Data source

We made a multicenter longitudinal observational study of a retrospective cohort of patients with information extracted from electronic medical records. The follow-up period was from January 1, 2006, to December 31, 2010. The study setting was the San Carlos Clinical Hospital of Madrid, which is a reference hospital for 22 basic PC areas located in the city of Madrid.

The study population was composed of subjects 24 years of age or older, with a single health insurance card, an open clinical record, and at least 1 medical visit to a PC center during 2006. Information was extracted from the PC health information system (OMI-AP). A case of HF was defined through the registration of a diagnosis of HF in the electronic medical record (codes K77 and K82 of the International Classification of PC 1).

The research protocol was conducted according to the Declaration of Helsinki guidelines, and the need to obtain written informed consent was waived due to the retrospective nature of the study. Patients’ records were anonymized and no identifiable personal data were available for the analysis. Research Ethics Committee at the San Carlos Clinical Hospital revised and approved the study protocol.

Variables

The primary outcome was mortality after the 5^th year of follow-up. Sociodemographic variables used in this study included age, sex, economic activity (active/pensioner), and the level of income (assigning each patient the income per capita of their health center). As clinical variables, the presence of cardiovascular risk factors and associated comorbidities (valvular heart disease, hypertension, ischemic heart disease, stroke, arrhythmias, diabetes, obesity or overweight, lipid metabolism disorder, smoking, and alcoholism), influenza vaccine administration in each of the years 2006-2010 (never vaccinated, partially vaccinated in some but not all years of the follow-up and fully vaccinated), and the prescribed pharmacological treatment (insulin, oral hypoglycaemic agents, antithrombotic agents, beta-blockers, calcium channel blockers, ACE inhibitors, angiotensin-receptor blockers (ARBs), nitrates, lipid reducers, and diuretics). As variables for the use of services, the complementary tests requested in PC have been incorporated, such as the number of blood tests (complete blood count and serum biochemistry), X-rays, electrocardiograms, as well as the number of referrals requested to cardiology (these tests were requested but there is no information on whether the patient went to perform them).

Statistical analysis

To evaluate the effect of hospitalizations an analysis based on propensity score matching (PSM) was conducted. PSM can be used to adjust for selection bias when assessing causal effects in observational studies as it permits to achieve balance between two groups, based on the conditional probability of receiving an event of interest (in this case, hospitalization due to worsening HF) given a series of measured covariates. The PSM was calculated through a binary logistic regression model based on confounding factors at baseline (Table 1), using the nearest neighbor algorithm without replacement, which consists of making 1: 1 pairings and the covariates selected as predictors. For the selection of covariates, a bivariate analysis was performed, including in the model that presents a statistically significant relationship with the dependent variable (presence or absence of hospitalization) or others that were of interest for the study objectives (according to the 2016 European Society of Cardiology [ESC] guidelines⁶).

A global imbalance χ² test was performed, which simultaneously evaluated whether any variable or any linear combination of the variables is unbalanced means after the pairing.

We also introduced a simple multivariate imbalance measure (Iacus, King and Porro, 2011)^7,8. This measure presents a perfect global balance if L1 = 0, and the larger values represent a greater imbalance between the groups, with a maximum of L1 = 1. Therefore, we had a good method if L1^matching ≤ L1.

The standardized mean differences were plotted, which quantify the bias in the means (or proportions) of covariates between the groups and should be close to zero after matching.

A Kaplan–Meier curve analysis was applied to determine survival at 5 years for variable hospitalizations. Differences between survival curves were assessed using the log-rank test.

Survival was estimated by taking the starting date as the date of registration of the disease in the medical history and the end date as the end of follow-up (censorship) or the date of the death of the patient. To study the mortality, death by any recorded cause in the medical record was included, and the date of death was considered as the date when the patient was removed from the health insurance card system.

The Cox proportional hazards model was used to estimate the relationship between the presence or absence of hospitalizations and mortality, adjusted for PSM and potential confounding variables. Hazard ratio and its 95% confidence interval were calculated for predictive variables. The validation of the model was carried out through the basic assumptions to verify the validation of the model: proportionality and log-linear.

For the statistical treatment and graphic representation of the data, the statistical package SPSS, v. 22.0 and Stata, v. 14.0 were used.

PSM

The analysis is focused on a cohort of 529 couples, with a total of 1058 patients between hospitalized and non-hospitalized (Fig. 1). After matching, patients with and without HF hospitalization were balanced in all covariates (Tables 1 and 2). Table 1 shows sociodemographic characteristics, cardiovascular risk factors and associated comorbidity and use of services before and after PSM. After matching, 84 (15.9%) patients died who were not hospitalized while 116 (21.9) hospitalized patients died. Table 2 displays pharmacological treatment.

Figure 1 Flow chart for the assembly of a matched cohort. 

The p-value of the χ² balance test of Hansen and Bowers was 0.998, showing good covariate balance after matching. The results of measure L1, before matching (0.999) and after matching (0.994), confirmed the balance achieved. The histograms in figure 2a and b represent the standardized mean differences of all covariates, quadratic terms, and interactions together with the estimated (overlapping) kernel density function, before and after the PSM, where it was observed that these differences were centered at zero after matching.

Figure 2 (A and B) Histograms with overlaid kernel density estimates of standardized differences before and after matching. 

Survival analysis

The survival of patients in the subpopulation after PSM was reduced from the 2^nd year of follow-up compared to the cohort before PSM, with survival of 95, 84, 76, 67, and 59%, respectively, at 1, 2, 3, 4, and 5 years, with an average survival of 47.6 months (46.2-48.9), which decreased 1.4 points after applying the PSM technique¹. Mean survival was 49.7 months (47.9-51.4) for the non-hospitalized and 45.6 months (43.5-47.6) for hospitalized patients. Figure 3 shows statistically significant differences between the curves (Log-rank test, p = 0.002). Therefore, the hospitalized group showed less survival rates.

Figure 3 Kaplan–Meier estimated survival curves. 

Prognostic factors for mortality

The crude effect of being hospitalized was heart rate (HR) = 1.56 (1.17-2.06). After adjustment for PSM and confounding factors, the effect of being hospitalized decreased (HR = 1.49, 1.12-1.98) (Table 3).

Having prescribed nitrates were associated with a higher risk of mortality after PSM (HR = 1.56, 1.08-2.24) (Table 3). Factors with a protective effect were having received influenza vaccination every year (HR = 0.04, 0.01-0.15) and X-rays having been requested by PC physicians (HR = 0.76, 0.67- 0.88) (Table 3).