ABIDE: An Accurate Predictive Model of Liver Decompensation in Patients With Nonalcoholic Fatty Liver-Related Cirrhosis

Data Source/Study Design and Participants

The derivation cohort for the model originated from the NAFLD progression consortium, which is an international initiative of six centers aiming to study the natural history and prognosis of biopsy-proven NAFLD with advanced fibrosis. The design and outcomes of this multicenter cohort study have been described.^{(9, 20)} Briefly, 458 patients with NASH and bridging fibrosis or cirrhosis were assessed from April 1995 to November 2013. From this cohort, only those with cirrhosis (n = 299) were included, originating from Spain (two centers [n = 184]), Australia (two centers [n = 116]), Hong Kong (one center [n = 82]), and Cuba (one center [n = 76]). Follow-up extended to December 2016 for the derivation cohort.

All patients in the derivation cohort had provided informed consent for their medical information to be used for research with ethics committee approval from the participating centers.

The validation cohort consisted of patients retrospectively identified from Indiana University Hospital with biopsy-proven NASH cirrhosis (n = 244) assessed between April 2004 and November 2016. Follow-up extended to December 2017 for the validation cohort. The study was approved by the local institutional review board with waiver of consent. The majority of patients in the validation cohort have been described in published papers.^{(21, 22)}

Inclusion and Exclusion Criteria

Patients were included on the basis of biopsy-proven NAFLD cirrhosis based on the NASH Clinical Research Network scoring system. The following exclusion criteria were applied to the derivation set: previous history of hepatic decompensation, Model for End-Stage Liver Disease (MELD) score ≥15, CTP ≥7, excess alcohol consumption (average >20 g/day for men or >10 g/day for women in the last 2 years or during follow-up), prior bariatric surgery, secondary causes of NAFLD or other chronic liver diseases, albumin <3.0 g/dL, total bilirubin > 2.0 mg/dL, international normalized ratio (INR) > 1.7, platelets <100,000 mm³, a diagnosis of HCC at inclusion or during first 6 months of follow-up, known human immunodeficiency virus, type 1 diabetes, and life-threatening diseases with reduced life expectancy of less than 12 months. The exclusion criteria for the validation set were similar apart from a higher cutoff for bilirubin (>2.5 mg/dL) and lower cutoff for platelet count (<40,000 mm³) to include a wider spectrum of patients with cirrhosis. All patients included in the study were CTP A with no previous episodes of clinical decompensation. Patients with varices at baseline were included; however, no patient had undergone variceal banding before inclusion.

Outcomes

The primary outcome was the first event of hepatic decompensation, defined by the occurrence of any of the following: ascites (identified or confirmed by abdominal ultrasound), upper gastrointestinal bleeding secondary to portal hypertension (confirmed by endoscopy in the presence of gastroesophageal varices or portal hypertensive gastropathy), or hepatic encephalopathy (established by clinical parameters, neuropsychological tests, or electroencephalogram). If multiple decompensated events occurred in the same patient, the patient was regarded as having one hepatic decompensation, and only the earliest event was adopted for the time-to-event analysis. Outcomes were assessed by local hepatologists after review of the patient clinical history, examination findings, and investigations.

Follow-Up

Patients were followed from the day of liver biopsy until the occurrence of death, liver transplant, or last visit. The outcome was evaluated by an experienced hepatologist in each center every 3-6 months. At each visit, a medical history, physical examination, and standard laboratory tests were performed. All patients underwent initial and subsequent upper gastrointestinal endoscopy following the recommended guidelines for patients with cirrhosis.⁽²³⁾ Additional endoscopies were conducted at the discretion of each investigator. Patients lost to follow-up (n = 13) were censored at the last day they were known to be alive.

Model Variables

The following variables were prospectively collected in the derivation and validation cohorts at baseline at the time of liver biopsy: demographics (age, sex, race/ethnicity), medical history (hypertension, type 2 diabetes, medications), anthropometric variables (weight, body mass index), serum biochemical parameters (fasting glucose, insulin, hemoglobin A1c [HbA1], triglycerides, total cholesterol, high-density and low-density lipoprotein cholesterol, aminotransferases, bilirubin, albumin, INR, creatinine), and others (alcohol consumption and cigarette smoking, presence of gastroesophageal varices). Alcohol consumption was classified into nondrinkers (no alcohol intake) or moderate drinkers (1-70 g/week in women, 1-140 g/week in men). The following scores were computed at baseline: MELD score, CTP score, NFS, fibrosis-4 (FIB-4) score, albumin-bilirubin (ALBI), and ALBI-FIB-4.^(24-28)

Statistical Analysis

Baseline characteristics in derivation and validation cohorts were summarized and compared. Values were reported using percentage for categorical variables and as mean and standard deviation for continuous variables. The chi-squared test was applied to categorical variables, and the Wilcoxon signed-ranks test and analysis of variance test were used to compare means between groups.

Model Development

To estimate the association of baseline variables with the first event of hepatic decompensation, competing risk regression models were performed according to the method of Fine and Gray in the derivation cohort. The occurrence of nonliver death was considered as a competing event for hepatic decompensation. Subhazard ratios (sHR) with 95% confidence intervals (CIs) were calculated for each predictor. Those variables with values of P < 0.10 in univariate analysis were entered into a multivariate competing risk regression analysis. Backward stepwise selection procedures were used for variable selection.

For each multivariate regression analysis, the concordance statistics and Bayes information criteria (BIC) and Akaike information criteria (AIC) were used for the selection and final criteria for inclusion of variables in the model, respectively. Both BIC and AIC are measures of likelihood in which lower values indicate a better fit and a penalty is paid for inclusion of a new variable. Thus, this method contributes to select and design not only the most plausible variables but the most parsimonious predictive model.

Model Accuracy

Time-dependent area under the curve (tAUC) in the presence of censored data were used to evaluate model accuracy for the prediction of hepatic decompensation.⁽²⁹⁾ Model accuracy was estimated at 5 and 10 years of follow-up. The optimal model cut point was selected using the Youden and Liu criterion.⁽³⁰⁾ After model cut point selection, the cumulative incidence of hepatic decompensation for each group was reported using the cumulative incidence function after competing risk regression analysis. Differences between groups were reported by the Fine and Gray method.

Model Validation

The accuracy of the model developed in the derivation set was assessed using time-dependent receiver operating curves at 5 and 10 years of follow-up. The cumulative incidence of hepatic decompensation at 5 and 10 years was reported according to decile cut points of the model and compared using the Fine and Gray method.

Model Calibration

Two methods were used to assess model calibration (the extent to which the predicted probabilities and actual probabilities agree): the Hosmer-Lemeshow (H-L) calibration statistic as modified by D’Agostino and Nam⁽³¹⁾ and the calibration slope.⁽³²⁾ H-L is a measure of calibration, and ideally no significant differences should exist between observed and predicted probabilities. Calibration slope is a robust method for censoring and ideally takes a value of 1. Both methods were calculated and reported in both derivation and validation cohorts.

Overall Model Performance

Overall model performance was tested using the Brier score, which integrates discrimination and calibration at the same time and reflects the mean squared difference between the real outcome and the predicted probability.⁽³²⁾ Additionally, the accuracy of the model was compared with MELD, CTP, NFS, FIB-4, ALBI, and ALBI-FIB-4 scores using tAUC values. To explore the performance of the model in different settings, subgroup analysis (diabetes vs. no diabetes, age <55 vs. age ≥55, female vs. male, presence or absence of gastroesophageal varices, use of beta blockers at baseline) was conducted in the whole cohort (n = 543). Age 55 was chosen as the median age of the entire cohort.

Statistical analyses were carried out using STATA software, release 14.2. All CIs, significance tests, and resulting P values were two-sided, with an alpha level of 0.05.

Derivation and Validation Cohorts: Participant Overview

Of 512 participants assessed for eligibility in the derivation cohort, 299 patients with compensated cirrhosis fulfilled the inclusion criteria, whereas 244 of 346 patients were eligible in the validation cohort. A flow diagram outlining the patient selection is shown in Fig. 1.

At baseline, patients in the derivation cohort were slightly older compared with those in the validation cohort (56.8 vs. 54.2 years, P < 0.001) and were less likely to be White (60.5% vs. 100%), as shown in Table 1. HbA1c levels and the prevalence of diabetes were similar between groups, whereas the validation cohort had higher baseline total cholesterol and low-density lipoprotein cholesterol levels. In comparison with the derivation cohort, patients in the validation cohort were more likely to have a history of hypertension (84% vs. 62%), esophageal varices (43% vs. 28%), and higher body mass index (37.2 vs. 32.1 kg/m²) and levels of fasting glucose (164 vs. 146 mg/dL). Liver function and MELD and CTP scores were similar between cohorts. Baseline characteristics by country of enrollment are shown in Supporting Table S1.

During a median follow-up of 5.1 (range 2.6-8.4) and 5.4 years (range 4.5-8.8), 81 patients (27%) and 132 (54%) developed an initial hepatic decompensation event in the derivation and validation cohorts, respectively. Three patients in the derivation cohort and 15 in the validation cohort died of non-liver-related causes and were analyzed as competing events. Specific types of liver complications were ascites (n = 60), variceal hemorrhage (n = 16), and encephalopathy (n = 5) in the derivation cohort and ascites (n = 69), encephalopathy (n = 40), and variceal hemorrhage (n = 23) in the validation cohort.

Predictors of Hepatic Decompensation

The following baseline variables (outlined in Table 2) were associated with future hepatic decompensation in the derivation cohort (n = 299) using univariate competing risk regression analysis: White and Latino ethnicity, presence of esophageal varices, type 2 diabetes, INR, albumin, total bilirubin, aspartate aminotransferase (AST)/alanine aminotransferase (ALT) ratio, and platelet levels. Multivariate analysis confirmed the presence of esophageal varices (sHR, 2.0; 95% CI, 1.18-3.61; P < 0.001), type 2 diabetes (sHR, 2.2; 95% CI, 1.20-3.58; P < 0.005), INR (sHR, 6.6; 95% CI, 2.87-10.4; P < 0.001), total bilirubin (sHR, 1.3; 95% CI, 1.14-1.67; P < 0.005), and AST/ALT ratio (sHR, 2.2; 95% CI, 1.50-3.45; P < 0.001) as independent predictors of hepatic decompensation.

Model Development

The contribution of each variable identified on multivariate analysis to predict hepatic decompensation was assessed using tAUC and information criterion (see Table 3). Each variable incrementally increased the model accuracy to a maximum tAUC of 0.79 when all five variables were included. The combination of five variables also showed the lowest information criterion (AIC 395 and BIC 406), demonstrating that the model was parsimonious and not overfitted. The final model (named ABIDE) consisted of AST/ALT ratio, Bilirubin, INR, type 2 Diabetes, and Esophageal varices, with a formula of (2.003 × INR + 0.824 × AST/ALT ratio + 0.821 × [Type 2 diabetes: 0 if absent, 1 if present] + 0.806 × [esophageal varices: 0 if absent, 1 if present] + 0.332 × total bilirubin).

Model Accuracy and Cut-Offs

The accuracy of the ABIDE model in the derivation cohort to predict events at 5 years was good (tAUC, 0.80; 95% CI, 0.73-0.84) and similar at 10 years (tAUC, 0.78; 95% CI, 0.71-0.83), shown in Table 4 and Fig. 2. Both Youden and Liu criterion identified the optimal cut point as 4.1 based on sensitivity and specificity of 79% and 85%, respectively. Patients with an ABIDE score ≥4.1 had a higher risk (sHR, 6.7; 95% CI, 4.0-11.2; P < 0.001; Fig. 3A) and higher 5-year cumulative incidence of hepatic decompensation compared with those patients with a score <4.1 (37% vs. 6%, P < 0.001). The 10-year cumulative incidence remained divergent (76% vs. 18%, P < 0.001), as shown in Fig. 3A. The mean time to liver-related complications in the derivation cohort was 3.8 years (95% CI, 3.5-4.1) and 6.7 years (95% CI, 5.1-5.9) in patients with ABIDE scores ≥4.1 and <4.1, respectively (P = 0.01). The cumulative incidence of hepatic complications increased with each ABIDE score decile with a 2.1% 5-year incidence in the lowest decile (<2.99), compared with 58.0% in the highest decile (>5.44), as outlined in Supporting Table S2 and Fig. 4.

Model Validation

In the validation cohort (n = 244), the accuracy of the ABIDE model remained satisfactory at 5 years (tAUC, 0.78; 95% CI, 0.73-0.84) and 10 years (tAUC, 0.76; 95% CI, 0.70-0.82), as shown in Table 4, Fig. 2, and Supporting Table S3. Patients with a score ≥4.1 had a higher risk of hepatic decompensation (sHR, 4.3; 95% CI, 2.5-7.4; P < 0.001), whereas those with a score <4.1 had minimal risk (sHR, 0.22; 95% CI, 0.13-0.39; P < 0.001), as shown in Fig. 3B. Time to hepatic decompensation was shorter in those with ABIDE score ≥4.1 (mean 4.2 years; 95% CI, 3.9-4.6) compared with those with ABIDE score <4.1 (mean 6.5 years; 95% CI, 4.6-5.3; P = 0.03).

Comparison With Other Models and Sensitivity Analysis

Other noninvasive fibrosis models (NFS, FIB-4) and prognostic cirrhosis models (MELD, CTP, ALBI, and ALBI-FIB-4) were also predictive of future decompensation (Supporting Figs. S2-S11). However, the ABIDE model had better discriminative power at 5 and 10 years compared with NFS, FIB-4 and also when compared with MELD, CTP, ALBI and ALBI-FIB-4 scores, in both the derivation and validation cohorts (Table 4). Notably, less than 10% of the validation cohort had low FIB-4 or NFS scores, demonstrating a limited ability to discriminate individuals with favorable prognosis, whereas <5% of the cohort had a high ALBI score, demonstrating limited ability to detect individuals who were at greatest risk of decompensation.

The accuracy of the ABIDE score was not impacted by the presence of type 2 diabetes, age, or sex. The accuracy of the model was marginally lower in the presence of varices (Supporting Table S4). A subanalysis in patients with varices showed equivalent accuracy of the model at 5 and 10 years in patients both with and without beta blockers at baseline (Supporting Table S5).

Model Calibration and Performance

Using the H-L goodness-of-fit test, the model calibrated well with no differences between the observed and predicted probability of developing hepatic decompensation at 10 years in both derivation (H-L 11.0; P = 0.23) and validation (H-L 6.73; P = 0.56) cohorts. Additional calibration (calibration slope, calibration plot) and overall accuracy of the model by means of the Brier score were also evaluated (Supporting Fig. S1). The ABIDE model demonstrated good calibration in both the derivation (calibration slope, 1.08) and validation cohorts (calibration slope, 1.00). Furthermore, the Brier score was 0.12 and 0.07 for derivation and validation cohorts, respectively, showing satisfactory overall performance of the model.