The patient population used to develop this risk score (development cohort) consisted of the eligible patients enrolled in Guangdong Provincial People’s Hospital between January 2010 and May 2016. All patients with STEMI underwent PCI via radial, femoral, or both approaches, while patients with both accesses were classified as those with femoral assess, and patients with implantation of intra-aortic balloon pump (IABP) were also included. According to the same criteria, the population used for the validation (validation cohort) included those enrolled between June 2016 and May 2018, with complete data compared to the development cohort. Additionally, the patients screened in accordance with the criteria in other research centers were included for external validation. To widen the extent of application, this risk model was also validated among patients with the non-ST-elevation acute coronary syndrome (NSTE-ACS) treated with PCI in our recent publication (9).

Other details of the patients, study design, treatment and procedure, and outcome definitions have been addressed in our previously published protocol (10). Clinical data including patient demographics, laboratory tests, procedural details, medical treatments, and clinical outcomes were collected from the medical records. Vital variables (such as clinical events) were double-checked, and inconsistent data were verified by a third researcher. All clinical adverse events were evaluated by an independent clinical events committee blinded to the research.

The primary endpoint was the occurrence of a P-AMI infection during hospitalization, which was defined as the infection requiring antibiotics (reflecting the clinical influence of infection compatible with the necessity for additional treatment) (11). Besides, the infection was determined using ICD-10-CM codes. The types of infections included pulmonary infection, urinary infection, or other infections including abdominal sepsis, primary bacteremia, and unidentified primary infection site based on the clinical records. Other details were reported in the published protocol (10). The secondary endpoints were in-hospital all-cause death, and major adverse cardiovascular events (MACE) including stroke, all-cause death, target vessel revascularization, and recurrent myocardial infarction.

As for sample size determination, the concept of events per variable (EPV) of 10 or greater was applied as previously reported (12). Considering that no more than 8 factors were included in the development model, at least 80 infections were required. Our development dataset comprised 296 infections, and the value of events per variable was 42.3 (296/7).

Categorical variables were expressed as number (percentage) and continuous variables were expressed as mean ± SD, as appropriate. We chose a backward stepwise logistic regression model to develop the risk score. All listed candidate predictors (Supplementary Table 1) were included to develop the risk score, while variables with ≥ 10% missing values were excluded. And for those with incomplete data, we assumed that patients with the missing data ≤ 5% occurred at random depending on the clinical variables and used multiple imputations. Besides, a bootstrap method was applied to pick out the best subset of risk factors.

The predictive accuracy was evaluated using both discriminations evaluated by the C-statistic, and calibration assessed by the Hosmer-Lemeshow χ² statistic and calibration plot. We performed the external and subgroup validation (older, sex, anemia, diabetes, and chronic kidney disease in the development and validation cohorts) to further measure the risk score’s stability. Additionally, we compared the discrimination of the present model with other risk scores previously reported using the area under the receiver operating characteristic (13–15). To assess the utility of the risk score, we conducted the decision curve analysis. A higher net benefit across the whole range of clinically applicable and acceptable probability thresholds indicated better clinical values (16). Other detailed information of the statistical analysis has been previously published (10).

All analyses were performed with SAS (version 9.4, SAS Institute, 210 Cary, NC, United States) and all figures were plotted with R (version 3.6.2, R Foundation for Statistical Computing, Vienna, Austria).

A total of 1,842 patients with STEMI who underwent PCI (mean age, 61.54 years; 17.5% female) were finally included in the development cohort (Figure 1). Among them, 217 patients (11.8%) experienced post-AMI infection during hospitalization. The in-hospital all-cause death was 17.5%, compared to 1.9% for patients without infections. Patients with post-AMI infection were more likely to be older, and have diabetes, chronic obstructive pulmonary disease (COPD), higher white blood cell (WBC) counts, and diuretics or insulin use. Moreover, these patients had lower levels of serum albumin, hemoglobin, estimated glomerular filtration rate, and left ventricular ejection fraction (Table 1).

Among all 72 variables recorded according to the clinical data, 56 candidates were enrolled for the risk model development, while 16 variables were excluded for ≥ 10% missing values (Supplementary Table 1). After multivariate logistic regression analysis, the following seven variables were ultimately included: age, Killip classification, WBC count, serum albumin, insulin use, diuretic use, and transfemoral approach (Table 2 and Supplementary Figures 1, 2). Subsequently, the risk score, named post-AMI infection score (PAMIIS), was developed attributing corresponding points to each variable (Figure 2), and achieved predicted probabilities for infection development ranging from 0.7 to 99.6%, with a score range between 0 and 24 points (Supplementary Table 2 and Supplementary Figure 3).

This risk score showed good discriminative power (C-statistic:0.851, 95% confidence interval [CI]:0.824–0.877), with good calibration ability in line with the Hosmer-Lemeshow test (χ² = 10.07; P = 0.345) (Figure 2), while the internal validation with bootstrapping also yielded a similar result (C-statistic:0.851, bias-corrected 95% CI:0.831–0.871). More notably, the calibration plots of predicted vs. observed incidences of post-AMI infection across risk scores confirmed an excellent calibration (Figure 3).

For external validation, 1,270 patients with STEMI undergoing PCI were enrolled, and 142 patients (11.2%) experienced post-AMI infection, with a similar infection rate compared to the development cohort (11.2% vs. 11.8%; P = 0.607) (Table 1). Multiple differences in baseline characteristics were observed between the validation and development cohorts (Supplementary Table 3). The validation dataset exhibited semblable distribution of factors (excluding serum albumin, and WBC count) incorporated in the current risk model, and mean risk scores in comparison with the development dataset.

In the validation cohort, this model had similarly good discriminative power and calibration as the development cohort (C-statistic: 0.851, 95% CI:0.818–0.884; Hosmer-Lemeshow test χ² = 5.974, P = 0.650), respectively (Figure 2). The observed incidences of post-AMI infection were highly consistent with the predicted ones (Figure 3).

Furthermore, after validation in the population from other two centers, in which 1,002 patients with STEMI undergoing PCI were included (Supplementary Figure 4) and 75 (7.5%) experienced post-AMI infection, this model remained a prominent predictor of infection (C-statistic:0.770, 95% CI:0.710–0.831; Hosmer-Lemeshow test χ² = 15.412, P = 0.118) (Supplementary Figure 5).

To extend the utility, we further validated the risk score among patients with NSTE-ACS after PCI. A total of 4,745 patients (mean age, 63.70 years, 24.4% female, mean risk score 3.03 ± 2.15) were included (Supplementary Figure 6), and 118 (2.49%) experienced infection. The incidence of all-cause death and MACE were 0.30 and 0.80%, respectively. The comparison of baseline data was summarized between patients with or without infection in Supplementary Table 4. This risk model demonstrated analogously favorable discriminative ability in patients with NSTE-ACS (C-statistic:0.800, 95% CI:0.754–0.845), but the predicted risk was higher than the observed one in these relatively low-risk patients (Supplementary Figures 7, 8). After times 0.684 to the predicted risk, the observed incidences of post-AMI infection were highly consistent with the predicted ones with Hosmer-Lemeshow test χ² value of 5.596 (P = 0.588, Supplementary Figure 7).

Furthermore, this risk score exhibited good predictive values for in-hospital all-cause death and MACE in both cohorts. Similar results were observed among patients with NSTE-ACS (Supplementary Figure 9).

As for IABP, there were 196 patients with implantation of IABP in the development cohort. To evaluate the potential influence of IABP on the risk score, we performed the subgroup of IABP analysis. And the result showed that the present risk score showed a similar predictive effect on infection between patients with (AUC:0.816, 95% CI:0.779–0.854) and without (AUC:0.789, 95% CI:0.724–0.854) use of IABP (Supplementary Table 5). Besides, subgroup analyses including age, gender, anemia, diabetes, and chronic kidney disease proved consistent predictive values for P-AMI infection in both datasets (Supplementary Table 5), with the best calibration in the aforementioned subgroups (calibration slope nearly 1) (Supplementary Figure 10).

Additionally, compared to previously established risk scores [age, estimated glomerular filtration rate, and ejection fraction (AGEF); age, creatinine, and ejection fraction (ACEF); and Canada Acute Coronary Syndrome (CACS) risk score], the present risk score, PAMIIS, achieved better discriminative power (Supplementary Figure 11).

The optimal cut-off identifying post-AMI infection was a score of 6.5 (sensitivity 81.5%; specificity 74.0%). As shown in Supplementary Figures 12, 13, increased risk of infection was obtained across risk score groups. Compared to those with a low-risk score, high-risk patients also presented elevated MACE and all-cause death in the development, validation, and NSTE-ACS cohorts (P < 0.001). Based on the types of infection, the incidences of pulmonary and urinary infection performed as the upward trend across groups in both datasets (Supplementary Figure 14). Decision curve analysis demonstrated that the present risk score displayed a better-standardized net benefit for infection than other risk scores, indicating better clinical usefulness (Supplementary Figure 15).