Prognostic value of serum lactate kinetics in critically ill patients with cirrhosis and acute-on-chronic liver failure: a multicenter study

Baseline characteristics of the included patients

In the derivation cohort (MIMIC cohort), most patients were male (65.0%) and white (70.9%). Mean age was 58 ± 13 years. Alcoholic liver disease was the main cause of liver cirrhosis, which accounted for 54.5% of cases. The main causes of ICU admission were infections/sepsis, organ failure, and gastrointestinal bleeding. The majority of patients presented with ACLF (69.7%), and ACLF stage 3 was most frequent (31.7%). The median MELD, CLIF-SOFA, and CLIF-C ACLF scores at admission were 25 (19–34), 11 (8–14), and 55 (48–62), respectively. Most patients received vasopressors (60.8%) and mechanical ventilation (77.6%). The median length of ICU stay was 5 (3–10) days. In-hospital mortality was 40.3%, all-cause 28-day mortality was 39.9%, 90-day mortality was 49.4%, and 1-year mortality was 56.2%.

There were two validation cohorts (eICU cohort and WMU cohort) in our study. Patients who came from the eICU cohort had similar baseline characteristics to patients in the MIMIC cohort. In the WMU cohort, all of the participants were Chinese. The main cause of liver cirrhosis was viral hepatitis (59.1%). All of the scores were lower in the WMU cohort, compared to the other two cohorts. Moreover, mortalities in the WMU cohort were also lower than in the other cohorts. There were four patients who underwent liver transplantation within 28 days, and there were nine patients who underwent liver transplantation within 90 days after ICU admission. More information about the baseline characteristics of the patients in the three cohorts is listed in Table 1.

Patients were stratified by Δ24Lac quartiles in the derivation cohort. As presented in Table 2, Δ24Lac levels were positively associated with MAP levels, body temperature levels, 24-hour urine output, and mechanical ventilation support and were inversely associated with bilirubin levels, vasopressor use, and mortality. Figure 1 shows Kaplan-Meier curves illustrating the cumulative survival rates stratified by Δ24Lac quartiles.

Figure 1. Kaplan-Meier curves stratified by Δ24Lac quartiles. The curves showed different the cumulative survival rates of patients with different Δ24Lac levels. (A): cirrhotic patients. (B): ACLF patients.

Independent effect of lactate clearance on mortality

In the derivation cohort, we saw significant inverse associations between increasing Δ24Lac and in-hospital mortality. The unadjusted OR for every ten percent increase in Δ24Lac was 0.85 (95% CI 0.81–0.90; P <0.01). After multivariable adjustment, the OR for every ten percent increase in Δ24Lac was 0.84 (95% CI 0.78–0.90; P <0.01). The OR for in-hospital death (vs. Δ24Lac ≤ 0) was 0.40 (95% CI 0.17–0.93) for Δ24Lac 0 to ≤ 0.3; 0.17 (95% CI, 0.07–0.41) for Δ24Lac 0.3 to ≤ 0.55; and 0.03 (95% CI, 0.01–0.10) for Δ24Lac > 0.55 (P for trend <0.001). Δ24Lac was also independently associated with in-hospital mortality in patients with ACLF (Table 3). Further, through Cox regression, we demonstrated that decreased Δ24Lac was significantly associated with increased 28-day and 90-day mortality rates (Table 4 and Supplementary Table 2). In the competing risk analysis (WMU cohort), we also found that Δ24Lac was independently correlated with increased mortality rates (Supplementary Table 3). Each 10% increase in Δ24Lac led to 10% decrease in 90-day mortality.

We repeated the lactate measurements at day 3-7 and calculated the lactate clearance (ΔLac_3-7). We demonstrated that ΔLac_3-7 was also an independent and significant predictor for poor prognosis in critically ill patients with cirrhosis and ACLF (Supplementary Table 4).

Development and validation of lactate clearance -adjusted LiFe (LiFe-Δ24Lac)

Lactate clearance was identified as a significant predictor of mortality, independent of the LiFe score (Tables 3 and 4). We hypothesized that the addition of lactate clearance instead of lactate to the LiFe score may improve the performance of the score for predicting mortality. The LiFe-Δ24Lac score was calculated by adding points for each of the following risk factors: total bilirubin 0–1.9, ≥2.0–3.9, ≥4.0–5.9, ≥6.0 mg/dL; INR 0–1.9, ≥2.0–3.9, ≥4.0–5.9, ≥6.0; and lactate clearance >0.55, >0.30–0.55, >0–0.30, ≤ 0.

In patients with cirrhosis, AUROCs for LiFe-Δ24Lac in the prediction of hospital, 28-day, 90-day mortality, and 1-year mortality (0.75, 0.74, 0.72 and 0.69) were significantly higher than those obtained for original LiFe score (0.69, 0.67, 0.66 and 0.61, all p < 0.01]). In predicting hospital and 1-year mortality, the addition of Δ24Lac to the LiFe score increased net reclassification improvements (NRIs) by 12.9% (p<0.01) and 16.4% (p<0.01), respectively, and increased integrated discrimination improvements (IDIs) by 8.1% (p=0.03) and 6.6% (p=0.04), respectively.

In patients with ACLF, the LiFe-Δ24Lac score also improved the prediction of mortality, as shown by the significant increase in AUROCs (in-hospital mortality: 0.75; 28-day: 0.74; 90-day: 0.70; 1-year: 0.69) compared with that of the LiFe score (in-hospital mortality: 0.68; 28-day: 0.66; 90-day: 0.63; 1-year: 0.60, all p<0.01). NRIs and IDIs also improved significantly after adding Δ24Lac (in-hospital mortality: NRI 16.4% [p<0.01]; IDI 8.8%, [p=0.04]; 1-year mortality: NRI: 20.2%, [p<0.01]; IDI 7.8%, [p=0.05]).

We also compared LiFe-Δ24Lac score with CLIF-C-ACLF score. LiFe-Δ24Lac score performed with a higher discrimination than CLIF-C ACLF score, especially for predicting short-term mortality (hospital and 28-day mortality, p<0.05). The prognostic performance of these scoring systems in the derivation and validation cohort are illustrated in Figure 2. The calibration plots showed good agreement between the LiFe-Δ24Lac score prediction and actual observation in both the primary and validation cohorts (Supplementary Figure 1).

Figure 2. AUROCs for LiFe-Δ24Lac, CLIF-C ACLF and LiFe scores in prediction of mortality in patients with cirrhosis and ACLF in the derivation and validation cohort. Cirrhosis patients, MIMIC cohort: 28-day mortality (A); 90-day mortality (B); eICU cohort: hospital mortality (C); WMU cohort: 28-day mortality (D); 90-day mortality (E). ACLF patients: MIMIC cohort: 28-day mortality (F); 90-day mortality (G); eICU cohort: hospital mortality (H). WMU cohort: 28-day mortality (I); 90-day mortality (J).

Outcome events

We re-grouped the LiFe-Δ24Lac scores into three categories (relatively low-risk: 0–3; intermediate-risk: 4–6; and high-risk: 7–9) using X-tile software. In the derivation cohort, the in-hospital, 28-day and 90-day mortality rates were respectively 21.7% (45/207), 21.7% (45/207) and 33.8%(70/207) for patients with low risk; 51.5% (86/167), 50.9%(85/167) and 58.7% (98/167) for patients with middle risk; and 76.4% (42/55), 74.5%(41/55) and 80.0% (44/55) for patients with high risk. More detailed information about outcome events is presented in Table 5.

Data source

We extracted the patients’ dataset from the Medical Information Mart for Intensive Care (MIMIC)-III and the eICU Collaborative Research Database (eICU) [26, 27]. The MIMIC-III is a database consisting of over 40000 ICU patients who stayed in intensive care units of the Beth Israel Deaconess Medical Center between 2001 and 2012. The eICU database comprises 200,859 ICU patients admitted to 208 hospitals located throughout the United States between 2014 and 2015. Medical history, clinical information, vital signs, and laboratory parameters were documented in these databases. We obtained access to these databases after completing the National Institute of Health’s training course ‘Protecting Human Research Participants’ (certificate number: 25557915).

Study design

We conducted an international multicenter retrospective cohort study. Adult cirrhotic patients with hyperlactatemia on admission were eligible for inclusion in this study. We excluded patients who stayed in the ICU for no more than 48 hours. The primary endpoint of the study was hospital mortality, and the secondary endpoint was determined considering the all-cause 28-day, 90-day and 1-year mortality rates. The diagnosis of cirrhosis was based on histopathology, ultrasonography or computed tomography findings, clinical evidence of liver dysfunction or portal hypertension. We diagnosed and graded ACLF according to the criteria established by the EASL–CLIF consortium [28]. Hyperlactatemia was defined by arterial lactate levels above 2.0 mmol/L.

In the derivation cohort, we included 429 cirrhotic patients (7,189 measurements) with hyperlactatemia from the MIMIC database. The flow chart is presented in Figure 3. Additionally, we analysed a validation cohort of 303 patients collected from the eICU database. Another validation cohort (WMU cohort) of 222 cirrhotic patients collected from ICUs of the First Affiliated Hospital of Wenzhou Medical University (Zhejiang, China) between January 2012 and January 2018 was also analysed.

Figure 3. A flow diagram of study participants (derivation cohort).

The MIMIC and eICU cohorts’ mortality information came from the social security database intheUnitedStates. The mortality information of the WMU cohort was collected by review of medical records or by contacting the patients. The ethics committee of the First Affiliated Hospital of Wenzhou Medical University approved this study. All study procedures followed the principles of the Declaration of Helsinki. No informed consent was required because all the data were anonymized.

Lactate-clearance and scoring systems calculation

The main focus of the study was arterial lactate kinetics over the first 48 hours of admission. To obtain Δ24Lac, we recorded the maximum lactate levels of day 1 and day 2. We calculated Δ24Lac based on the formula: Δ24Lac (%) = (Lactate _{day1 max} – Lactate _{day2 max})/ Lactate _{day1 max}[8]. Time-based protocols of lactate concentration determination were not prespecified. This reflected the real-world scenario regarding the timing and frequency of lactate measurements. Although there was no time-based protocol regarding lactate measurements, the lactate values were measured on a relatively regular basis in enrolled patients (at least one time during each 6-h shift on admission day, and at least one time per day during the ICU stay).

We also repeated the measurement on day 3–7 to assure our prediction was right. We calculated the lactate clearance at day 3–7 according to the formula:

$\Delta La{c}_{3-7}\left(\%\right)=\left(La{c}_{day1max}–La{c}_{day3-7max}\right)/La{c}_{day1max}.$

The Child-Pugh, model of end-stage liver disease (MELD), sequential organ failure assessment (SOFA), chronic liver failure–SOFA (CLIF-SOFA), and CLIF-C ACLFs scores were calculated according to the published formulae [28–32]. The LiFe score was calculated by adding points for each of the following risk factors: total bilirubin 0–1.9, ≥2.0–3.9, ≥4.0–5.9, ≥6.0 mg/dL; INR 0–1.9, ≥2.0–3.9, ≥4.0–5.9, ≥6.0; arterial lactate 0–1.9, ≥2.0–3.9, ≥4.0–5.9, ≥6.0 mg/dL[17].

Statistical analysis

We categorized lactate clearance as ≤ 0, >0–0.3, >0.3–0.55, and >0.55. Continuous variables were expressed as median with interquartile range (IQR) and compared by Kruskal-Wallis test. Categorical variables were expressed as percentages (%) and compared by Chi-Square test. To maximize data availability, we used multiple imputations, based on 5 replications and a chained equation approach method in the R MI procedure, to account for missing albumin, INR, bilirubin, creatine, and urine output date. We also performed sensitivity analyses using a complete-case analysis. Univariate and multivariate logistic regressions were used to identify the association between lactate clearance and hospital mortality. Odds ratios (OR) were reported with a 95% confidence interval (CI). The association between 28-day and 90-day mortality and lactate clearance was analysed by univariable and multivariable Cox proportional hazards models, and hazard ratios (HR) were calculated. We performed tests for linear trend by entering the median value of the variable. We conducted a multivariate competing risk regression (Fine-Gray model) in the WMU cohort to assess the effect of lactate clearance on mortality. Liver transplantation was taken into account as a competing risk.

We plotted cumulative survival curves by the Kaplan-Meier method and compared the survival curves using the log-rank test. The performances of the scoring systems were assessed by calculation of the area under the receiver- operating characteristic curve (AUROC) and assessed using the DeLong test [33]. Furthermore, we used the net reclassification index (NRI) and integrated discrimination improvement (IDI) to evaluate the additive predictive value of lactate clearance instead of lactate over the LiFe score in assessing the improvement of prognostic value [34]. A calibration curve was used to compare the predicted probability of survival versus actual, using 500 bootstraps resamples to reduce overfit bias. All of the tests were two sided, and a p value of <0.05 was considered statistically significant. The statistical packages R (version 3.4.3, The R Foundation; http://www.r-project.org) and MedCalc (version 15.2.2 Ostend, Belgium) were used for statistical analyses.

The LiFe-Δ24Lac scores were re-grouped into three categories (relatively low-risk, intermediate-risk and high-risk) using X-tile software (Version 3.6.1, Yale University, USA) to calculate the optimal cut-off values[35].