Impact of obesity on short- and long-term mortality in patients with sepsis: A retrospective analysis of the large clinical database MIMIC-III

Background The purpose of our study was to explore the relationship between body weight and short-and long-term clinical outcomes in patients with sepsis. Methods We retrospectively analyzed 11,499 patients with sepsis at the Beth Israel Deaconess Medical Center (Boston, MA, USA) registered in the Medical Information Market Intensive Care (MIMIC-III) database from 2001 to 2012. Cox proportional hazards regression assessed the relationships between body mass index and 30-day and 1-year mortality. Results Patients were divided into four groups according to body mass index (underweight: 336 [6.0%]; normal weight: 1,752 [31.4%]; overweight: 1,563 [28.1%]; and obese: 1,920 [34.5%]), 30-day mortality (42.3%, 36.6%, 32.2%, and 29.6%; p<0.001), 1-year mortality, (64.6%, 56.8%, 52.5%, and 46.7%; p<0.001), and in-hospital mortality (35.4%, 34.3%, 31.6%, and 29.9%; p=0.018). In addition, obese patients had notably longer mechanical ventilation periods and intensive care unit and hospital lengths of stay. The Cox proportional hazards regression analysis conrmed that underweight patients had a 13% and 24% increased risk of death within 30 days and 1 year, respectively, compared with normal-weight patients. For overweight patients, these risks were 17% and 14% lower, respectively, than those reported for normal-weight patients. For obese patients, these risks were 22% and 21% lower than those observed in normal-weight patients. Conclusion This retrospective analysis showed that overweight or obese patients showed improved survival within 30 days and 1 year after admission to the intensive care unit.

Thus far, the relationship between obesity and survival in patients with sepsis has been controversial. Our goal was to gain insight into the relationship between obesity, and short-and long-term outcomes in patients with sepsis based on large-sample data. We aimed to determine whether obesity is an independent risk factor for sepsis outcome, provide an indicator for the prediction of the risk of sepsisrelated mortality and identify modi able targets for the reduction of this risk.

Database
We conducted a large-scale, single-center, retrospective cohort study using data collected from the  [12]. MIMIC-III was included in the ICU admission at Beth Israel Deaconess Medical Center from June 1, 2001 to October 31, 2012, containing > 46,000 independent electronic medical records from the ICU. The database is freely accessible, and any researcher who accepts the data usage agreement and completes the "protecting human subjects" training can apply for access to the data [13]. Patient informed consent or ethical approval was not required as identi cation details had been removed from all data. One author (DW) was allowed to access the database (certi cation number: 27714078) and was responsible for data extraction.

Participants
All patients in the database were selected. The inclusion criteria were: (1) identi cation of sepsis in the MIMIC-III database based on the International Classi cation of Diseases, 9th revision (ICD-9) code; (2) adults (≥ 18 years old) admitted to the ICU; and (3) complete medical records, including records of weight and virtual identi ers, which can be linked to their clinical data. For patients admitted to the ICU multiple times, we only included information recorded during the rst admission.

Data extraction and management
Considering that it was more likely we would obtain patient height and weight indicators for assessing their obesity status in the clinic, we calculated the BMI (weight in kilograms divided by height in meters squared [kg/m 2 ]). We categorized the BMI as follows: underweight (< 18.5 kg/m 2 ); normal weight (18.5-<25 kg/m 2 ); overweight (25.0-<30 kg/m 2 ); and obese ( > = 30 kg/m 2 ) [14]. We recorded the weight and height of patients on the rst day of admission to the ICU.
We considered factors that may confound the relationship between obesity and sepsis. Other variables extracted from the MIMIC-III database included demographic characteristics (i.e., age, sex, ethnicity, marital status, insurance, admission type, ICU rst service), Elixhauser comorbidity conditions [15] and severity scores. Severity scores included the Acute Physiology Score III (APS III) [16], the Simpli ed APS II (SAPS II) [17], the Sequential Organ Failure Assessment (SOFA) [18,19], Additionally, data regarding the use of vasopressors (e.g., dopamine, epinephrine, and norepinephrine), mechanical ventilation, renal replacement therapy, and length of ICU stay and hospitalization were extracted from the database.
The primary outcome was mortality 30 days and 1 year after admission to the ICU. The secondary outcomes were in-hospital mortality, and length of stay in the ICU and the hospital. For patients who expired outside the hospital, the Social Security Death Index was associated with the database for investigations related to mortality.
All scripts used for demographic characteristics, severity score calculation, and comorbidity were obtained from the github website (https://github.com/MIT-LCP/mimic-code/tree/master/concepts, date of access: May 2018). Data extraction was performed using structured query language (SQL) in PostgreSQL tools (v9.6; PostgreSQL Global Development Group).

Statistical analysis
The Kolmogorov-Smirnov test was used to test the normality of continuous variables. The distribution data were expressed as mean ± standard deviation. The non-normal continuous variables were represented by the median (interquartile range), while the categorical variables were represented by numbers (%). Patients were separated into four groups based on the BMI (underweight, normal, overweight, and obese).
Quanti able data were comparable between the four groups. Continuous variables were compared using the non-parametric Kruskal-Wallis H test, while categorical variables were compared using the χ 2 test. Kaplan-Meier survival curves were produced according to the BMI classi cation to show the probability of survival after 30 days and 1 year, and compared using the log-rank test. Clinical data were compared between survivors and non-survivors following 30 days and 1 year. Continuous variables were compared using the non-parametric Mann-Whitney U test, while categorical variables were compared using the χ 2 test. Cox proportional hazards regression analysis was performed to assess the factors associated with 30-day and 1-year mortality. The variables signi cantly associated with 30-day or 1-year mortality in the univariate analysis were employed in the Cox proportional hazards regression analysis. Variables satisfying the proportional hazards assumption were integrated into the Cox proportional risk regression model to determine the factors affecting the 30-day and 1-year survival rates.
We conducted sensitivity analyses to examine whether these missing data impacted our results. First, we imputed BMIs for the 1,120 patients without heights data whom we had excluded from the primary study population and rerunning the multivariate regression model to check whether our method of estimating height to handle missing height records distorted the conclusion. Second, we strati ed the sample by potential confounders such as age, gender, and ICU types to assess whether differences by these characteristics were observed. Finally, we performed a subgroup analysis of severely obese patients as a separate group, de ned as BMI ≥ 40 kg/m2, And re-run the nal model to see if morbidly obese (high mortality) patients are biased towards results and whether morbidly obese has reduced mortality in patients with sepsis compared to normal patients.

Results
Demographic and clinical characteristics Of the 5,907 patients registered in the MIMIC-III database, 5,571 met all inclusion criteria (Fig. 1). Table 1 summarizes the demographic data for each BMI category. According to the BMI, the patients were classi ed into the following four groups: underweight (336 patients, 6.0%); normal weight (1,752 patients, 31.4%); overweight (1,563 patients, 28.1%); and obese (1,920 patients, 34.5%). Obese patients were younger and more likely to be married compared with those in the normal-weight group (p < 0.001). There were signi cant differences between the BMI categories in terms of race, insurance category, admission type, and type of the ICU ward for the rst admission (p < 0.001). Without adjusting for any clinical covariates, obese patients had a higher incidence of chronic health conditions (including diabetes, hypertension, and coronary heart disease) than normal-weight patients (p < 0.001). However, the incidence of acquired immunode ciency syndrome (AIDS) was lower in the obese patient population (p < 0.001). The severity scores for the BMI categories were very similar. In addition, obese participants had worse SAPSII scores than normal-weight participants, which means they were lighter and had a better prognosis (P < 0.001). As expected, overweight and obese patients were more likely to receive mechanical ventilation and be treated with vasoactive drugs. In overweight and obese patients, the original hospital mortality, 30-day, and 1-year mortality rates were signi cantly lower than those reported in normal-weight patients (p < 0.001). However, the patients with higher BMI had a longer hospital stay and intensive care unit (ICU) stay than did the underweight patients.

Univariate analysis of interventions in the ICU
We investigated the need for mechanical ventilation, dialysis, and vasoactive drugs in patients with sepsis according to the BMI during their stay in the ICU. The results are shown in Table 1. While in the intensive care unit, a wide range of vasoactive drugs are used during the hospitalization of obese ICU patients. In addition, the duration of mechanical ventilation and dialysis in patients with sepsis increased in parallel with the BMI. There was difference between the four groups in the need for mechanical ventilation, dialysis or dopamine, or the number of patients who received treatment with norepinephrine.

Construction of a Kaplan-Meier survival curve
The probability for 30-day and 1-year survival, according to the BMI category, was compared using a shared log-rank test. The analysis indicated that a higher BMI was linked to better prognosis (p < 0.001) (Figs. 2).

Cox proportional hazards analyses of 30-day mortality
A multivariate cox regression model was constructed as follows: variables with a P value < 0.1 or clinically signi cant variables were included in the model in the univariate test (Table S1). In addition, the variables satisfying the proportional hazard assumption were integrated into the Cox proportional hazards regression model to determine the factors affecting the 30-day survival rate. Finally, age, ethnicity, marital status, type of admission, ICU rst services, BMI, severity score and clinical interventions were incorporated in the Cox proportional hazards regression model (  The results, as shown in Table 2, indicate that chronic pulmonary, liver disease, kidney failure, and metastatic cancer, higher SAPS-II and APSIII scores, more aggressive use of vasoactive drugs (i.e., epinephrine, norepinephrine, dopamine), and organ intervention support therapy (i.e., ventilation, dialysis) were associated with an increased risk of death within 30 days (p < 0.05 for each). The striking observation to emerge from the data comparison was that hypertension reduced the risk of death by 16% (HR: 0.84, 95% CI: 0.76-0.92; p < 0.001), this risk in diabetes patients was 19% lower than those who without underlying diseases (HR: 0.81, 95% CI: 0.68-0.97; p = 0.02).

Cox proportional hazards analyses of 1-year mortality
Multivariate proportional hazards regression models were established by adjusting for the simultaneous impact of potential confounders which were associated with survival rates in the univariate analyses (Table S1). Finally, age, sex, ethnicity, type of admission, ICU rst services, BMI, and clinical interventions were included in the Cox proportional hazards regression model ( In addition, advanced age, more emergency admission type, higher SAPS-II and APSIII scores, more aggressive use of vasoactive drugs (i.e., epinephrine, norepinephrine, dopamine), and organ intervention support therapy (i.e., ventilation, dialysis) can increased risk of death within 1 year (p < 0.001). As compared with patients who initially treated in the CCU, those who entered the surgical intensive care unit for the rst time had a 26% lower mortality risk (HR: 0.74, 95% CI: 0.63-0.87; p < 0.001). This risk in cardiac surgery recovery unit patients was 30% lower than that recorded in CCU patients (HR: 0.70, 95% CI: 0.54-0.90; p = 0.006). Given that the latter is more likely to accept critically ill patients who come from elective surgery. Most of the patients were young, not suffering from chronic underlying diseases. While those in the CCU are generally admitted as an emergency with more critical diagnoses.
As expected, high-mortality conditions, including congestive heart failure, liver disease, kidney failure, lymphoma, solid tumors, and metastatic cancer, were associated with an increased risk of death within 1 year (p < 0.05 for each). Hypertension was associated with lower risk of death (p < 0.001). Interestingly, chronic pulmonary patients had higher risk at 30 days (HR: 1.16, 95% CI: 1.04-1.31; p = 0.01), but withdrew from the nal regression model of the 1-year mortality risk.

Sensitivity analyses
Furthermore, we conducted multiple sensitivity analyses, which did not appreciably change our results.
When we removed the patients with imputed heights, we had 4451 patients, the results of the nal multivariable regression model remained consistent with the primary analysis in which those with imputed heights were included. Forest plots presents the results obtained from the Sensitivity analysis of age, gender, ICU type (Fig. 3).The sensitivity analysis was performed to explore potential sources of heterogeneity. We found that at the age of (65-80), the 30-day mortality rate of overweight and obese people decreased by 32% and 28%, respectively. The 30-day mortality rate of overweight female patients deducted by 36% (p < 0.01 for each). Overweight and obese people bene t signi cantly from 30-day survival. What emerges from the results reported here is that the subgroup of morbidly obese patients, with BMIs ≥ 40 kg/m2, also had a considerable survival advantage. The 30-day and 1-year mortality rates in the morbidly obese cohort decreased by 15.0% and 26.0%, respectively, which were lower than that recorded in normal-weight patients. Subgroup results for morbidly obese people were largely con rmed by the nal regression model.

Discussion
Sepsis is a life-threatening organ dysfunction caused by an infection. The incidence of sepsis is increasing, these sepsis survivors may suffer from additional complications, such as higher risk of readmission, cardiovascular disease, cognitive impairment, and death in the subsequent years. Study showed that in the rst year after the onset of sepsis, approximately 60% of sepsis survivors had at least one rehospitalization episode, and one in six sepsis survivors expired [21]. Some retrospective and prospective studies have focused only on short-term sepsis outcomes in patients, showing that nearly 60% of deaths attributed to sepsis occur within 30 days. Hence, it is necessary to analyze the short-and long-term prognosis in patients with sepsis. Our study nally adjusted for a number of potential confounding factors (i.e., demographic factors, comorbidities, underlying diseases, disease severity, and clinical interventions) and found that the BMI and mortality in patients with sepsis are independently related. Short-and long-term mortality in overweight and obese patients is lower than that observed in normal-weight BMI patients. In contrast, a lower BMI predicts a relatively high risk of death.
The results of current studies on the impact of obesity on the prognosis of patients with sepsis are controversial [22]. Obesity is a high-risk factor for death in the general population [23]. The BMI remains a useful indicator of overall health because it is highly correlated with the body surface area, which is often used as a proxy for the classi cation of obesity. Lee et al. reported that underweight is associated with patient mortality [24]. However, through multivariate analysis, the BMI was not identi ed as an independent factor for clinical outcomes. Instead, our study revealed that the BMI is an independent predictor of prognosis in patients with sepsis. In > 1,000 nationally representative large-sample hospitals in the USA, research has found that obesity is signi cantly associated with a 16% reduction in the risk of death for hospitalized sepsis patients [25]. The meta-analysis concluded that overweight or obese individuals had lower adjusted mortality rates when entering the ICU due to sepsis or septic shock [26]. In addition, overweight and obese patients are more likely to have comorbidities, including chronic heart failure, chronic obstructive pulmonary disease, and diabetes, and tend to receive more aggressive clinical interventions. Previous studies reported that obese patients have prolonged ventilation and prolonged length of stay in the ICU [27,28]. Respiratory failure in patients with sepsis often manifests as acute respiratory distress syndrome. Previous studies have shown that patients with high BMI are more likely to develop acute respiratory distress syndrome and stay longer in hospital than individuals with normal weight [29], while those with higher BMI have more respiratory support. These results are consistent with our ndings.
The mechanism involved in the association of BMI and sepsis-related mortality is unknown. Firstly, sepsis involves an acutely abnormal metabolic state in which body fat can be used as energy to respond to the body's response to acute illness. Studies have shown that weight gain provides an indirect nutritional reserve that plays a vital role in survival during acute life-threatening diseases [30]. A recent study conducted by Alberda et al. focused on the importance of nutritional supplementation for critically ill patients. They found that the positive effects of increased nutrition mainly occurred in underweight and normal weight patients and a small number of moderately obese patients [31]. In addition, the protective effect during critical illnesses may be attributed to the higher levels of proin ammatory cytokines in obese healthy individuals versus normal-weight healthy individuals. This effect may be promoted by M1 in ammatory activation switches to alternative M2 anti-in ammatory activation [32]. Obese patients with sepsis may have less severe in ammatory response, less tissue damage, less septic shock, and consequently better survival. Finally, higher BMI results in increased deposition of adipose tissue. Adipose tissue is increasingly recognized as a functional endocrine organ and is associated with increased activity of the renin-angiotensin system. 33 It appears to exert a hemodynamic protective effect during sepsis and may reduce its effect on uids or the need for vasopressor support [34,35].
According to the clinical diagnosis and clinical intervention check-in analysis, each of the four types of ICU has a different patient population. However, when we separately analyze patients in each ICU, our results on the protective effect of obesity remain valid. Despite the inherent differences among patients in medical ICU, surgical ICU, coronary care unit, and cardiac surgery recovery unit, obese and overweight patients have a lower risk of death. Conversely, patients with a low BMI have a higher risk of death than normal-weight patients. Patients in the medical ICU had a disadvantage compared with those in the cardiac unit, while the cardiac surgery recovery unit and surgical ICU were effective in improving patient outcomes.
We assessed the baseline health of patients prior to admission by determining the presence of AIDS, lymphoma, solid tumor, or metastatic cancer. The study revealed that the diagnosis of lymphoma was evenly distributed among normal, overweight, and obese patients, which is consistent with the distribution observed in our overall study population. Overweight and obese patients have a markedly lower prevalence rate of human immunode ciency virus compared with normal-weight patients. However, in our study, patients with AIDS accounted for only 1.2% of the population. The incidence of metastases in underweight patients is markedly higher than that observed in overweight and obese patients. This is in line with the protective role stated in our hypothesis. The consequences of solid tumors and lymphomas did not show a correlation and were not statistically signi cant. However, consistent with previous studies [35], we found that obese patients admitted to the ICU are usually younger than those who are underweight. We also noted that obese patients constitute the majority of private insurance coverage rather than health insurance or Medicaid, which are state or federal programs in the USA providing coverage for older, low-income, or critical patients with chronic health problems. Differences in age and insurance can re ect the inherent differences in overall health or access to health care, and may explain

Conclusions
Our study shows that higher BMI is a protective factor in the prognosis of patients diagnosed with sepsis in the ICU. A higher BMI is associated with a lower risk of mortality. The exact mechanism of this association is unclear. Therefore, we conducted extensive analysis to ensure that our ndings were not in uenced by confounding factors or other biases. In the future, well-designed studies are warranted to assess the impact of obesity on sepsis and gain insights into the underlying mechanisms of this

Availability of data and materials
The data were available on the MIMIC-III website at https://mimic.physionet.org/ Authors' contributions DW conceived the idea, performed the analysis, and drafted the manuscript. LD participated in the study design. ZP interpreted the results and helped to revise the manuscript. LS helped to frame the idea of the study and helped to analyze the data. All authors read and approved the nal manuscript.
Ethics approval and consent to participate The study was an analysis of a third-party anonymized publicly available database with pre-existing institutional review board (IRB) approval.

Consent for publication
Not applicable.

Declaration of Con icting Interests
The authors declare that there is no con ict of interest.  Figure 1 Selection of study cohort. Figure 2