The design of the China PEACE-Retrospective AMI Study has been published previously[25, 26]. In brief, a nationally representative sample of AMI hospitalizations was obtained following a two-stage sampling design: First, we identified hospitals using a simple random sampling procedure within each of the 5 study strata: Eastern-rural, Central-rural, Western-rural, Eastern-urban, and Central/Western-urban regions, since hospital volumes and clinical capacities differ between urban and rural areas as well as among the three official economic-geographic regions (Eastern, Central, and Western) of Mainland China. We considered Central and Western urban regions together given their similar per capita income and health services capacity. In the 3 rural strata, the sampling framework consisted of the central hospital in each of the predefined rural regions (2010 central hospitals in 2010 rural regions). In the 2 urban strata, the sampling framework consisted of the highest-level hospitals in each of the predefined urban regions (833 hospitals in 287 urban regions). Since the majority of hospitals in China are publicly owned and administered, hospital closure is rare. We selected representative hospitals from 2011 to reflect current practices and traced this cohort of hospitals backward to 2006 and 2001 to describe temporal trends. Second, we drew hospitalizations from the selected hospitals based on the local hospital database for patients with a definite discharge diagnosis of AMI in each year using random sampling procedures. Patients with AMI were identified using International Classification of Diseases - Clinical Modification codes, including versions 9 (410.xx) and 10 (I21.xx), when available or through principal discharge diagnosis terms. Information on the patient characteristics, in-hospital treatments, and outcomes were extracted from the medical records. Hospital characteristics were derived from a standardized survey to all the selected hospitals, as well as information from medical records.

The central ethics committee at the China National Center for Cardiovascular Diseases approved the China PEACE-Retrospective AMI Study. All collaborating hospitals accepted the central ethics approval except for five hospitals, which obtained local approval. The Chinese government, who provided financial support for the study, had no role in the design or conduct of the study; in the collection, management, analysis, and interpretation of the data; or in the preparation or approval of the manuscript. The study is registered at www.clinicaltrials.gov (NCT01624883).

Our study sample included 16,100 patients who had AMI on presentation to the hospital. We excluded those who died within hospital, who withdrew treatment by request of the patient or the family due to deteriorating clinical condition, who transferred into or out of the hospital, or who had a coronary artery bypass graft (CABG) during hospitalization (Figure 1). In a secondary analysis, we considered only the subgroup of patients without major complications during hospitalization, for whom an extended LOS maybe particularly unnecessary. Uncomplicated patients were defined as those without recurrent myocardial infarction or angina, cardiogenic shock, cardiac arrest, new-onset heart failure, atrial fibrillation, ventricular tachycardia or fibrillation, stroke, bleeding, acute renal failure, or infection during hospitalization.

LOS in days was primarily determined by the admission and discharge dates documented on the front page of the medical records, and complemented by other sections of the record when either date was missing, uninterpretable, or mis-documented on the front page.

Hospitals were divided into tertiles based on the median LOS across all years. To assess statistical differences and take into account the clustering of patients within hospitals, patient characteristics were compared across the tertiles of hospitals using the Chi-square test or ANOVA for clustered data. We also examined the trends in LOS over the study period among all patients and in uncomplicated patients only. Mann-Kendall trend test was used based on three time points (2001, 2006, and 2011). When deriving the patient-level LOS statistics, the weight for each patient was the inversed sampling fraction of patients from the hospital multiplied by the inverse sampling fraction of hospitals from the geographic region for each study year.

Our subsequent analytic cohort further excluded patients who had LOS lower than the 1st percentile value (i.e. who were discharged on the same day or overnight) or greater than the 99th percentile value (>42 days). LOS was considered a log-normal distribution. In the overall study sample, two-level generalized linear mixed model with patients nested within hospitals was used to determine the risk-standardized LOS (RS-LOS). The “risk profile” consisted of factors that may be associated with the hospitalization LOS, including patient’s sociodemographics, cardiac and noncardiac comorbidities, indicators of the severity of AMI, and year. RS-LOS for each patient was defined as the ratio of predicted to expected LOS, multiplied by the average LOS of the cohort. The expected LOS for each patient was estimated by applying the estimated regression coefficients to the characteristics of the patient and adding the average of the 160 hospital-specific intercepts. The predicted LOS of each patient was calculated with similar methods but using individual hospital-specific intercepts instead of the average value. Therefore, although RS-LOS was technically calculated for each patient based on the model, it was really a hospital-specific indicator (i.e. patients within the same hospital had identical RS-LOS).

Furthermore, we assessed inter-hospital variation in RS-LOS across the study years, among all hospitals as well as stratified by secondary and tertiary hospital levels. We did not adjust for patient in-hospital treatments and complications as they could act as mediators, and adjustment for these factors may result in underestimates of variation.

Lastly, we examined the association of year-specific hospital RS-LOS with the selected hospital characteristics, some of which might change over the years for the same hospital. Year-specific hospital RS-LOS was derived similarly as described above, except it was estimated separately in patients from each of three study years. We fit a generalized estimating equation (GEE) linear model, weighted by hospitals’ AMI patient volume for each study year in the study sample to account for the uncertainty in the estimates of RS-LOS, and took into account the correlations of different year-specific observations for the same hospital. Absolute differences in RS-LOS and 95% confidence intervals (CIs) were reported for each hospital characteristic, adjusting for temporal (i.e. year) effect.

All comparisons were 2-sided, with a p-value less than 0.05 considered statistically significant. Statistical analysis was performed using the SAS software (version 9.3, SAS Institute, Cary, NC).

The median and interquartile range (IQR) of LOS was 13 (8–19) days in 2001, 11 (7–16) days in 2006, and 11 (7–15) days in 2011 (Table 1); the rankings of the average LOS of the three tertiles of hospitals were consistent throughout the three study years. Compared with hospitals in the low tertile of median LOS, high-tertile hospitals contributed more patients to our study sample in 2001 (16.2% vs. 10.4%) and fewer in 2011 (55.4% vs. 61.1%). With respect to patient characteristics by tertiles of hospitals, high-tertile hospitals had a greater proportion of current smokers (38.4%) and a lower proportion of patients with chronic renal disease (16.7%); middle-tertile hospitals had a greater proportion of patients with a history of MI (12.5%); and low-tertile hospitals had fewer patients with estimated glomerular filtration rate > 90 ml/(min*1.73 m²) (26.0%). Other patient characteristics were not statistically different across the three tertiles of hospitals. Of note, with regard to the distribution of outliers in LOS, low-tertile hospitals had more patients who had LOS of 0 or 1 day (4.0%), while high-tertile hospitals had more patients who had LOS of >42 days (1.4%).

As Table 2 shows, 8.2% of the patients in high-tertile hospitals received no biomarker testing in contrast to 17.0% in low-tertile hospitals. Percentages of medication use, reperfusion therapies, and overall cardiac procedures during the hospitalization were similar between patients in hospitals from the three tertiles. In-hospital outcomes were also similar (Table 3), except for in-hospital infection – 12.6% of the patients had infections acquired during hospitalization in high-tertile hospitals, whereas the proportions were 10.8% and 8.7% for middle- and low-tertile hospitals, respectively.

Compared with 2001, LOS for patients with AMI decreased in 2006 and 2011 (p for trend <0.001) (Figure 2). After weighting, among the 1,901 patients from 2001 the mean LOS was 14.6 days. The corresponding LOS for 2006 (n = 3,553) and 2011 (n = 7,252) were 12.6 and 11.9 days, respectively. Among the 8,049 uncomplicated patients, the weighted mean LOS was 13.4 days for the 1,196 patients from 2001, 11.2 days for the 2,243 patients from 2006, and 10.8 days for the 4,610 patients from 2011.

In the multivariable risk-standardization model, after adjusting for other patient characteristics, study year was significantly associated with LOS for AMI. Compared with a patient with the same underlying risk in 2011, a patient in 2001 had a 25% longer LOS (ratio, 1.25; 95% CI, 1.22-1.29) and a patient in 2006 had an 8% longer LOS (ratio: 1.08; 95% CI, 1.06-1.10). Further information about the model coefficients and the residual plot of the risk-standardization model are provided in the Additional file1.

Overall mean hospital RS-LOS was 12.5 (s.d. 1.8 days). Among all 160 hospitals, there was a substantial variation in hospital RS-LOS, ranging from 9.2 to 18.1 days (Figure 3). The patterns of variations were similar between secondary and tertiary hospitals. Among 95 secondary hospitals, the mean hospital RS-LOS was 12.0 days, with a s.d. of 1.7 days and a range of 9.2 to 18.1 days (see Additional file2 for the graphic pattern of variation). Among 65 tertiary hospitals, the mean hospital RS-LOS was 12.8 days, with a s.d. of 1.7 days and a range of 9.5 to 17.8 days (see Additional file3 for the graphic pattern of variation).

After adjusting the standard errors of estimates using GEE, geographic region was the only factor independently significantly associated with hospital RS-LOS (Table 4). Compared with hospitals in the Eastern region, Central-region hospitals on average had 1.6 days (p = 0.02) shorter RS-LOS. All other hospital characteristics relating to capacity for AMI treatment were not significantly associated with LOS.