Modeling Outcomes in Children With Biliary Atresia With Native Liver After 2 Years of Age .

Study Population

Two prospective registries of infants and children with BA are included in ChiLDReN: A Prospective Database of Infants with Cholestasis (PROBE; clinicaltrials.gov NCT00061828) and the Biliary Atresia Study in Infants and Children (BASIC; clinicaltrials.gov NCT00345553). PROBE has enrolled infants ≤180 days of age with neonatal cholestasis since June 1, 2004, as described.⁽14, 17⁾ Baseline data and biospecimens are collected at the time of KPE; follow‐up visits occur periodically until 18 months of age and then annually from age 2 to 20 years, unless LT, death, or loss to follow‐up occur. Since May 1, 2006, BASIC has enrolled participants with BA who are over 6 months of age, with BA confirmed after review of biopsy, radiographic, and surgical reports, as described.⁽14⁾ Baseline data are collected at enrollment and annually until age 20 years, LT, or loss to follow‐up.

Institutional Review Board approval was obtained from each ChiLDReN site and the Data Coordinating Center; parents or legal guardians of infants provided written informed consent. The current study analyzed data from participants with BA with KPE in PROBE and BASIC with enrollment at or before age 2 years, with SNL at age 2, and follow‐up data collected thereafter.

Outcome Variables

We analyzed two time‐to‐event outcomes: (1) time from age 2 years to LT or death (LTD) and (2) time from age 2 years to a composite outcome of a sentinel event (SE), defined as the first report of ascites, GI bleeding, or hepatopulmonary syndrome (HPS), whichever happened first.

LTDs are recorded on data collection forms at the time of the event. SE data are collected at protocolized follow‐up visits. Study definitions of SEs are standardized within the PROBE and BASIC protocols. The presence of ascites was determined by either physical exam or patient receipt of diuretics. HPS required documentation of hypoxia plus evidence of intrapulmonary shunting by bubble contrast echocardiography with agitated saline.⁽18⁾ GI bleeding was defined as hematemesis, hematochezia, or melena with endoscopic documentation of actively bleeding esophageal or gastric varices or visualization of esophageal varices in the absence of any other identifiable cause of hemorrhage.

Candidate Predictors

Candidate predictors analyzed for the first outcome of LTD included 14 variables: platelet count (10³/mm³), total bilirubin (TB, mg/dL), aspartate aminotransferase (AST, U/L), alanine aminotransferase (ALT, U/L), prothrombin time international normalized ratio (INR), albumin (g/dL), gamma‐glutamyltransferase (GGT, U/L), AST to platelet ratio index (APRI), height failure, weight failure, and history of ascites, cholangitis, GI bleed, and splenomegaly.

Baseline laboratory and anthropologic measurement values were obtained from the closest visit within 6 months before or after 2 years of age. Medical history variables were reported before age 2 years. Height and weight failure were defined as age‐ and sex‐adjusted z score <−2, calculated using SAS macros provided by the Centers for Disease Control and Prevention (https://www.cdc.gov/nccdphp/dnpao/growthcharts/resources/sas.htm). Splenomegaly was defined as palpable >2 cm below the costal margin. Diagnosis of cholangitis required presence of fever of >38°C without other obvious clinical source of infection as well as a combination of clinical findings. Clinical findings included new onset of acholic stools; right upper quadrant pain or tenderness; both elevation of direct bilirubin by 25% and at least >1 mg/dL above previous baseline; rise in 2 or more of AST, ALT, and alkaline phosphatase; GGT to 1.5 times the upper limit of normal or >25% above baseline values if previously elevated; and clinical and biochemical improvement after treatment with antibiotics. History of ascites before 2 years of age was defined as described above and limited to reports after 6 months of age.

For developing the prognostic model of SE, the list of candidate predictors included all the above predictors except for history of GI bleed or ascites.

Statistical Analysis

We used Cox proportional hazard models and cause‐specific hazard competing risk models for modeling the risk of LTD and SE, respectively.⁽19⁾ In the model for SE, LT and death were treated as competing events for SE. Time 0 (baseline) is 2 years of age for both models. Those participants with an SE before age 2 years were excluded from the SE model development because the focus was on new onset after age 2.

Before modeling, distributions of all predictors were examined for extreme values. We used P‐spline functions to explore the potential nonlinear effect of continuous predictors. Covariates with nonlinear effect were transformed to obtain the best model fit. A random forest survival model and a random forest competing risk model were used to explore potential interactions between candidate predictors.⁽20, 21⁾ Twenty‐two multiple imputed data sets were generated to fill in missing covariate values, using a sequence of regression models implemented in IVEWARE, incorporating laboratory measures from the visit closest to 1 year of age.⁽22⁾

To select the best prediction model, we used a stepwise selection procedure, with entry criteria P < 0.10 and stay criteria P < 0.05 on each of the 22 imputed data sets. The selected variables were ordered according to how frequently they were included in the 22 final models. We then selected our final model based on the combined results of the 22 imputed data sets, using Rubin’s rule and the algorithm described as follows.⁽23⁾ We added predictors into the model one at a time in the order of frequency obtained above, starting with the variable with the highest frequency. We stopped this process when the newly added variable had a stay P > 0.05.

In addition to calculating the apparent Harrell’s C‐index, cross‐validation assessed internal validity of the prognostic models.⁽24⁾ For each imputed data set, a 5‐fold cross‐validation was conducted by partitioning the study sample into five equal‐size subsamples. Of these, four were used as the training set and one was used as the test set. Median values of Harrell’s C‐index for the tests set across all imputed data sets were calculated.

Study Population

Between June 2004 and August 2017, 1,151 participants with BA with KPE were enrolled (543 in PROBE; 608 in BASIC) (Fig. 1). Of these, we identified a final cohort of 240 participants with BA (196 from PROBE; 44 from BASIC) enrolled before age 2 years, had an age 2‐year study visit, and had follow‐up after age 2 years. The median age at last observation was 5.1 years (range, 2.0‐13.3 years).

Demographic and baseline laboratory/clinical characteristics of the cohort are listed in Table 1. The median age at KPE was 59 days (interquartile range [IQR], 42‐74 days). The median age at enrollment into PROBE and BASIC was 1.8 months (range, 0.5‐5.0 months) and 16.0 months (range, 4.5‐23.9 months), respectively. None of the included participants in BASIC had been referred to a ChiLDReN site for LT evaluation.

Polysplenia/asplenia was documented in 3.9% of cases, cardiac anomaly in 15.1%, and GI anomalies in 11.2%. A history of cholangitis, ascites, and GI bleed was documented in 43.3%, 17.5%, and 4.6%, respectively. Baseline laboratory values were available from 187 to 229 individuals (depending on the missingness of each collected variable) and notable for a median TB of 0.5 mg/dL (IQR, 0.3‐0.9 mg/dL), platelet count of 205 10³/mm³ (IQR, 134‐271 10³/mm³), GGT of 130 U/L (IQR, 43‐327 U/L), ALT of 83 U/L (IQR, 42‐155 U/L), AST of 84 U/L (IQR, 56‐154 U/L), and APRI of 1.2 (IQR, 0.7‐2.2) (Table 1).

Forty‐eight participants had a reported SE before age 2 years (n = 42 ascites, n = 1 HPS, n = 11 GI bleed, n = 6 with more than one SE) and were excluded from SE modeling, resulting in a total of 192 participants in ChiLDReN used in the SE model. Demographic and baseline characteristics of participants in the SE model compared with those excluded due to a history of SE are listed in Supporting Table S1.

Risk Factors and a Prognostic Model for LTD

LTD was reported in 38 (37 LT and 1 death) study participants, with a cumulative incidence of 23.7% (95% confidence interval (CI), 16.2%‐32.0%) by 10 years of age (Fig. 2A).

A slowing of the rate of LTD was observed over the study follow‐up by comparing the time to event from ages 2 to 5 years (~15% cumulative incidence by age 5 years) to ages 6‐10 years (additional incidence of 8% by age 10 years). Only 5 participants were followed until age 12 years, and no events were observed thereafter, limiting the ability to predict events beyond age 10 years (29 participants). Exploratory random forest analysis found no significant interactions between candidate predictors.

Clinical variables associated with risk of LTD in univariate analysis (Table 2) were history of GI bleed (hazard ratio [HR], 6.56; 95% CI, 3.00‐14.35), history of ascites (HR, 3.38; 95% CI, 1.75‐6.56), splenomegaly (HR, 6.46; 95% CI, 2.29‐18.20), and height growth failure (HR, 2.91; 95% CI, 1.12‐7.56).

All laboratory variables and APRI (Table 2) were significantly associated with risk of LTD (P < 0.05). Of note, doubling the TB (for example, TB rise from 0.5 mg/dL to 1.0 mg/dL) was associated with an HR increase of 2.64 (95% CI, 2.11‐3.29), and doubling the APRI (e.g., 1.2 vs. 0.6) was associated with an HR increase of 1.96 (95% CI, 1.57‐2.44) for the occurrence of LTD.

The stepwise selection approach developed a model (BA LTD) with five variables: TB, platelet count, albumin, and history of ascites or cholangitis (Table 3). The clinical applicability of the BA LTD model is illustrated in the corresponding nomogram (Supporting Fig. S1).

The BA LTD risk equation was evaluated by examination of Kaplan‐Meier curves for LTD, stratified by quartile of the estimated risk in the cohort (Fig. 2B,C). Groups 1 to 4 are ranked by lowest to highest risk. Transplant‐free survival is substantially lower at all ages in group 4, reaching just 53% at age 7 years. In group 3, predicted transplant‐free survival deviates from groups 1 and 2 (89% vs ≥97%) at around 7 years of age. The distribution of each of the variables that contribute to the risk of LTD by quartile provides the characteristics of participants in our cohort at the baseline age of 2 years. Approximately half the subjects in groups 2‐4 (having had a history of cholangitis and the median albumin of all groups) were in a normal range. The median platelet count of <150 10³/mm³ and a median TB of 1.8 mg/dL were noteworthy in the highest risk group 4. A visual representation of predicted probability of LTD with preselected laboratory values along with the clinical variables in the model is shown in Supporting Fig. S2.

The median of apparent Harrell’s C‐index among the 22 imputed data sets for this model was 0.88 (range, 0.86‐0.89), indicating very good discrimination. Internal validation using the jackknife method on the 22 imputed data sets resulted in a median Harrell’s C‐index of the test sets of 0.87 (range, 0.40‐0.99), indicating a very small magnitude of overfitting to the data set.

Risk Factors and a Prognostic Model for SEs

Of 192 participants, a first SE occurred in 27 participants (14 GI bleed, 13 ascites, and 1 HPS; n = 1 with ascites and GI bleed reported on the same date). The cumulative incidence of first SE was 21.5% (95% CI, 14.2%‐29.8%) by 10 years of age (Fig. 3A). Exploratory analysis using random forest analysis did not reveal significant interactions between candidate predictors. Of these 27 participants, 11 underwent LT during the follow‐up period.

The only clinical variable associated with the risk of an SE in the univariate analysis (Table 4) was a history of splenomegaly, with an HR of 2.96 (95% CI, 1.25‐7.01). All laboratory variables, except INR, were associated with risk of SE in the univariate analysis (P < 0.05). Doubling the TB level was associated with an HR of 1.69 (95% CI, 1.30‐2.19). Doubling the APRI level was associated with an HR of 1.88 (95% CI, 1.48‐2.40).