Acute coronary syndrome (ACS), which is mainly triggered by the rupture, erosion, or calcified nodules of atherosclerotic plaques and results in acute myocardial ischemia, myocardial cell necrosis, and inflammation, is a leading cause of morbidity and mortality in developed and developing countries [1, 2, 3]. Increasing data have revealed that early and timely detection of ACS and the administration of effective interventions could significantly improve patient prognosis. Therefore, doctors need to prioritize the pursuit of user-friendly biomarkers for the prompt detection of those at risk of ACS. Several risk scores, such as the Global Registry of Acute Coronary Events (GRACE) and thrombolysis in myocardial infarction (TIMI) scores, have been developed mainly for Caucasian populations with chest pain. However, these scores vary in performance, especially among the Chinese population, and might perform better in identifying patients with ACS at high risk of major adverse cardiovascular events [4, 5, 6]. Moreover, few studies have focused on the risk of ACS in hospitalized patients, regardless of whether they have chest pain.

Dyslipidemia is a notable independent risk factor for the onset and advancement of ACS and has been acknowledged as a focal point for therapeutic intervention [7, 8, 9]. Unlike the widely accepted traditional lipid indicators, a growing number of studies have found that some lipid-related biomarkers, including residual cholesterol (RC), triglyceride-glucose (TyG) index, and uric acid–to–high-density lipoprotein cholesterol (HDL-C) ratio (UHR), could serve as predictors of the incidence and prognosis of cardiovascular diseases [10, 11, 12]. However, the precision of these biomarkers or models remains relatively inadequate. Hence, there is an urgent need for further improvement to accurately stratify the risk of ACS in patients. Machine learning (ML), which is gaining prominence in various fields, can overcome these limitations [13, 14]. In recent years, several ML models have been developed to identify individuals at risk for ACS and acute myocardial infarction [15, 16]. However, most studies rely on bedside electrocardiogram detection or cardiac troponin concentrations, and some models are primarily designed based on lipid-related biomarkers. Therefore, this study aimed to develop a risk score to identify ACS risk groups in hospitalized patients using ML methods and visually interpret the model using SHapley Additive exPlanations (SHAP) methods to assist clinicians in identifying and managing high-risk groups.

In this study, we developed and evaluated an explainable prediction model based on lipid-related indicators to predict ACS in hospitalized patients. Using data from a large multicenter study, we demonstrated that our risk stratification score achieved considerable accuracy in predicting ACS. To the best of our knowledge, this is the first and most extensive study to use both conventional and ML approaches for this therapeutic assignment. Furthermore, our study had the largest sample size to date, encompassing over 10,000 hospitalized patients from three major hospitals.

ACS is a complex cardiovascular disease with multifactorial pathogenesis, and lipid abnormalities are among the most critical risk factors [24]. Traditionally, LDL-C has been the primary target of lipid-lowering therapy and is regarded as the “gold standard” for assessing cardiovascular risk [25, 26]. However, despite achieving the target LDL-C levels in recent years, the incidence of ACS remains high, suggesting that relying solely on LDL-C may be insufficient to fully capture ACS risk. Thus, identifying new lipid metabolic markers to improve the prediction of ACS has become a priority.

In this study, we systematically explored the role of two emerging lipid-related markers, UHR and TyG, in predicting ACS risk while validating the established role of Lp(a). Incorporating these markers into predictive models showed that UHR and TyG significantly enhanced the performance of the models, underscoring their clinical value in assessing ACS risk. UHR, a novel composite metabolic marker representing the uric acid–to–HDL ratio, has been used to evaluate metabolic health [27]. Elevated uric acid levels are strongly associated with various metabolic disorders such as hypertension, obesity, and diabetes [28, 29], whereas HDL-C is known for its protective cardiovascular effects [30]. Our findings showed that higher UHR values were strongly associated with an increased risk of ACS, suggesting that UHR not only reflects lipid metabolism but may also interact with other risk factors, such as metabolic syndrome, to serve as an effective tool for cardiovascular risk assessment.

Lp(a) is a lipoprotein particle structurally similar to LDL-C but contains a unique protein component, apolipoprotein(a), which has long been recognized as an independent risk factor for atherosclerosis and thrombosis [31, 32]. Our findings further support its role in ACS, with elevated Lp(a) levels significantly associated with higher ACS risk. However, the primary focus of our study was TyG, an indicator of insulin resistance that combines triglyceride and fasting glucose levels. Insulin resistance is a pivotal contributor to atherosclerosis and cardiovascular diseases. Our results demonstrated that TyG was strongly associated with ACS occurrence. Given the increasing prevalence of obesity and metabolic syndrome, TyG—a simple and accessible marker—has substantial potential for predicting cardiovascular events. Collectively, UHR, TyG, and Lp(a) offer valuable insights for improving ACS risk assessment in clinical practice.

We employed several machine learning models to evaluate the predictive effects of these novel lipid markers. Compared with traditional models, the Random Forest and LightGBM models significantly improved the prediction accuracy after incorporating UHR, Lp(a), and TyG. This improvement reflects a substantial enhancement in the predictive power from including these markers. Notably, LightGBM outperformed the other models in terms of sensitivity and accuracy, demonstrating its strength in handling complex, high-dimensional data. This data-driven ML approach offers new possibilities for the early detection of cardiovascular diseases. This enables a clearer understanding of the relative contributions of various risk factors to disease progression. More importantly, an easy-to-use risk score based on the top 10 features derived from the ML models performed relatively well in both the training and validation sets. Patients in the high-risk group had the highest likelihood for ACS, and this trend remained robust across all patients, regardless of comorbidities. ML models, which can handle large amounts of data and complex relationships by extracting important metrics, are widely used for early diagnosis and prognosis assessment in patients with ACS. The primary benefits of ML are its strong performance, improved stability in selecting key metrics, and significantly reduced computation time, making it suitable for handling large datasets. Our study involved a concurrent analysis of three hospital databases. Given the large sample size and abundance of clinical variables, ML techniques are optimal for identifying important ACS risk factors in hospitalized patients. Several ACS risk scores have been identified in previous studies; among them, the GRACE and TIMI scores may be the most widely accepted for predicting the incidence and prognosis of ACS. However, their performance varies across studies, particularly in China. Using data from a prospective Chinese registry, Wang et al. [33] reported that the Biomarker-based Prognostic Assessment for Patients with Stable Angina and Acute Coronary Syndromes (BIPass) score outperformed the GRACE and TIMI risk scores in predicting cardiovascular events in patients with ACS. Unlike the above-mentioned studies that examined prognosis in patients with ACS, we focused only on the risk of ACS in hospitalized patients and did not explore the association between this risk score and prognosis in all patients. Moreover, our risk score has relatively good predictive power for ACS in both cohorts, enabling all clinicians, not just cardiologists, to identify hospitalized patients at a high risk of ACS using this score.

Another key innovation of this study is its use of a multicenter design with extensive data analysis. We analyzed large-scale datasets from different regions and populations to conduct stratified analyses and validation. This multicenter design improved the external validity and generalizability of the model, ensuring stability and consistency across subgroups and reducing biases caused by regional and population differences. By integrating data from multiple centers, our findings provide a more robust basis for clinical generalization and hold significant potential for widespread application in the global prevention and management of cardiovascular diseases.

This study had several limitations. First, as a retrospective analysis, causality cannot be established between novel lipid-related markers (UHR, Lp(a), TyG) and ACS, only associations. Biases in data collection, differences in participating hospitals, and patient selection may affect the generalizability of the findings, despite the efforts to minimize them through a multicenter analysis. Second, the predictive models may have been overfitted, as indicated by the discrepancy in the AUC values between the training and test sets. Third, we did not compare the performance of our risk score for ACS with other established scores, such as the GRACE and TIMI scores, which were primarily designed for prognostic assessment in confirmed ACS patients. Future prospective studies were advocated to validate our score alongside these tools in a cohort where both diagnostic and prognostic data are available. Moreover, we did not investigate the association between this score and short- or long-term outcomes in patients with ACS. Some traditional risk factors (e.g., smoking, alcohol use, and medical history) were not included in this risk score, which may explain its relatively weaker performance in certain aspects. Furthermore, potential confounders such as lifestyle, medication use, and genetic factors were not fully controlled. Prospective cohort studies are required to validate these markers and confirm their clinical utility.

Our study effectively designed and internally validated a predictive risk stratification model for ACS in hospitalized patients using lipid-related indicators. By leveraging explainable ML approaches such as SHAP, we gained valuable insights into the correlations between crucial parameters and ACS, while the final model demonstrated strong predictive accuracy. This risk stratification model could help doctors more effectively pinpoint patients at increased risk of ACS, implement suitable preventive measures, and ultimately lower morbidity while improving quality of life. This study demonstrates the prospective relevance of a unique risk stratification model based on lipid-related indicators for developing personalized therapy and improving clinical decision-making for ACS.

ACS, acute coronary syndrome; AUC, area under the curve; BIPass, Biomarker-based Prognostic Assessment for Patients with Stable Angina and Acute Coronary Syndromes; GRACE, Global Registry of Acute Coronary Events; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; LightGBM, Light Gradient Boosting Machine; Lp(a), lipoprotein(a); ML, machine learning; SHAP, SHapley Additive exPlanations; TIMI, thrombolysis in myocardial infarction; TyG, triglyceride-glucose index; UHR, uric acid-to-HDL-C ratio; RC, residual cholesterol; ROC, receiver operating characteristic; WBC, white blood cells; HGB, hemoglobin; ALT, alanine aminotransferase; ALB, albumin.

The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request.

ZBL and QXW designed the research study. JJW, YHLuo, BS, YHLi, SQC, TH, ZC, FFY, YYH, and ZZ performed the research. JJW and YHLuo analyzed the data. JJW drafted the manuscript. All authors contributed to critical revision of the manuscript for important intellectual content. All authors read and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

This study was approved by the Medical Ethics Committee of Zhongnan Hospital of Wuhan University (2021053), the Medical Ethics Committee of China Resources Wisco General Hospital (CRWG2024R031J), and the Ethics Committee of Enshi Central Hospital (2024-052-01). This study complies with the Declaration of Helsinki. Due to the retrospective nature of the study, the patient informed consent statement has been waived.

Not applicable.

This study was supported by a grant from the National Natural Science Foundation of China (No.82270402, 82200974).

The authors declare no conflict of interest.

Supplementary material associated with this article can be found, in the online version, at https://doi.org/10.31083/RCM44578.