A total of 180 patients were examined for eligibility, and 172 were included in the study. The reasons for non-participation were: 3 patients with HCC, 3 with repeated alcohol use, and 2 with severe infection disease. The period of recruitment lasted from May 2003 to February 2004. One hundred and forty one patients completed the follow-up period. Thirty one patients died during the study, 29 liver-related and 2 unrelated to liver disease (myocardial infarction). One hundred and seventy patients were included in the outcome analyses.
The patients’ clinical and serological features are summarized in Table .
In the derivation data set, the median follow-up period was 56 wk (range, 4-104 wk). The CP median score was 7 (range, 5-14) with 61% of the patients being CP class B and C. The MELD and BioCliM median scores were 17 (range, 8-42) and 7.7 (range, 5.7-13.6), respectively. During follow-up, 29 patients (17%) died. The 4-, 12-, 24-, 52- and 104-wk survival rates were 98%, 98%, 90%, 89% and 83%, respectively.
The patients of the validation group were followed for a median of 58 wk (range, 8–104 wk) during which 13 died. The 4-, 12-, 24-, 52- and 104-wk survival rates were 96%, 95%, 88%, 84% and 83%, respectively. The CP median score was 7 (range, 5-14) with 65% of the patients being CP class B and C. The MELD and BioCliM median scores were 18 (range, 9-43) and 7.9 (range, 6-13.8), respectively.
None of the patients in the derivation or validation groups underwent liver transplantation during the follow-up period.
Overall survival according to single prognostic factors
Univariate analysis for 104-wk overall survival: Univariate analysis using Cox proportional hazards models showed that serum levels of creatinine, bilirubin, cholesterol, albumin, prothrombin time, partial thromboplastin time, ascites, spontaneous bacterial peritonitis, encephalopathy and bleeding esophageal varices were significantly associated with survival (Table ).
Association of baseline characteristics with mortality in 170 cirrhotic patients, results from univariate Cox proportional hazards models
Multivariate analysis for 104-wk overall survival: Multivariate Cox regression analysis included those variables independently related to survival resulting from univariate analysis. The selected variables were available in all patients that entered the forward stepwise model. Of the candidate variables, only ascites, encephalopathy, bleeding esophageal varices and serum creatinine were independently predictive of survival (Table ).
Contributions of different variables to survival prediction at 104 wk, results from multivariate Cox regression models1
The estimated hazard risk for ascites suggested that the risk of death for uncontrolled ascites was 10.2 times greater than for those with absent or controlled ascites. The risk of death in those patients with relapsing bleeding and uncontrolled encephalopathy increased 3.25 times compared to those without bleeding or with non-relapsing bleeding, and 2.5 times compared to those with absent or controlled encephalopathy. In terms of impact in prognosis, the ascites (hazard ratio (HR), 10.2) and serum creatinine (HR, 3.99) were the most important prognostic factors.
Model derivation and development
In the model derivation cohort, 11 potential variables selected from the univariate analysis (P < 0.05) were calculated for model inclusion. Of these only 5 were included in the model. The likelihood ratio statistic showed the significance of the addition of each variable separately to a predictive model that included ascites only (Table ). The χ2 statistic was progressively increased with the addition of creatinine, bleeding esophageal varices, hepatic encephalopathy, and bilirubin. The c-statistic in the model that included only ascites was 0.76, based on the c-statistic for censored data. When creatinine, BEV, HE and bilirubin were added to the model, the c-statistic was improved to 0.83, 0.85, 0.89, and 0.90, respectively. In the same context, the combination of ascites, creatinine, BEV, HE and bilirubin revealed the smallest BIC value (1951.77), thus, in the derivation set, the model with the combination of clinical and biochemical variables appeared to improve the risk prediction.
Computational formula for 104-wk risk using best-fitting model
The regression coefficients of the formula for calculating the new risk score (biochemical and clinical model) were selected from a Cox regression model[16
] and are reported in Table .
The risk scores for individual patients were calculated using the following equation: [1.370 × loge (creatinine mmol/L) + 0.349 × loge (bilirubin mmol/L) + 2.310 × (ascites: 0 if absent or medically controlled and 1 if uncontrolled) + 0.909 × (encephalopathy: 0 if absent or medically controlled and 1 if uncontrolled) + 1.195 × (bleeding esophageal varices: 0 if absent or present without relapses and 1 if present with relapses). The clinical variables were coded depending on the clinical response to medical treatment. The variables grouped together as “absent or medically controlled” (ascites and encephalopathy) and “absent or present without relapses” (bleeding esophageal varices) have been so grouped because their survival was similar in each one of them (Figure ). The missing values were imputed for survival modeling.
Kaplan-Meier estimated survival curves for clinical variables. A: Ascites; B: Encephalopathy; C: Bleeding esophageal varices.
Survival probabilities were derived from the Cox proportional hazards model: S(t) = S0(t)exp(R-Ro). S(t) is the survival probability in wk, S0(t) the baseline survival function, R the individual risk score and R0 the risk score of the average patient in the series. For example, the 12-wk survival probability is calculated as: S(12 wk)= 0.981exp(BioCliM score-7), where 0.981 is the 12-wk baseline survival and 7 is the reference BioCliM score. To ease its use, the score was multiplied by 100.
Predictive models for 12-, 52- and 104-wk survival
Comparison of the c-statistic values among the CP, MELD and BioCliM scores was performed. All scoring systems were found to have diagnostic accuracy in predicting survival. The BioCliM score, however, showed to have better discriminative power in predicting short- (12 wk), intermediate- (52 wk) and long-term survival (104 wk) than the rest of the scores (Figure ).
Comparison of the c-index values of the MELD, Child-Pugh and BioCliM scores for 12- (A), 52- (B) and 104-wk (C) survival. SE indicates standard errors. The different values were compared with BioCliM score using the bootstrap method.
The c-statistic for the CP and MELD scores were almost identical for 12-wk survival (0.82 and 0.82), and slightly higher for CP as compared with MELD for 52-wk (0.84 and 0.82) and 104-wk (0.86 and 0.82) survival.
We used an alternative way of computing the concordance probability for a censored outcome to estimate the true concordance probability in samples with a high censored proportion. The concordance probability estimates for the CP (CPE, 0.71; SE, 0.042), MELD (CPE, 0.74; SE, 0.043) and BioCliM (CPE, 0.78; SE, 0.050) models were lower at 12 wk in comparison with those obtained using the standard c-statistic value. Finally, the CPE at 12 wk was consistently higher for BioCliM as compared with CP and MELD scores.
Discrimination and model validation
The Hosmer-Lemeshow statistic (H-L) is a measure of the discrepancy between the observed and predicted risk. A better calibrated model would have a smaller discrepancy between the observed and predicted and thus a smaller H-L statistic.
A significant P value for the H-L statistic indicates a significant deviation between predicted and observed outcomes. Figure compares the calibration of the BioCliM, MELD and CP scores in predicting the probability of death at 104 wk. The H-L statistic was 4.69 for the BioCliM score, 17.06 for the MELD score and 14.23 for the CP score, indicating a good calibration for all models; however, this analysis clearly shows that BioCliM is better calibrated.
Figure 3 Observed and predicted probability of events at 104 wk. A, B and C shows the observed and predicted probability of death according to BioCliM, MELD and Child-Pugh scores in 10 groups (deciles) of patients, respectively. A significant P-value for the Hosmer-Lemeshow (more ...)
Figure illustrates the observed and expected Kaplan-Meier survival curves for each score in 2 patient subgroups divided according to risk score as low risk (R ≤ 8) and high risk (R > 8), selected from the “Calixto Garcia” Hospital. Using a cutoff value of 8 (risk score) to predict probability of survival within 104 wk, the sensitivity and specificity of the BioCliM score was 90% and 87%, respectively. Median survival was 104 wk and 47 wk for low- and high-risk groups, respectively. There were no significant differences between the observed and expected survival curves in the stratified risk groups (low risk, P = 0.61; high risk, P = 0.77). Thus, the BioCliM score allowed accurate prediction of survival in the cirrhotic patient validation group.
Figure 4 Observed and expected 104-wk survival curves for the BioCliM score. Survival of 85 independent patients from the “Calixto Garcia” Hospital who were stratified according to their risk score into two risk groups (low-risk ≤ 8 and (more ...)
Survival according to the BioCliM score
The differences in the short-, intermediate- and long-term survival between patients with low risk (≤ 8), and high risk (> 8) scores were compared (Figure ).
Kaplan-Meier estimated survival curves for the BioCli Model score.
Overall survival rates were significantly different between low-risk and high-risk patients (P < 0.0001). The 12-wk survival rates were 98% and 64% for low and high risk, respectively. For low and high risk, 1-year survival rates were 97% and 3%, and 2-year survival rates were 95% and 0%, respectively. Patients with a high risk score had the highest risk of mortality compared to patients with low values. Patients with a BioCliM score of ≥ 8 had a median survival of < 47 wk in comparison to patients with a median survival of 104 wk for patients with a BioCliM score of < 8.