Predicting Early Mortality After Acute Variceal Hemorrhage Based on Classification and Regression Tree Analysis

Background & Aims

Available prognostic models for mortality after an acute variceal hemorrhage have limitations that restrict their clinical value. We assessed the performance of a novel prognostic approach based on classification and regression tree (CART) analysis.

Methods

Logistic regression (LR) and CART analyses were performed to identify prognostic models for mortality at 6 weeks in a single-center cohort of 267 consecutive patients with acute variceal bleeding. Receiver operating characteristic (ROC) curves were constructed to assess the performance of the models. Prognostic models were fitted and validated by split-sample technique (training set, 164 patients, 2001–2005; test set, 103 patients, 2006–2008).

Results

After 6 weeks, 21% of patients experienced rebleeding and 24% died. The best LR model was based on Child–Pugh score, creatinine level, bacterial infection, and hepatocellular carcinoma. CART analysis provided a simple algorithm based on the combined use of just 3 variables (Child–Pugh score, creatinine level, and bacterial infection), allowing accurate early discrimination of 3 distinct prognostic subgroups with 8% (low risk), 17% (intermediate), and 50% to 73% (high) mortality. Its accuracy was similar to the LR model (area under the ROC curves, 0.81 vs 0.84; P = .17) and better than that of Child–Pugh (0.75; P = .05) and model for end-stage liver disease (0.74; P = .05). The prognostic accuracy of both LR and CART models was validated in the test set (area under the ROC curve values, 0.81 and 0.83, respectively).

Conclusions

A simple CART algorithm based on Child–Pugh score, creatinine level, and infection allowed an accurate predictive assessment of 6-week mortality after acute variceal bleeding.