Machine learning model predicted in-hospital ACS mortality with AUROC 0.90 internally and better stratified risk than GRACE, reducing low-risk mortality by 50% and increasing high-risk detection by 10
Does a machine learning-based prediction model improve in-hospital mortality risk stratification compared to the GRACE score in Chinese ACS patients?
A novel machine learning-based model incorporating clinical and meteorological variables accurately predicts in-hospital mortality risk in Chinese ACS patients, outperforming the traditional GRACE score.
Abstract Background Acute Coronary Syndrome (ACS), a severe life-threatening cardiovascular disease, has long been the focus of the medical field. Tools for predicting and stratifying in-hospital mortality risk of ACS have gradually been developed and applied in clinical practice to assist in the management and treatment of ACS patients. However, the tools, such as GRACE have shown limitations when applied to Chinese populations. It requires specific formulas for calculations, and the AUROC does not exceed 0.80. Purpose We have explored the in-hospital mortality risk prediction and stratification tools tailored for Chinese ACS patients, aiming to provide auxiliary tools for clinical diagnosis and treatment. Methods This study utilized the data of over 940,000 ACS patients from 2019 to 2022, collected by the Chinese Cardiovascular Association-Chest Pain Center, which covers 5,417 healthcare institutions in China, and collected meteorological data. By constructing four variable scenarios and applying five machine learning models, and using various validation methods such as internal validation, time and space external validation, an in-hospital mortality risk prediction model for ACS was developed. Further development of risk stratification methods was carried out, and compared with GRACE. Ultimately, an in-hospital mortality risk prediction and stratification tool tailored for the Chinese population was developed. Results The final ACS in-hospital mortality risk prediction model includes 12 variables: age, heart rate, systolic blood pressure, ST segment change, cardiac arrest at admission, elevated myocardial enzyme levels, Killip classification, creatinine level, Creatine kinase isoenzyme (CK-MB), annual average PM2.5 concentration, winter mean temperature, and summer mean temperature. The internal validation AUROC is approximately 0.90, while the time and space external validation AUROCs are 0.88 and 0.91, respectively. Stratification of low, medium and high risks was performed according to the predicted risk of 1%, 1%-10%, and 10%. Compared to the GRACE method, the actual mortality among the low-risk group classified by this method is reduced by 50%, while the actual mortality among the high-risk group is increased by 109%. Conclusion The ACS in-hospital risk prediction and stratification method established in this study performs well in the Chinese population and has advantages over GRACE. It can promptly assess the in-hospital mortality risk and stratify the risk for ACS patients upon admission. Targeted medical interventions can then be provided based on different risk levels. This approach will reduce unnecessary medical procedures for low-risk patients and provide timely and appropriate medical interventions for high-risk patients, optimizing the allocation of medical resources and enhancing the overall quality and efficiency of medical services. It offers more precise, efficient, and personalized medical care for patients.Model verification and method evaluation Diagram of the tool
Li et al. (2025) studied this question. Machine learning model predicted in-hospital ACS mortality with AUROC 0.90 internally and better stratified risk than GRACE, reducing low-risk mortality by 50% and increasing high-risk detection by 10.