Primary sclerosing cholangitis (PSC) is a chronic

Transcription

1 Predicting Clinical and Economic Outcomes After Liver Transplantation Using the Mayo Primary Sclerosing Cholangitis Model and Child-Pugh Score Jayant A. Talwalkar, * Eric Seaberg, W. Ray Kim, * and Russell H. Wiesner * Issues in the selection and timing of liver transplantation for primary sclerosing cholangitis (PSC) remain controversial. Although the Child-Pugh classification (CP) score and Mayo PSC model have similar abilities to estimate pretransplantation survival, a comparison of these 2 scores in predicting survival after liver transplantation has not been conducted. The aim of this study is to compare the Mayo PSC model and CP score in predicting patient survival and related economic outcomes after liver transplantation. Data from 128 patients with PSC, identified from the NIDDK database, were used to calculate patientspecific Mayo PSC and CP scores before transplantation. Levels reflecting a poor outcome were defined a priori. Receiver operating characteristic (ROC) curves and regression methods (Cox proportional hazards and linear regression models) were used to assess the relationship between these 2 scores and 5 post liver transplantation outcome measures. CP score was found to be a significantly (P <.05) better predictor of death 4 months or less after liver transplantation than: (a) length of hospital stay >21 days (or death before discharge) and (b) resource utilization >200,000 units (measured by area under the ROC curve). The Cox model identified statistically significant (P <.05) associations between CP score and each outcome after adjusting for the Mayo PSC risk score. Similar results were not observed for the Mayo PSC model when adjusted for CP score. Among patients with PSC undergoing liver transplantation, CP score was a better overall predictor of both survival and economic resource utilization compared with the Mayo PSC model. (Liver Transpl 2000;6: ) Primary sclerosing cholangitis (PSC) is a chronic cholestatic liver disease of unknown cause frequently associated with inflammatory bowel disease (IBD) and, specifically, chronic ulcerative colitis. The obliterative destruction of both large and small bile ducts frequently observed on histological examination often leads to the development of ductopenia, cholestasis, fibrosis, and biliary cirrhosis. 1 PSC is often observed after an insidious onset to follow a slow but progressive course that leads ultimately to hepatic failure if liver transplantation is not performed. However, the variability and fluctuating nature of the disease have made it difficult to determine when liver transplantation may be most effective before the onset of debilitating disease among individual patients. 2 Several investigators have developed prognostic models derived from individual patient data to accurately estimate survival with PSC and to aid in the selection for and timing for liver transplantation. 3 Recently, the United Network for Organ Sharing (UNOS) decided to use the Child-Pugh classification (CP) score to determine the major criteria necessary in listing patients for liver transplantation. 4 Although supporting evidence in the literature defends the use of the CP score, 5 another investigation has shown that the most recent Mayo PSC model was superior to the CP score in estimating survival without transplantation among patients with PSC. 6 As posttransplantation survival continues to improve over time with the development of improved surgical techniques and immunosuppressive regimens, information regarding clinical-based outcomes, like effective resource utilization or health-related quality of life, has been of increasing importance. Economic considerations for the overall process of liver transplantation must now also be a consideration because its effectiveness has become established. 7 The use of both clinical and institutional resources has been reported in monetary terms to be greater than $150,000 on average for a single liver transplant procedure and its attendant costs of preoperative and postoperative care. 8 Although descriptive studies in the literature have primarily focused on costs based on retrospective data, there has been little evidence to date examining the value of existing prognostic models for predicting economic outcomes after liver transplantation. The aim of this study is to compare the Mayo PSC model and CP score in predicting the following outcomes after liver transplantation: patient survival, hos- From the *Division of Gastroenterology and Hepatology, Mayo Clinic and Foundation, Rochester, MN; and the Graduate School of Public Health, University of Pittsburgh, Pittsburgh, PA. For the National Institutes of Diabetes and Digestive and Kidney Diseases Liver Transplantation Database Group. Address reprint requests to Russell H. Wiesner, MD, Division of Gastroenterology and Hepatology, Mayo Clinic and Foundation, 200 First St SW, Rochester, MN Telephone: ; FAX: ; rwiesner@mayo.edu Copyright 2000 by the American Association for the Study of Liver Diseases /00/ $3.00/0 doi: /jlts Liver Transplantation, Vol 6, No 6 (November), 2000: pp

2 754 Talwalkar et al Table 1. New Mayo Risk Score Model for PSC Model Revised Mayo PSC pital length of stay, intensive care unit (ICU) length of stay, intraoperative packed red blood cell (PRBC) requirement, and overall resource utilization measured by monetary indices. Patients and Methods Formula R 0.03 (age in years) 0.54 log (bilirubin in milligrams per deciliter) 0.54 log (aspartate aminotransferase in units per liter) 1.24 (history of variceal bleeding) 0.84 (albumin in grams per deciliter) Patient Population and Data Collection The data used for this analysis were obtained from a cohort of 128 liver transplant recipients enrolled onto the National Institutes of Diabetes and Digestive and Kidney Diseases Liver Transplantation Database (LTD) who had PSC listed as the primary cause of native liver failure. The LTD, which has been described previously, 9 is an ongoing prospective study of 916 liver transplant recipients for whom detailed demographic, clinical, and outcome measures are documented and analyzed to identify factors associated with various outcomes after liver transplantation. The 3 participating clinical centers are the University of Nebraska, the University of California at San Francisco, and the Mayo Clinic. This project database is maintained and administered by the Epidemiology Data Coordinating Center in the Graduate School of Public Health at the University of Pittsburgh. This study cohort represents 14% of the 916 liver transplant recipients enrolled onto the LTD between April 15, 1990, and June 30, At the time of this analysis, all patients had complete follow-up through June 30, CP and Mayo PSC scores were computed using data collected within 48 hours of the initial liver transplant procedure. In the rare instance in which pretransplantation information was missing, data obtained during a prior evaluation were considered, provided they were obtained no more than 30 days before transplantation. Prediction Models The original Mayo PSC natural history model was derived from clinical data of 174 patients in whom the diagnosis of PSC was based on established criteria. 10 Refinement of this prognostic model subsequently included patient age, serum bilirubin level, splenomegaly status, and histological stage. 11 A revised version was subsequently developed to enhance the clinical usefulness of the model by eliminating histological stage and thus obviating the need for a liver biopsy. Five variables, including patient age, serum bilirubin level, serum albumin level, serum aspartate aminotransferase level, and a history of esophageal variceal bleeding, were derived retrospectively from data related to the original 174 PSC patients (Table 1). Validation of the revised Mayo PSC model has been performed in a least 1 independent testing set to date. 12 CP score is a modification of the scheme originally proposed by Child and Turcotte with the variation developed for patients affected by cholestatic liver disease. Five clinical variables, including hepatic encephalopathy, ascites, serum bilirubin level, serum albumin level, and prolonged prothrombin time, are assigned a numeric score from 1 to 3 signifying the clinical severity of each predictor (Table 2). 13,14 These 5 individual scores are then summed to yield the overall CP score, which ranges from 5 to 15 points. Outcomes Five outcome measures of the success of liver transplantation were used in this analysis. These include patient survival after liver transplantation, hospital and ICU stays, PRBC use dur- Table 2. Description of Child-Pugh Classification Score Parameter Hepatic encephalopathy None Stage 1-2 Stage 3-4 Ascites Nil Mild-moderate Moderate-severe Bilirubin (mg/dl) PBC/PSC Other Albumin (g/dl) Prolonged prothrombin time (s) NOTE. CP class: 5-6 points A; 7-9 points B; points C. Abbreviation: PBC, primary biliary cirrhosis. Points

3 Outcomes After Liver Transplantation for PSC 755 ing the procedure, and resource use between the time of transplantation and the initial hospital discharge. Information about the first 4 outcomes (survival, hospital and ICU stays, and PRBC use) were documented prospectively on LTD data collection forms. Resource utilization was collected as an ancillary study to the LTD, the details of which have been reported previously. 7 Statistical Analysis All statistical analyses were performed using SAS for OS/ (SAS Institute, Cary, NC). The correlation between Mayo PSC and CP scores was estimated using Pearson s correlation coefficient. Cox proportional hazards 15 and multiple linear regression models were used to assess the relative strength of association between each score and outcome. The analyses of survival after liver transplantation, time to ICU discharge, and time to hospital discharge were performed using the Cox proportional hazards model, whereas the analyses of intraoperative PRBC use and overall resource use were performed using multiple linear regression models. Because the PRBC and resource use data were markedly skewed, these data were log-transformed for the multiple regression analyses. To directly compare the prognostic abilities of these 2 scores, receiver operating characteristic (ROC) curves 16 were generated and the areas under the curve (AUC) were compared. An ROC curve is constructed by plotting sensitivity (y axis) versus 1-specificity (x axis) for all possible cut points of each predictive model. The AUC generated by connected ROC points represents a measure of the ability of each model to predict the given outcome. An AUC equal to 1 indicates perfect ability to predict an outcome or event, whereas an AUC equal to 0.5 signifies that a prediction rule is no better than that expected by chance. The statistical comparison of the 2 models was performed by comparing the areas under each of the ROC curves in which the larger AUC represents the stronger predictor of a given outcome. We used the nonparametric approach based on the generalized U statistic described by DeLong et al. 17 The ROC curve approach required that we define cut points representing good versus poor outcomes. The following cut points were developed a priori. For each outcome, a definition used for ROC curve analysis is given. Early death is defined as death during the first 120 days after the initial liver transplantation. Prolonged ICU stay is defined as an ICU stay greater than 3 days or death before discharge from the ICU. Prolonged hospital stay indicates a hospital stay greater than 21 days or death before discharge from the hospital. High PRBC use indicates the use of more than 3,500 ml of PRBCs during the transplantation procedure. High resource use indicates resource utilization of more than 200,000 units between the initial transplantation procedure and the earlier of death or hospital discharge. All statistical analyses were performed using 2-sided tests, with an level of 0.05 to identify statistical significance. Table 3. Clinical Characteristics of the 128 Patients With PSC Undergoing Liver Transplantation No. % Mean (SD) Range Subject age (yr) (10.7) Patient sex Men Women CP score at OLT (1.9) 5-13 Karnofsky score at OLT (1.8) 2-10 Abbreviation: OLT, orthotopic liver transplantation. Results Selected clinical characteristics of the 128 patients with PSC in this study population are listed in Table 3. Both Mayo PSC and CP scores for all patients in the study cohort are strongly correlated ( 0.81; Fig. 1). Because of missing data, CP score could be computed for 125 patients, and the Mayo PSC score was computed for 126 patients. Bivariate analyses showed CP score to be significantly associated with time to hospital discharge (P.0006), time to ICU discharge (P.002), intraoperative PRBC use (P.0001), overall resource utilization (P.0001), and 4-month survival (P.02). The Mayo PSC score is associated with time to ICU discharge (P.03), intraoperative PRBC use (P.0001), and overall resource utilization (P.04). Both 4-month survival (P.37) and time to hospital discharge (P.09) are not significantly associated with the Mayo PSC score. Four-month survival (P.004), time to hospital discharge (P.0002), time to ICU discharge

4 756 Talwalkar et al Figure 1. Correlation between CP score and Mayo PSC score for all patients in the study cohort. (P.01), intraoperative PRBC use (P.0001), and overall resource utilization (P.0001) are significantly associated with CP score when adjusted for the Mayo PSC score. Relationships between CP score and each outcome remain in the expected direction as well. However, the Mayo PSC score is noted to have a negative association with 4-month survival (P.08), time to hospital discharge (P.02), time to ICU discharge (P.32), intraoperative PRBC use (P.28), and overall resource utilization (P.002) when adjusted for CP score. In comparing the 2 scores using ROC curves, specific cut points of each outcome were identified for analysis. These included 4-month patient survival after liver transplantation, length of hospital stay greater than 21 days, length of ICU stay greater than 3 days, intraoperative PRBC transfusion requirements greater than 3,500 ml, and overall resource utilization greater than 200,000 units from transplantation to death or discharge. Calculation of an AUC for both prediction scores (with each outcome of interest) showed the CP score to be a stronger predictor of 4-month patient survival after transplantation (0.784 v 0.619; P.02), length of hospital stay greater than 21 days (0.652 v 0.562; P.006), and overall resource utilization greater than 200,000 units from transplantation to death or hospital discharge (0.723 v 0.627; P.008). Predicted intraoperative PRBC use greater than 3,500 ml (CP, v Mayo, 0.628; P.06) was noted to be marginally significant (Fig. 2). The 2 scores performed similarly in predicting ICU stay greater than 3 days (CP, v Mayo, 0.685; P.85). Each predictive model was then examined at specific levels of interest for the outcomes of death and resource utilization for each AUC. In applying these findings, for example, subjects with a CP score of 10 or greater were observed with a 1-year survival rate of 81% compared with a 91% 1-year survival rate among patients in the highest quartile using the Mayo PSC model. Similarly, the cost of liver transplantation exceeded 200,000 units in 63% of patients with a CP score of 10 or greater compared with only 39% of patients who were in the Mayo PSC model s highest quartile. Discussion In determining the optimal timing for liver transplantation in patients with PSC, consideration must be given to a minimum of 3 factors: (1) the estimated Figure 2. Comparisons between the predictive abilities of the CP and Mayo PSC scores with specific outcomes of interest using ROC curve methods: (A) patient survival less than 4 months after transplantation; (B) length of hospital stay greater than 21 days; (C) overall resource utilization greater than 200,00 units; (D) intraoperative PRBC use.

5 Outcomes After Liver Transplantation for PSC 757 survival of affected patients who do not receive liver transplantation, (2) the estimated survival of patients subsequent to liver transplantation, and (3) the magnitude of overall resource utilization required to manage allograft recipients. 18 The CP score has been adopted by UNOS as the basis for determining when patients are eligible for liver transplant waiting list selection. 3,4 Although recent studies have focused on the CP and Mayo PSC model scores for their ability to estimate survival without liver transplantation, 5,6 the examination of estimated survival after liver transplantation has not been performed with specific reference to CP score. In this study, we have shown that CP score is superior to the Mayo PSC model in estimating patient death less than 4 months after liver transplantation, length of hospital stay greater than 21 days, intraoperative PRBC use greater than 3,500 ml, and overall resource utilization greater than 200,000 units. Both the Mayo PSC model and CP score appear to perform similarly with respect to length of ICU stay in this patient cohort. Previous investigations have attempted to develop both simple and complex models to predict patient and allograft survival after liver transplantation. Doyle et al 19 examined the preoperative clinical data from 148 adult liver transplant recipients to develop prognostic models and ROC curves for predicting early death or the need for retransplantation. Although focusing specifically on the outcome of graft function, the investigators did not examine the association between CP score and outcome as a primary aim. Although CP score has been observed as a stronger predictor of patient survival independent of Mayo PSC risk score, previous comparisons with the revised Mayo model containing serum aspartate aminotransferase level and history of variceal bleeding have not been performed. 4 In addition, the patients undergoing liver transplantation for PSC at the Mayo Clinic represent a wider spectrum of disease severity than the cohort used by Shetty et al. 5 This may render the CP score an insensitive method of outcome prediction when applied to the Mayo patient cohort. More recently, pretransplantation variables predicting survival after transplantation were identified among 118 patients with PSC. 20 Multivariate analyses showed a history of IBD, ascites, previous upper abdominal surgery, serum creatinine level, and history of biliary tree malignancy to be significant predictors of posttransplantation survival. Validation of this prediction model in an independent data set of 30 patients who underwent transplantation for PSC showed no statistically significant difference between observed and expected outcomes in this cohort (P.67). Although notable for the observation of a previous history of IBD with Crohn s disease imparting a negative influence on posttransplantation survival, this finding requires validation in independent studies. Furthermore, the wide 95% confidence intervals calculated for survival probabilities at the extremes of outcome (0% or 100% survival) suggest that use of this model in decision making for an individual patient may be limited. The finding of 24 deaths within the first year of transplantation among this cohort also casts doubt on the applicability of this model. Increases in the standard error for all predictor variables in the final model (except a history of IBD, in which more than 50% of subjects were affected) are reflective of diminished model precision. Our study is notable for the examination of outcomes other than patient survival. Although traditionally an accepted end point for therapeutic efficacy, the singular use of mortality to measure the effectiveness of liver transplantation is limited in scope. Because medically accepted but costly procedures have been exposed to increasing scrutiny, such outcomes as length of hospital stay and intraoperative PRBC use are valuable and discrete measures of resource utilization relevant to the process of liver transplantation. The ability to predict resource utilization for liver transplantation based on preoperative severity of disease can dramatically influence the burden placed on other sectors of the health care arena. 7,21,22 The association with CP score and each of the outcomes measured in this study was observed to be in the direction of expected results. Higher CP scores were seen to reflect increasing risks for death, length of ICU and overall hospital stays, intraoperative PRBC use, and overall resource utilization. Conversely, once CP score is accounted for, our analyses suggest that a higher Mayo PSC risk score would be associated with a lower risk for developing a negative outcome. The observation of this negative relationship is most likely related to the very high correlation between CP score and Mayo PSC risk score. With both scores in the model, it is unlikely to encounter significance in the expected direction for both variables because they are measuring the same recipient status. A greater stability in the CP score for predicting the outcomes selected in this investigation compared with the Mayo PSC risk score would also contribute to this finding. ROC curve analysis has been used as a technique to determine the performance of a test or prediction rule independent of the prevalence associated with an outcome. 16,19 An ROC curve with an AUC of 0.5 is associated with a model that is unable to provide predictive

6 758 Talwalkar et al information about a selected outcome. A number of ROC curves generated for specific outcomes in this study were noted to have 95% confidence intervals that cross the boundary of AUC equal to 0.5 (not shown). Reasons for this may include a reduced number of events (outcomes) in relation to the entire cohort of patients who underwent transplantation for PSC. For a predictive model to strongly predict an individual subject s outcome based on specific clinical characteristics (discrimination), this may be at the expense of incurring an undesirable number of false-positive results for an entire cohort (calibration). Goodness-of-fit statistics as a measure of applicability for a prediction model are representative of sufficient calibration, yet evidence to support the presence of excellent discrimination cannot be inferred. 21 Reduced sample sizes and the failure to collect information on all possible predictive variables that influence the outcome of interest can be responsible for prognostic models with poor calibration. The Mayo PSC model and CP score have been used to predict pretransplantation patient survival and posttransplantation outcomes among patients with PSC undergoing liver transplantation. An examination of the prognostic ability for each index indicates that the CP score is a better overall predictor of health care outcomes in this study population. We have also shown the utility of ROC analysis as a method in comparing prognostic models that may be recommended for the pursuit of continued process improvements in liver transplantation. References 1. LaRusso NF, Wiesner RH, Ludwig J, MacCarty RL. Primary sclerosing cholangitis. N Engl J Med 1984;310: Wiesner RH, Porayko MR, Dickson ER, Gores GJ, LaRusso NF, Hay JE, et al. Selection and timing of liver transplantation in primary biliary cirrhosis and primary sclerosing cholangitis. Hepatology 1992;16: Wiesner RH. Liver transplantation for primary biliary cirrhosis and primary sclerosing cholangitis: Predicting outcomes with natural history models. Mayo Clin Proc 1998;73: Lucey MR, Brown KA, Everson GT, Fung JJ, Gish R, Keeffe EB, et al. Minimal criteria for placement of adults on the liver transplant waiting list: A report of a national conference organized by the American Society of Transplant Physicians and the American Association for the Study of Liver Diseases. Liver Transpl Surg 1997;3: Shetty K, Rybicki L, Carey WD. The Child-Pugh classification as a prognostic indicator for survival in primary sclerosing cholangitis. Hepatology 1997;25: Kim WR, Poterucha JJ, Wiesner RH, LaRusso NF, Lindor KD, Petz J, et al. The relative role of the Child-Pugh classification and the Mayo natural history model in the assessment of survival in patients with primary sclerosing cholangitis. Hepatology 1999; 29: Showstack J, Katz PP, Lake JR, Brown RS Jr, Dudley RA, Belle S, et al for the NIDDK Liver Transplantation Database Group. Resource utilization in liver transplantation: Effects of patient characteristics and clinical practice. JAMA 1999;281: Evans RW, Manninen DL, Dong FB. An economic analysis of liver transplantation: Costs, insurance coverage, and reimbursement. Gastroenterol Clin North Am 1993;22: Wei YL, Detre KM, Everhart JE. The NIDDK Liver Transplantation Database. Liver Transpl Surg 1997;3: Wiesner RH, Grambsch PM, Dickson ER, Ludwig J, MacCarty RL, Hunter EB, et al. Primary sclerosing cholangitis: Natural history, prognostic factors and survival analysis. Hepatology 1989;10: Dickson ER, Murtaugh PA, Wiesner RH, Grambsch PM, Fleming TR, Ludwig J, et al. Primary sclerosing cholangitis: Refinement and validation of survival models. Gastroenterology 1992; 103: Kim WR, Therneau TM, Wiesner RH, Poterucha JJ, Benson JT, Malinochoc M, et al. A revised natural history model for primary sclerosing cholangitis. Mayo Clin Proc 2000;75: Child CG III, Turcotte JG. Surgery and portal hypertension. In: Child CG III (ed). The liver and portal hypertension. Philadelphia: Saunders, 1964: Pugh RN, Murray-Lyon IM, Dawson JL, Pietroni MC, Williams R. Transection of the oesophagus for bleeding oesophageal varices. Br J Med 1973;60: Cox DR. Regression models and life tables. J R Stat Soc B 1972;34: Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982; 143: DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: A non-parametric approach. Biometrics 1988:44; Kim WR, Dickson ER. The role of prognostic models in the timing of liver transplantation: Application in cholestatic liver diseases. Clin Liver Dis 1997;1: Doyle HR, Marino IR, Jabbour N, Zetti G, McMichael J, Mitchell S, et al. Early death or retransplantation in adults after orthotopic liver transplantation. Can outcome be predicted? Transplantation 1994;57: Neuberger J, Gunson B, Komolmit P, Davies MH, Christensen E. Pretransplant prediction of prognosis after liver transplantation in primary sclerosing cholangitis using a Cox regression model. Hepatology 1999;29: Kim WR, Therneau TM, Dickson ER, Evans RW. Preoperative predictors of resource utilization in liver transplantation. Clin Transpl 1995; Ricci P, Therneau TM, Malinchoc M, Benson JT, Petz JL, Klintmalm GB, et al. A prognostic model for the outcome of liver transplantation in patients with cholestatic liver disease. Hepatology 1997;25: