Use of plasma lactate level to predict 28-day mortality in non-elderly and elderly sepsis patients based on the MIMIC-III database

Purpose: We co mpared the use of lactate level for predicting 28-day mortality in nonelderly (<65 years) and elderly (≥65 years) sepsis patients who were admitted to an intensive care unit (ICU). A multivariate logistic regression model was established to predict 28-day mortality for each group. Methods: This retrospective study used the Medical Information Mart for Intensive Care Ⅲ, a publicly available database of ICUs. Eligible sepsis patients were at least 18 years-old, hospitalized for at least 24 h, and had lactate levels measured in the ICU. Univariate logistic regression analysis and step-wise multivariable logistic regression models were used to identify factors associated with 28-day mortality. Results: The 28-day mortality was 30.9% among the 2482 patients, and was significantly greater in elderly than non-elderly patients. Within each age group, the lactate level was greater for non-survivors than survivors. Among non-survivors, the lactate level was significantly higher for the non-elderly than the elderly. Adjusted logistic regression analysis showed that non-elderly patients with lactate levels of 2.0–4.0 mmol/L and above 4.0 mmol/L had greater risk of death than those with normal lactate levels. For all patients, the stepwise logistic regression model had an area under the receiver operating curve (AUROC) of 0.752; for non-elderly patients, the model had an AUROC of 0.793; for elderly patients, the model had an AUROC of 0.711. The Hosmer-Lemeshow test indicated acceptable goodness-of-fit for each group (P=0.206, P=0.646, and P= 0.482, respectively). Conclusion: In our population of sepsis patients, the lactate level was about 0.9 mmol/L lower in elderly non-survivors than non-elderly survivors. A plasma lactate level above 2.0 mmol/L was an independent risk factor for death at 28-days among non-elderly patients. Our logistic regression models effectively predicted 28-day mortality of sepsis patients in

3 different age groups.

Background
The clinical mortality rate of sepsis is now higher than that of myocardial infarction and, except for heart disease, sepsis is the main cause of death in the ICU [1]. The ultimate cause of death from sepsis is organ dysfunction caused by the patient's reaction to the infection [2]. Because of their reduced immune responses and resistance, sepsis more common among the elderly. The prevalence and mortality of severe sepsis have increased significantly over time [3].
The plasma lactate level is an important biomarker that reflects the oxygen metabolism of tissues. The root cause of septic shock is tissue hypoxia, and this leads to an increased level of glucose, anaerobic glycolysis, and lactate production [4]. Persistent hyperlactatemia, which suggests that tissue hypoxia has not been corrected, is associated with adverse outcome. An elevated level of plasma lactate during sepsis is associated with more severe disease and poor prognosis [5]. However, there have been questions about the use of lactate as a risk marker for sepsis and the ability of lactate level to predict 28day mortality in patients with sepsis.
The rapid development of medical information resources has made a large number of electronic health records available [6]. Analysis of these records has been a focus of significant research in medical research and related fields [7]. The Medical Information Mart for Intensive Care (MIMIC-Ⅲ), which we used for the present study, is a publicly available database developed by the Laboratory of Computational Physiology at the Massachusetts Institute of Technology. This database integrates comprehensive clinical data from ICU patients who received care at the Beth Israel Medical Center from 2001 to 2012. It has de-identified data on demography, vital signs, laboratory tests, medical records, imaging reports, drug use, and other clinically significant information [8]. 4 Researchers and institutions around the world have published numerous studies based on analysis of the MIMIC-Ⅲ database [9][10][11][12][13].
The present study is the first to use the MIMIC-Ⅲ database to compare the predictive value of plasma lactate level on 28-day mortality of non-elderly and elderly patients with sepsis in the ICU.

Study population
The MIMIC-Ⅲ database was used to identify all adults diagnosed with sepsis, severe sepsis, or septic shock with a first ICU admission. All included patients were at least 18 years-old and were hospitalized for at least 24 h. Patients whose initial lactate levels and chart events were not recorded were excluded.
The gold standard used for diagnosis of sepsis was the 2001 consensus definition [14], which defines sepsis as infections consisting of 2 or more systemic inflammatory response syndrome (SIRS) criteria (temperature above 38°C or below 36°C, heart rate greater than 90/min, respiratory rate greater than 20/min or PaCO2 below 32 mmHg, and white blood cell count greater than 12,000 or less than 4000 cells/mL or more than 10% band forms) [14].

Study design
Independent variables (including demographic characteristics, major complications, major infection sites, laboratory data, vital signs, mortality prediction scores, and 28-day prognosis) were extracted from MIMIC-Ⅲ via PostgreSQL, a structured query language with Navicat Premium 12.
Plasma venous lactate levels that were recorded upon ICU admission were analyzed.

Methods of analysis
SPSS 17.0 software was used for data analysis. Continuous variables are expressed as medians and inter-quartile range, and the Mann-Whitney U test was used for comparisons.
Categorical variables are expressed as numbers and percentages, and were compared using the chi-square test or Fisher exact test. Multivariate logistic regression was used to identify factors significantly and independently associated with lactate and the prognosis of sepsis. The Hosmer-Lemeshow test was used to evaluate the suitability of the model.
Prognosis was evaluated using receiver operating characteristic (ROC) analysis, and the ability of the regression model to predict 28-day mortality was assessed by calculation of the area under the ROC (AUROC). Youden's index was used to assess the performance of the diagnostic test, and the maximum point of Youden's index was used as the cut-off point (sensitivity + specificity -1). For all analyses, a P value below 0.05 was considered significant.

Results
There were 46,476 patients who were first admitted to the ICU, and 3512 had diagnoses of sepsis (ICD 995.91), severe sepsis (ICD 995.92), or septic shock (ICD 785.52; Figure 1).
After exclusion of 5 patients who were younger than 18 years-old, 309 patients who were discharged from the ICU within 24 h, 4 patients who did not have chart event data, and 710 patients whose initial lactate levels were not measured, there were 2482 patients.
Among these 2482 patients, 1100 were younger than 65 years and 1382 were 65 years or older. Table 1 shows the baseline clinical and demographic characteristics of the 2482 patients overall and of the elderly and non-elderly groups. The overall 28-day mortality rate was 30.9%, and the rate in elderly patients was 64.8%. Most patients in the non-elderly and elderly groups were male. The average duration in the ICU was 126.07 h for the non-6 elderly and 101.80 h for the elderly. There were significant differences in the racial composition of the two age groups. In particular, there were higher percentages of black and Hispanic/Latino patients in the non-elderly group. The two age groups also had significant differences in many clinical characteristics, including multiple vital signs, mortality prediction scores, major comorbidities, and major source of infection.
Analysis of major complications indicated the incidence of hypertension, congestive heart failure, chronic renal insufficiency, cerebrovascular disease, and diabetes were greater in elderly patients, but the incidence of cirrhosis was greater in non-elderly patients. The elderly group had more respiratory tract infections (34.2% vs. 29.7%), but the non-elderly group had a greater incidence of skin and soft tissue infections (10.9% vs. 6.9%). The 28day mortality was significantly greater in the non-elderly group (36.0% vs. 24.5%). The two groups had no significant difference in plasma lactate level. (Table 2) indicated the mortality rate increased with patient age within each age group, and that time in the ICU had a positive association with survivorship only in the non-elderly group. Sex and race had no significant effect in either age group. The incidence of cirrhosis, chronic renal insufficiency, malignancy, and lactate level were significantly greater among nonsurvivors in each age group. There were significant differences in mortality prediction scores (SOFA and SAPS) of the two age groups. Further analysis (Table 3) showed that the lactate level was similar for elderly and non-elderly survivors (1.8 vs. 1.8 mmol/L, P = 0.571), but was greater in non-elderly non-survivors than elderly non-survivors (2.2 vs. 3.1 mmol/L, P < 0.001).

Analysis of survivors and non-survivors
We initially used univariate logistic regression analysis to identify variables related to 28day mortality. The subsequent multivariate logistic regression analysis, in which patients with normal levels of lactate (<2 mmol/L) were used as the reference group, indicated multiple factors were significantly and independently associated with 28-day mortality: age, SOFA, SAPS, SpO 2 , and malignancy ( Table 4). The crude and adjusted ORs indicated the risk of death at 28 days in the non-elderly group increased with increased lactate level, but there was no such correlation in the elderly group.
We also used multivariate analysis to determine the impact of other clinical factors on mortality among patients overall (Figure 2), the non-elderly group (Figure 3), and the elderly group (Figure 4). The stepwise logistic regression analysis indicated that 28-day mortality correlated with age, lactate level, SOFA score, SAPS score, SpO 2 , and malignancy among all patients; with lactate level, SOFA score, and malignancy in the nonelderly group; and with SOFA score, SAPS score, SpO 2 , and malignancy in the elderly group. Table 5 shows the regression equations and Homser-Lemeshow test results for each group.
We performed ROC analysis to evaluate the diagnostic performance of the logistic regression models for the three different groups ( Table 6). The AUROC for 28-day

Discussion
Sepsis is associated with a high mortality rate, but there are limited objective and effective clinical markers of prognosis. More than half of patients with sepsis in the United States are more than 65 years-old. Relative to the non-elderly, the elderly have a 13.1-8 fold greater risk of sepsis and 1.56-fold greater risk of death from sepsis [15].
In this study, we retrospectively analyzed the clinical characteristics, 28-day mortality rate, and the relationship of plasma lactate level with the prognosis of sepsis patients in the MIMIC-Ⅲ database, and established a multivariate logistic regression model to predict 28-day mortality in different age groups. Most of these patients were from an emergency department, so those with high lactate levels were treated soon after admission. This may have contributed to the lower overall lactate levels in our study than in a previous study [16].
The elderly patients are more likely than the nonelderly to develop sepsis due to Gramnegative bacteria, especially among patients with pneumonia and fungal infections.
Respiratory tract infections are also more common causes of sepsis in elderly patients [15]. Relative to the non-elderly, we found that elderly sepsis patients had a higher 28day mortality rate and that sepsis was more likely to be caused by a respiratory tract infection, in agreement with previous studies [16,17]. In contrast, we found that sepsis in non-elderly patients was more likely to be caused by skin and soft tissue infections.
The major indicators of poor prognosis in elderly sepsis patients are shock, elevated plasma lactate, and organ failure (especially of the respiratory system or heart) [18]. In addition, previous research indicated that advanced age is an independent risk factor for severe sepsis and death from sepsis [19]. Our multivariate adjusted logistic regression analysis showed that lactate level was an independent risk factor for 28-day mortality for non-elderly sepsis patients, but this relationship was not significant for the elderly. This might be a result of blunted inflammatory responses in the elderly.
Our study is the first to use the MIMIC database to analyze the effect of plasma lactate level on 28-day mortality among sepsis patients in the ICU. Our results indicated that elevated lactate level was associated with increased 28-day mortality in non-elderly and elderly patients, but it was a more reliable prognostic indicator for the non-elderly. A previous study reported that elevated lactate level was associated with poor prognosis for ICU patients after ruptured abdominal aortic aneurysm repair [20]. However, plasma lactate concentration reflects overall changes of the body's metabolism, so its sensitivity often low.
The models established in this study to predict 28-day mortality considered multiple factors (age, lactate level, comorbidities, and mortality prediction scores). The results of the Hosmer and Lemeshow Test (P>0.05) indicated that the information in the current data was fully extracted, and the established regression models had good statistical fits.
The AUROC was 0.752 for all patients, 0.711 for elderly patients, and 0.793 for non-elderly patients, indicating that the models could be used to reliably predict 28-day prognosis for each group. In clinical settings, these models may therefore be useful for predicting the probability of 28-day mortality in patients with sepsis and in deciding which patients should be closely monitored and provided with necessary interventions to prevent death.
A limitation of this study is that we only examined a relatively small number of patients.
In addition, although we examined the effect of ethnicity, patients in the MIMIC-Ⅲ database were mainly white, and there were very few patients from other ethnic groups.
This points to the need for further forward-looking, large sample validation and risk grading studies. More simple and effective prediction methods should be used in clinical practice to achieve targeted interventions and reduce the incidence of death.

Conclusions
Measurement of plasma lactate level is a simple and inexpensive method that clinicians can use to assess the risk of mortality in patients with sepsis. The lactate level among elderly non-survivors was about 0.9 mmol/L lower than among non-elderly survivors. Our results indicated that lactate level is an independent risk factor for 28-day prognosis in 10 non-elderly patients with sepsis. The ability of elevated plasma lactate (>2.0 mmol/L) to predict 28-day prognosis was better in non-elderly than elderly sepsis patients. The multivariate logistic regression models established in this study reliably predicted 28-day mortality in sepsis patients from different age groups.

Acknowledgments
The authors would like to thank MIMIC Ⅲ program for access to the database.

Funding
No funding was obtained for this study.

Availability of data and materials
The datasets analyzed during the current study are available in https://github.com/MIT-LCP/mimic-code/tree/master/concepts.

Authors' contributions
YHD designed the methods and experiments, and contributed to the writing of manuscript.

11
XYM and YFH cleaned the data. YYH, JC and YRL provided guidance and reviewed the manuscript critically. JYP supervised the study and revised the paper. All authors read and approved the final manuscript.

Ethics approval and consent to participate
Ethical consent was not required in this study, since the MIMIC Ⅲ data were analyzed namelessly.