COLGE is a single-center prospective registry of consecutively admitted patients with acute coronary syndrome (ACS). All patients with MI were included. MI was defined according to the third universal definition of type 1 MI.28

Data were collected by cardiology residents and specialists via an application embedded in the ATRIUM software of the Madeira Regional Health Authority and are subject to the data protection regulations of the regional healthcare system. Patients’ demographic, clinical management and clinical outcome data were collected. Previous diagnoses were defined, if known, according to specific medication or to the most recent European guidelines at the time of data collection. Patient names were hidden during data management. Therapeutic decisions were left to the attending physician and site-specific protocols.

This manuscript was written in accordance with the TRIPOD checklist29 and the study complies with the 1975 declaration of Helsinki.30

The primary endpoint was all-cause mortality during the index hospitalization. The secondary endpoint was cardiovascular mortality during the index hospitalization.

Cardiovascular mortality was defined as death due to the immediate consequences of MI, such as progressive heart failure, mechanical complications or malignant arrhythmias, or death due to complications during coronary procedures to treat myocardial ischemia.

Angioplasty was defined as balloon angioplasty with or without stent deployment.

Bleeding was defined as any sign or symptom of bleeding that requires medical intervention, leads to prolonged hospitalization or increased level of care, requires adjustment of anticoagulation or antiplatelet therapy, leads to a fall in hemoglobin ≥2.0 g/dl, requires transfusion of whole blood or red cells, is located in a critical area or organ (intracranial, intraspinal, intraocular, retroperitoneal, intra-articular or intramuscular with compartment syndrome), or is fatal.

Clinical variables available at first medical contact were assessed as predictors of the primary endpoint. Only the first recorded value was counted for continuous variables. A history of clinical conditions such as diabetes or hypertension was not included since they are not measurable and may not be available at first medical contact. Patients with missing data in any of these variables were excluded from the analysis.

A minimum sample of 385 participants was calculated to be required for a 5% margin of error with 95% confidence level (CI).

Categorical variables were presented as percentages and compared using the chi-square test. Continuous variables were tested for normality with the Kolmogorov-Smirnov test. Those that displayed a normal distribution were presented as means and standard deviation and compared using analysis of variance. Continuous variables without normal distribution were presented as medians and interquartile range and compared with the Kruskal-Wallis test.

Associations of the selected predictor variables with the primary endpoint were first tested using any of the above-mentioned tests. Variables associated with in-hospital mortality were further assessed using multivariate backward Wald logistic regression, isolating the strongest predictors of the primary endpoint. Using a simple multiplicative model, the selected variables were placed as numerator or denominator according to their direct or indirect relationship with the primary endpoint, respectively.

The model's robustness was assessed in multivariate enter logistic regression, adjusted to other confounding variables associated with in-hospital mortality that were not available at patient admission or not selected in the first regression.

The model's predictive power was assessed using receiver operating characteristic (ROC) curves and the respective area under the curve (AUC). The model was compared with other validated risk scores (GRACE, TIMI, PURSUIT, SRI, EMMACE, CHA2DS2-VASc-HS and ProACS). ROC curves were compared using the DeLong test.31

The model was then categorized into four subgroups using IBM SPSS® Statistics Optimal Binning, further adapted by the authors to increase the clinical utility of the resulting subgroups. The predictive power of the categorized model was compared with the classic Killip classification on admission. Patient reclassification was assessed using continuous net reclassification improvement (cNRI).

Statistical analyses were performed in IBM SPSS Statistics® version 23.0.0.0. ROC curve analysis and the DeLong test were performed in MedCalc® version 14.8.1 and cNRI was developed in R software, version 1.0.136, using the PredictABEL package.32 All statistical tests were two-sided with a cut-off value of 0.05 for significance.

A total of 1783 patients were consecutively admitted for ACS between October 2009 and September 2016. Patients with unstable angina (n=210) and patients with missing data in the variables available on admission (n=69) were excluded (Figure 1), resulting in a population of 1504 patients consecutively admitted for MI (Table 1). This sample size provides a margin of error of 2.53% with 95% CI.

Figure 1.

Flowchart of the patient inclusion process. ACS: acute coronary syndrome; MI: myocardial infarction.

(0.09MB).

The clinical variables available at first medical contact were isolated and their associations with in-hospital mortality were tested (Supplementary Table 1). Variables with a statistically significant association with mortality were age, gender, chest pain, Killip classification at admission, systolic and diastolic blood pressure and heart rate at admission.

The isolated variables were studied in a multivariate backward Wald logistic regression model. The regression model correctly predicted 92.8% of the primary endpoint events. The most robust variables were Killip class on admission, age, systolic blood pressure and heart rate (Table 2).

Using a basic multiplicative structure, the four selected variables were placed as numerator or denominator depending whether their odds ratio for mortality was above or below 1.0, respectively. The score was named KAsH, after the initials of each variable and their position as numerator or denominator.

In ROC curve analysis, KAsH demonstrated excellent predictive power for in-hospital mortality, with an AUC of 0.861, 89% sensitivity and 70% specificity (Supplementary Figure 1).

KAsH displayed excellent robustness, maintaining predictive power even after adjustment for confounding variables including gender, diabetes, previous MI, heart failure or chronic kidney disease, no chest pain on admission, any clinically significant bleeding, cardiogenic shock, mechanical complications, no invasive coronary angiography and STEMI (odds ratio 1.004; 95% CI 1.001-1.008; p=0.012) (Supplementary Table 3).

Addition of the continuous variables of age, systolic blood pressure and heart rate to Killip class significantly improved classification: 81.5% of patients were classified into a better risk stratum with all four variables compared to the Killip classification alone (Table 3).

Compared with previously validated scores, KAsH displayed the highest predictive power for both all-cause and cardiovascular mortality (Figure 2). In the subgroup analysis, KAsH showed excellent consistency, being the highest-rated score in the majority of subgroups for both primary and secondary endpoints (Table 4 and Supplementary Table 4).

Figure 2.

Comparison of receiver operating curves for KAsH and other risk scores for predicting all-cause (A) and cardiovascular (B) mortality.

(0.16MB).

Using cut-offs derived from the IBM SPSS® Optimal Binning protocol, KAsH was categorized into four risk groups. The cut-offs were adjusted for natural numbers (multiples of ten), both for greater simplicity and to adequately discriminate four clinically relevant patient subgroups. The groups displayed excellent mortality discrimination (Table 5).

After categorization, KAsH retained very good predictive power for in-hospital all-cause and cardiovascular mortality (AUC 0.839 and 0.834, respectively). It displayed a highly significant improvement in predictive power compared to the Killip classification alone (AUC 0.775, p<0.0001) (Supplementary Figure 2).

Firstly, although COLGE is a prospective continuous registry, some patients were excluded due to missing data. However, they were only a small proportion of the overall population (n=4.3%), and this is unlikely to have significantly changed the analysis.

Secondly, KAsH was created based on the experience of a single center, and thus requires further external validation in other populations.