Data source

We performed a cohort study using data extracted from the Medical Information Mart for Intensive Care III (MIMIC III) clinical database (version 1.4), which contained over 58 000 hospital admission data entries of adult patients and neonates admitted to various critical care units between 2001 and 2012.²⁵ One of the study investigators (WZ) was allowed to download data from the database, having completed the “Data or Specimens Only Research” course (record identity, 25222342). The requirement for individual patient consent was waived, as the project neither contained any protected health information nor impacted clinical care.²⁵

Patient records for the external validation of our findings were obtained from the First Affiliated Hospital of Wenzhou Medical University (Wenzhou, Zhejiang, China) after approval from that institution’s ethical committee. All study participants provided written informed consent and their data confidentiality was protected.

Study participants

A flowchart of the inclusion and exclusion procedure for MIMIC III is presented in figure 1. We adopted The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis‑3; a diagnosis flowchart is available in Supplementary material, figure S1) to extract 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 (eg, blood, urine, sputum).¹ In other words, if culture was obtained, we required that an antibiotic was administered within 72 hours, whereas if the antibiotic was administered first, culture was required within 24 hours.¹ Moreover, we defined the period of suspected infection as ranging between 24 hours before and 24 hours after admission to an intensive care unit (ICU). Patients in the CareVue and MetaVision information systems of MIMIC III were admitted before and after 2008, respectively. Only patient data stored in the MetaVision system were collected for analysis. Antibiotic prescription data were available only after 2002; thus, there was a fraction (1/7) of the CareVue patients who were missing data to meet our definition of suspected infection. It was the simplest option for us to limit the cohort to the MetaVision system, because the resulting sample size was sufficient.

To minimize the effect of potential confounding variables in our analysis, patients aged 16 years or younger and those with repeated admissions to the ICU, a history of alcohol abuse, overt chronic liver disease (including chronic viral hepatitis, autoimmune liver disease, alcoholic liver disease, overt liver cirrhosis, or liver transplant), hematologic or solid malignancies, or chronic kidney disease were excluded from the initial study cohort. Furthermore, the exclusion criteria for the sepsis cohort were as follows: current treatment relating to cardiac, vascular, or thoracic surgery. We assumed that these subpopulations had physiological abnormalities yet caused by factors unrelated to sepsis.

The data for 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 for 90 days after admission.

Data extraction

The data were extracted from MIMIC III and our hospital system and included information on patients’ sex, age, race, body mass index (BMI), laboratory investigations, ICU interventions, vital statistics, comorbidities, and length of hospital stay. Sores for the evaluation of illness severity, including SAPS II and SOFA scales, were calculated based on their predefined criteria.^6,7 The mean values of BMI, laboratory parameters, and vital statistics during the first 24 hours of ICU stay were regarded as baseline data. The SAPS II and SOFA scores as well as the necessity to perform interventions with vasopressors and mechanical ventilation were evaluated during the first 24 hours of ICU stay.

Exposures and outcomes

We calculated 3 liver fibrosis indexes (APRI, FIB‑4, and NFS) using their existing formulas (figure 2).^16-18 These indexes were evaluated at baseline with factors assumed to reflect patients’ initial condition on ICU admission, and we categorized the patients by the quartiles of their index values at baseline. Diabetes was defined according to the International Classification of Diseases, Ninth Revision codes or hemoglobin A_1C level of 6.5% or greater, and prediabetes, as a hemoglobin A_1C level ranging between 5.7% and 6.5%.

The primary endpoint of the present study was 28‑day mortality. Secondary endpoints included 90‑day mortality, in‑hospital mortality, and renal replacement therapy (RRT). Mortality in MIMIC III was calculated based on the dates of admission and death obtained from social security records.

Statistical analysis

The Kolmogorov–Smirnov test was used to check the normality assumption for numerical variables. Normally and non‑normally distributed variables were compared using the unpaired t test and the Wilcoxon rank sum test, respectively. Comparisons for categorical variables were performed using the Pearson χ² test and the Fisher exact test. Normally distributed data were expressed as mean (SD), and non‑normally distributed data, as median (interquartile range [IQR]). Categorical variables were presented as frequency and percentage.

We assessed the associations of the 3 indexes with the primary and secondary outcomes using logistic regression analysis. The results were expressed as odds ratios (ORs) with 95% CIs. Septic patients were categorized according to the quartiles of their index values at baseline, and quartile 1 was considered the reference for all subsequent analyses.

A 2‑tailed P value less than 0.05 was considered significant. Statistical analyses were performed using the SPSS software, version 20.0 (SPSS, Chicago, Illinois, United States), the MedCalc software, version 19.0.5 (MedCalc Software, Ostend, Belgium), and the MATLAB software, version R2018b (MathWorks, Natick, Massachusetts, United States).

Multivariable analysis, sensitivity analysis, and external validation

Due to the influence of missing data and potentially relevant confounding factors, several additional analyses were performed to further verify the predictive ability of the liver fibrosis indexes.

First, we attempted to adjust the potential confounding variables through multivariable logistic regression analysis. The following variables were adjusted in the multivariable model: sex, race, laboratory parameters (white blood cell count, hemoglobin, lactate, creatinine, international normalized ratio, partial thromboplastin time, sodium, and potassium levels), vital statistics (heart rate, mean blood pressure, respiration rate, body temperature, and pulse oxygen saturation), comorbidities (congestive heart failure, cardiac arrhythmias, hypertension, chronic pulmonary disease, and diabetes), SOFA and SAPS II scores, and length of hospital stay. Forward likelihood ratio selection was used to filter the included variables.

Second, subset analyses based on 2 liver function indexes were performed to determine whether patients with abnormal baseline liver function distorted the results. Albumin and bilirubin, representing synthesis and metabolism in the liver, were used to divide the patients into groups with normal and abnormal levels according to reference ranges.

Third, we conducted a comparative analysis between the septic and nonseptic patients according to the indexes. Moreover, we performed an additional analysis to establish whether similar results also applied for nonseptic patients.

Fourth, some patients were excluded in the primary analysis, because their index data were not complete during the first 24 hours of ICU stay. Thus, sensitivity analysis was performed for patients in whom baseline index values could not be used, but data from their ICU stay were available.

Fifth, we conducted separate analyses to determine whether liver fibrosis indexes combined with SOFA or SAPS II scores could improve the predictive performance regarding patient outcomes. Performance discrimination was assessed by calculating the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUROC). The DeLong test was used to assess differences in AUROC among the different models. Additionally, we calculated the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) to evaluate improvement associated with the liver fibrosis indexes relative to the SOFA or SAPS II score.

Sixth, we repeated the primary analysis using the Kaplan–Meier and Cox regression analyses instead of logistic regression analysis to evaluate the impact of various analytical methods. The results were presented in the form of a survival curve and the hazard ratio with 95% CI, respectively.

Finally, external validation was introduced to verify whether similar results can be observed in the East Asian population.

Baseline characteristics of the study cohort

The baseline characteristics of the APRI, FIB‑4, and NFS sepsis cohorts are summarized in table 1. The median (IQR) APRI, FIB‑4, and NFS values were 0.537 (0.296−1.339), 2.365 (1.305−4.837), and 0.791 (–0.198 to 1.858), respectively.