Patients diagnosed with ACLF, as proposed by the Asian Pacific Association for the Study of the Liver (Sarin et al., 2019), were recruited from the general clinic population from May 6, 2017, to April 6, 2022. Inclusion criteria were as follows: (a) 18–80 years of age; (b) chronic hepatitis B infections; (c) PE or PE+ double plasma molecular absorb system (DPMAS) or PE+ hemofiltration ALSS support; (d) no medical history of LT. Exclusion criteria were as follows: (a) hepatocellular carcinoma, severe extrahepatic diseases, or other tumors; (b) cardiopulmonary disease or bleeding; (c) hypersensitivity reactions to plasma, human albumin, or heparin. This study also included patients who presented without ACLF at admission but later developed ACLF before starting ALSS therapy.

Many artificial liver treatment modes are often used in combination with clinical practice. To achieve the purpose of this study while complying with real-world evidence, we enrolled HBV-ACLF patients who received PE-centered ALSS therapy (PE alone or in combination with hemofiltration, DPMAS, or plasma perfusion). On the other hand, there were often multiple ALSS procedures in one patient. Under these circumstances, we used propensity score-match (PSM) to eliminate selection bias. PSM between the death group and the survivor group was developed according to the estimated propensity scores, which were calculated using a logistic regression model for the existence of ALSS as a function of the following parameters: sex, number of ALSS sessions, and modes of ALSS sessions. These confounding factors can be a primary source of heterogeneity in predicting in-hospital outcomes (Deidda et al., 2019). The matching was performed using nearest-neighbor matching within 0.2 standard deviations of pooled propensity scores.

All procedures involving human participants accorded with the ethical standards of the institutional and national research committee and with the 2013 Declaration of Helsinki and its later amendments or comparable ethical standards. The Bioethics Committee of the Third Central Hospital of Tianjin approved the study. Appropriate approvals were obtained from patients or their legal surrogates before enrollment.

During hospitalization, all patients received standard medical therapy. Complications were also treated (Xie et al., 2020), including a high-calorie diet, nucleoside analogs for HBV DNA-positive patients, sodium restriction (when necessary), diuretics and paracentesis with albumin infusion for ascites, and prophylactic antibiotics for bacterial infections. After the patient’s condition was stabilized, ALSS was performed every two to three times per week and was discontinued if bleeding or circulatory complications occurred (Chen et al., 2016).

Coagulation assays prothrombin activity (PTA), international normalized ratio (INR), activated partial thromboplastin time (APTT), fibrinogen (FBG), thrombin time(TT), antithrombin (AT) and fibrinogen degradation products (FDP) were performed using a Stago automated coagulation analyzer (STA Compact Max^®, France) according to the manufacturer’s instructions. D-dimer was measured using a fluorometric method on a GP1600 (Getein Biotech). A Roche Cobas 8000 c701 automated chemistry analyzer (Roche Diagnostics, Mannheim, Germany) was used to measure total protein (TP), albumin (ALB), globulin (Glb), total bilirubin (TBil), direct bilirubin (DBil), indirect bilirubin (IBil), alanine aminotransferase (ALT), aspartate aminotransferase (AST), mitochondrial aspartate transaminase (m-AST), alkaline phosphatase (ALP), gamma-glutamyl transpeptidase (GGT), total bile acid (TBA), cholinesterase (CHE), prealbumin (PA), creatinine (CR), Na, Urea and uric acid (UA). Neutrophils (NEUT), neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR) and platelets (PLT) were measured using the Advia 2120 hematology system (Siemens).

Data were collected according to case report forms. All clinical parameters and laboratory biomarkers were divided into four subgroups (baseline; before ALSS; after ALSS; change ratio) according to the patient’s state. Parameters were measured early in the day after admission to generate the baseline group. Laboratory biomarkers tested in the early morning before each session of ALSS treatment were the before-ALSS group; those tested after each ALSS treatment were the after-ALSS group. The change ratio = (post-treatment - pre-treatment level)/pre-treatment. The predictive models (baseline, pre-model, post-model, change model) were built using the data in each subgroup according to the in-hospital outcomes.

IBM SPSS Statistics for Windows software package, version 25 (IBM Corp., Armonk, N.Y., USA) and GraphPad Prism version 9.3.1 (GraphPad Software, Inc., San Diego, California) were used for statistical analysis. Metric data were expressed as mean values with standard deviation or median (p25, p75). The qualitative data were expressed as frequency and composition. Because ALSS sessions for a single patient might correlate with one another more closely than the sessions between patients, GEE models were used to manage the clustering. Univariate and multivariate GEE analyses were performed. A forward stepwise selection method was used to select variables for the multivariate GEE model. A quasi-likelihood under the independence model criterion (QIC) was calculated; the model with the smallest QIC was the most parsimonious (Pan, 2001). The standardized regression coefficient β was calculated as a tracking coefficient (Twisk et al., 1996). Discrimination and calibration are critical aspects characterizing the performance of a prediction model (Steyerberg et al., 2010; Moons et al., 2015). Calibration and discrimination were performed to determine whether the GEE models were adequate for their purpose. The discrimination was assessed by comparing the receiver operating characteristic (ROC) curves using the z-test (Delong’s method (DeLong et al., 1988). The discriminative power is excellent if the area under the ROC curve is > 0.80, very good if > 0.75, and good (acceptable) if > 0.70 (Moons et al., 2015). Using GraphPad Prism software, calibration plots were used to compare the mean predicted probability and the mean observed outcome. Results were displayed graphically with predicted outcome probabilities (Y-axis) plotted against observed outcome frequencies (X-axis). The goodness-of-fit test measures the differences between observed and expected outcomes over deciles (ten groups of patients) of risk (Moons et al., 2015; Nezic, 2020). Perfect predictions should be on the 45° line (Steyerberg et al., 2010). A P-value < 0.05 was considered statistically significant.

The design is summarized in Figure S1. We included 110 consecutive patients who died in the hospital (death group) and 110 PSM patients who survived to discharge (survivor group). After PSM, in the death or survivor groups, 22 patients underwent one session of PE-centered ALSS treatment; 18 underwent two; 26 underwent three; 11 underwent four; 15 underwent five; 18 underwent six sessions. There were 726 sessions of PE-centered ALSS treatment for 220 patients, with an average of 3.3 sessions per patient (1–6 sessions per patient).

One hundred ten patients survived in-hospital, and 110 did not. A comparison of baseline characteristics between the death and survivor group density is displayed in Table 1. As shown in the table, HBV-ACLF patients in the survivor group were significantly younger and had substantially lower levels of pulse (P=0.001), INR (P<0.0001), APTT(P<0.0001), D-dimer (P=0.001), TBil (P=0.02), DBil (P=0.04), IBil (P=0.025), CR (P=0.022), MELD scores (P=0.002), and 3-month mortality based on MELD scores (P=0.001) than non-survivors; they also had higher levels of PTA (P<0.0001), ALT (P<0.0001), AST (P=0.008), m-AST (P=0.006), and CHE (P<0.0001) at admission than non-survivors.

APTT(P<0.0001), Age (P<0.0001), D-dimer (P=0.011), ALT(P<0.0001), pulse (P=0.001), and TBil (P=0.016) were entered into a multivariate GEE model (Table 2), which was the most parsimonious and had the smallest QIC (211.55).

The ROC curves and calibration plots are shown in Figure 1. The discriminative power of the multivariate GEE model was excellent (AUC=0.872) (Figure 1A), better than APTT (AUC=0.707), Age (AUC=0.711), D-dimer (AUC=0.742), ALT(AUC=0.650), pulse(AUC=0.672) or TBil (AUC=0.593) alone (Table S1). Calibration of the multivariate model (Y = 0.9653*X + 0.01733, (R2 = 0.9677, P<0.0001)) (Figure 1B) showed better agreement between the predicted and observed probabilities than the univariate models (Figure S2).

The laboratory biomarkers before ALSS treatment were compared between the 363 survivors and 363 non-survivors. Relationships between laboratory biomarkers before ALSS treatment and in-hospital outcomes is shown in Table S2, HBV-ACLF patients in the survivor group had substantially lower levels of INR (P=0.006), APTT(P<0.0001), D-dimer (P=0.006), TBil (P<0.0001), DBil (P<0.0001), IBil (P<0.0001), DBil/TBil (P=0.001), CR (P=0.032), Urea (P=0.004), UA (P=0.002), NEUT (P=0.007), MELD Sore (P<0.0001), and 3-month mortality based on MELD scores (P<0.0001), but had higher levels of PTA (P<0.0001), AT (P=0.004), TP (P<0.0001), Glb (P<0.0001), GGT (P=0.004), CHE (P<0.0001), PA (P<0.0001), Na (P=0.006) and PLT (P<0.0001) before ALSS treatment than patients who did not survive.

TBil (P<0.0001), Glb (P<0.0001), APTT (P=0.01), GGT(P=0.003) and UA (P<0.0001) were entered in multivariate GEE model (Table 2), which was the most parsimonious with the smallest QIC (622.11).

The ROC curves and calibration plots are shown in Figure 1. The discriminative power of the multivariate GEE model was excellent (AUC=0.885) (Figure 1C), better than TBil (AUC=0.808), Glb (AUC=0.716), APTT (AUC=0.709), GGT(AUC=0.658), or UA (AUC=0.614) alone (Table S3). Calibration of the multivariate model (Y = 0.9878*X + 0.006093 (R2 = 0.9715, P<0.0001)) (Figure 1D) showed better agreement between the predicted and observed probabilities than the univariate models (Figure S3).

The laboratory biomarkers after ALSS treatment were compared between the 363 survivors and 363 non-survivors. Relationships between laboratory biomarkers after ALSS treatment and in-hospital outcomes is shown in Table S4, HBV-ACLF patients in the survivor group had substantially lower levels of INR (P<0.0001), APTT(P=0.001), D-dimer (P=0.017), TBil (P<0.0001), DBil (P<0.0001), IBil (P<0.0001), DBil/TBil (P<0.0001), CR (P<0.0001), Urea (P<0.0001), UA (P=0.001), NEUT (P<0.0001), NLR (P<0.0001), MELD scores (P<0.0001), and 3-month mortality based on MELD scores (P<0.0001), but had higher levels of PTA (P=0.001), TP (P<0.0001), Glb (P<0.0001), GGT (P=0.007), PA (P<0.0001), Na (P<0.0001) and PLT (P<0.0001) after ALSS treatment than HBV-ACLF patients who did not survive.

MELD Sore (P<0.0001), TBil (P<0.0001), Glb (P<0.0001), PA (P<0.0001) and UA (P<0.0001) were entered into multivariate GEE model (Table 2), which is the most parsimonious with the smallest QIC (530.88).

The ROC curves and calibration plot are shown in Figure 2. The discriminative power of the multivariate GEE model was excellent (AUC=0.888) (Figure 2A), better than the MELD Sore (AUC=0.851), TBil (AUC=0.803), Glb (AUC=0.797), PA (AUC=0.784) or UA (AUC=0.659) alone (Table S5). Calibration of the multivariate model (Y = 1.036*X + 1.670e-005 (R2 = 0.9693, P<0.0001)) (Figure 2B) showed better agreement between the predicted and observed probabilities than the univariate models (Figure S4).

Relationships between the rate of change ((Post-ALSS-Pre-ALSS)/Pre-ALSS) of each laboratory biomarker and in-hospital outcomes are shown in Table S6. The change ratio was compared between the 363 survivors and 363 non-survivors, HBV-ACLF patients in the survivor group had substantially lower levels of INR (P=0.002), DBil/TBil (P=0.036), ALT(P<0.0001), AST(P=0.002), m-AST(P=0.046), m-AST/AST(P<0.0001), GGT(P=0.001), CR (P<0.0001), Urea (P<0.0001), NEUT (P=0.007), MELD scores (P<0.0001), and 3-month mortality based on MELD scores (P<0.0001), but had higher levels of PTA (P=0.040), D-dimer (P=0.013), ALB (P<0.0001), TP (P<0.0001), Glb (P<0.0001) and PLT (P=0.039) than HBV-ACLF patients who did not survive.

MELD Sore (P<0.0001), CR (P<0.0001), TP (P<0.0001), AST (P=0.024), m-AST/AST(P<0.0001), D-dimer (P=0.015), NEUT (P=0.041) and DBil/TBil (P=0.014) were entered into multivariate GEE model (Table 2), which is the most parsimonious with the smallest QIC (843.09).

The ROC curves and calibration plots are shown in Figure 2. The discriminative power of the multivariate GEE model was very good (AUC=0.779) (Figure 2C), better than the MELD Sore (AUC=0.683), CR (AUC=0.622), TP (AUC=0.613), AST (AUC=0.617), m-AST/AST (AUC=0.501), D-dimer (AUC=0.503), NEUT (AUC=0.592) or DBil/TBil (AUC=0.550) alone (Table S7). Calibration of the multivariate model (Y = 0.9381*X + 0.03090 (R2 = 0.9546, P<0.0001)) in Figure 2D showed better agreement between the predicted and observed probabilities than the univariate models (Figure S5).