Introduction
Acute-on-chronic liver failure (ACLF) is a common type of clinical syndrome with rapid deterioration of liver function, organ failure(s) and high short-term mortality.1 Hepatitis B virus (HBV) poses a serious threat to human health, due to its devastating effect on liver function.2 In the Asia-Pacific region, HBV is the leading cause of chronic liver disease.3
At present, liver transplantation (LT) is still the most beneficial and feasible therapy for patients with ACLF.4–5 However, 20–30% of patients remain at risk to be delisted from the transplant list, and wait-list mortality is high due to patients being too sick for LT and succumbing to the condition. Hence, it is a significant unmet need to accurately distinguish ACLF patients who are suitable for LT therapy, and seize the best chance for LT.6 Therefore, an accurate prognostic scoring system is needed to guide and optimize the therapeutic strategy for patients with ACLF.7 At present, the model of end-stage liver disease (MELD) score is the most commonly used tool for designating patients to the wait-list for LT.8 However, among the candidates listed for LT, the MELD score may not capture the ACLF severity and adequately evaluate the outcome in the ACLF. Meanwhile, due to differences in patient background queues, the MELD score may not be reasonably applied for HBV-ACLF. Furthermore, although some prognostic scoring systems have been developed to predict the HBV-ACLF short-term (such as 30-day and 90-day) mortality, including the 30-day HBV-ACLFD model previously developed by the investigators9, the efficacy is still scanty to predict the mid-term (such as 180-day) mortality of patients with HBV-ACLF.
The present study aimed to derive novel predictive models to evaluate the 180-day mortality of patients with HBV-ACLF based on the backward stepwise logistic regression (LR) and classification and regression tree (CART) analysis, and to evaluate whether these new models are superior to the MELD scores, providing guidance for clinical treatment decision making.
Methods
Study design
A total of 445 patients, who were diagnosed with HBV-ACLF at Beijing You’an Hospital, Capital Medical University, from June 2014 and December 2018, were selected for the present study. Among these patients, merely 171 entered the final selection. The selection process for HBV-ACLF patients inclusion in the present study is presented in Figure 1.
The enrolment criteria for these patients corresponded to the Asian Pacific Association for ACLF.10 The inclusion criteria were as follows: (a) patients who were at least 16 years-old; (b) patients who were HBV surface antigen (HBsAg)-positive for at least 6 months; (c) patients with a total bilirubin (TBIL) of >171 µmol/L and a sudden exacerbation of liver disease; (d) patients with an international normalized ratio (INR) of >1.5; (e) patients who had ascites within 4 weeks and/or had an onset of hepatic encephalopathy (HE). The exclusion criteria were as follows: (a) pregnant or lactating patients; (b) patients co-infected with human immunodeficiency virus; (c) patients with severe diseases, such as heart dysfunction, previous renal failure, cancer, etc.; (d) patients with infection upon admission to the hospital; (e) patients compounded by other causes of liver damage, such as hepatitis A, C, or E, autoimmune hepatitis, alcohol consumption, or hereditary liver diseases.
The study protocol was approved by the Ethics Committee on Clinical Trials of Beijing You’an Hospital, Capital Medical University. All methods and procedures related to the present study were morally accorded with the laws of the Declaration of Helsinki.
A total of 27 parameters were retrospectively collected as potential risk factors. The parameters included sex, age, serum creatinine level, blood urea nitrogen (BUN) level, aspartate transaminase (AST) level, aspartate alanine transaminase (ALT) level, albumin level, TBIL level [normal reference range: 5–21 μmol/L], serum sodium level, serum potassium level, ammonia level, prothrombin activity (PTA), INR, white blood cell (WBC), hemoglobin (HGB), red blood cell (RBC) count, platelet (PLT) count, lymphocytes (L), neutrophils (N), monocytes (M), time begin, HBV DNA, HBsAg, and complications such as hepatorenal syndrome, ascites, infection, pleural effusion, cirrhosis, and HE. The result (survival or death) for each subject with HBV-ACLF was recorded. The MELD equation was applied to calculate the score for severity as: 9.57 × ln (creatinine, mg/dL) + 3.78 × ln (bilirubin, mg/dL) + 11.20 × ln (INR) + 6.43. The minimal values were forced to 1.0 for calculation purposes.11
LR analysis and nomogram generation
A multivariable LR analysis was performed for the prediction of HBV-ACLF. The candidate predictors were as follows: sex, age, creatinine, BUN, AST, ALT, albumin, TBIL, serum sodium level, serum potassium level, ammonia level, PTA, INR, WBC, RBC, HGB, PLT, L, N, M, time begin, HBV DNA, HBsAg, hepatorenal syndrome, ascites, infection, pleural effusion, cirrhosis, and HE.
In order to identify the significant predictors, 1,000 random samples were generated from the 171 patients through bootstrap resampling with replacement, and backward stepwise LR was conducted for each patient. Then, the predictors selected by the backward stepwise regression were included in the final model. Next, a 10-fold cross-validation was used to calculate the C-index and generate the calibrated statistics. Finally, the parameters for the final model were generated. Based on the results of the logistic regression, the 95% confidence interval (CI) and odds ratio (OR) were calculated. The performance of the model was assessed by sensitivity, and by evaluating the discriminative capacity via the area under the receiver operating characteristic curve (AUROC). A nomogram was created based on the LR analysis, and the nomogram was constructed using the rms package.
Analysis of the CART
A CART analysis was performed for the 171 patients, and it was verified whether this method could calculate more useful clinical results, when compared to the LR model. The candidate predictors were the same as those used in the LR model. The CART analysis divided the data (parent node) into two subsets (child nodes) through the function of the predictor variables. These two subsets were the new parent nodes, which were further split into two child nodes. This process was continued until all patients were classified. After finding the best split for each variable, the CART algorithm used the best overall split to divide the data, and assigned a prediction category for each subgroup. The CART recursively proceeded in this manner, until a predetermined stopping criterion was reached. The algorithm was allowed to go on indefinitely, enabling the model to identify the entirely or almost entirely homogeneous splits.
In the present study, the CART analysis was used to predict the 180-day mortality of patients with HBV-ACLF. The mortality rate, 95% CI and OR were determined. The 10-fold cross-validation was used to trim and optimize the tree, and minimize the relative misclassification. The C-index and the receiver operating characteristic curve were generated to evaluate the performance of the final decision tree.
Statistical analysis
Continuous variables were presented as mean ± standard deviation (SD), and compared using the Mann-Whitney test, and unpaired or two-tailed t-test. Categorical variables were compared using the chi-square test. The predictive accuracy of the LR model was calculated with the concordance statistic, which ranged from 0.5 (no discrimination) to 1.0 (perfect discrimination). The calibration was assessed using the calibration plot, which was implied by a 45° diagonal line with the 1,000 bootstrap samples, in order to decrease the overfit bias.12 The ROC curve analysis was performed using the MedCalc 17.0 software (Mariakerke, Belgium). The nomogram and CART analysis were performed using the R statistical software, version 4.0.2 (http://www.Rproject.org ). The additional statistical analysis was analyzed using the SPSS 25 software (IBM Corp., Armonk, NY, USA). The reported statistical significance levels were all two-sided, and the statistical significance was set at 0.05.
Results
Baseline characteristics
A total of 171 patients who were diagnosed with HBV-ACLF were involved in the present study. The comparison of the clinical characteristics of HBV-ACLF patients stratified by mortality are presented in Table 1. There were no significant differences in sex distribution, potassium/sodium, cirrhosis, HBV DNA, and time begin between the non-survivor (death) group and survivor group (p>0.05). However, the differences in age, PTA, INR, TBIL and L were statistically significant between these two groups (p<0.0001).
Table 1Baseline characteristics of the patients, stratified by mortality
Variable | Overall, n=171 | Non-survivors, n=62 | Survivors, n=109 | p |
---|
Age in years | 45.17 (12.49) | 48.74 (12.54) | 43.14 (12.05) | <0.0001 |
Men, n(%) | 151(88.3) | 51(82.3) | 100(91.7) | 0.064 |
Ascites, n(%) | 103(60.2) | 46(74.2) | 57(52.3) | 0.005 |
HE, n(%) | 20(11.7) | 13(20.9) | 7(6.4) | 0.004 |
Infection, n(%) | 92(53.8) | 38(61.3) | 54(49.5) | 0.138 |
K/Na, n(%) | 17(9.9) | 5(8.1) | 12(11) | 0.536 |
HBeAg, n(%) | 91(53.2) | 32(51.6) | 59(54.1) | 0.751 |
HRS, n(%) | 5(2.9) | 4(6.4) | 1(0.9) | 0.111 |
Pleural effusion, n(%) | 7(4.1) | 5(8.1) | 2(1.8) | 0.115 |
Cirrhosis, n(%) | 137(80.1) | 49(79) | 88(80.7) | 0.789 |
lgHBV DNA | 4.76(1.93) | 4.61(2.06) | 4.84(1.87) | 0.491 |
HBsAg | 3,948.19 (5,194.35) | 4,541.64 (7,356.64) | 3,610.64 (3,403.74) | 0.944 |
ALT | 464.02 (577.17) | 325.65 (305.42) | 542.73 (674.12) | 0.047 |
AST | 377.83 (413.63) | 342.99 (286.41) | 397.66 (471.05) | 0.393 |
TBIL | 353.83 (138.49) | 408.7 (146.06) | 322.62 (124.20) | <0.0001 |
BUN | 4.86 (2.49) | 5.56 (2.78) | 4.46 (2.22) | 0.002 |
Cr | 75.16 (36.97) | 81.50 (42.99) | 71.55 (32.72) | 0.174 |
WBC | 7.34 (3.52) | 7.89 (4.28) | 7.022 (2.99) | 0.422 |
L | 20.59 (8.58) | 16.77 (6.74) | 22.76 (8.78) | <0.0001 |
M | 9.35 (3.78) | 10.20 (4.38) | 8.86 (3.32) | 0.025 |
N | 93.6 (65.55) | 69.2 (18.06) | 63.47 (7.38) | 0.002 |
PTA | 35.97 (9.3) | 31.50 (8.58) | 38.51 (9.62) | <0.0001 |
INR | 2.11 (0.56) | 2.36 (0.65) | 1.97 (0.45) | <0.0001 |
RBC | 3.88 (0.85) | 3.82 (0.90) | 3.91 (.82) | 0.336 |
HGB | 124.3 (21.0) | 122.33 (21.95) | 125.42 (20.46) | 0.402 |
PLT | 104.77 (51.73) | 96.12 (52.58) | 109.68 (50.83) | 0.036 |
Time begin | 22.72 (19.01) | 23.44 (16.22) | 22.31 (0.49) | 0.081 |
ALB | 31.06 (4.13) | 30.85 (4.31) | 31.18 (4.04) | 0.615 |
LR analysis and nomogram
In order to deeply identify the independent predictors of mortality in the present study, multivariate backward stepwise LR analysis was performed. It was found that age, TBIL, PTA, L, M and HE were significantly associated with the 180-day mortality (Table 2). The C-index for the LR model with these predictors was 0.878. In the 1,000 bootstrap data, the calibration plot for the prediction indicated a good fit (Fig. 2), and the Brier score was 0.1898. Based on the results of the LR analysis, a nomogram was drawn to predict the patient’s mortality rate (Fig. 3). A higher score calculated based on the sum of assigned points of each predictor in the nomogram corresponded to a higher probability of death.
Table 2Multivariable predictors of mortality of HBV-ACLF
Variable | β-coefficient | OR(95% CI) | p |
---|
HE | 1.635 | 5.13 (1.282,20.512) | 0.021 |
TBIL | 0.006 | 1.006 (1.002,1.009) | 0.001 |
PTA | −0.115 | 0.892 (0.845,0.941) | 0.0001 |
L | −0.130 | 0.878 (0.825,0.935) | 0.0001 |
M | 0.215 | 1.240 (1.087,1.414) | 0.001 |
Age | 0.049 | 1.050 (1.014,1.087) | 0.006 |
CART analysis
In the CART model, TBIL was identified as the variable for the initial split, with an optimal value of 381.10 μmol/L, and L was selected as the variable for the second split, with a discrimination level of 13.78%. When L was >13.78%, the next best predictor for HBV-ACLF was PTA, with an optimal cut-off value of 33.2. For the node of patients who have a TBIL level of >381.1 μmol/L, an L of >13.78% and a PTA level higher than 33.2, M was selected as the additional significant variable, and this was dichotomized at a level of 10.96%.
Finally, a total of nine subgroups of patients were generated through the seven predictive variables chosen via the CART analysis: subgroup 1 (TBIL ≥381.10 μmol/L and L <13.78%), subgroup 2 (TBIL <381.10 μmol/L and BUN ≥7.915 mmol/L), subgroup 3 (TBIL <381.10 μmol/L, BUN <7.915 mmol/L, and age <56.00 years-old), subgroup 4 (TBIL ≥381.10 μmol/L, L ≥13.78%, PTA <33.20, and age <43.50 years-old), subgroup 5 (TBIL ≥381.10 μmol/L, L ≥13.78%, PTA <33.20, and age ≥ 43.50 years-old), subgroup 6 (TBIL ≥381.10 μmol/L, L ≥13.78%, PTA ≥33.20, and M ≥10.96%), subgroup 7 (TBIL ≥381.10 μmol/L, L ≥13.78%, PTA ≥33.20, and M <10.96%), subgroup 8 (TBIL <381.10 μmol/L, BUN <7.915 mmol/L, age ≥56.00 years-old, and N ≥65.10%), and subgroup 9 (TBIL <381.10 μmol/L, BUN <7.915 mmol/L, age ≥56.00 years-old, and N <65.1%) (Fig. 3). Each patient was sorted to subgroups based on flow chart of the derived CART. The mortality rates for each subgroup are presented in Figure 4. The C-index for the CART model with these predictors was 0.878.
Comparison among the LR, CART and MELD score
As shown in the Figure 5, the predictive power for the 180-day mortality for HBV-ACLF among the LR, CART and MELD score was determined. The CART analysis had an AUROC of 0.878 (95% CI: 0.810–0.923). The performance of the LR analysis was high, with an AUROC of 0.878 (95% CI: 0.820–0.923). However, there was no significant difference between the CART and LR model (p=0.9659). In Table 3, the MELD score had an AUROC of 0.728 (95% CI: 0.655–0.793), which was significantly lower than that for the LR and CART model (p<0.0001).
Table 3The predictive value of mortality of the CART score and other models
Models | AUROC | 95% CI | p | Youden’s index | Sensitivity, % | Specificity, % |
---|
CART | 0.878 | 0.819–0.923 | 0.0001 | 0.6280 | 90.32 | 72.48 |
LR | 0.878 | 0.820–0.923 | 0.0001 | 0.6255 | 85.48 | 77.06 |
MELD | 0.728 | 0.655–0.793 | | 0.4553 | 75.81 | 69.72 |
Discussion
HBV-ACLF is defined as a hazardous syndrome with multiorgan failure.3–5 Worldwide, it has been demonstrated that LT brings survival profit for selected patients with ACLF. Due to the rapid progression and unpredictable results, accurate prognostic scoring systems are the precondition for optimizing the clinical therapeutic strategy for HBV-ACLF patients. Although the MELD score has been verified to promote the allocation of donor livers, this is still not an ideal indicator for HBV-ACLF patients.13 Although several prognostic models have been developed to predict the HBV-ACLF short-term (such as 30-day and 90-day) mortality,14–17 including the 30-day HBV-ACLFD model previously developed by the investigators,9 there is still a lack of a prognostic model to predict the mid-term (such as 180-day) mortality of patients with HBV-ACLF. In the present study, the LR and CART models were developed to predict the 180-day mortality of patients with HBV-ACLF. Both the LR and CART models could be used as medical decision-managing tools by clinicians.
In the present study a new LR model was established, which included age, TBIL, L, M, HE and PTA as prognostic factors for the 180-day mortality. The AUROC for this prognostic model was significantly higher than that for the MELD score. Except for LR, a novel CART model was also developed to predict the 180-day outcome of HBV-ACLF patients. In the present study, the CART model included age, TBIL, PTA, L, M, N and BUN. These seven potential variables were the important predictors for the survival of HBV-ACLF patients. Both the LR and CART models appeared to perform better than the MELD score. Meanwhile, the investigators also made the LR models easier to use in clinic by drawing a nomogram.
Compared to traditional models, the CART model has many advantages. First, the CART can conduct highly biased clinical data, and reveal the complicated relationships among different variables. This generates a clearly visible decision tree that contains many binary splits, which are more accessible and convenient for clinical applications. Second, in the present study, the CART model had better predictive accuracy, when compared to the MELD score. At present, some organ function-based scoring systems, including the chronic liver failure-sequential organ failure assessment score, the CLIF Association ACLF score,18 the chronic liver failure-sequential organ failure assessment score (, the Acute Physiology and Chronic Health Assessment II score,19 and the sequential organ failure assessment score, have also been used to make predictions for the mortality of ACLF. Compared to these scoring systems, the CART model is much easier to apply. Third, the CART model is more convenient for LT patients, in terms of estimating the risk stratification. Shi et al.20 used a CART model to validate the 3-month mortality of patients with HBV-ACLF. This revealed the profit of the CART model to predict the HBV-ACLF risk stratification.
However, there were some limitations in the present study. The present study was a single-center retrospective study that mostly involved male patients. However, it was not easy to collect more data of the mid-term outcome of HBV-ACLF patients. Hence, further validation is needed through a larger study.
Conclusions
The LR and CART model was derived to predict the 180-day clinical outcomes in HBV-ACLF patients. These models can be helpful for doctors who need to make vital clinical decisions for patients with HBV-ACLF. However, larger multicenter studies and further evaluations are needed.
Abbreviations
- ACLF:
acute-on-chronic liver failure
- ALT:
alanine transaminase
- AST:
aspartate transaminase
- AUROC:
area under the receiver operating characteristic curve
- BUN:
urea nitrogen
- CART:
classification and regression tree
- CI:
confidence interval
- HBsAg:
hepatitis B virus surface antigen
- HBV:
hepatitis B virus
- HE:
hepatic encephalopathy
- HGB:
hemoglobin
- INR:
international normalized ratio
- L:
lymphocyte
- LT:
liver transplantation
- LR:
logistic regression
- M:
monocyte
- MELD:
model for end-stage liver disease
- N:
neutrophil
- OR:
odds ratio
- PLT:
platelet
- PTA:
prothrombin activity
- RBC:
red blood cell
- SD:
standard deviation
- TBIL:
total bilirubin
- WBC:
white blood cell
Declarations
Data sharing statement
No additional data are available.
Funding
The study was supported by the National Natural Science Foundation of China (No. 81470888) and (No. 82002461), the Medjaden Academy and Research Foundation for Young Scientists (No. MJR20211110), and the Fund for Fostering Young Scholars of Peking University Health Science Center (No. BMU2021PY010).
Conflict of interest
The authors have no conflict of interests related to this publication.
Authors’ contributions
Conception and design of the study (QM), collection of the data (JW, ZW), analysis of the data (RX, JY), and writing of the paper (RX, JY).