Acute myocardial infarction (AMI) is a major cause of hospitalization and mortality in China, while ST-segment elevation myocardial infarction (STEMI) accounts for over 80% of myocardial infarctions [1-3]. It is critical to accurately predict the risks of in-hospital mortality for patients with STEMI to improve prognosis. Traditionally, most risk prediction models have been based on generalized linear regression methods [4,5]. Although straightforward to understand and apply, these models require parametric assumptions [6,7]. For example, using the logistic regression (LR) method, the Global Registry in Acute Coronary Events (GRACE) [4] and Thrombolysis in Myocardial Infarction (TIMI) risk scores [5] oversimplified the complexity of the real association among variables and outcome, resulting in poor predictive accuracy [8,9]. Recently, machine learning (ML) techniques have been increasingly used for predicting different clinical events in cardiovascular disease [10-12] and have achieved higher accuracy than traditional models. However, ML models, often built on a large number of variables, are difficult to use in clinical practice due to the need for extensive input data and the challenge of identifying specific therapeutic targets. The complexity and ambiguity of ML models require a shift toward explainable artificial intelligence (XAI) methods to guarantee that the model outputs are comprehensible for end users [13]. Moreover, ML models tend to use a large number of variables [14-16]. However, in clinical practice, where many scenarios are unknown, a significant challenge is how to apply the model more flexibly when some variables are missing. Therefore, we aimed to develop an ML risk prediction model for in-hospital mortality in patients with STEMI that is not only highly accurate but also explainable and flexible with the number of input variables (tolerant to the missing variables), making it easy to use in clinical practice.

In this study, we proposed a risk model that predicted in-hospital mortality for patients with STEMI by incorporating the ML method XGBoost and the model interpretation approach SHAP. The model we constructed had excellent performance in terms of high predictive accuracy, high tolerance to missing values (flexibility), and good clinical interpretability. Importantly, we identified the top 7 clinical factors affecting in-hospital mortality as age, LVEF, Killip class, heart rate, creatinine, glucose, and WBC. Among these, LVEF, glucose, and WBC were not included in the current traditional predictive models. Although creatinine is also included in the GRACE score, its relationship with mortality is not a simple linear one. The predictive value of glucose and WBC exceeds that of other variables in traditional predictive models, such as blood pressure, weight, and medical history (hypertension, diabetes, and angina). We believe that these findings can help doctors understand the value of ML models and uncover the pathophysiological significance of certain clinical variables in myocardial infarction.

While traditional statistical models such as TIMI and GRACE, as recommended by current guidelines [26], are useful and user-friendly, their overly simplified nature may result in inadequate predictive accuracy for risk classification and decision-making [8]. First, these models are developed based on a limited number of variables and may not encompass comprehensive information. Second, the LR method used by these models requires strong assumptions, including a linear relationship under the logit function, independence of observations, and no multicollinearity among variables [7,8,25,27]. This results in underestimating the complexity of the real association among variables and outcomes.

In contrast, ML methods can handle a larger number of variables, require no parametric assumptions, and can learn the complex relationships hidden in the data automatically [9]. The XGBoost method overcomes these limitations by generating a series of classification and regression trees (CARTs) with each one learning the residuals of its predecessors. The boosting mechanism gives the model a strong predictive power. As observed, the XGBoost model achieved an impressive AUC of 0.896 (95% CI 0.884-0.909) on the CAMI validation set, outperforming the other methods and proving to be a more powerful and effective tool for clinical risk prediction.

The XGBoost model’s ability to tolerate missing values makes it well-suited for clinical applications, where incomplete variables are frequent [28-30]. While most ML methods achieve accuracy and precision by learning from a large number of variables, they often lose practicality because it is usually difficult to collect all the predictors used in the model in clinical practice. In such cases, missing values must be imputed if clinicians still want to apply the model. The proposed XGBoost model overcomes this weakness thanks to its ability to deal with missing values. We demonstrated that the XGBoost model’s performance is relatively robust when faced with incomplete data compared to the traditional LR model. Even with only the top 10 important variables, the XGBoost model achieved an AUC of 0.877(95% CI 0.862-0.891) on the CAMI validation set. On the independent China PEACE set with only the top 10 important variables available, XGBoost gained an AUC of 0.840 (95% CI 0.829-0.852) compared to TIMI 0.762 (95% CI 0.748-0.776) and GRACE 0.789 (95% CI 0.776-0.803). These results demonstrated the XGBoost model’s flexibility and generalization ability, which could alleviate concerns about the feasibility of applying complex ML models in clinical practice.

Another concern about the complex ML approaches applied in clinical practice is their lack of transparency. Unlike the widely employed LR method, whose coefficients clearly indicate the effect of predictive factors on the outcome, the black-box nature of complex ML algorithms applied in medical tasks has been seriously criticized and doubted in recent years [8,9]. To address this issue, our study used SHAP to interpret how the predicted risk was determined for individual patients and uncover the complex relationship between predictors and outcomes embedded in the XGBoost model.

Our results showed that HDL-C displayed a U-shaped relationship with in-hospital mortality among patients with STEMI. In the previous studies, Madsen et al [31] reported a U-shaped association between HDL-C and mortality, using data from 52,268 men and 64,240 women enrolled in 2 prospective population-based studies. Similarly, Bowe et al [32] found a U-shaped relationship between HDL-C and the risk of all-cause mortality in patients with kidney disease. For the variable potassium, our result showed that the patients with STEMI with potassium levels ranging from 4.13 to 4.49 mmol/L had the lowest in-hospital mortality risk, while levels greater than 4.5 mmol/L increased the mortality risk. Clinical practice guidelines recommend maintaining serum potassium levels between 4.0 and 5.0 mmol/L in patients with acute myocardial infarction (AMI) [33,34]. However, recent studies have challenged these guidelines, reporting that potassium levels greater than 4.5 mmol/L are associated with increased mortality [35-37]. Our study found that creatinine >1.1mg/dl (94.5/L) contributed to a higher in-hospital mortality risk. A previous study [38] reported that an elevated serum creatinine level (defined as creatinine ≥1.2 mg/dl) predicted a higher long-term mortality risk in patients with AMI.

For the variable blood glucose, our results showed that levels less than 8.15 mmol/L were safer for patients with STEMI. Another study reported that the best cutoff values for 30-day mortality among patients with STEMI were 149 mg/dL (8.27 mmol/L) for those without diabetes, 231 mg/dL (12.82 mmol/L) for those with diabetes, and 169 mg/dL (9.38 mmol/L) for all patients [39]. For the variable WBC, our result showed that a higher WBC count was associated with higher in-hospital mortality risk, with a safer threshold being less than 10.77/L. Cannon et al [40] reported that mortality at 30 days showed a curvilinear increase with increasing WBC count, with mortality rising in patients with WBC count >10,000 /dL (P<.0001). Previous studies often investigated this relationship by categorizing or binning continuous variables and regressing the outcome on the categorical variables. However, this approach is heavily influenced by predefined cutoffs and cannot provide a continuous picture of the relationship. In contrast, our model offered more thorough and quantitative insights into the exact change in risk induced by specific patient characteristics. By interpreting how each variable contributed to in-hospital mortality, our study could help clinicians identify specific therapeutic targets and further guide patient management.

Our research has a certain guiding significance for clinical implementation. First, the new model is significantly superior to traditional GRACE and TIMI models, helping doctors predict patient prognosis. Second, ML has identified several variables not included in past models, which may serve as potential targets for clinical intervention or provide further understanding of the pathophysiology of disease development, such as WBC and blood glucose. Third, while clinicians often find it difficult to understand the variables selected by ML, adopting the XGBoost model and model interpretation approach SHAP further increases accuracy by capturing nonlinear relationships among the predictors and outcomes. This offers a clear explanation for why ML can improve predictive efficiency, thus enhancing clinicians’ understanding of the performance improvement of ML. Methodologically, we used internal validation and a large sample size of independent external validation, all leading to consistent conclusions.

However, despite the superior performance of the proposed XGBoost model, several limitations still exist. First, the proposed XGBoost model was derived and validated on the Chinese STEMI patient cohort. Further validation is needed to confirm its efficiency on more general cohorts. Second, the study was designed prospectively, but this research is a retrospective analysis, so the variables recruited in our study may be limited. The model may be more powerful if more informative variables were added.

In conclusion, the proposed ML model in our paper demonstrated strong advantages in predictive ability, flexibility, and interpretability. Although some results need further study and verification, we have shown the benefits of complex models in the field of disease predictions. We offered a web calculator for convenient application, and we hope our study can help augment and extend the effectiveness of cardiologists to improve patient care and promote incorporating ML into daily practice.