This was a retrospective observational study in which data were extracted from an online international database, the Medical Information Mart for Intensive Care III (MIMIC-III) . The MIMIC-III is a large, single-center database that comprises information related to patients admitted to critical care units at a large tertiary care hospital located in Boston. The database contains 53,423 distinct hospital admissions for adult patients (age 16 years or above) admitted to critical care units between 2001 and 2012. It is possible to access this database by passing an examination and obtaining the certification. One author (HC) obtained access and was responsible for the data extraction (certification number 27252652).
Selection of participants
Septic patients with an initial lactate level > 2.0 mmol/L after ICU admission were eligible for inclusion in our study. The diagnoses of sepsis were consistent with the third sepsis definition , that is, patients with documented or suspected infection and an acute change in total SOFA score ≥ 2 points. Infection was identified from the ICD-9 code in the MIMIC-III database. Patients who were younger than 18 years or spent fewer than 48 h in the ICU were excluded. For patients who were admitted to the ICU more than once, only the first ICU stay was included for analysis. Included patients were subsequently divided into two groups: the early lactate group (EL group, the initial lactate level was measured within 1 h after ICU admission) and the late lactate group (LL group, the initial lactate level was measured more than 1 h after ICU admission).
Data from the MIMIC-III database that were considered baseline characteristics within the first 24 h after ICU admission included the following: gender, weight, race, admission type, admission period, severity at admission measured by Sequential Organ Failure Assessment (SOFA) score, quick Sequential Organ Failure Assessment (qSOFA) score, Simplified Acute Physiology Score II (SAPS II) score, Overall Anxiety Severity and Impairment Scale (OASIS) score, Elixhauser comorbidity score, use of mechanical ventilation, use of renal replacement therapy (RRT), and administration of vasopressors. Initial lactate levels and vital signs, including mean arterial pressure (MAP), heart rate, temperature (°C), and respiratory rate, were also extracted. If a variable was recorded more than once in the first 24 h, we used the value related to the greatest severity of sepsis.
Comorbidities including congestive heart failure (CHF), atrial fibrillation (AFIB), chronic renal disease, liver disease, chronic obstructive pulmonary disease (COPD), stroke, and malignant tumor were identified on the basis of the recorded ICD-9 codes. The site of infection and blood culture data were also collected for analysis.
Outcomes and therapeutic interventions
The primary exposure was the early lactate measurement, which was defined as an initial lactate level measured within 1 h after ICU admission. The primary outcome of the present study was 28-day mortality. Secondary outcomes included mechanical ventilation-free days and vasopressor-free days within 28 days after ICU admission, AKI stage, and the duration of ICU and hospital stays.
Therapeutic interventions in our study might have transmitted the effect of early lactate measurement to the primary outcome, including the time to initial vasopressor administration (hours), time to initial antibiotic therapy (hours), time to initial intravenous fluid (IVF) treatment (hours), and volume (L) of IVF administered within 6 h and 24 h of ICU admission.
The details of missing data are summarized in Additional file 1: Table S1. More than 20% of the missing data were removed from our analysis, and the remaining missing data were obtained with the multiple imputation method.
Causal mediation analysis
Causal mediation analysis (CMA)  is a method for separating the total effect of a treatment into direct and indirect effects. The indirect effect on the outcome is mediated via a mediator. The analysis reports consist of the average causal mediation effect (ACME), average direct effect (ADE), and total effect. In our study, we used the early lactate measurement as the treatment and the time to initial IVF, time to initial antibiotic treatment, and time to initial vasopressor administration as mediator variables to explore whether the effect of early lactate measurement on the primary outcome is mediated by the mediator variables mentioned above.
Continuous variables are expressed as the mean ± standard deviation or median (interquartile range), as appropriate. Categorical variables are shown as proportions. Student’s t test, analysis of variance, and the Mann-Whitney U test were used as appropriate. Categorical variables were compared using the χ2 test.
Multivariate modeling of the association between early lactate measurement and 28-day mortality was performed with logistic regression. Baseline variables that were considered clinically relevant or that showed a univariate relationship with the outcome (p < 0.10) were entered into a multivariate logistic regression model as covariates, and included age, gender, weight, admission type, admission period, severity scores, use of mechanical ventilation, use of RRT, use of vasopressors, comorbidities, site of infection, MAP, and initial lactate level. The variance inflation factor (VIF) method was used to examine multicollinearity, and VIF ≥ 5 suggested multicollinearity in our model. Subgroup analyses according to gender, admission period, comorbidities, vasopressor use, and site of infection were performed. We also investigated the relationship between a delay in initial lactate measurement and a delay in remeasurement in the EL group and 28-day mortality by multivariate logistic regression.
Propensity score matching (PSM)  was used to minimize the imbalance of covariates between the EL and LL groups. A multivariate logistic regression model was used to estimate the patient’s propensity scores for early lactate measurement. A 1:1 nearest neighbor matching was applied with a caliper width of 0.02 in our study. The standardized mean differences (SMDs) and p values were calculated to evaluate the effectiveness of the PSM. We then used the PSM model and the doubly robust model [16, 17]—the combination of the multivariate logistic regression model and PSM model—to further clarify the relationship between early lactate measurement and 28-day mortality. Outcomes and therapeutic interventions were generated from a matched cohort. CMA was employed to explore the association described above.
All statistical analyses were performed using the R package (version 3.6.0), and p < 0.05 was considered statistically significant.