Development and Validation of a 90-day Mortality Prediction Nomogram for AMI Patients: A Retrospective Cohort Study

Background: The purpose of this study was to identify the factors inuencing the 90-day mortality of acute myocardial infarction(AMI) patients, and to establish a prognostic model for these patients based on the MIMIC-III database. Methods: Retrospective study methods were used to collect AMI patient data that met the inclusion criteria from the MIMIC-III database. Variable importance selection was determined using the random forest algorithm. Multiple logistic regression was used to determine AMI-related risk factors, with the results represented as a nomogram. Results: The baseline scores for the training and validation groups were very at, and indicators for developing risk-model nomograms were obtained after random forest and multiple logistic regression. The AUC of the risk model was the highest (0.826 and 0.818 in the training and validation groups, respectively) . The Hosmer-Lemeshow goodness-of-t test and standard curve both produced very consistent results. Both the NRI and IDI values indicated that the risk model had signicant predictive power, and DCA results indicated that the risk model had good net benets for clinical application. Conclusions: The results of this study indicated that age, troponinT, VT, VFI, MI_his, APS-III, bypass, and PCI were risk factors for 90-day mortality in AMI patients. Interactive nomograms could provide intuitive and concise personalized 90-day mortality predictions for AMI patients. R.Y.. Data collection or processing: R.Y., W.M.. Analysis or interpretation: R.Y., W.M., L.Z., T.H. and D.H.. Literature search: J.L., Z.D., R.Y. and S.Z.. Writing: R.Y. and all authors controls.


Introduction
Acute myocardial infarction (AMI) is a serious type of coronary heart disease that is a serious threat to human health and survival. Reports suggest that more than 8 million lives are threatened by AMI each year [1,2]. These reports highlight the signi cance of studying the prognosis of AMI patients, which can promote timely medical interventions and improve the prognosis of these patients, reducing the economic burden and other negative impacts of AMI on patient's families and wider society.
The current prognostic scoring system for AMI patients is not comprehensive. For example, the current scoring system of coronary heart disease is the Global Registry of Acute Coronary Events (GRACE) [3,4], which analyzes the unstable angina pectoris in acute coronary syndrome and acute non-ST segment elevations to evaluate risk of myocardial infarction. The GRACE Score cannot clearly predict all types of AMI, since prognosis prediction of AMI patients cannot be applied completely. Other medical scores, such as the Acute Physiology Score (APS-III) [5] and the Sequential Organ Failure Assessment (SOFA) score [6,7], have not been adapted for AMI patients.
The present study was based on a large sample from the Medical Information Mart for Intensive Care III (MIMIC-III) database [8,9], a large critical medical database, which was used to develop a new predictive nomogram to independently predict the 90-day mortality, and provide treatment guidance and prognosis improvement for AMI patients. Study population and data extraction Data were collected from patients who were rst admitted to the ICU (if there were multiple admissions for the same patient) and diagnosed with AMI according to the ICD-9. Those younger than 18 years or for whom more than 5% of the required data were missing were excluded from this study.

Statistical analyses
Univariate analyses were applied to all variables in our study. The Shapiro-Wilk test was used to assess the distributions of variable. Continuous variables that did not conform to a normal distribution were represented by median values and interquartile ranges, and Kruskal-Wallis rank-sum or Mann-Whitney U tests were used to compare them. All classi ed variables were expressed as numerals or percentages and compared using chi-square or Fisher's exact tests.
Random forests were used to select or exclude variables for the model. Multivariate logistic regression analysis was used to develop predictive models. The aim of the nomogram was to predict the probability of 90-day mortality in AMI patients. To develop and verify the model, AMI patients were randomly divided into training and validation groups, at a ratio of 7:3.
The discriminability of the nomogram was evaluated using the area under the receiver operating characteristic curve (AUC) , which was also used to identify which model was better at predicting AMI. The Hosmer-Lemeshow goodness-of-t test (P>0.05) and calibration plots were used to calibrate the nomogram. Integrated discrimination improvement(IDI) and net reclassi cation improvement(NRI) values were calculated to estimate model differentiation. Decision curve analysis(DCA) was used to predict the clinical effectiveness and net bene ts of the model.
During logistic regression and model development, missing values were identi ed using multiple interpolation. R software (version 4.0.3) was used for all statistical analyses. P values less than 0.05 (in two-sided tests) were considered to be signi cant.

Characteristics of AMI patients
After screening of inclusion and exclusion criteria, the nal sample for the study was 4610 patients, of which 894 had died within 90 days ( Table 1) . The only index that differed signi cantly between the two groups was AG. Among the 1383 patients in the modeling group, 76 (5.5%) had a previous history of myocardial infarction and 488 (35.3%) had undergone PCI.
Variable selection and nomogram development Figure 1 shows the process and results of the random forest feature selection, which identi ed that the following 19 risk factors affected the prognosis of AMI patients: APS-III, bypass, age, PCI, AG, PT, INR, LDH, AST, troponinT, CK, platelets, congestive heart failure, APTT, MI_his, CK_MB, AFI, VT, and VFI. Multivariate logistic regression analysis was conducted on these risk factors, and this results are listed in Table 2. We developed our new risk model by identifying eight signi cant risk factors using multivariate logistic regression analysis. An interactive nomogram was developed based on this model for predicting the 90-day mortality of AMI patients ( Figure 2) . Based on the GRACE Score table, we extracted and processed all variables except ST segment reduction to produce a similar GRACE Score and this results are listed in Table 3.

Nomogram performance
AUCs were used to evaluate the differences between our risk model and similar GRACE Score, APS-III, and SOFA as shown in Figure 3. Our risk model had the highest AUC values (0.826 and 0.818 in the training and validation groups, respectively) .
Pair-wise AUC comparisons of the different models revealed all P values to be less than 0.001, indicating that there were statistically signi cant differences in the predictive ability between the different models.

Nomogram calibration
The Hosmer-Lemeshow goodness-of-t test revealed a high consistency between the prediction and observation probabilities of the training (chi-square=16.91, P=0.051) and validation groups (chi-square=11.72, P=0.230) . The calibration diagram also revealed good consistency between the predicted and observed results of the training and validation groups (Figure 4) . These results indicate that the nomogram developed in this study exhibits improved prediction ability compared with the other models.

Clinical application
The DCA plots indicated the high clinical net bene t of our risk model and demonstrated its clinical applicability and impact on clinical decision-making ( Figure 5) .

Discussion
AMI is a serious threat to human health and mortality worldwide, and has a high incidence rate with an ongoing tendency to increase [10][11][12]. All patients with AMI should receive short-term risk assessments as soon as possible after admission, which we aimed to provide with 90-day mortality risk prediction models that guide the treatment and prognosis of patients.
The most widely used scoring system is the GRACE Score [3,5,13], but there are many limitations to this system. First, it was established and validated between 1999 and 2003, and the characteristics of AMI patients and treatment regimens are constantly changing. Second, the GRACE Score also considers patients with ST-segment depression, but the type of AMI is far more than this. Similarly, other widely used clinical scoring systems such as APS-III and SOFA are not speci c to AMI, and cannot be generalized to the risks and prognosis predictions of all AMI patients.
Based on the lack of an accurate scoring system for AMI patients, 4610 patients and 30 risk factors relating to AMI were selected using sample data from the MIMIC-III database [14][15][16][17][18]. Using a random forest [19] selection method identi ed 19 risk factors that had been identi ed in the multivariate logistic regression analysis [20]. We concluded that there were eight independent risk factors for AMI, and our new risk prediction model was developed using these eight factors, which were then drawn into a nomogram that provides accurate, simple, and intuitive predictions [21][22][23]. The new model indicated that there were six variables in the following order of increasing risk: age, APS-III, troponin T, MI_his, VT, and VFI. This suggests that AMI patients with advanced age, poor acute physiological status, elevated troponin T, previous history of myocardial infarction, VT, and VFI had a higher risk of death at 90 days after diagnosis, and should be a priority for timely interventions and treatments. The application of early PCI and bypass surgery can somewhat reduce the risk of 90-day mortality in AMI patients, which is conducive to their treatment and for improving their prognosis [24,25]. In addition, the factors used to develop our risk model are readily available and can be routinely collected from historical records, strengthening the clinical application value of our prediction model.
We conducted a comprehensive evaluation of our developed prognostic model. We extracted most of the risk factors in the GRACE Score, constructed our own GRACE Score, and extracted additional APS-III and SOFA scores to compare the four models to better assess the predictive power of the new model. This analysis revealed that many variables with similar GRACE Scores were not statistically signi cant after the multiple logistic regression analysis, such as age and meanHR after GRACE scale classi cation, AST, and CK, also indicating that the GRACE Score is not applicable to AMI patients.
The AUC values of the new model before and after internalization veri cation were 0.826 and 0.818, respectively, which are far higher than the AUC values of the other three comparison models. The P values from AUC pairwise comparisons were all less than 0.001, indicating that the predictive e cacy varied between the different models, and that the new model had improved ability in distinguishing the 90-day mortality of AMI patients. The Hosmer-Lemeshow goodness-of-t test and calibration plots further indicated that the model was consistent with the data.
We also used NRI, IDI, and DCA to assess whether the newly developed prognostic model performed well and its clinical usability [26]. Our results indicated that the new model was better than the other three models for both groups. DCA indicated that the new model also had greater clinical net bene ts than the other three models, and can be better applied to clinical decision-making.
As expected, this study had some limitations. First, it had a retrospective design, and so prospective con rmatory studies are required. Second, only internal validation was used in this study, with external validation from other institutions providing further validation of our nomogram. Third, multiple logistic regression was used in this study, and it may be necessary to add time variables to further verify the accuracy of the results.

Conclusions
Through random forests and multivariate logistic regression analysis on our large data sample from the MIMIC-III database, we identi ed eight independent risk factors for 90-day mortality in AMI patients. We

Funding
The authors have not declared a speci c grant for this research from any