This retrospective observational study investigated patients admitted for STEMI to SNS hospitals and treated with PCI. The data source was the Minimum Basic Data Set (MBDS)11 of the Ministry of Health, Consumer Affairs, and Social Welfare. The usefulness of this database for the study of acute coronary syndrome in Spain was recently validated.12 The following events occurring between January 1, 2003, and December 31, 2005, were selected: patients with a principal diagnosis of STEMI who were treated with PCI and discharged alive (ICD-9-CM codes; table 1 of the supplementary data). Patients treated with thrombolysis or cardiac surgery were excluded to reduce analytical biases. Also excluded were patients with any of the following complications associated with the index event: heart failure, acute pulmonary edema, ventricular arrhythmias, cardiac arrest, and cardiogenic shock (table 1 of the supplementary data).

To improve the consistency and quality of the data, a) events corresponding to patients admitted to one hospital and transferred to another for PCI were concatenated and the lengths of the stay in the referring hospital were attributed to the second hospital, and b) events corresponding to patients younger than 35 years and older than 94 years were excluded, as well as discharges against medical advice or due to transfer to a social health center or an unknown reason, patients with events not causing hospitalization who were discharged to home, patients transferred to another hospital who returned to the referring hospital, and events classified in Major Diagnostic Category 14 (pregnancy, childbirth, and the puerperium) of the All Patient Refined Diagnosis-Related Groups.13

The length of hospital stay, measured in days, was calculated as the difference between the discharge and admission dates and its changes over time were analyzed between 2003 and 2015. The impact of a short length of hospital stay (≤ 3 days) was evaluated for the 2012 to 2014 period because the MBDS enabled reliable identification of readmission events during this period. Outcome variables were unscheduled readmission 30 days after discharge from the index event for cardiovascular diseases (rheumatic heart disease, hypertensive heart disease, ischemic heart disease, pulmonary vascular diseases, other heart diseases, and aortic dissection and aneurysm, as well as other admissions to cardiology and cardiac surgery, regardless of the principal diagnosis) and the in-hospital mortality during these readmissions (because the MBDS does not include out-of-hospital information).

Multilevel mixed-effects Poisson regression was applied to adjust the length of stay, given its skewed distribution.14 In addition to year of discharge and patient age and sex, the risk factors included in the models designed by the Centers for Medicare and Medicaid Services (CMS methodology) were used to adjust for the risk of mortality and readmission for AMI.15,16 We adapted the models to the structure of the MBDS, after grouping the secondary diagnoses according to the clinical condition categories proposed by Pope et al.,17 updated annually by the Agency for Healthcare Research and Quality.18 The expected length of stay was derived from the individual predictions obtained from the estimated model.

Taking into account the existence of characteristics specific to the patients and to the treating centers that are independent of the quality of the health care provided,19 the 30-day readmission and mortality rates were adjusted for risk according to the CMS methodology. The independent variables were those included in the models for readmission and mortality from AMI, respectively, and a dichotomous variable was included to identify whether the stay was short or long.

Multilevel logistic regression models20 were constructed that, in addition to clinical and demographic variables, included a specific random effect related to the hospital21,22; using backward elimination, significance levels of P <.05 and P ≥ .10 were applied for factor selection and elimination, respectively. The discrimination of the definitive models was assessed using the area under the receiver operating characteristic curve (AUC).

In all multilevel models, if the treating hospital could not be identified, the corresponding events were eliminated; in addition, the incidence rate ratios or odds ratios (ORs) were calculated, as appropriate, as well as their 95% confidence intervals (95%CIs).

The risk-standardized readmission and mortality ratios (RSRRs and RSMRs, respectively) were calculated as the ratios between the expected outcome (which individually considers the functioning of the hospital treating the patient) and the expected outcome (which considers standard functioning according to the average of all hospitals) multiplied by the crude readmission or mortality rate of the study population.18,23 If the RSRR (or RSMR) of a hospital is greater than the crude readmission rate (or crude mortality rate), the probability of readmission (or mortality) in that hospital is higher than the average of the hospitals studied.

To minimize selection bias, the impact of a short length of stay on readmission and mortality during these readmissions was assessed with propensity score matching (k-nearest neighbors matching option, psmatch2, Stata). To do this, we selected long-stay events that, based on the patients’ demographic characteristics and comorbidities, had a similar short-stay probability to that of each short stay event. The matching was performed using the risk adjustment models at a 1:1 ratio and with a maximum propensity score difference (caliper) of 0.05 standard deviations and without replacement. Also calculated were the probability of readmission or death, the effect of the differences between the 2 groups (average treatment effect on the treated [ATT]), and the ORs with their 95%CIs.

To discriminate between high- and low-volume hospitals (according to the number of events attended), a k-means clustering algorithm was used in two thirds of the dataset to obtain the maximum intracluster and minimum intercluster densities, which were validated with the remaining third of the dataset.

The RSRR and RSMR of 30-day readmissions were compared between the 2 length of stay groups classified according to their complexity using the RECALCAR24 typology (table 2 of the supplementary data) and the volume of events recorded.

As sensitivity analysis, the impact of transfers between hospitals was evaluated by excluding index events with discharge due to transfer.

Quantitative variables are presented as means±standard deviations and categorical variables as frequencies and percentages. Correlations among quantitative variables were analyzed with the Spearman rank coefficient (ρ). The t test was used to compare 2 categories, whereas analysis of variance (ANOVA) with the Bonferroni correction was used for 3 or more. Comparisons among discrete variables were made using the chi-square test or Fisher exact test. All comparisons were 2-sided and differences were considered significant at P <.05. All statistical analyses were performed with STATA 13 or SPSS 21.0.

In total, 205 016 hospitalization events with STEMI as the principal diagnosis and treatment with PCI were identified; of these, 190 078 patients were discharged alive. Once transfer events were concatenated across hospitals, 188 854 events remained. After exclusions, 134 002 events comprised the study population for the analysis of trends (figure 1).

Figure 1.

Flowchart of patients’ events. MSCBS-MBDS, Minimum Basic Data Set provided by the Spanish Ministry of Health, Consumer Affairs, and Social Welfare; PCI, percutaneous coronary intervention; SNS, Spanish National Health System; STEMI, ST-segment elevation myocardial infarction. *Exclusions were not mutually exclusive.

(0.47MB).

The crude mean length of hospital stay was 7.2±5.5 days and ranged between an annual maximum of 9.0±7.2 days in 2003 and a minimum of 6.1±4.7 days in 2015. The length decreased each year vs the previous year by 3.2%. The adjusted length of hospital stay was significantly reduced (incidence rate ratio <1; P <.001) each year of the period analyzed. With 2003 as reference (table 1), the expected mean lengths of stay showed significant annual differences (figure 2) and the percentage of short lengths of stays increased (from 14.3% in 2003 to 19.3% in 2015; P <.001) (figure 3).

Figure 2.

Annual changes in mean expected lengths of hospital stay (adjusted for the length the stay using Poisson multilevel regression). The bars represent the 95%CI of the mean adjusted length of stay. 95%CI, 95% confidence interval.

(0.19MB).

Figure 3.

Annual changes in the proportions of discharges from short (≤ 3 days) hospital stays.

(0.11MB).

Between January 1, 2012, and December 31, 2014, 38 137 index events (28.5%) were identified from all patients with STEMI index events treated with PCI recorded in the MBDS. Of these, 6486 (17%) were short. Short lengths of stay were more likely in women, younger people, and patients with fewer comorbidities (table 2).

The most frequent causes of readmission are shown in table 3 of the supplementary data. The crude rate of 30-day readmission was 2.59% (2.69% for short hospital stays vs 2.61%) and the crude rate of mortality during these readmissions was 0.11% (0.15% vs 0.10%).

The adjusted risk models of 30-day readmissions and mortality during these admissions are shown in table 3. A short hospital stay was not an independent predictor in either of the models: OR=1.10 (95%CI, 0.92-1.32; P=.30) and OR=1.94 (95%CI, 0.93-14.03; P=.077), respectively.

Discrimination for 30-day readmission was low (AUC=0.65) but was appreciably higher (AUC=0.84) for mortality during these readmissions (figure 1 of the supplementary data). The intraclass correlation coefficients of both models were 0.02 and 0.11, respectively, and their median ORs were 1.3 and 1.8, indicating considerable variability among hospitals.

To analyze the effects of a short hospital stay, propensity score matching was used to obtain 2 cohorts of 12 966 events for readmissions and of 12 972 events for mortality (99.99% and 100% of the events with short hospital stays). The characteristics of these groups are shown in table 4. After matching, there were no significant differences between the 2 groups in 30-day readmissions (ATT=0.024 vs 0.019; P=.204; OR=1.26; 95%CI, 0.98-1.62) or 30-day mortality (ATT=0.001 vs 0.001; P=.52; OR=1.5; 95%CI, 0.48-5.13).

There were also no significant differences between the 2 groups in the RSRR (2.54% vs 2.55%; P=.020). Although the mean RSMR at 30 days was significantly higher in the short hospital stay group, the difference was probably not clinically significant (0.103% vs 0.109%; P <.001) (table 4 of the supplementary data).

From 2012 to 2014, hospitals with higher volume (> 415 events) had a larger proportion of short hospital stays (18.10% vs 17.10%; P=.02), but no significant differences were found in the mean lengths of stay, the 30-day RSRR, or the 30-day RSMR according to volume (all P > .05). Neither were there significant correlations of the RSMR or RSRR with the patient volume nor were their differences significant according to hospital type (all P> .05).

In the sensitivity analysis, events corresponding to 4192 patients (11%) from 2012 to 2014 were excluded due to their transfer to another hospital. As in the original models, neither a short hospital stay nor their respective ORs or ATTs obtained from matching were significant.

Although this retrospective analysis is based on administrative data, the validity of administrative registries is comparable to that of medical registries12,22 and the reliability of these studies allows public comparison of hospitals in terms of outcomes.30

In addition, in contrast to the CMS methodology, only readmissions due to cardiovascular disease were analyzed (and mortality during these readmission) due to a lack of events with a principal diagnosis other than that of cardiovascular disease (not included in the MBDS provided by the Spanish Ministry of Health, Consumer Affairs, and Social Welfare). However, cardiac events are probably a better indicator of performance related to length of stay than all-cause readmissions.31 Moreover, the difference in the mortality of patients with STEMI from cardiovascular diseases observed in other studies32,33 vs our findings is likely due to the absence of patient selection in those studies. For example, from our population with STEMI, we excluded patients with heart failure, acute pulmonary edema, ventricular arrhythmias, cardiac arrest, and cardiogenic shock. In addition, readmissions were not considered a recurrent outcome variable. However, the risk of adverse events was low or very low in our population, and a low burden of recurrences would be expected.

Another limitation lies in the inability of ICD-9-CM coding to accurately determine whether primary PCI was performed. However, according to the Spanish Cardiac Catheterization and Coronary Intervention Registry,34 86% of PCIs in STEMI are primary procedures and 5% are rescue PCIs (after fibrinolysis, which were excluded from our study population). Thus, 91% of the PCIs would be primary in our study.