Development of an optimized risk score to predict short‐term death among acute myocardial infarction patients in rural China

Abstract Background Risk stratification of patients with acute myocardial infarction (AMI) is of great clinical significance. Hypothesis The present study aimed to establish an optimized risk score to predict short‐term (6‐month) death among rural AMI patients from China. Methods We enrolled 6581 AMI patients and extracted relevant data. Patients were divided chronologically into a derivation cohort (n = 5539), to establish the multivariable risk prediction model, and a validation cohort (n = 1042), to validate the risk score. Results Six variables were identified as independent predictors of short‐term death and were used to establish the risk score: age, Killip class, blood glucose, creatinine, pulmonary artery systolic pressure, and percutaneous coronary intervention treatment. The area under the ROC curve (AUC) of the optimized risk score was 0.82 within the derivation cohort and 0.81 within the validation cohort. The diagnostic performance of the optimized risk score was superior to that of the GRACE risk score (AUC 0.76 and 0.75 in the derivation and validation cohorts, respectively; p < .05). Conclusion These results indicate that the optimized scoring method developed here is a simple and valuable instrument to accurately predict the risk of short‐term mortality in rural patients with AMI.

(ACC/AHA) 4 and the European Society of Cardiology (ESC) 5 recommend that the most appropriate pharmacological and interventional management should be determined after comprehensive risk assessment.
Many risk models of in-hospital or short-term mortality have been developed among patients with acute coronary syndrome. [6][7]8,9,10,11 Among them, the Global Registry in Acute Coronary Events (GRACE) score is the most commonly used to predict short-term death. 12 However, the GRACE score was developed at a time when patient characteristics and management differed significantly from current practice, and few participants were from Asia. Therefore, and attending to the lack of research and the need to update the existing models in this specific population, the purpose of our study was to develop a multivariable COX regression model to better predict short-term mortality risk among patients with AMI in rural China. We hope that our results will be very valuable to assist clinicians in early identification of highrisk patients, to reduce in turn the mortality associated with acute coronary syndrome in this underserved population.

| Study subjects
This observational, retrospective study was conducted at Linyi People's Hospital from January 2013 to December 2018. It included patients with AMI and collected data on patient's demographics, clinical presentations, medical history, risk factors, treatment, and clinical outcomes.
The study protocol was approved by the Linyi People' Hospital Ethics Committee. The study included patients with ST-segment elevation myocardial infarction (STEMI) as well as those presenting with non-STelevation myocardial infarction (NSTEMI), in accordance with the third universal definition of myocardial infarction. 13 Exclusion criteria: patients with active inflammation, liver failure or renal failure on admission; patients with critical data missing; patients lost to follow-up.
A total of 7533 patients were enrolled in this study: Patients enrolled in 2013-2017 were assigned into a derivation cohort (n = 6303) to establish the multivariable COX regression model, and patients enrolled in 2018 were assigned into a validation cohort (n = 1230) to validate the risk score. The patients in the derivation cohort and the validation cohort were proved to be homogeneous and comparable (shown in Table S1).
As per the exclusion criteria, 184 patients with active inflammation, liver failure or renal failure, 743 patients for which critical data were missing, and 25 patients lost to follow-up were excluded from analysis. We finally included 6581 AMI patients, of which 562 died over the course of the study. As shown in Figure 1.

| Outcome assessment and clinical definitions
The primary endpoint was all-cause short-term death, defined as cardiac or non-cardiac death from admission to 6-month follow-up. None of the deaths recorded over the course of this study were caused by accidental injuries such as trauma or car accidents.
Medical history and vital signs were determined at the time of first hospital presentation. Standard definitions of clinical history and physical examination parameters were applied as described in the ACC/AHA Task Force on clinical Data Standards [14][15][16][17]  All the variables assessed in this study were collected from electronic medical records. After hospital discharge, clinical end point information was acquired by telephone follow-up.

| Statistical analysis
Statistical analysis was performed using IBM SPSS Statistics for Windows version 21 (IBM Corp., Armonk, New York, United States of America). The distribution pattern of the variables was analyzed using the Kolmogorov-Smirnov test. Continuous variables are presented as mean ± SD or median (25th and 75th percentiles). Parametric and non-parametric continuous variables were compared using Student's t test and Mann-Whitney U test, respectively. Categorical variables were compared using Pearson χ 2 test, and the results expressed as percentages. All tests were two-sided; p ≤ .05 or a 95% confidence interval (CI) that did not include 1.0 indicated significance.
Univariate Cox regression was performed to examine the association between individual baseline variables and short-term mortality, described as hazard ratio (HR) and 95% confidence interval (CI). For prediction of short-term mortality risk, the optimized risk model was created by fitting a multivariable COX model to clinical, laboratory analysis, medical history, and treatment variables. All variables that achieved significance (p ≤ .05) on univariable selection were selected to fit the multivariable COX regression model. Then, a stepwise forward selection process was used to identify independent predictors of short-term death. After selection, the variables with p ≤ .05 were retained in the final model.  23 Receiver operating characteristic (ROC) curves were constructed to assess the discrimination of the model. Z test was applied to compare the differences between the two scoring methods.
After the optimized risk score was established, we compared it with the GRACE score, 12 which was evaluated before discharge to predict the risk of 6-month mortality.

| Baseline characteristics
Baseline characteristics for all the patients are shown in Table 1.Compared with survivors, patients who died were older, more often female, and had extensive anterior myocardial infarction. They showed also a greater incidence of polymorphic ventricular arrhythmia, and had a higher Killip class than survivors. Among the dead patients, 95 (16.9%) received Percutaneous coronary intervention (PCI) treatment. Among survivors, 2774 (46.1%) were treated with PCI. The short-term death group had higher heart rate and lower systolic and diastolic blood pressure than survivors. Patients who died also had more comorbidities, that is, higher prevalence of cerebral infarction, atrial/ventricular arrhythmia, chronic kidney disease, diabetes mellitus, and previous heart failure.
On laboratory analysis, patients who died had higher white blood cell (WBC) count, blood glucose, and serum creatinine. On cardiac color doppler ultrasound, non survivors had higher left atrium anteroposterior diameter (LAD), higher PASP, and lower LVEF values. The utility of antibacterial drugs and vasoactive drugs was higher, while the utility of aspirin, low-molecular-weight heparin, lipid-lowering drugs, nitrate drugs, β-receptor antagonists, and Angiotensin-converting enzyme inhibitor / Angiotensin Receptor Blocker (ACEI/ARB) drugs was lower for dead patients.

| Independent predictors of short-term death
Univariate Cox regression was performed to examine the association between individual baseline variables and short-term mortality

| Optimized scoring method
We developed a simplified risk score by attributing an integer number to each variable according to their estimated coefficients ( Table 2).
The optimized score thus obtained ranges from 0 to 254, and the corresponding short-term death risk ranges from 0.3% to 97.7%.
We then divided all participants within the derivation cohort into quartiles based on the corresponding risk scores. Each quartile contained approximately one fourth of the population (

| DISCUSSION
In this large-scale contemporary retrospective study of patients with AMI from rural China, we identified six independent predictors of short-term death and used these variables to develop and validate a risk prediction tool for short-term death among AMI patients. The optimized score, which was designed for rapid risk assessment after presentation, had high discrimination ability in both the derivation and validation cohorts. A significant gradient of increasing short-term mortality risk was identified as the optimized score increased.
Many risk prediction tools have been developed to assess shortor long-term death risk for acute coronary syndrome (ACS) patients.
Among these, the GRACE risk score, proposed in 2003, remains the most popular and validated model. Therefore, and since our novel scoring method shares a similar study object and primary endpoint event as the GRACE risk model, we used the latter to contrast and validate the predictive power of the optimized risk score model herein described.
From 1999 to 2008, the profile of AMI patients changed over time, with a slight increase in NSTEMI and a decrease in STEMI cases. 24 In turn, the incidence and short-and long-term mortality rates of AMI were impacted by an updated diagnostic criteria introduced in 2014. [25][26][27] In parallel, mortality rates for both STEMI and NSTEMI patients have declined significantly due to improved medication management and optimization of invasive treatments. 28  ies have shown that blood glucose levels are commonly elevated in early AMI and represent an independent risk factor for increased inhospital or short-term mortality. 18,19 However, the impact of blood glucose in the context of MI is complex, involving both protective and deleterious effects on the myocardium. 29 This may explain why blood glucose was not included in the GRACE or CAMI risk scores. 10,30 Pulmonary hypertension is frequently observed following AMI, 31 and elevated PASP after AMI is often associated with poor prognosis. 32,33 Indeed, PASP was proved to be a strong independent predictor of short-term death in a previous study. 34 These findings imply that special attention must be paid to elevated PASP during comprehensive evaluation of cardiac function in AMI patients.
In our multivariable COX model, the HRs of blood glucose and PASP were 1.046 and 1.023 respectively, indicating that neither of them can be used individually to determine the risk of adverse events.
Nonetheless, the diagnostic performance of the risk score is improved by addition of blood glucose and PASP into the risk model. were also older than 80 years, or had complications such as stress ulcers, abnormal liver and kidney function, and ventricular tachycardia and hypotension. Thus, for non-survivors the utility of PCI, aspirin, low-molecular-weight heparin, lipid-lowering drugs, nitrate drugs, β-receptor antagonists, and ACEI/ARB drugs was lower.
In recent years, a rise in the morbidity and mortality of the rural population in China has been apparent. We hope that our study can provide a reference for larger studies in the future and serve as a model to identify causative factors driving increased disease incidence and mortality in rural China. The risk score system presented here is simple, easy to implement, and includes objective indicators. We thus believe that adopting our scoring method will significantly improve prediction of short-term risk of death in AMI patients.
Our study has some limitations. First, this is a retrospective study and all the patients included were assessed at a single research center in China. Second, although the discrimination ability of the optimized risk score was confirmed in a separate cohort of patients, its predictive accuracy for short-term death should be further validated using a larger population sample. In addition, whether this risk score can be extrapolated to different ethnicities requires further validation.
In summary, we developed and validated a risk score model to predict short-term death risk in patients with AMI. Our optimized risk score proved to be superior, in our large AMI cohort, to the widely used GRACE risk model. We expect that this optimized risk model will provide clinicians with a simple and useful tool to accurately assess short-term death risk and to select appropriate treatment and level of care for AMI patients.

ACKNOWLEDGMENTS
We thank our study group for their contributions to the research design, study execution, and data analysis. We thank all the investigators and coordinators involved for their active participation and outstanding work.