Time-Dependent Cox Regression Model Is Superior in Prediction of Prognosis in Primary Sclerosing Cholangitis Kirsten Muri Boberg,1 Giuseppe Rocca,2 Thore Egeland,1 Annika Bergquist,3 Ulrika Broome´ ,3 Llorenc Caballeria,4 Roger Chapman,5 Rolf Hultcrantz,6 Stephen Mitchell,5 Albert Pares,4 Floriano Rosina,7 and Erik Schrumpf 1 More precise prognostic models are needed for prediction of survival in patients with primary sclerosing cholangitis (PSC), particularly for the selection of candidates for liver transplantation. The aim of this study was to develop a time-dependent prognostic model for the calculation of updated short-term survival probability in PSC. Consecutive clinical and laboratory follow-up data from the time of diagnosis were collected from the files of 330 PSC patients from 5 European centers, followed for a median of 8.4 years since diagnosis. Time-fixed and time-dependent Cox regression analyses, as well as the additive regression model, were applied. The reliability of the models was tested by a cross-validation procedure. Bilirubin (on a logarithmic scale), albumin, and age at diagnosis of PSC were identified as independent prognostic factors in multivariate analysis of both the time-fixed and the time-dependent Cox regression models. The importance of bilirubin was more pronounced in the time-dependent model (hazard ratio [HR], 2.84) than in the time-fixed analysis (hazard ratio, 1.51). The additive regression model indicated that once the patients survive beyond the first 5 years, the impact on prognosis of albumin at diagnosis ceases. The time-dependent prognostic model was superior to the time-fixed variant in assigning low 1-year survival probabilities to patients that actually survived less than 1 year. In conclusion, a time-dependent Cox regression model has the potential to estimate a more precise short-term prognosis in PSC compared with the traditional time-fixed models. (HEPATOLOGY 2002;35:652-657.)
P
rognostic models are complex tools that combine patient characteristics to predict clinical outcomes.1 They are of potential value in difficult clinical decision making, such as selecting patients for liver transplantation and optimal timing of such interventions.2 Primary sclerosing cholangitis (PSC) is a chronic cholestatic liver disease that usually progresses to cirrhosis.3 Liver transplantation is the only curative treatment in this disease.4 Several prognostic models for PSC have been published in the past,5-9 but they have had limited application in clinical routine. In particular, the ability of these models to discriminate between patients needing and not needing transplantation is not optimal.10
Abbreviations: PSC, primary sclerosing cholangitis; PBC, primary biliary cirrhosis; IBD, inflammatory bowel disease; HR, hazard ratio; ALP, alkaline phosphatases; ALT, alanine transaminase; AST, aspartate transaminase; PI, prognostic index. From the 1Medical Dept. and Dept. of Epidemiology, Rikshospitalet, Oslo, Norway; 2Dept. of Gastroenterology, Molinette Hospital, Torino, Italy; 3Unit of Gastroenterology and Hepatology, Karolinska Institute, Huddinge Hospital, Stockholm, Sweden; 4Liver Unit, Hospital Clinic I Provincial, University of Barcelona, Barcelona, Spain; 5Dept. of Medicine, John Radcliffe Hospital, Oxford, UK; 6Dept. of Medicine, Karolinska Hospital, Stockholm, Sweden; and 7Dept. of Gastroenterology, Gradenigo Hospital, Torino, Italy. Received July 19, 2001; accepted December 21, 2001. Supported by the European Commission (BMH4-CT96-0779). Address reprint requests to: Kirsten Muri Boberg, Medical Dept., Rikshospitalet, 0027 Oslo, Norway. E-mail:
[email protected]; fax: (47) 23-07-08-31. Copyright © 2002 by the American Association for the Study of Liver Diseases. 0270-9139/02/3503-0021$35.00/0 doi:10.1053/jhep.2002.31872 652
The great variability of the disease course in PSC most likely is one factor contributing to the low precision of prognostic models. All previous prognostic models for this disease have been time-fixed Cox regression models.5-9 These models assume that the variables noted at one single time point for each patient are sufficient to predict survival. Moreover, time-fixed models may not be appropriate for updating prognosis at later time points.2 In PSC, as in most chronic diseases, the clinical situation changes with time. Conceivably, the weight of prognostic indicators changes in different phases of the disease, and estimates of prognosis should improve if such time-dependent changes are taken into account. A prognostic index based on a time-dependent model may be estimated repeatedly during the course of the disease in the individual patient. A time-dependent Cox regression model has been applied to primary biliary cirrhosis (PBC), which is another cholestatic liver disease, giving more precise estimates on short-term prognosis compared with a time-fixed procedure.11 The time-dependent model does not apply for long-term survival. Because PBC is characterized by a more steadily progressive course than PSC, it cannot be inferred that a prognostic model is equally suitable in both conditions. Prognostic models that can improve our clinical decisions in the care of PSC patients are needed. The purpose of the present study was to construct a timedependent prognostic model for PSC, based on serial follow-up data from a large group of patients, and subsequently to discuss the reliability of the model. The practical use of the model is also illustrated in a specific example based on patient data.
HEPATOLOGY, Vol. 35, No. 3, 2002
Patients and Methods Patient Population and Data Collection. The study population consisted of 330 PSC patients recruited from the 5 centers participating in the European Study Group of Primary Sclerosing Cholangitis, including Barcelona (n ⫽ 17), Oslo (n ⫽ 150), Oxford (n ⫽ 60), Stockholm (n ⫽ 92), and Torino (n ⫽ 11). All of the units serve as both basic-care and referral centers for liver transplantation. PSC was diagnosed according to accepted criteria, including typical findings of bile duct irregularities, strictures, and dilatations at cholangiography in all cases.12-14 The time of diagnosis was defined as the first cholangiography consistent with PSC. Inflammatory bowel disease (IBD) was diagnosed by colonoscopy with biopsies.15 Each center contributed its longitudinal follow-up files on consecutive PSC patients. Defined data from the patient’s visits and major disease–related events were associated to a date and collected in a database. Data collection started from diagnosis of PSC and extended until death, liver transplantation, or last visit. The original database comprised 394 PSC patients. Among these, 9 were small-duct PSC patients who were excluded. Data were insufficient (neither albumin nor bilirubin recordings at diagnosis were available) for a time-dependent analysis in 55 cases, leaving 330 patients for the present study. The median follow-up time was 8.4 years from diagnosis of PSC. The patients were followed until death or last visit up to October 1, 1997, for the Swedish patients and May 1, 1998, for the other centers. Besides jaundice and pruritus, symptoms like abdominal pain, fever, weight loss, and fatigue were recorded when considered attributable to the liver disease. The definition of an episode of cholangitis included antibiotic treatment. Jaundice, hepatomegaly, splenomegaly, ascites, encephalopathy, and esophageal varices were noted as clinical findings. The biochemical analyses were standard procedures at each center. Statistical Methods. Potential prognostic variables were evaluated as predictors of survival in both time-fixed and timedependent Cox models.16,17 In the time-fixed model, only the initial records at the time of diagnosis of PSC were applied. The time-dependent model used the repeated measurements of the potential prognostic variables during the entire disease duration. Death and liver transplantation were combined to one end-point in the main analyses. In addition, liver transplantations were censored in alternative calculations. Plots were used to check the validity of the assumptions underlying the models, and a test based on the method described by Therneau18 was performed. The additive regression model19 is presented partly to supplement the time-fixed Cox analysis, partly as an alternative. As opposed to the Cox model, this model does not require proportional hazards to be valid. To discuss this important point in some detail, a general example is helpful. Assume that age at diagnosis is a prognostic factor. It may happen that the short-term survival, e.g., survival up to 1 year, is independent of age at diagnosis, whereas long-term death risk from the disease is considerably lower for young patients. In this case, the standard Cox model is inappropriate, whereas the additive model applies. Some problems of inconsistency in the Cox model are alleviated using the additive model. Cases in which the Cox model and the additive model produce essentially similar results are reassuring, and cases of differences should be scrutinized.
BOBERG ET AL.
653
The main assumption of the additive model is that the intensities of an event (death from PSC in this study) are related linearly to the prognostic variables. From a practical point of view, this implies that the additive approach gives the additional risk in absolute terms that are induced by a certain covariate value, and this may be as important from a medical point of view as the relative risk.19 An unfortunate feature of the additive model is that the intensities could come out as negative values, which does not make sense. If this happens, it indicates that the model does not fit the data. Further tests of goodness of fits are based on plots of residuals.19 We have used informal methods based on the initial univariate analyses and clinical judgement to develop multivariate models. This procedure is preferred to alternative automatic procedures like stepwise methods,20 which, in particular, may be problematic when there are missing observations. The time-fixed and the time-dependent Cox regression models were compared using a cross-validation procedure.21 With the time-dependent model, estimation of prognosis is only possible for short time spans.11 When a decision on enlisting for liver transplantation must be reached, in particular, the short-term prognosis of, for example, 1 year is relevant. We therefore chose the 1-year survival probability as a parameter in the test procedure. The test set consisted of all deaths within 1 year after the diagnosis of PSC. We implemented the classical “leave-one-out” procedure to compare and validate the models. The basic idea is that an observation is temporarily removed from the data set. The remaining observations are used to estimate the model and to predict the removed observation. This procedure is repeated for all observations in the test data set. For our data set, this general idea was implemented as follows: we temporarily omitted the first of the deaths from the data and estimated time-fixed and time-dependent models based on the remaining patients. The 1-year survival probability for the omitted patient was then calculated according to the 2 models. The same procedure was repeated for all deaths. The models were then evaluated by comparing predicted survival probability with the survival actually observed. For that purpose, we divided the deaths into 3 categories: those occurring within the first year, between years 1 and 5, and after 5 years, respectively. A good prognostic model should assign a low 1-year survival probability to the first category. In fact, the 1-year survival probability for this group would be zero in the perfect model. We chose a 1-year survival probability of 0.5 (50%) or less for that group as a criterion for a satisfactory model. On the other hand, we considered that the 1-year survival probability should not be below 0.5 for the patients surviving 5 years. The statistical analyses were performed with S-plus 4.5 (MathSoft Inc., Seattle, WA). An S-plus add on performing the additive regression model is freely available from http://www.med.uio.no/imb/stat/stat/projects/addreg.html.
Results Patients. The 330 PSC patients included 107 (32%) females and 223 (68%) males. During follow-up, 68 (21%) patients died and 57 (17%) underwent liver transplantation. Cholangiocarcinoma was diagnosed in 44 (13%) patients. The median survival time from diagnosis of PSC was 11.7 years (95% CI, 10.1-14.8 years). Table 1 summarizes the characteristics of the symptoms and
654
BOBERG ET AL.
HEPATOLOGY, March 2002
Table 1. Descriptive Statistics for the Categorical Variables at Diagnosis of PSC Variable (no. of patients with data, among the total of 330)
n (%)
IBD (328) Symptom Jaundice (318) Pruritus (314) Fatigue (315) Abdominal pain (313) Weight loss (309) Fever (316) Variceal bleeding (316) Clinical finding Jaundice (306) Hepatomegaly (283) Esophageal varices (214) Splenomegaly (278) Ascites (304)
272 (82.9) 90 (28.3) 78 (24.8) 72 (22.9) 69 (22.0) 44 (14.2) 26 (8.2) 7 (2.2) 91 (29.7) 73 (25.8) 22 (10.3) 26 (9.4) 12 (3.9)
clinical findings at diagnosis of PSC (categorical variables). IBD was diagnosed in 272 (83%) patients. The biochemical parameters and age at diagnosis (continuous variables) are presented in Table 2. Univariate Statistical Analyses. The results of univariate analyses of the categorical and the continuous variables are given in Tables 3 and 4, respectively. The tables present the hazard ratios (HR) and 95% CI based on univariate Cox regression analysis. The P values are calculated according to both the Cox regression analysis and the additive model of Aalen.19 Bilirubin, alkaline phosphatases (ALP), alanine transaminase (ALT), and aspartate transaminase (AST) were treated on a logarithmic scale (Table 4). Table 3 illustrates, for example, that hepatomegaly at diagnosis of PSC is associated with a doubled risk (HR ⫽ 2.04) of reaching the end-points death or liver transplantation, and that this is significant, with a 95% CI of 1.37-3.04 (P ⫽ .0005). In addition to hepatomegaly, the following categorical variables were significantly associated to an adverse outcome in the univariate models: jaundice, pruritus, abdominal pain, fatigue, weight loss, ascites, splenomegaly, and the presence of esophageal varices (Table 3). For the continuous variables, higher values of ln(bilirubin), ln(ALP), and age at diagnosis of PSC, as well as lower values of hemoglobin, albumin, and Normotest, were significant indicators of a worse prognosis (Table 4). The P values obtained from the Table 2. Descriptive Statistics for the Continuous Variables at Diagnosis of PSC Variable (no. of patients with data, among the total of 330)
Age (yr) (330) Hemoglobin (g/dL) (297) Platelet count (⫻106/L) (268) Bilirubin (mol/L) (303) ALP (U/L) (322) AST (U/L) (316) ALT (U/L) (266) Albumin (g/L) (266) Normotest* (%) (196)
Median (range)
36.6 (13.0-82.3) 13.4 (7.2-17.9) 271 (17-801) 17 (4-612) 745 (114-7,990) 86 (10-1,478) 126 (13-2,416) 39 (18-51) 100 (15-200)
*Prothrombin time as Normotest (Nycomed Pharma, Oslo, Norway).
Table 3. Univariate Statistical Analysis of Categorical Variables Variable
HR
95% CI
Cox Regression P
Additive Model P
Sex IBD Jaundice (symptom) Pruritus Fever Abdominal pain Fatigue Weight loss Variceal bleeding Jaundice (finding) Ascites Hepatomegaly Splenomegaly Esophageal varices
0.841 1.164 3.449 1.866 1.84 2.123 3.268 2.061 1.869 3.679 5.545 2.039 2.103 2.813
0.577-1.224 0.696-1.947 2.41-4.936 1.256-2.774 1.008-3.359 1.408-3.202 2.215-4.822 1.302-3.262 0.591-5.911 2.556-5.296 2.838-10.834 1.367-3.039 1.164-3.801 1.577-5.02
.3652 .5627 0 .002 .0472 .0003 0 .002 .2873 0 0 .0005 .0138 .0005
.3863 .5821 0 .0085 .1212 .0039 0 .0154 .42 0 .0099 .0027 .0636 .0141
additive regression model closely followed those of the Cox model for all variables, except splenomegaly (Tables 3 and 4). Figure 1 illustrates the additive regression of life times for albumin. The solid line declines roughly linearly with time for the first 5 years, and thereafter continues as a more flattened curve. This curve indicates that the impact of albumin on prognosis only lasts for 5 years, i.e., the decreased albumin concentration is associated with an increased risk of reaching the end-point within 5 years. If a patient survives beyond 5 years, the initial albumin concentration apparently no longer is of prognostic importance. The symptoms, fatigue and weight loss, were associated with a pronounced risk of reaching the end-points within 6 to 9 months, but not after that, given survival beyond this initial time period. Multivariate Statistical Analyses. The multiple models derived from the time-fixed and the time-dependent approaches are given in Tables 5 and 6, respectively. Both models include ln(bilirubin), albumin, and age at diagnosis of PSC. The most noticeable difference between the 2 models is that the importance of ln(bilirubin) is more pronounced in the time-dependent model, i.e., the relative importance of bilirubin compared with the other statistical variables is more pronounced. The results were only marginally altered in the case that liver transplantations were considered as censored observations instead of being combined with death as one common end-point. The P values of the additive regression model followed those of the time-fixed Cox analysis quite closely (Table 5). Table 4. Univariate Statistical Analysis for Continuous Variables Variable
HR
95% CI
Cox Regression P
Additive Model P
Age at diagnosis Hemoglobin Platelet count ln (bilirubin)* ln (ALP) ln (AST) ln (ALT) Albumin Normotest
1.02 0.81 0.998 1.63 1.39 1.23 0.862 0.894 0.987
1.009-1.031 0.731-0.897 0.996-1.0 1.42-1.88 1.1-1.76 0.987-1.54 0.689-1.08 0.867-0.923 0.98-0.995
.0005 .0001 .0694 0 .0062 .065 .19 0 .0008
.0004 .0006 .0751 0 .0076 .0336 .1338 0 .0008
*One-unit increase on the natural logarithmic scale, corresponding to an increase of 2.71 on the original scale, increases the risk by 1.63.
HEPATOLOGY, Vol. 35, No. 3, 2002
BOBERG ET AL.
655
Table 6. Multivariate Model Based on the Time-Dependent Approach Variable
ln (bilirubin) Albumin Age at diagnosis of PSC
Coefficient
HR
Cox Regression P
1.0436 ⫺0.1166 0.0125
2.840 0.913 1.005
.0000 .0000 .0055
resulting in an estimated 1-year survival probability of approximately 35% from Fig. 3.
Discussion Fig. 1. Additive regression of life times for albumin. The solid line declines until approximately 5 years, indicating that low values of albumin are of negative prognostic value during the first 5 years after diagnosis of PSC. Dotted lines represent 95% CIs.
Eighteen patients died within 1 year from diagnosis of PSC and were used in the cross-validation procedure in which the 1-year survival probabilities according to the time-fixed and time-dependent models were calculated. The time-dependent model performed better than the time-fixed model from several points of view. One of our criteria for a satisfactory prognostic model was that it should assign a 1-year survival probability of 0.5 or less to those patients actually reaching end-point within the first year. The choice of 0.5 is somewhat arbitrary, and we add some further information below to aid in the comparison of the models. The average 1-year survival probability was lower when calculated from the time-dependent than the time-fixed model (0.65 vs. 0.73). More importantly, 7 (39%) of the probabilities were 0.5 or less for the time-dependent model compared with only 2 (11%) for the time-fixed model. A more complete picture is provided in Fig. 2. For the long-term survivors, on the other hand, the 1-year survival probabilities should be high. The time-dependent model assigned a probability of less than 0.5 to one patient who survived more than 5 years, whereas this was never the case for the time-fixed model. In this regard, the time-fixed model performed better. Prognostic Indices. From the multiple time-dependent prognostic model, a prognostic index (PI) can be calculated: PI ⫽ 1.04(ln[bilirubin] ⫺ 3.31) ⫺ 0.12(albumin ⫺ 37.27) ⫹ 0.013(age at diagnosis ⫺ 36.04). To exemplify, considering a patient with a bilirubin concentration of 359 mol/L, albumin level of 31 g/L, and age of 31.90 years at diagnosis of PSC, we find: PI ⫽ 1.04(ln[359] ⫺ 3.31) ⫺ 0.12(31 ⫺ 37.27) ⫹ 0.013(31.90 ⫺ 36.04) ⫽ 3.4,
Table 5. Multivariate Model Based on the Time-Fixed Approach Variable
ln (bilirubin) Albumin Age at diagnosis of PSC
Coefficient
HR
Cox Regression P
0.4112 ⫺0.0886 0.0271
1.509 0.915 1.027
0 0 .0001
Additive Model P
0 0 .0007
In this study, we present a time-dependent proportional-hazards model for the estimation of prognosis in PSC. The timedependent model performed markedly better than the time-fixed analysis by assigning a low 1-year survival probability to a higher proportion of patients who actually survived less than 1 year. This observation is in accordance with the concept that time-fixed models may tend to overestimate survival if they are applied to follow-up data.11,22 PSC patients frequently present and are diagnosed during periods of exacerbation, and experience prolonged remissions thereafter. Overestimation of survival by models based on the initial status may therefore be problematic, particularly for this disease. Multivariate Cox regression analysis with time-dependent variables has previously been constructed for patients with cirrhosis of various etiologies23 and for patients with PBC.11,24,25 Such models are calculated from follow-up data and can be used to update the patient’s short-term prognosis according to a change in the clinical condition.2,26 Time-fixed models, on the other hand, do not take
Fig. 2. Comparison of the 1-year survival probabilities assigned by the time-fixed and time-dependent prognostic models, respectively, to 18 PSC patients who died within 1 year of diagnosis. The mark corresponding to predicted 1-year survival 0.92 according to the time-fixed model and 0.98 according to the time-dependent model represents 3 patients. The timedependent model assigns a 1-year survival probability of ⱕ0.5 to 7 patients, whereas this is the case for only 2 patients by the time-fixed model.
656
BOBERG ET AL.
Fig. 3. Prognosis of PSC patients based on the time-dependent model. The curve shows the relationship between the prognostic index (PI) calculated from the multiple time-dependent prognostic model and the estimated 1-year survival probability.
into account changes of the covariates at times after entry. In the study by Christensen et al.,11 the time-dependent model predicted the survival probability for up to 6 months in PBC better than a time-fixed model for this disease. In the case of PSC, efforts until now have only focused on improving the applicability of established time-fixed models.27 The incorporation of updated covariates represents a new approach to assess prognosis in this disease. Optimal timing of liver transplantation is one of the major clinical challenges in the care of PSC patients. Because the time-dependent model is expected to give better short-term predictions of the individual outcomes than a corresponding time-fixed method,26 this new approach may assist in the selection of candidates for liver transplantation. ln(bilirubin), albumin, and age were independent prognostic factors in multivariate analyses in both the time-fixed and the time-dependent Cox regression models based on data from our population of PSC patients, but the relative importance of bilirubin compared with the other statistical variables was more pronounced in the latter model. For each unit increase on the natural logarithmic scale, corresponding to an increase of 2.71 on the original scale, bilirubin increased the risk by a factor of 2.840 in the updated covariates model compared with 1.509 in the time-fixed model. Serum bilirubin concentration is a prognostic factor in the majority of the previous prognostic models for PSC.5,7,8,28 The present model also uses the information in bilirubin levels associated with changes in the clinical condition. All of the traditional prognostic models include age and histologic stage as variables.5-8 Recently, a revised model that avoids the need for a liver biopsy and that only includes more reproducible variables (age, bilirubin, albumin, AST, and history of variceal bleeding) was proposed by the Mayo Clinic.9 The Child-Pugh score has also been suggested as a useful predictor of survival in PSC patients, but only as a variable in a time-fixed model.29 Comparing the Child-Pugh classification with the Mayo natural-history model, Kim et al.30 concluded that the Mayo model is superior at least in early stages of PSC. As in the revised Mayo model,9 we did not enter liver histology as a potential prognostic variable. In clinical routine, and particularly for current
HEPATOLOGY, March 2002
updating of prognosis, it is impractical to rely on a liver biopsy. Age and serum bilirubin and albumin levels that we identified as significant prognostic variables in the multiple models are parameters that have the advantage of being easily available and reproducible. Ideally, a time-dependent prognostic model should be derived from follow-up of patients at regular intervals without missing observations.11,26 Missing observations could potentially bias results. However, this is not likely to be a major problem in our case, in which the focus was on the comparison of models. The median survival time for the patients remains virtually the same if the 55 patients with incomplete data are included. We have included information from hospital admissions as a result of disease exacerbations as well as from follow-up in the steady-state condition for our patients, but with the limitations given by a retrospective data collection. Missing values and irregular intervals between the follow-ups have been handled by assuming that the value of a variable remains constant until the next observation.11 The additive regression model of Aalen has not previously been implemented in studies of prognosis in PSC. This model has the ability to describe variations in the significance of prognostic variables over time.19 Furthermore, we re-emphasize that the assumption of proportional hazards required for the Cox model to be true does not always hold, and violations are not always easily detected. It is therefore important to introduce and study alternative approaches. Obviously, credibility is given to the results when different methods deliver essentially similar conclusions. There are several ways to validate a prognostic model.11,26 The optimal approach is to test prognosis on an independent, but related, group of patients. Collection of multiple data sets over a long follow-up time from a large number of patients requires considerable efforts. To obtain as much power as possible in the new model, we chose to include all of the available patients in the calculations. The multivariate models (Tables 5 and 6) passed the tests we performed using plots and Martingale-based residuals.18 In a few cases, the proportionality assumption of the Cox model was violated, but the significance was confirmed by the additive model in these instances. Ideally, if a prognostic model is designed to be used as input to a decision regarding transplantation, the model should be developed based on a representative sample of such patients. A new, similarly selected group of patients should then serve to validate and compare the various models. Such data sets have not been available for us or others, as far as we know. The closest we come to mimicking such data is the cross-validation procedure that we chose to test the reliability of the models.21 This procedure is designed to check the prognostic features of a model by omitting one patient at a time from the calculation of the model, and subsequently comparing the observed survival probability with that estimated from the model. This is performed to avoid the risk of overfitting. The cross-validation method also allowed a comparison between the time-fixed and time-dependent models. It is important to note that although the time-dependent model performed better than the time-fixed variant in this test, both models are hampered by imprecision in terms of assigning a correct survival probability to the individual patient. This is a general limitation to prognostic models for a number of reasons.2,9,22 The prognosis of a patient also depends on biological factors other than
HEPATOLOGY, Vol. 35, No. 3, 2002
those we can include as variables in statistical models. Important prognostic variables may be unknown, or they may interact in unknown ways.2 Still, we consider the current time-dependent prognostic model to represent an important new aspect of disease modeling in PSC that should also be validated in independent groups of patients.
References 1. Johnston ME, Langton KB, Haynes RB, Matthieu D. A critical appraisal of research on the effects of computer based decision support systems on clinical performance and patient outcomes. Ann Intern Med 1994;120:135-142. 2. Christensen E. Prognostic models in chronic liver disease: validity, usefulness and future role. J Hepatol 1997;26:1414-1424. 3. Wiesner RH. Diagnostic criteria, clinical manifestations and natural history of primary sclerosing cholangitis. In: Krawitt EL, Wiesner RH, Nishioka M, eds. Autoimmune Liver Diseases. 2nd ed. Amsterdam: Elsevier Science B.V., 1998:381-412. 4. Boberg KM, Schrumpf E. Treatment of primary sclerosing cholangitis. In: Krawitt EL, Wiesner RH, Nishioka M, eds. Autoimmune Liver Diseases. 2nd ed. Amsterdam: Elsevier Science B.V., 1998:529551. 5. Wiesner RH, Grambsch PM, Dickson ER, Ludwig J, MacCarty RL, Hunter EB, Fleming TR, et al. Primary sclerosing cholangitis: natural history, prognostic factors and survival analysis. HEPATOLOGY 1989; 10:430-436. 6. Farrant JM, Hayllar KM, Wilkinson ML, Karani J, Portmann BC, Westaby D, Williams R. Natural history and prognostic variables in primary sclerosing cholangitis. Gastroenterology 1991;100:17101717. 7. Dickson ER, Murtaugh PA, Wiesner RH, Grambsch PM, Fleming TR, Ludwig J, LaRusso NF, et al. Primary sclerosing cholangitis: refinement and validation of survival models. Gastroenterology 1992; 103:1893-1901. 8. Broome´ U, Olsson R, Lo¨o¨f L, Bodemar G, Hultcrantz R, Danielsson Å, Prytz H, et al. Natural history and prognostic factors in 305 Swedish patients with primary sclerosing cholangitis. Gut 1996;38:610615. 9. Kim WR, Therneau TM, Wiesner RH, Poterucha JJ, Benson JT, Malinchoc M, LaRusso NF, et al. A revised natural history model for primary sclerosing cholangitis. Mayo Clin Proc 2000;75:688-694. 10. Broome´ U, Eriksson LS. Assessment for liver transplantation in patients with primary sclerosing cholangitis. J Hepatol 1994;20:654659. 11. Christensen E, Altman DG, Neuberger J, de Stavola BL, Tygstrup N, Williams R, the PBC1 and PBC2 trial groups. Updating prognosis in primary biliary cirrhosis using a time-dependent Cox regression model. Gastroenterology 1993;105:1865-1876. 12. Chapman RWG, Arborgh BÅM, Rhodes JM, Summerfield JA, Dick R, Scheuer PJ, Sherlock S. Primary sclerosing cholangitis: a review of its clinical features, cholangiography, and hepatic histology. Gut 1980;21:870-877.
BOBERG ET AL.
657
13. Wiesner RH, LaRusso NF. Clinicopathologic features of the syndrome of primary sclerosing cholangitis. Gastroenterology 1980;79: 200-206. 14. Schrumpf E, Elgjo K, Fausa O, Gjone E, Kolmannskog F, Ritland S. Sclerosing cholangitis in ulcerative colitis. Scand J Gastroenterol 1980;15:689-697. 15. Fausa O, Schrumpf E, Elgjo K. Relationship of inflammatory bowel disease and primary sclerosing cholangitis. Semin Liver Dis 1991;11: 31-39. 16. Cox DR. Regression models and life-tables (with Discussion). JR Stat Soc 1972;34:187-220. 17. Therneau TM. Extending the Cox Model. Technical Report Number 58. Mayo Foundation, Rochester, MN, 1996. 18. Therneau TM, Grambsch PM, Fleming TR. Martingale-based residuals for survival models. Biometrika 1990;77:147-160. 19. Aalen OO. Further results on the non-parametric linear regression model in survival analysis. Stat Med 1993;12:1569-1588. 20. Orlandi F, Christensen E. A consensus conference on prognostic studies in hepatology. J Hepatol 1999;30:171-172. 21. Verweij PJM, Houwelingen HC. Cross validation in survival analysis. Stat Med 1993;12:2305-2314. 22. Klion FM, Fabry TL, Palmer M, Schaffner F. Prediction of survival of patients with primary biliary cirrhosis. Examination of the Mayo Clinic model on a group of patients with known endpoint. Gastroenterology 1992;102:310-313. 23. Christensen E, Schlichting P, Kragh Andersen P, Fauerholdt L, Schou G, Vestergaard Pedersen B, Juhl E, et al. Updating prognosis and therapeutic effect evaluation in cirrhosis with Cox’s multiple regression model for time-dependent variables. Scand J Gastroenterol 1986;21:163-174. 24. Murtaugh PA, Dickson ER, Van Dam GM, Malinchoc M, Grambsch PM, Langworthy AL, Gips CH. Primary biliary cirrhosis: prediction of short-term survival based on repeated patient visits. HEPATOLOGY 1994;20:126-134. 25. Hughes MD, Raskino CL, Pocock SJ, Biagini MR, Burroughs AK. Prediction of short-term survival with an application in primary biliary cirrhosis. Stat Med 1992;11:1731-1745. 26. Altman DG, de Stavola BL. Practical problems in fitting a proportional hazards model to data with updated measurements of the covariates. Stat Med 1994;13:301-341. 27. Wiesner RH. Liver transplantation for primary biliary cirrhosis and primary sclerosing cholangitis: predicting outcomes with natural history models. May Clin Proc 1998;73:575-588. 28. Schrumpf E, Abdelnoor M, Fausa O, Jenssen E, Kolmannskog F. Risk factors in primary sclerosing cholangitis. J Hepatol 1994;21: 1061-1066. 29. Shetty K, Rybicki L, Carey WD. The Child-Pugh classification as a prognostic indicator for survival in primary sclerosing cholangitis. HEPATOLOGY 1997;25:1049-1053. 30. Kim WR, Poterucha JJ, Wiesner RH, LaRusso NF, Lindor KD, Petz J, Therneau TM, 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; 29:1643-1648.
