Consecutive patients with suspected acute myocardial infarction (ST-segment or non-ST-segment elevation), admitted to this hospital and who underwent coronary angiography, were retrospectively enrolled from January 1, 2012, to September 30, 2015. AMI was diagnosed according to the third universal definition of myocardial infarction [17]. Exclusion criteria included: (1) incomplete key clinical data; (2) history of coronary artery bypass grafting; (3) congenital heart disease and other heart diseases requiring surgical intervention; (4) poor CAG images; and (5) lost to follow-up. A total of 45 clinical variables and angiographic variables were collected from the electronic medical record system of the First Affiliated Hospital of Soochow University, according to the CatLet angiographic scoring system. Variables included patient demographics, medical and personal histories, clinical examination findings, coronary anatomy, and lesion-related adverse angiographic features.

The CatLet angiographic scoring system has been described in full elsewhere. In brief, this comprehensive angiographic scoring system has been developed based on the model of 17 myocardial segments, the law of competitive supply, and the law of flow conservation. A total of 54 (6 $\times$ 3 $\times$ 3) types of coronary circulation are defined based on the CatLet angiographic scoring system. Specifically, six right coronary artery (RCA) types are classified as: posterior descending artery (PDA) zero, PDA only, small RCA, average RCA, large RCA, and super RCA. Three left anterior descending artery (LAD) types are classified as short LAD, average LAD, and long LAD. The three diagonal branches (Dx) classifications included small Dx, inter. Dx, and large Dx. The coronary artery segments are weighted by the number of myocardial segments they perfuse to explain and semi-quantify coronary variation. The weights of the coronary segments are obtained using an online calculator or by mental calculation. The lesion score is the product of the weighting factor of the coronary segment and its degree of stenosis (2.0 for nonocclusive lesions and 5.0 for occlusive lesions). The scores of individual lesions are then summed to give a total score. In addition to scoring the lesions, the system also qualitatively records those adverse angiographic characteristics of the lesions, such as severe calcification, culprit vessels, severe distortion, angulation, bifurcation lesions, trifurcation, and thrombus burden. The CatLet score calculator is available at http://www.catletscore.com.

The primary outcome was cardiac death, defined as death due to AMI, heart failure, arrhythmia, other cardiovascular causes, or unexpected sudden death without an apparent noncardiac cause [18]. All patients were followed up for this outcome for 4 years. The clinical variables were age, sex, smoking and drinking status, body mass index (BMI), serum albumin, serum creatinine, left ventricular ejection fraction (LVEF), history of stroke, history of diabetes, and hypertension. Angiographic variables were heavy calcification and culprit vessels.

The dataset was divided into training and testing sets at a ratio of 7:3. The training dataset was used to build the model, and the testing dataset was used for validation. Initially, we applied the least absolute shrinkage and selection operator (LASSO) regression, a method that performed variable selection and coefficient shrinkage through regularization [19]. LASSO regression performed feature selection by reducing the coefficients of less significant features to zero, thereby effectively eliminating them from the model, and utilized 10-fold cross-validation to determine the best lambda value, thereby minimizing the average cross-validation error [20]. Subsequently, we implemented Boruta feature selection, a random forest-based algorithm that identified all relevant variables by comparing the importance of original features with randomly generated “shadow features” [21]. To ensure a robust and concise model, we chose the intersection of features identified by LASSO regression and the Boruta algorithm as our final set of predictor variables. We constructed and tested six ML algorithms, including k-nearest neighbors (KNN), support vector machine (SVM), adaptive boosting (Adaboost), light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), and logistic regression (LR). The best performing model was selected to predict cardiac death in AMI patients. Model development utilized 10-fold cross-validation to improve reliability and generalization. The performance evaluation of the model included multiple indicators: area under the curve (AUC), accuracy, sensitivity, specificity, and F1 score. For all these performance measures, values range from 0 to 1, with higher scores indicating better model performance. After a thorough evaluation of model performance, we selected the model that showed the highest stability across all performance metrics as our final prediction model. To improve the interpretability of this model, we performed SHAP analysis, which provided insight into the importance and ranking of each variable [22]. SHAP values clearly showed the positive or negative impact of each variable on the model prediction with a screening threshold of 0.05. To address the issue of data imbalance, we adopted the random oversampling techniques and used the ROSE package in R software (version 4.4.3, The R Foundation for Statistical Computing, Vienna, Austria) to build new balanced training data.

Participants were divided into three groups based on the tertiles of the CatLet score. Normally distributed continuous variables were expressed as mean $\pm$ standard deviation, while non-normally distributed continuous variables were expressed as the median and interquartile range (IQR). Categorical variables were expressed as counts (percentages). For the analysis of continuous variables, analysis of variance (ANOVA) was used for normally distributed data, while the Kruskal-Wallis test was applied for non-normally distributed data. Categorical variables were analyzed using the chi-square test. Multivariable Cox proportional hazards regression models were used to examine the association between CatLet score and cardiac death. Model 1 was not adjusted; Model 2 was adjusted for age and sex; Model 3 included all the covariates from Model 2 as well as primary hypertension, type 2 diabetes mellitus, BMI, prior stroke, smoking, drinking, serum albumin, serum creatinine, LVEF, and heavy calcification. Restricted cubic spline analysis was used to investigate the linear or nonlinear association between the CatLet score and cardiac death. All missing data were removed. Statistical analysis was performed using R software (version 4.4.3, The R Foundation for Statistical Computing, Vienna, Austria), and a two-tailed p 0.05 was considered a statistically significant difference.

A total of 1018 patients were identified for potential analysis, of which 386 patients were excluded, and 767 patients finally met the inclusion criteria, as shown in Fig. 1. Table 1 shows the baseline characteristics stratified by the tertiles of CatLet score. The average age was 63.54 years. Notably, individuals with higher CatLet scores tended to be older, predominantly male, and had lower serum albumin levels, ejection fraction, and BMI. Furthermore, they exhibited a higher prevalence of hypertension, stroke, or heavy calcification of the lesion.

Fig. 1.

The whole study workflow. Abbreviations: KNN, k-nearest neighbors; SVM, support vector machine; Adaboost, adaptive boosting; LightGBM, light gradient boosting machine; XGBoost, extreme gradient boosting; LR, logistic regression; SHAP, Shapley Additive Explanations.

Multivariable Cox proportional hazards models were used to assess the impact of CatLet score on cardiac death after adjusting for all covariates (model 3). CatLet score was significantly associated with the risk of cardiac death (hazard ratio (HR): 1.05, 95% CI: 1.02–1.09, p = 0.002, per 1 point increase). On the categorical scale, as compared with Q1, the HRs for cardiac death were 2.06 (0.68–6.23) for Q2 and 3.71 (1.36–10.08) for Q3 (p for trend = 0.004), respectively (Table 2).

Using the restricted cubic spline analysis, the linear trend between the CatLet score and cardiac death was revealed (p for overall = 0.036; p for nonlinear = 0.741) (Fig. 2). The threshold effect analysis determined the critical level of the CatLet score to be 13.5. The Kaplan-Meier survival curves showed that the survival probability decreased with the increase of the tertiles of CatLet score (p 0.001) (Fig. 3).

Fig. 2.

Distribution plot of the CatLet score and the relationship between the CatLet score and cardiac death using a restricted cubic spline regression model.

Fig. 3.

Kaplan-Meier survival curves for cardiac death stratified by the tertiles of CatLet score.

LASSO regression was used to select features in the training set, and the characteristics of variable coefficients were shown in Fig. 4a. As shown in Fig. 4b, the iterative analysis was carried out by a ten-fold cross-validation method. The LASSO regression identified 11 variables. The Boruta algorithm recognized 12 variables, as shown in the green and yellow boxes in Fig. 4c. By intersecting the variables derived from the two algorithms, a total of 7 variables were finally used to build the ML model. These variables included serum creatinine, LVEF, age, serum albumin, CatLet score, heavy calcification, and culprit LAD.

Fig. 4.

Feature selection by LASSO regression and Boruta’s algorithm. (a) The variation characteristics of the LASSO coefficient. The different colored lines represented the different variables. (b) Optimization parameters ($\lambda$) of the LASSO model were selected by 10-fold cross-validation. The left dashed line represented $\lambda$ min (minimum cross-validated error), while the right dashed line indicated $\lambda$1se (the largest $\lambda$ within one standard error of $\lambda$ min). (c) Feature identification via the Boruta algorithm. The X-axis represented all features, and the Y-axis was the Z-value of each feature. The green boxes represented significant variables, while the yellow ones denoted tentative, and the red ones indicated unimportant. The circles represent the fluctuation of feature importance in multiple random iterations, indicating exceptionally high or low individual importance score.

As shown in Table 3 and Fig. 5, KNN, SVM, and LightGBM models all performed well on the training set, but their performances declined on the test set. The XGBoost model demonstrated superior performance on the training set, with an AUC of 0.900, accuracy of 0.903, sensitivity of 0.941, specificity of 0.860, and F1 score of 0.912. However, its performance on the test set showed a significant drop, particularly in AUC (0.772), indicating a potential over-fitting. Adaboost had the highest F1 score (0.916), suggesting that it achieved the best balance between precision and recall on the test set. The AUC, accuracy, and sensitivity of the Adaboost model were higher than those of LR in the test set. Therefore, the Adaboost model was selected as the best model to predict cardiac death in AMI patients. The results of the statistical comparison of ROC curves of each model were shown in the Supplementary Tables 1,2.

Fig. 5.

ROC of the six ML models. (a) ROC of the training dataset. (b) ROC of the test dataset. Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; KNN, k-nearest neighbors; SVM, support vector machine; Adaboost, adaptive boosting; LightGBM, light gradient boosting machine; XGBoost, extreme gradient boosting; LR, logistic regression.

We conducted an interpretable analysis of the Adaboost model. The importance of each feature was ranked from highest to lowest according to the SHAP analysis (Fig. 6a). SHAP values represented the influence of each feature on the final prediction result. Fig. 6b showed the influence of each feature on the model output. Each dot in a row symbolized a patient, and its color denoted the feature value: yellow for larger values and purple for lower values. The more scattered the points of the graph, the greater the influence of the variables on the model. In our study, older age, higher CatLet scores, culprit LAD, and heavy calcification were associated with increased risk of cardiac death, while elevated levels of albumin and LVEF served as protective factors.

Fig. 6.

The SHAP analysis of the Adaboost model. (a) A bar plot displaying the mean SHAP value for variables. (b) The beeswarm plots displayed the distribution of variables. (c) SHAP waterfall plot for case 1. (d) SHAP waterfall plot for case 2. Abbreviations: SHAP, Shapley Additive Explanations, and Adaboost, adaptive boosting.

To elucidate the decision-making process at the individual level, SHAP waterfall plots were utilized to conduct local interpretable analysis on two representative cases randomly selected from the test set. For one case (Fig. 6c), the model significantly reduced the mortality risk for a 66-year-old patient (from baseline $f$(x) = 0 to final prediction $f$(x) = –0.163). The most significant protective factor was non-culprit LAD (0, SHAP value: –0.0358), a lower CatLet score (9.56, SHAP value: –0.0406), and no heavy calcification (0, SHAP value: –0.0204), indicating relatively mild disease severity. Creatinine level (69.2 umol/L, SHAP value: –0.0583) and ejection fraction remained stable (0.532, SHAP value: –0.104). The adverse factor was slightly below-normal levels of albumin (32.9 U/L, SHAP value: +0.065). In contrast, case 2 (Fig. 6d) presented an 83-year-old patient, where the model indicated an increased risk of death (compared to the baseline E[$f$(x)] = 0, the final predicted $f$(x) = 0.254). Significant risk factors included older age (83, SHAP value: +0.133), higher CatLet score (31.3, SHAP value: +0.0874), and severe calcification (1, SHAP value: +0.0831). The SHAP value quantified the contribution of each feature, with positive values (yellow bars) indicating risk factors and negative values (magenta bars) indicating protective effects.