Basic characteristics
A total of 271 patients accepted LT at our institution between March 1st, 2012 and December 31st, 2019, of which 132 patients achieving the APASL criteria of ACLF were eligible for this study. The flowchart for inclusion is shown in Figure 1. Nineteen ACLF patients (14.4%) died within 90 days after liver transplantation. The causes of death were upper gastrointestinal bleeding (n=7, 36.8%), multiple organ dysfunction syndrome (MODS) (n=6, 31.5%), severe pneumonia (n=3, 15.8%), abdominal infection (n=1, 5.3%), heart arrest (n=1, 5.3%) and cerebral herniation (n=1, 5.3%).
Comparison of clinical data between the survival group and the death group displayed no statistical difference in terms of age, sex, etiology, total bilirubin (TBiL), creatinine (Cr), international normalized ratio (INR), hepatic encephalopathy (HE), prothrombin time (PT), urea, albumin, direct bilirubin (DBIL), hemoglobin, platelet count, white blood cells (WBC), neutrophil count (NEUT) and lymphocyte count (LYM) (Table 1).
Patient survival analysis
Based on traditional formulae scores, the cut-off values of the important score-related parameters (Cr, INR, TBiL, Plt and WBC) were determined, which were applied for Kaplan-Meier (KM) survival analysis. The results indicated that the post-transplant mortality among ACLF patients was significantly associated with higher values of Cr (Cr≥132. 6 μmol/L) and INR (INR≥2.0) (P<0.05). The differences among other factors were not statistically significant (Figure 2).
The model of Cox regression was applied to identify the independent risk factors for short-term outcome. Univariate Cox regression analysis revealed that Cr (P=0.001) and INR (P=0.034) were poor prognostic indicators for ACLF patients following LT. Factors with P < 0.15 were further analyzed in multivariate cox regression. The results of multivariate analysis displayed that Cr (HR, 1.006; 95% CI, 1.001–1.011; P = 0.030) and INR (HR, 1.454; 95% CI, 1.100–1.921; P = 0.009) were independent prognostic markers of short-term outcome (Table 2).
Predictive value of conventional models
In comparison to those in the survival group, the scores of conventional models, including the MELD score, ABIC, CLIF-C OFs, CLIF-SOFAs, and CLIF-C ACLFs were higher in the death group. However, only MELD score (P=0.01) and CLIF-C ACLFs (p=0.04) showed significance between the survival group and death group (Figure 2 and Table 3). According to the ROC analysis, the area under a receiver operating characteristics (AUROC) of MELDs (AUROC: 0.704) was higher than those of ABIC (AUROC: 0.607), CLIF-C OFs (AUROC: 0.606), CLIF-C ACLFs (AUROC: 0.653), and CLIF-SOFAs (AUROC: 0.633) for prediction of the 90-day outcome in ACLF patients following LT (Figure 3).
Predictive value of ML models
Four ML models (SVM, LR, MLP and RF) were trained and compared to improve the prediction performance. All ML models had good performance in terms of AUROC (Figure 3), higher than those of conventional models. Among the ML models, the RF model had the highest AUROC of 0.94. The final result of RF model was derived from majority voting by the ten trees. The AUROCs of SVM, LR and MLP were 0.81, 0.83 and 0.89, respectively. Cr, INR, etiology, DBiL, LYM and NEUT were chosen to develop the SVM and LR models. The coefficients of parameters in the models are described in Table 4. In the two models, the coefficients of Cr and INR are negative, indicating a negative correlation with the prognosis in ACLF patients following LT. The other parameters including etiology of liver disease, DBIL, LYM and NEUT, are positively associated with transplant outcome.