.0% of cases, in contrast to Centers 1 and 3, accounting for ~45.0%. The incidence of ventricular fibrillation or ventricular tachycardia was disproportionately higher in Center 4 (10.9%). Left ventricular ejection fraction (LVEF) was comparable across Centers 1, 3, and 4, ranging between 42.0% and 45.0%, but was notably higher in Center 2, at 51.9%.
The use of ticagrelor was relatively limited in Centers 3 (26.7%) and 4 (1.7%), where clopidogrel remained the antiplatelet agent of choice. By contrast, Centers 1 and 2 demonstrated a significantly higher adoption of ticagrelor (~64.0%). These differences can be attributed to the timeline of patient recruitment; Centers 3 and 4 primarily enrolled patients in 2017–2018, when clopidogrel was widely prescribed in Vietnam. Following subsequent updates to guidelines highlighting ticagrelor’s superior efficacy over clopidogrel in the primary PCI setting, its use saw a significant increase in Centers 1 and 2.
Radial artery access was preferred in Centers 1 and 3 (over 90.0%), whereas femoral artery access was predominantly used in Centers 2 and 4 (93.0%–100.0%). Left main coronary artery involvement was observed in approximately 15.0% of cases in Centers 2 and 4, compared to around 5.0% in Centers 1 and 3.
These demographic and clinical variations among centers likely influenced in-hospital mortality rates, ranging from 10.0% to 12.0% across most centers but reaching a peak of 17.9% in Center 2.
Pre-PCI models
Feature selection
The performance metrics of various ML models using the complete set of features are outlined in Supplementary Table S1. The Catboost model demonstrated strong performance, with an accuracy of 90.67% and an AUC (Area Under the Curve) of 91.66%, indicating its effectiveness in differentiating diagnostic outcomes. In contrast, while the Random Forest model achieved an AUC of 91.88%, its lower sensitivity of 43.64% suggested challenges in identifying true positive cases. These results highlighted the need to improve model sensitivity without compromising accuracy or specificity. Feature selection methods help optimise models by reducing complexity and improving computational efficiency, making them more interpretable for clinical use.
The application of feature selection methods demonstrated that reducing the feature set to 10 still maintained robust model performance. Specifically, the CatBoost model, optimised with features selected through ElasticNet regression, achieved an improved AUC of 92.16%, an increased sensitivity of 59.74%, an enhanced G-Mean of 75.33%, and an F1 score of 90.21%. These models outperformed traditional risk-scoring systems, such as GRACE 2.0 (AUC of 83.48%) and CADILLAC (AUC of 87.01%), as shown in Table 2. DeLong’s test was conducted using CatBoost (ElasticNet) as the reference model to assess the statistical significance of these differences. The analysis confirmed that the improvements in AUC over all traditional scores were statistically significant (all p values < 0.0001), further supporting the superior discriminative performance of the proposed ML approach. The complete performance comparisons of pre-PCI ML models with various feature selection methods are presented in Supplementary Table S2.
Key features selected by each method are detailed in Supplemental Table S3. Notably, the Killip class, the CFS, hemoglobin level, and heart rate were among the most chosen features, underscoring their importance across various predictive models.
Evaluation and interpretation of the top-performing model
The top-performing ML model underwent a comprehensive evaluation to ensure a thorough understanding of its predictive capabilities and the factors driving its performance. This evaluation included an analysis of feature importance metrics, SHAP values, calibration plots, and causal effect assessments to investigate the relationships between features and outcomes. The evaluation incorporated SHAP summaries, causal effect analyses, and dependence plots, providing detailed insights into how individual features influenced model predictions and highlighting the alignment between Feature Importance rankings and SHAP values. Additionally, the causal effect analyses elucidated the direct and indirect relationships between features and outcomes, reinforcing the robustness and interpretability of the model.
Feature importance
As illustrated in Fig. 1, the CatBoost model, refined through ElasticNet regression for feature selection, highlighted the CFS as the most pivotal predictor of patient outcomes. This was closely followed by glucose levels, the Killip class, systolic blood pressure, and the time interval from symptom onset to hospital admission, which emerged as other critical determinants influencing the model’s predictive accuracy. Additional variables, including heart rate, creatinine levels, hemoglobin levels, patient age, and ventricular tachycardia/fibrillation at admission, also demonstrated substantial significance, albeit with a lesser impact.
Feature importance of CatBoost model after applying ElasticNet regression method for feature selection.
Performance of the top-ranking model
The comparative analysis of the CatBoost model with ElasticNet feature selection, the Logistic Regression model enhanced with RFE feature selection, and traditional risk models provides valuable insights into the balance between model complexity and interpretability in predictive modeling. Advanced ML models like CatBoost excel at capturing intricate, non-linear patterns in data, yielding superior predictive accuracy. However, these benefits come at the cost of reduced interpretability and greater computational demands compared to simpler models such as Logistic Regression.
Logistic Regression, particularly when paired with RFE feature selection, strikes a compelling balance by offering enhanced interpretability, computational efficiency, and competitive performance, making it well-suited for specific clinical applications. Among the Logistic Regression approaches, the RFE-selected model stood out as the top performer. Nonetheless, the CatBoost model with ElasticNet feature selection demonstrated the highest AUC, outperforming all other models, including the RFE-enhanced Logistic Regression.
Performance differences were apparent in several evaluation metrics, including ROC curves, calibration plots, and precision-recall (PR) curves. The CatBoost model exhibited the most reliable calibration (Fig. 2b), aligning closely with the ideal diagonal across most probability ranges, with only minor overestimations at higher probabilities. In contrast, the Logistic Regression model with RFE displayed moderate calibration, exhibiting a tendency to overestimate risk across all levels of observed probabilities.
CatBoost model with ElasticNet feature selection versus Logistic Regression model with RFE feature selection versus traditional models. a Receiver operating characteristic (ROC) curves, b calibration plots, c precision-recall curves. AUC area under the curve, CI confidence interval, PR precision-recall.
The CatBoost model consistently maintained superior precision across all recall levels in the PR domain (Fig. 2c), particularly excelling in high-recall scenarios. These findings underscore the CatBoost model’s superior discrimination, calibration, and precision when applying ElasticNet feature selection, highlighting its potential for robust application in clinical settings.
SHapley additive exPlanations and causal random forests analysis
The CatBoost model, refined with features selected through ElasticNet regression, was evaluated using SHAP analysis (Figs. 3a and 4) and Causal Random Forests (Fig. 3b), highlighting their distinct yet complementary contributions to elucidating model predictions.
SHAP summary plot and causal effect of CatBoost model with ElasticNet feature selection. a SHAP summary plot, b causal effect.
SHAP dependence plot of the 10 features of the CatBoost model with ElasticNet feature selection. bpm, beat per minute.
The SHAP summary plot (Fig. 3a) demonstrates individual features’ relative importance and contributions to the model’s predictions. Among these, the Killip class emerged as the most critical predictor, followed by the CFS, glucose, systolic blood pressure, and the time interval from symptom onset to hospital admission. Higher feature values (depicted in purple) were associated with an increased mortality risk, whereas lower values (shown in green) corresponded to a decreased risk. Other variables, including creatinine levels, heart rate, patient age, hemoglobin levels, and ventricular tachycardia/fibrillation at admission, also played significant roles. This visualisation encapsulates the magnitude and direction of each feature’s impact while capturing the complex, non-linear, and interactive effects that influence the model’s predictions.
SHAP dependence plots (Fig. 4) offer deeper insights into the relationships between individual features and their contributions to the model. These plots illustrate several critical patterns: a strong positive association of the Killip class and the CFS with mortality risk, non-linear glucose and creatinine level effects, and an inverted U-shaped impact of systolic blood pressure and heart rate. Furthermore, these plots underscore the critical role of timely admission, revealing a strong correlation between delayed admission and increased mortality risk.
The Causal Random Forests analysis (Fig. 3b) complements these findings by quantifying the direct impact of each feature on patient outcomes, thereby providing a causal perspective. The presence of ventricular tachycardia/fibrillation at admission exhibited the most substantial causal influence, followed by the Killip class, corroborating its prominent role in the SHAP analysis. Similarly, the CFS, glucose, and creatinine levels also displayed significant causal effects. This alignment reinforces the conclusions drawn from the SHAP analysis, affirming that the most influential features identified by SHAP also hold substantial causal relevance.
SHAP analysis explains the magnitude and direction of feature contributions to individual predictions, accounting for both linear and non-linear interactions. In contrast, causal effect analysis focuses on direct relationships between features and outcomes. The strong agreement between SHAP-identified influential features—such as the Killip class, the CFS, glucose, and creatinine levels—and their corresponding causal effects highlight these variables’ robustness and clinical significance. This synergy enhances the model’s interpretability, reliability, and clinical relevance, providing a comprehensive understanding of patient mortality risk.
Post-PCI models
Feature selection
In the post-PCI assessment, ML models with a complete set of features exhibited marked improvements in performance metrics compared to pre-PCI models (Supplementary Table S4). These improvements were achieved without feature selection, most likely attributable to the inclusion of additional post-PCI features. The consistently improved metrics across models highlight their capacity to more accurately differentiate clinical outcomes, paving the way for further exploration of feature selection methods to refine these results while balancing computational efficiency.
As shown in Table 3, the Random Forest with Ridge feature selection achieved the highest AUC of 95.10%, outperforming its pre-PCI performance, where the same method yielded an AUC of 91.36%. Additionally, compared to traditional risk-scoring systems like GRACE 2.0 and CADILLAC, the observed improvements in accuracy and AUC highlighted the superior predictive capabilities of post-PCI ML models, especially when combined with optimal feature selection methods. These improvements were further supported by statistical comparison using DeLong’s test, which confirmed that the observed differences in AUC were highly significant (all p values < 0.0001). The complete performance comparisons of post-PCI ML models with various feature selection methods are presented in Supplementary Table S5.
Key features selected by each method are detailed in Supplemental Table S6. Once again, the Killip class and the CFS were the top two most frequently chosen features across various types of models.
Evaluation and interpretation of the top-performing model
Feature importance
A comparison of Feature Importance between the pre-PCI top-performing model (Fig. 1) and the post-PCI top-performing model (Fig. 5) reveals distinct shifts in the key predictors of patient outcomes. In the pre-PCI model, primary drivers of mortality risk included the CFS, glucose levels, the Killip class, and systolic blood pressure, underscoring the predictive value of baseline clinical and laboratory parameters before intervention. In contrast, the post-PCI model retained the Killip class, glucose levels, the CFS, and creatinine levels as pivotal predictors but introduced the significance of post-procedural factors, such as LVEF, post-PCI TIMI blood flow, and the interval from symptom onset to wire crossing. This shift highlights the evolving role of procedural and recovery-related variables in shaping outcomes.
Feature Importance of Random Forest model after applying Ridge regression method for feature selection. PCI percutaneous coronary intervention.
Performance of the top-ranking model
Consistent with the pre-PCI analysis, the Random Forest model with Ridge feature selection emerged as the top performer, achieving the highest AUC and outperforming the Logistic Regression model with Lasso feature selection and established risk models like GRACE 2.0 and TIMI. However, the performance gap between the Random Forest and Logistic Regression models narrowed slightly in the post-PCI context, suggesting that incorporating procedural and outcome-related features enhances predictive capability across complex and simpler models.
Regarding calibration (Fig. 6b), the Random Forest model exhibited robust reliability, aligning closely with the ideal diagonal line over most probability ranges and demonstrating minimal underestimation at higher probability levels. While this pattern was consistent with pre-PCI findings, the Logistic Regression model’s calibration in the post-PCI scenario revealed more significant deviations, overestimating risk across all levels of observed probabilities.
Random Forest model with Ridge feature selection versus Logistic Regression model with Bayesian feature selection versus traditional models. a Receiver operating characteristic (ROC) curves, b calibration plots, c precision-recall curves. AUC area under the curve, CI confidence interval, PR precision recall.
The PR curves (Fig. 6c) further underscored the Random Forest model’s advantage, maintaining superior precision across all levels of recall, particularly in high-recall scenarios. Although this trend echoed the pre-PCI results, the PR performance gap between the models was slightly less pronounced post-PCI.
In summary, the Random Forest model with Ridge feature selection consistently performed better in predicting post-PCI risk. However, the reduced AUC performance gap relative to the Logistic Regression model indicates that the enriched post-PCI feature set enhances predictive performance across a broader range of model complexities.
SHapley additive exPlanations and causal random forest analysis
These analyses unveiled critical predictors and their roles in understanding patient outcomes. Compared to the pre-PCI analysis, the post-PCI findings highlight the additional influence of procedural factors alongside baseline clinical characteristics.
The SHAP summary plot (Fig. 7a) demonstrates that the Killip class, the CFS, and glucose consistently ranked among the most influential predictors in pre- and post-PCI models. However, the post-PCI model further underscored the significance of post-procedural variables, including LVEF, post-PCI TIMI blood flow, and the time from symptom onset to wire crossing.
SHAP summary plot and causal effect of Random Forest model with Ridge feature selection. a SHAP summary plot, b causal effect. PCI percutaneous coronary intervention.
Dependence plots (Fig. 8) reveal nuanced relationships, such as the linear increase in risk associated with higher Killip class and CFS scores, the non-linear associations of glucose and creatinine levels with outcomes, and a U-shaped relationship for systolic blood pressure, where deviations from an optimal range elevated risk. Delays in intervention were strongly linked to poorer outcomes, while high post-PCI TIMI blood flow correlated with improved survival rates.
SHAP dependence plot of the 10 features of the Random Forest model with Ridge feature selection. bpm beat per minute, PCI percutaneous coronary intervention.
Causal effect analysis (Fig. 7b) reinforced these findings, with the Killip class exhibiting the most significant causal impact on mortality. Other variables, such as the CFS, glucose, and creatinine, also displayed notable causal contributions. Procedural metrics, particularly post-PCI TIMI blood flow, further highlighted the critical role of successful intervention in shaping patient outcomes.
Both pre- and post-PCI models consistently identified the Killip class, the CFS, and glucose as pivotal predictors, underscoring their central role in mortality prediction. However, the post-PCI model expanded the pre-PCI framework by integrating procedural features, offering a more nuanced and comprehensive understanding of post-intervention risk. This enhanced perspective bridges baseline clinical factors with the impact of procedural success, providing valuable insights into optimizing post-PCI patient care.
