The UK-PBC risk scores: Derivation and validation of a scoring system for long-term prediction of end-stage liver disease in primary biliary cholangitis

Abstract

The biochemical response to ursodeoxycholic acid (UDCA)--so-called "treatment response"--strongly predicts long-term outcome in primary biliary cholangitis (PBC). Several long-term prognostic models based solely on the treatment response have been developed that are widely used to risk stratify PBC patients and guide their management. However, they do not take other prognostic variables into account, such as the stage of the liver disease. We sought to improve existing long-term prognostic models of PBC using data from the UK-PBC Research Cohort. We performed Cox's proportional hazards regression analysis of diverse explanatory variables in a derivation cohort of 1,916 UDCA-treated participants. We used nonautomatic backward selection to derive the best-fitting Cox model, from which we derived a multivariable fractional polynomial model. We combined linear predictors and baseline survivor functions in equations to score the risk of a liver transplant or liver-related death occurring within 5, 10, or 15 years. We validated these risk scores in an independent cohort of 1,249 UDCA-treated participants. The best-fitting model consisted of the baseline albumin and platelet count, as well as the bilirubin, transaminases, and alkaline phosphatase, after 12 months of UDCA. In the validation cohort, the 5-, 10-, and 15-year risk scores were highly accurate (areas under the curve: >0.90).

Conclusions: The prognosis of PBC patients can be accurately evaluated using the UK-PBC risk scores. They may be used to identify high-risk patients for closer monitoring and second-line therapies, as well as low-risk patients who could potentially be followed up in primary care.

Fig. 1

Relationship between the hazard ratio for a liver event (liver death or liver transplantation) and each variable within the final model together with the best-fitting polynomial line in the UK-PBC Research Cohort (A. ALP12; B. BIL12; C. TAl2; D. Albumin; E. Platelets). To allow for inter-operator variability, BIL12, TAl2 and ALP12 were analysed as multiples of their upper reference levels whereas Albumin and Platelets were analysed as multiples of their lower reference levels. The best-fitting polynomial line to describe the relationship between risk and each variable is shown. Note in particular that for each variable, risk increases or decreases as a continuum and there are no points at which the trajectory of risk suddenly changes, which suggests they are best modelled as continuous variables. Note also that some variables do not have a linear relationship with risk and for this reason, they were transformed using multivariable fractional polynomials (see text). Abbreviations: ALP12, alkaline phosphatase after 12 months of UDCA; BIL12, bilirubin after 12 months of UDCA; LLN, lower limit of normal; ULN, upper limit of normal; TA12, transaminases after 12 months of UDCA; UDCA, ursodeoxycholic acid.

Fig. 2

Receiver operating characteristic (ROC) curves for the prediction of death or liver transplantation (LT) according to the UK-PBC risk scores at 5 (A), 10 (B) and 15 years (C). Note: Area under the ROC curve (AUROC) can be interpreted as a summary index of classification performance. An AUC value of 0.5 indicates a ’random call,’ whereas an AUC value of 1.0 indicates perfect separation of events and non events.

Fig. 3

Receiver operating characteristic (ROC) curves for the prediction of death or liver transplantation (LT) using the Barcelona, Paris 1, Toronto and Paris II definitions of treatment response, in the UK-PBC cohort at 5 (A), 10 (B) and 15 years (C). Each plot in Figure 3 shows only one point because the Barcelona, Paris I, Toronto and Paris II definitions of treatment response are dichotomous, having a pre-defined threshold and only two levels: responder and non responder. The ROC curve is therefore plotted using the single, pre defined threshold. In contrast, the UK-PBC Risk Scores are continuous. The ROC curve is therefore plotted by incrementally varying the threshold and measuring the sensitivity and specificity at each threshold. Comparison to the Rotterdam criteria was not possible because serum albumin after 12 months of treatment was not available for all participants in the cohort.

Fig. 4

Predicted versus observed risk of an event across each decile of the 5 (A), 10 (B) and 15 year (C) UK-PBC risk scores. There is close correspondence between the predicted and observed risks, suggesting that the risk scores are well calibrated.

Fig. 5

Receiver operator characteristic curves for the 5 (A), 10 (B) and 15 years (C) UK-PBC risk scores in the subgroup of untreated participants (n = 754)* of the UK-PBC cohort* This subgroup includes only participants who were not treated with UDCA and had been followed up for at least 12 months.