Impact of post-sepsis cardiovascular complications on mortality in sepsis survivors: a population-based study

The present study used the National Health Insurance Research Database (NHIRD) of Taiwan. Taiwan’s National Health Insurance (NHI) program is a single-payer mandatory health insurance system that covers over 99% of the 23,000,000 people residing in Taiwan. The database used a systematic sampling strategy to select one million participants who were representative of the demographic and geographic region distribution of Taiwan in 2001. This sample, hereafter referred to as the one million longitudinal sample, was followed from 2001 to 2012 to form a closed cohort for research purposes. The database collects outpatient and inpatient electronic records on demographics, eligibility, vital status, diagnoses (International Classification of Diseases, ninth revision, Clinical Modification [ICD-9-CM]), operations, and prescriptions. All claims can be linked in chronological order to provide a temporal sequence of all health services utilization. Patient consent was not required for this study as this was an anonymized electronic database study. This study was approved by the Institutional Review Board of National Taiwan University Hospital. This observational study was performed in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for reporting observational studies [16].

The study was a population-based cohort study, consisting of all emergency department (ED) or hospital-treated sepsis patients between 2001 and 2011.

Sepsis patients were identified based on the coding system used by Angus et al., which is considered to most closely resemble the latest Sepsis-3 definition among the published algorithms [17]. Operationally, sepsis patients were identified using ICD-9-CM codes for the presence of either a bacterial or fungal infection (Additional file 4: Appendix 1) plus dysfunction of one or more organ systems (Additional file 4: Appendix 2) [1]. The ICD-9-CM codes for the identification of infections in this study were the same as those used by Angus et al. (including 1286 distinct infection codes). The ICD-9-CM codes for acute organ dysfunctions were supplemented with NHI procedure codes to increase specificity. In this manuscript, MI refers to myocardial infarction alone, stroke refers to cerebrovascular events alone, and MI/stroke stands for the combined outcome of MI and stroke.

To ascertain the independent association between sepsis and incident myocardial infarction (MI)/stroke after sepsis hospitalization, we constructed a non-sepsis comparison cohort using the risk set sampling and propensity score matching technique. The non-sepsis comparison cohort was constructed using a two-stage procedure. In the first stage, we used the risk set sampling method to sample 100 non-sepsis patients for each sepsis case, matching on admission date, 5-year age group, sex, and quartile Charlson Comorbidity Score (0, 1–2, 3–4 and ≥ 5) [18]. In the second stage, we created a propensity score (PS) consisting of a comprehensive set of covariates associated with sepsis (Additional file 1: Table S1). We performed 1:1 PS matching using a greedy algorithm to construct the final non-sepsis comparison cohort. The cohort construction process is shown in Fig. 1.

As the primary aim was to analyze the impact of post-sepsis stroke/MI on mid- or long-term survival, the primary “exposure” in this study was incident MI/stroke after sepsis hospitalization. Our previous study found that the critical period for MI/stroke is 70 days after discharge from a sepsis hospitalization. We therefore assessed for MI/stroke within the 70-day risk period as the primary exposure of this study. Patients with stroke were identified by the presence of an ICD-9-CM diagnosis code of 433.xx or 434.xx, which have a positive predictive value (PPV) of 0.96 and 1.00, respectively, in the NHIRD database [19]. Patients with incident acute MI were identified by any primary or secondary admission diagnosis containing ICD-9-CM code of 410.xx together with a prescription for antiplatelet therapy, such as aspirin, clopidogrel, dipyridamole, and ticlopidine. Newer classes of antiplatelets were not available in Taiwan during the study period. The above search algorithm yields a PPV of 0.84 for acute MI in the NHIRD database [20]. The primary outcomes of interest were 180-day and 365-day all-cause mortality.

For both the primary sepsis cohort and non-sepsis comparison cohort, we anchored the 71st day after hospital discharge as the index day of cohort entry. Patients were followed up for three outcomes: death, termination of health insurance coverage, and end of study (the 365th day from the index date), whichever came first. To control for unbalanced covariates between patients with/without post-sepsis cardiovascular events, we collected information on demographics, urbanization level, insurance premium level, chronic comorbidities, and risk factors for sepsis. In order to remove the effect of covariates that could develop after incident cardiovascular events, we collected all covariate information from the beginning of the study (year 2001) to the discharge date of the index hospitalization. The level of urbanization of the cities/towns was stratified into four levels based upon a composite score obtained by calculating population density (people/km²), population percent of people with an educational level of college or above (%), percent of people over 65 years (%), percent of agriculture workers (%), and the number of physicians per 100,000 people. To further control for general health conditions that are not reflected by ICD-9 CM codes [21], we further collected the frequency of healthcare facility utilization in the 1-year period before sepsis admission as a proxy indicator of general health. The timeline of the study design and periods of data collection can be seen in Fig. 2.

The analysis was implemented in two stages. In the primary analysis, we analyzed the impact of incident MI/stroke on the mid- or long-term outcome of sepsis survivors. In the secondary analysis, we compared the incidence and outcome of incident stoke/MI after hospital discharge between the sepsis and non-sepsis cohorts. In the primary analysis, we first compared the baseline characteristics between sepsis survivors who developed incident acute MI/stroke and those who did not develop incident acute MI/stroke. Categorical variables were presented as frequency and percentage and compared using chi-squared tests. Continuous variables were presented as mean or median and compared by non-parametric Mann-Whitney U tests. We then plotted the cumulative mortality curves of patients with/without post-sepsis MI/stroke. The differences in cumulative mortality between the groups were compared using the Wilcoxon signed-rank test. The survival impact of post-sepsis MI/stroke was assessed by multivariable Cox regression analyses. The proportional hazards assumption was checked by plotting curves of the log of the negative log of the survival function against log time for patients with/without post-sepsis MI/stroke. Potential confounders were determined based on the potential relevance given the literature, including the following variables: demographics, urbanization level, insurance premium, comorbidity, and healthcare service utilization. In the secondary analysis, we constructed a non-sepsis comparison cohort using the PS matching method mentioned above. To evaluate the success of the matching process, we calculated the standardized difference of matching covariates between the sepsis cohort and the non-sepsis comparison cohort. To evaluate the independent association between sepsis and post-sepsis MI/stroke, we calculated the incidence and 180-day mortality of MI/stroke among patients for the sepsis cohort and the PS-matched non-sepsis comparison cohort. We used univariate conditional logistic regression to evaluate the risk of incident MI/stroke and stratified Cox regression analysis to evaluate mortality risk. All statistical analyses were performed in SAS 9.4 (SAS Inc. Cary NC). In all analyses, a P value of < 0.05 was deemed significant.

From one million NHIRD participants, we identified 42,316 patients who were hospitalized with sepsis (Fig. 1). There were 591 events observed in the 70-day risk period after discharge, of which 486 (82.2%) were stroke, 108 (18.3%) were MI, and 3 (0.5%) were concomitant stroke and MI. The preadmission characteristics of sepsis patients with/without post-sepsis MI/stroke are summarized in Table 1. Compared with patients without MI/stroke, patients who developed MI/stroke were older, associated with lower socioeconomic status, had a higher burden of preadmission comorbidities, and had a higher frequency of healthcare service utilization.

The 180-day mortality for sepsis survivors without incident MI/stroke was 4.61% (1450/31,423). Patients who developed post-sepsis MI/stroke had twofold higher mortality compared to patients without post-sepsis MI/stroke. The mortality of MI/stroke-naïve sepsis survivors rose to 7.35% (2309/31,423) at 365 days. The risk of mortality remained twofold higher for patients who developed post-sepsis MI/stroke. The crude mortality results are summarized in Table 2. To evaluate the survival impact of post-sepsis MI/stroke 180 days following hospital discharge, we plotted the cumulative hazard curve in Fig. 3. We observed that patients who developed MI/stroke had a significantly higher cumulative mortality than patients without cardiovascular complications (Wilcoxon test P = 0.003) within the 180-day period. Adjusting for demographic variables and potential confounders (e.g., comorbidities) in the Cox proportional hazard model, we found a significant increase in 180-day mortality for patients with post-sepsis MI (hazard ratio [HR] 2.55, 95% confidence interval [CI] 1.53, 4.27), post-sepsis stroke (HR 2.19, 95% CI 1.67, 2.87), and composite post-sepsis MI/stroke (HR 2.02, 95% CI 1.65, 2.47) as compared with patients without post-sepsis MI/stroke (Table 2). At 365 days, the increased risk persisted for patients with post-sepsis MI (HR 2.10, 95% CI 1.34, 3.32), post-sepsis stroke (HR 2.02, 95% CI 1.61, 2.52), and composite post-sepsis MI/stroke (HR 2.02, 95% CI 1.66, 2.49). The full Cox model and effect estimates of the adjusted covariates are summarized in Table 3.

Of the 42,316 identified sepsis patients, 41,251 were matched to a non-sepsis cohort by PS matching. The standardized difference of matching covariates between sepsis and non-sepsis patients was less than 6%, indicating a successful match (Additional file 1: Table S1). We compared the incidence of post-sepsis MI/stroke between the sepsis and non-sepsis cohorts. For the sepsis cohort (n = 41,251), the incidence for MI, stroke, and composite MI/stroke, was 0.6%, 2.9%, and 4.2%, respectively. In contrast, for the non-sepsis cohort (n = 41,251), the incidence for MI, ischemic stroke, and composite MI/stroke, was 0.6%, 1.9%, and 2.7%, respectively. Compared to patients hospitalized without sepsis, patients hospitalized with sepsis had an increased risk for stroke (PS-matched OR 1.75, 95% CI 1.60–1.92) and an increased risk for composite MI/stroke (PS-matched OR 1.72, 95% CI 1.60–1.85), but not for MI alone (OR 1.01, 95% CI 0.82–1.25) (Additional file 2: Table S2). Finally, we compared mortality between the two groups of patients. For the sepsis cohort (n = 41,251), the 180-day mortality for MI, stroke, and composite MI/stroke was 13%, 9%, 10%, respectively. On the other hand, for the non-sepsis cohort (n = 41,251), the 180-day mortality for MI, stroke, and composite MI/stroke was 13%, 10%, and 11%, respectively. Cox regression showed sepsis was not associated with an increased hazard of mortality after MI, stroke, or MI/stroke (Additional file 3: Table S3).