Data source
We performed a cohort study using data extracted from the Medical Information Mart for Intensive Care Ⅲ (MIMIC Ⅲ) clinical database (v1.4) which integrated deidentified and comprehensive clinical data of patients admitted to the Beth Israel Deaconess Medical Center (BIDMC) in Boston, Massachusetts [25]. MIMIC III database contains over 58000 hospital admissions data for adult patients and neonates admitted to various critical care units between 2001 and 2012. The Institutional Review Board of the BIDMC (Boston, MA, USA) and Massachusetts Institute of Technology (Cambridge, MA, USA) have approved the use of the MIMIC III database for authorized users. Wei Zhou has been allowed to download the database after completed the “Data or Specimens Only Research” course (Record ID: 25222342). Requirement for individual patient consent was waived for the reason that the project neither contained the protected health information nor impacted clinical care [25].
The patient records of additional external validation were obtained from the First Affiliated Hospital of Wenzhou Medical University (Wenzhou, Zhejiang, China) after approval from the Ethical Committee. All study participants provided written informed consent, and their data confidentiality were protected.
Study participants
A flowchart of the inclusion and exclusion procedure for the MIMIC III database was depicted in Figure 1. We adopted the third international consensus definitions for sepsis (Sepsis-3, diagnosis flowchart was presented in Figure S1) to extract the septic patients from the database [1]. Based on the Sepsis-3 criteria, patients with suspected infection and evidence of organ dysfunction (SOFA score ≥ 2) were identified as septic patients [1]. Suspected infection was defined as the concomitant administration of antibiotics and sampling of body fluid cultures (blood, urine, sputum, etc) [1]. In other words, if the culture was obtained, the antibiotic was required to be administered within 72 hours, whereas if the antibiotic was first, the culture was required within 24 hours [1]. Moreover, we set a suspected infection period during 24 hours before and after intensive care unit (ICU) admission. Patients in the CareVue and MetaVision information systems of MIMIC were admitted before and after 2008, respectively. Only patient data stored in the MetaVision system were collected for future analysis. The antibiotic prescription data were only available after 2002, thus, there was a fraction (1/7) of the CareVue patients who had missing data for the suspected infection definition. It was easiest to limit the cohort to the MetaVision system, as the sample size was sufficient.
To minimize the contribution of potential confounding for future analysis, patients with aged 16 years or younger, repeat admissions to the ICU, alcohol abuse, overt chronic liver disease, haematological and solid malignancies, and chronic kidney disease were excluded from initial cohort. Furthermore, the exclusion criteria for sepsis cohort were as follows: current service related with (1) cardiac surgery, (2) vascular surgical and (3) thoracic surgical. Because we believed that these sub-populations were physiologically abnormal but due to non-sepsis reasons.
The data of external validation were prospectively collected between October 12, 2017 and January 16, 2020 according to the same inclusion and exclusion criteria. The clinical outcomes were followed up to 90 days after admission.
Data extraction
The data were extracted from the MIMIC III database and our hospital including gender, age, race, body mass index (BMI), laboratory data, ICU interventions, vital statistics data, comorbidities and length of stay (LOS). Severity of illness scores including SAPS Ⅱ and SOFA were computed on the basis of prescribed criteria [6,7]. The averages of BMI, laboratory data and vital statistics data during the first 24 hours of ICU admission were regarded as baseline data. The scores of SAPS Ⅱ and SOFA, as well as interventions of vasopressor and mechanical ventilation were evaluated during the first 24 hours of ICU stay.
Exposures and outcomes
Three liver fibrosis indexes (APRI, FIB-4 and NFS) were calculated with previous formulas (Figure 2) [16,17,18]. These indexes were evaluated at baseline on factors believed to reflect the initial condition of ICU admission, and patients were categorized by quartile of index values at baseline. Diabetes was defined as diagnosis of International Classification of Diseases-9 (ICD-9) codes or hemoglobin A1c (HbA1c) ≥ 6.5%, and prediabetes as HbA1c 5.7% – 6.5%. As HbA1c represented the blood glucose of patients for 2 to 3 months, the value of HbA1c within 7 days after ICU admission can be regarded as baseline data.
The primary end point for the present study was 28-day mortality. The secondary end points were 90-day mortality, in-hospital mortality and renal replacement therapy (RRT). Mortality information of MIMIC III database was calculated based on dates of admission and death from the Social Security Death records.
Statistical analysis
Kolmogorov-Smirnov normality test was used to check the normality assumption for numerical variables. Differences in the normally and non-normally distributed variables were compared using the unpaired Student’s t test and Wilcoxon rank-sum test, respectively. Comparisons for categorical variables were performed by Pearson χ2 test and Fisher exact test. Normally distributed data were expressed as mean ± standard deviation, and non-normally distributed data were expressed as median with inter-quartile range. Categorical variables were expressed as frequency with percentage.
We assessed the associations of three indexes with primary and secondary outcomes using logistic regression analysis. The results were presented in form of odds ratio (OR) with 95% confidence interval (CI). The septic patients were categorized according to quartile of index values at baseline, and Quartile 1 was considered as the reference group for all subsequent analyses.
A two-sided P-value < 0.05 was regarded as representing statistical significance. Statistical analyses were performed using SPSS software 20.0 (SPSS, Chicago, IL, USA), MedCalc software 19.0.5 (MedCalc Software, Ostend, Belgium) and MATLAB software R2018b (MathWorks, Natick, MA, USA).
Multivariable analysis, sensitivity analysis and external validation
Due to the influences of missing data and potentially relevant confounding factors, several additional analyses were performed to further verify the predictive abilities of liver fibrosis indexes.
First, we attempted to adjust the clinically potential confounding variables through multivariable logistic regression analysis. The following variables were adjusted in the multivariable model: gender, race, laboratory data (white blood cell, hemoglobin, lactate, creatinine, international normalized ratio, partial thromboplastin time, sodium and potassium), vital statistics data (heart rate, mean blood pressure, respiration rate, temperature, pulse oxygen saturation), comorbidities (congestive heart failure, cardiac arrhythmias, hypertension, chronic pulmonary and diabetes), scores of SOFA and SAPS Ⅱ, and LOS. The forward LR selection was used to filter included variables.
Second, subset analyses on the basis of two different liver function indexes were performed to determine whether patients with abnormal baseline liver function distorted the results. Albumin and bilirubin, representing synthesis and metabolism of liver, were divided to normal and abnormal groups according to their normal ranges.
Third, we did a comparative analysis between sepsis and non-sepsis patients with index values. Moreover, we performed an additional analysis to see whether the similar results also existed in non-sepsis patients.
Fourth, a number of patients in the primary analysis were excluded because of no complete index data during the first 24 hours of ICU admission. Thus, a sensitivity analysis was performed for patients who did not have baseline index values but had available values during ICU stay.
Fifth, we made separate analyses to determine whether liver fibrosis indexes combined with SOFA or SAPS Ⅱ could improve the predictive performance of outcomes. Performance discrimination was assessed by calculation of the receiver operating characteristic (ROC) curve and the area under the receiver operating characteristic curve (AUROC). The DeLong test was used to assess differences in the AUROC among the different models. Additionally, net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were calculated to evaluate the improvement of liver fibrosis indexes to SOFA or SAPS Ⅱ.
Sixth, we repeated the primary analysis using Kaplan-Meier (K-M) and Cox regression analysis instead of logistic regression analysis to evaluate the impact of different analytical methods. The results were presented in form of survival curve and hazard ratio (HR) with 95% CI, respectively.
Finally, additional external validation was introduced to further verify whether the similar results also existed in the East Asian population.