The study population was identified from hospital discharge records (HDRs), and included all patients with primary or secondary ICD-9-CM codes of AMI (ICD-9 codes 410.xx) that were discharged between January 1, 2009 and November 30, 2009 from any hospital in Tuscany.
Data on hospital discharges in Tuscany region are routinely entered at hospital level into the HDR database. The hospital information system collects demographic information and ICD-9-CM diagnoses and procedures. These data are sent every 3 months for quality check to the Regional Health Information System Office that gives feedback to the hospitals on logical inconsistencies and missing data. Data are then corrected, completed and sent back the Regional Health Information System Office. As a results of this process, only a limited percentage of records are discarded, given that hospital reimbursement is based on these routine data.
The study was carried out in compliance with the Italian law on privacy (Art. 20–21, DL 196/2003) and the regulations of the Health Authorities of Tuscany Region on data management. Data were anonymized at the Regional Health Information System Office where each patient was assigned a unique identifier that is the same for all administrative databases. This identifier does not allow to trace the patient’s identity and other sensitive data. When anonymized administrative data are used to inform health care planning activities, the study is exempt from notification to the Ethics Committee provided written consent is obtained to use patient’s information stored in the hospital databases.
Records were excluded from the analysis on the basis of the following criteria:
Admissions preceded by a diagnosis of AMI in the preceding 8 weeks. This was done to exclude sequelae of a previous episode, according to the ICD-9-CM classification which defines "8 weeks" as the limit for a single "episode of care";
Admission lasting less than 2 days with discharge home;
Transfers from other hospitals;
Patients not resident in Tuscany;
Patients aged under 18 years or more than 100 years;
A diagnostic code 410.9 (myocardial infarction of unspecified site).
If patients were transferred, mortality was attributed to the hospital to which the patient was initially admitted
The outcome of interest was 30-day all-cause in-hospital mortality, defined as a death occurring for any reason within 30 days of the admission date during the index hospitalization or in any subsequent hospitalization. The follow-up was performed through a deterministic record linkage procedure of hospital discharge records from all Tuscan hospitals, using the unique patient identifier. The period of follow-up was determined as 30 days from the index admission. Our outcome differs from 30-day mortality used by other authors that also include deaths occurring out of the hospital, within 30 days of the index admission
[13, 14]. We could not analyse out-of-hospital deaths because they can be obtained from the mortality registry database available to researchers only after a time lag of several years.
Hospitals were categorised according to a presence or absence of the cardiac catheterisation laboratory, defined as “a laboratory operating in a hospital with in-house cardiovascular surgical support, in which both diagnostic and therapeutic procedures are performed on the heart and great vessels for a wide variety of cardiovascular diseases”
Information on the following variables, that might influence the outcome of interest and whose distribution may differ across hospitals, was retrieved from the HDRs: patient age, sex and comorbidities. The ICD-9-CM codes were used to define the presence or absence of specific comorbidities (hematologic diseases, previous AMI episodes, cerebrovascular diseases, vascular diseases, chronic nephropathies, tumours, diabetes, hypertensive disease, other forms of ischemic heart disease, conduction disorders and cardiac dysrhythmias, and chronic obstructive pulmonary disease) at the index admission and in the previous 2 years.
According to the ICD-9-CM classification that was adopted in Italy in 2007, the type of ST-segment elevation was determined as follows: patients with 410.1x–410.6x or 410.8x were defined as STEMI; patients with 410.7x were defined as NSTEMI. Codes 410.9 were excluded from the analyses
The validity of the ICD-9-CM coding in the Tuscany region compared with clinical data was examined in separate samples collected in the framework of AMI-Florence registry data
 and in the IN-ACS study
. Sensitivity, specificity and PPV were 90.4%, 71.9%, 69.5% in AMI-Florence (2009 data) and 98%, 90.4% and 96% in IN-ACS study, supporting the validity of the codes.
We determined the differences in patient characteristics according to the two hospital types (presence/absence of a cardiac catheterisation laboratory) using Pearson’s χ
test or the Student’s t-test as appropriate. Comparison of 30-day in-hospital mortality rates between patients with STEMI and NSTEMI was performed using Pearson’s χ
Because of the hierarchical structure of our data, with patients clustered into hospitals, we fitted random-effects (also known as multilevel or hierarchical) logistic regression models. The models included patient risk factors and random intercepts for each hospital, and allowed analysis of hospitals with a low case load
[10, 11, 19, 20]. We built several two-level models. The first model included only the random intercept to estimate inter-hospital variability in overall mortality rate. The second model contained relevant patient comorbidities (identified in a preliminary stepwise logistic regression model), age, gender and STEMI/NSTEMI categorisation. The significance level of entry and removal was set at 0.05. Age and gender were forced into the model. In the final model, we added the presence of a cardiac catheterisation laboratory in the hospital. In addition, we tested the interactions between outcome, AMI phenotype and gender, and then carried out a multilevel analysis on patients with STEMI and NSTEMI separately.
We estimated the odds ratios (ORs) with 95% confidence intervals (CIs) for patient and hospital-related characteristics and the random variance (σ
), which is a measure of inter-hospital variation. We used the area under the ROC curve to assess the discriminative ability of the model in predicting 30-day in-hospital mortality. This area (alternatively named c-index) varies from 0.5 to 1, with larger values denoting better model performance. We also reported other goodness of fit indexes, including pseudo R
, the Wald χ
test, Akaike information criterion (AIC) and Bayesian information criterion (BIC). Then, we calculated the adjusted hospital-specific mortality rates with 95% CIs to identify hospitals with mortality rates significantly different from the adjusted overall mortality rate. Hospital-specific adjusted mortality rates were calculated as the antilogit function of the random intercepts derived from the multilevel logistic regression model which included patient characteristics. The adjusted overall mortality rate was obtained by taking the antilogit function of the intercept estimated from the same model. We expressed these results in terms of RSMRs (risk-standardised mortality rates)
. RSMRs with 95% CIs were calculated as hospital-specific adjusted mortality rates over the adjusted overall mortality rate, multiplied by the unadjusted overall mortality rate.
Finally, we used funnel plots as an alternative method for inter-hospital comparison. Funnel plots are scatter plots of unadjusted 30-day in-hospital mortality rates against the number of patients discharged from each hospital
. We superimposed on the plot 95% (≈2 standard deviations) and 99.8% (≈3 standard deviations) control limits around the crude overall 30-day in-hospital mortality rate. We chose to use the crude overall rate as the target value in order to make the results from funnel plots comparable to the RSMRs from multilevel analyses.
Statistical analyses were performed using Stata software, version 12 (StataCorp LP, College Station, TX, USA).