Predicting prognosis in hepatocellular carcinoma after curative surgery with common clinicopathologic parameters

Abstract

Background: Surgical resection is one important curative treatment for hepatocellular carcinoma (HCC), but the prognosis following surgery differs substantially and such large variation is mainly unexplained. A review of the literature yields a number of clinicopathologic parameters associated with HCC prognosis. However, the results are not consistent due to lack of systemic approach to establish a prediction model incorporating all these parameters.

Methods: We conducted a retrospective analysis on the common clinicopathologic parameters from a cohort of 572 ethnic Chinese HCC patients who received curative surgery. The cases were randomly divided into training (n = 272) and validation (n = 300) sets. Each parameter was individually tested and the significant parameters were entered into a linear classifier for model building, and the prediction accuracy was assessed in the validation set

Results: Our findings based on the training set data reveal 6 common clinicopathologic parameters (tumor size, number of tumor nodules, tumor stage, venous infiltration status, and serum alpha-fetoprotein and total albumin levels) that were significantly associated with the overall HCC survival and disease-free survival (time to recurrence). We next built a linear classifier model by multivariate Cox regression to predict prognostic outcomes of HCC patients after curative surgery This analysis detected a considerable fraction of variance in HCC prognosis and the area under the ROC curve was about 70%. We further evaluated the model using two other protocols; leave-one-out procedure (n = 264) and independent validation (n = 300). Both were found to have excellent prediction power. The predicted score could separate patients into distinct groups with respect to survival (p-value = 1.8e-12) and disease free survival (p-value = 3.2e-7).

Conclusion: This described model will provide valuable guidance on prognosis after curative surgery for HCC in clinical practice. The adaptive nature allows easy accommodation for future new biomarker inputs, and it may serve as the foundation for future modeling and prediction for HCC prognosis after surgical treatment.

Figure 1

Kaplan-Meier survival curves of HCC patients in the training set. Relative hazard (h) was predicted for cancer survival (A) and disease-free survival (B) using a leave-one-out procedure. Patients were equally divided into two groups based on h, and their Kaplan-Meier survival functions were compared by log-rank test. Alternatively, we divided patients into three equal-sized groups based on their h, and observed excellent separation. Such results suggest the prediction is rather robust, and not sensitive to choice of grouping. The vertical bars on the survival curve denote censored patients.

Figure 2

Kaplan-Meier survival curves of HCC patients in the validation set. We first fit a multivariate Cox model using the initial training set focusing on overall cancer survival (A) and disease-free survival (B). This model was used to predict the relative hazard (h) for an independent testing set. Next, the testing patients were equally divided into two groups based on predicted h, and their Kaplan-Meier survival functions were compared by log-rank test.

Figure 3

ROC curve analyses of HCC patients in the training and testing datasets. Multivariate Cox model built on the initial training set was used to predict cancer prognosis in the training set and testing set. Time-dependent ROC and AUC were computed in the intervals of (A) 60 months and (B) 30 months to quantify the prediction accuracy.