Introduction

Sepsis is a life-threatening illness caused by dysregulated host immune response due to infection [1]. It occurs in 6% of all hospitalized adults in the United States [2], and its incidence is continuously increasing [3]. Further, a 2011 retrospective in-patient study in the United States reported an all-cause mortality rate of sepsis of 14.8% [4]. With respect to economic burden, sepsis has also been considered a significant concern [5], with the median hospital cost of sepsis treatment per patient estimated at approximately $16,000 [4]. Therefore, efforts for preventing sepsis and lowering mortality in sepsis patients have been emphasized.

Metformin is a biguanide drug that is most commonly prescribed for the management of type 2 diabetes mellitus [6]. In addition to its well-known hypoglycemic effect, some previous studies have suggested that metformin might inhibit the expression of pro-inflammatory factors in vitro and ameliorate inflammatory injuries in vivo [7,8,9,10]. Although the mechanism is unclear and controversial, metformin has been reported to inhibit the activity of the mitochondrial respiratory complex, which decreases the generation of ATP and activates adenosine 5′-monophosphate (AMP)-activated protein kinase (AMPK) [11]. AMPK activation have been confirmed to have a protective effect for sepsis in animal studies [12,13,14], indicating that AMPK plays a significant role in the pathogenesis of sepsis. Moreover, a recent meta-analysis of five cohort studies concluded that preadmission metformin use is associated with lower mortality, as evidenced by an odds ratio (OR) of 0.59 (95% confidence interval (CI) 0.43–0.79) [15]. However, the five cohort studies had a relatively small sample size, of which four were retrospective observational studies from a single institution [16,17,18,19], and only one was a prospective observational study [20]. Additionally, the meta-analysis focused only on mortality in sepsis patients and did not investigate the protective effect of metformin against the development of sepsis.

Therefore, this study aimed to investigate whether metformin helped prevent sepsis, using a sample cohort from South Korea. Additionally, we aimed to investigate the association between prior metformin therapy and 30-day mortality in sepsis patients.

Methods

Ethical statement

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee (X-1904-532-901) and the Health Insurance Review and Assessment Service (NHIS-2019-2-142).

Data source

The sample cohort database of the National Health Insurance Service (NHIS) was developed to provide data for academic research of the general population in South Korea. The database comprises a stratified random sample of one million people who were registered with the NHIS since 2002. It was designed to be representative of the national population with regard to demographic and socioeconomic information. The cohort was dynamic and followed up until the end of 2015. It was supplemented with additional cohort data, including data on infants, to allow for attrition due to death and loss to follow-up. Each year subjects were added to the cohort to replace people who had died or emigrated in the previous year, using stratified extraction methods to ensure that the cohort was representative of the national population [21].

Population

We evaluated adult diabetes patients (age, ≥ 18 years) registered in the 2010 cohort database of the NHIS. Diabetes was identified according to the International Classification of Disease (ICD)-10 diagnostic system (E10–E14). Individuals who died in 2010 and those who emigrated between 2011 and 2015 were excluded, as they could not be followed up.

Prior metformin therapy as exposure variable

The diabetes patients were divided into the metformin user group, defined as those who had been prescribed continuous oral metformin over a period of ≥ 90 days, and the control group, defined as all other subjects. The classification of metformin use was based on the metformin prescription data from October 2009 to December 2010. This was done to remove immediate exposure to metformin (< 90 days) before outcome evaluation (i.e., development of sepsis) from January 2011 as lag time approach in this study [22]. In addition, the metformin users were divided into two groups according to daily dosage, namely the high daily dosage group (> 1 g/day) and the low daily dosage group (≤ 1 g/day).

Study endpoints

The primary endpoint of this study was development of sepsis between 2011 and 2015, which was defined as registered sepsis (including septic shock: A40*, A41*, R65.2*, R572, and B377) during this period. The secondary endpoint was 30-day mortality among diabetes patients who developed sepsis between 2011 and 2015. The 30-day mortality was defined as any mortality within 30 days after the diagnosis of sepsis.

Confounders

The following variables were considered as confounders in this study: (1) demographic data (age and sex); (2) socioeconomic data (income level in deciles and place of residence in 2010 [Seoul, metropolitan cities, and other]); (3) the 2010 Charlson comorbidity index, which was calculated using registered ICD-10 diagnostic codes between 2009 and 2010 in the NHIS database (e-Appendix 1); (4) total number of hospital visit days in 2010; and (5) use of other anti-diabetic medications (sulfonylureas, alpha-glucosidase inhibitors, thiazolidinediones, and insulin). In the classification of place of residence, Seoul, the capital city, was assigned a separate category, and the cities of Incheon, Kwangju, Busan, Ulsan, Daegu, and Daejeon were classified as metropolitan cities. The number of hospital visit days included the number of hospital outpatient clinic visits and days spent on admission, but it did not include outpatient visits to primary care physicians. For example, an individual who visited a hospital outpatient clinic five times and was admitted to the hospital for 3 days would be considered to have had 8 hospital visit days. In the analysis, the number of hospital visit days was categorized into five groups (0, 1–7, 8–29, 30–90, and > 90 days).

Statistical analysis

The baseline characteristics were presented as means with standard deviations for continuous variables or frequencies with percentages for categorical variables. We performed propensity score (PS) matching to reduce confounders in observational studies [23] using the nearest neighbor method with a 1:1 ratio, without replacement, and a caliper width of 0.1. Logistic regression analysis was performed to calculate PSs as a logistic model, and all covariates were included in the PS model. The absolute value of standardized mean difference (ASD) was used to determine the balance between the metformin user group and the control group before and after PS matching. ASDs between the two groups were set to below 0.1 to determine whether the two groups were well balanced through PS matching. After confirming good balance between the two groups through PS matching, we performed a univariable logistic regression analysis for the development of sepsis between 2011 and 2015 in the PS-matched cohort. As a first sensitivity analysis, because observational studies often suffer from immortal bias because of the time period from the treatment to event (i.e., development of sepsis in this study) [24], we used the landmark method to determine the development of sepsis in the PS-matched cohort [25]. In the landmark method, we performed four logistic regression analysis for the development of sepsis in four different periods (i.e., 2011–2015, 2012–2015, 2013–2015, and 2014–2015), considering that metformin users or the control patients might die due to another disease (e.g., heart failure or trauma) before the development of sepsis.

Next, we performed multivariable logistic regression analysis for the development of sepsis between 2011 and 2015 in the entire 2010 sample cohort to (1) determine whether the results obtained from the PS-matched cohort were generalizable to the entire sample cohort in 2010 and (2) investigate the risk for development of sepsis among metformin users with other important covariates in context, not isolated. All covariates were included in the multivariable model for adjustment, and there was no multicollinearity in all multivariable models of the entire cohort with a variance inflation factor of < 2.0. Hosmer–Lemeshow statistics was used to confirm goodness of fit of multivariable models as P > 0.05. Additionally, we performed two subgroup analyses in the multivariable logistic regression analysis: (1) subgroup analysis according to the daily dose of metformin was performed to investigate whether the daily dosage of metformin influences the development of sepsis, (2) subgroup analysis according to severity of diabetes mellitus, because it might affect the use of anti-diabetic medication. On the basis of ICD-10 codes in 2010, the total cohort was divided into two groups, namely, diabetes mellitus with chronic complication and diabetes mellitus without chronic complication, as shown in e-Appendix 1.

For the secondary endpoint (30-day mortality) among patients who were diagnosed with sepsis during 2011–2015, we performed both PS matching and multivariable logistic regression analysis again using the same method. The results of the logistic regression models were presented as odds ratio (OR) with 95% confidence interval (CI). All statistical analyses were performed using R version 3.6.2 (R Foundation for Statistical Computing, Vienna, Austria), and P < 0.05 was considered statistically significant.

Results

In total, 78,876 adult diabetes patients were initially screened. Of these, 52 patients who emigrated between 2011 and 2015 and 1487 patients who died in 2010 were excluded from analysis. Thus, 77,337 patients (34,041 patients in the metformin user group and 43,296 patients in the control group) were included in the analysis. After the first PS matching, 41,608 patients (20,804 patients in each group) were included in the final analysis. Of the 77,337 diabetes patients, 2512 patients (3.2%; 1,045 patients in the metformin user group and 1,467 patients in the control group) were diagnosed with sepsis between 2011 and 2015. After the second PS matching, 1344 patients (672 patients in each group) were included in the secondary analysis. The patient selection flow chart is presented in Fig. 1. The results of the comparison of characteristics between the metformin user group and the control group in the entire cohort and in patients who developed sepsis are presented in Tables 1 and 2, respectively. All ASDs were below 0.1, indicating that all covariates between the two groups were well balanced through PS matching.

Fig. 1
figure 1

Patient selection flow chart. PSM, propensity score matching

Full size image
Table 1 Comparison of characteristics between metformin users and control group among diabetic patients before and after PS matching
Full size table
Table 2 Comparison of characteristics between metformin users and control group who were diagnosed as sepsis among diabetic patients before and after PS matching
Full size table

Development of sepsis according to prior metformin therapy

Table 3 shows the results of analysis of development of sepsis between 2011 and 2015 in diabetes patients before and after PS matching. After PS matching, the incidence rate of sepsis in the metformin user group and the control group was 2.9% (598/20,804) and 3.1% (649/20,804), respectively. In the logistic regression analysis, metformin use was not significantly associated with the risk of developing sepsis (OR: 0.92, 95% CI 0.82–1.03; P = 0.143). In sensitivity analysis according to daily metformin dosage, both low daily dosage (P = 0.775) and high daily dosage (P = 0.437) were not significantly associated with the risk of sepsis. The sensitivity analyses in both the multivariable logistic regression model in the entire cohort showed a similar tendency (Table 5) and in the PS-matched cohort according to survival time showed a similar tendency (e-Appendix 2).

Table 3 Diagnosis of sepsis during 2011 to 2015 in diabetic patients before and after PS matching
Full size table

Thirty-day mortality after diagnosis of sepsis in diabetes patients

Table 4 shows the results of analysis with respect to the 30-day mortality after diagnosis of sepsis before and after PS matching. After PS matching, the 30-day mortality rate in the metformin user group was 35.9% (241/672), while it was 37.4% (251/672) in the control group. In the logistic regression analysis, metformin use was not significantly associated with the risk of 30-day mortality (OR: 0.94, 95% CI 0.75–1.17; P = 0.571). The sensitivity analysis of the multivariable logistic regression model in the entire cohort showed a similar tendency (Table 5).

Table 4 Thirty day mortality after diagnosis of sepsis in diabetic patients before and after PS matching
Full size table
Table 5 Multivariable logistic regression analysis in entire cohort
Full size table

Subgroup analysis

Table 5 also shows the results of subgroup analyses according to daily dosage of metformin, and severity of diabetes mellitus. Both low and high daily dosage of metformin user group were not associated with development of sepsis, compared to control group (P = 0.084 and P = 0.428, respectively). Metformin user group was not associated with development of sepsis compared to control group in both groups: diabetes without and diabetes with chronic complication (P = 0.966 and P = 0.099, respectively). A similar trend was observed on subgroup analyses with regard to the secondary endpoint (30-day mortality) among sepsis patients.

Discussion

This population-based cohort study showed that prior metformin therapy in diabetes patients was not significantly associated with the development of sepsis. Additionally, prior metformin therapy was not associated with 30-day mortality in diabetes patients diagnosed with sepsis. This association was coincident regardless of the daily dosage of metformin and severity of diabetes mellitus.

Importantly, while most studies focus on mortality after sepsis diagnosis [16,17,18,19,20], this study focused on the incidence of sepsis among diabetes patients who use metformin in a nationwide sample cohort. Metformin is known to be associated with AMPK activation [11], which is known to have a protective effect against sepsis-induced organ injury [12]. Moreover, a recent study reported that disruption of the AMPK-dependent immunometabolism pathway might contribute to the development of sepsis [26]. As such, metformin therapy might have a potential benefit for preventing sepsis, but this has not been investigated previously. To our knowledge, our study is the first to show that metformin might have no significant preventive effect against sepsis in diabetes patients. In addition, we also found that the daily dose of metformin was not associated with the risk of sepsis.

Because of the observational study design, an important issue in this study was controlling for bias. First, immortal time bias, which is a well-known bias in observational studies [24], might be present in our study. For example, the risk of sepsis in the control group might be higher if patients in the metformin user group died in 2011 due to any other disease (e.g., cardiovascular disease) before the development of sepsis and the control group survived beyond 2011. To minimize this bias, we performed landmark analysis for the development of sepsis in the PS-matched cohort [25] and also included many confounders that might be closely related to death due to other diseases before sepsis development, including comorbidities and socioeconomic factors. Second, there might be a time-lag bias, and this might also affect the results of this observational study [27]. For instance, the results might be biased because of inadequate duration of exposure for the event to occur if a patient received a prescription for metformin from December 2010 within 30 days. To control for this bias, we defined metformin users using a lag-time approach as those who received prescriptions for metformin over 90 days until December 2010 [22]. Through this approach we excluded immediate exposure to metformin among diabetes patients before evaluating the development of sepsis from January 2011. In addition, we performed both PS adjustment and multivariable adjustment in the entire cohort to enhance the robustness of our findings.

Furthermore, aside from the main findings, we also found no significant association between 30-day mortality and prior metformin therapy in diabetes patients who developed sepsis. The association between prior metformin therapy and 30-day mortality is controversial. While some studies showed no significant association between prior metformin therapy and mortality in sepsis patients [18,19,20], others reported that prior metformin therapy was significantly associated with lower mortality in sepsis patients [16, 17]. Although a recent meta-analysis reported that prior metformin therapy was associated with lower mortality in sepsis patients [15], it involved a relatively small sample size, thus possibly limiting the robustness of its findings.

Our study has several limitations. First, some important variables, including body mass index, were not included in both PS adjustment and multivariable adjustment because they were not included in the NHIS data set. Second, we defined the comorbidities using ICD-10 codes to calculate Charlson comorbidity index, but the diseases specified by the ICD-10 codes might differ from the actual underlying diseases in the diabetes patients. Third, PS adjustment and multivariable adjustment could control only for known confounders; thus, there might be residual confounders that could have influenced the study findings. Fourth, our analysis was based on metformin prescription data in the NHIS database; it did not assess compliance among those classified as metformin users. Fifth, it is possible that patients in the control group may have been prescribed metformin, while patients in the metformin group might have stopped the treatment since 2011. Lastly, regarding analysis for 30-day mortality in sepsis patients, we did not adjust for the severity of sepsis (such as acute physiology and chronic health evaluation II score or simplified acute physiology score II) in this study. Considering these limitations, the results of this study should be interpreted cautiously, and further prospective, large population-based cohort studies are needed to confirm these findings.

In conclusion, prior metformin therapy had no significant association with sepsis risk and 30-day mortality after sepsis diagnosis. This association was coincident regardless of daily dosage of metformin, and severity of diabetes mellitus. Our study suggests that the association of prior metformin therapy with both risk and outcome of sepsis remains controversial, and future studies on this topic are needed.