The purpose of our study was to develop a systematic model according to universally recognized clinicopathologic parameters for improved accuracy of prognostic outcome prediction in HCC patients after curative surgery. Therefore, all acknowledged factors associated with outcomes were evaluated. Notably, certain important parameters, such as tumor staging and venous infiltration status, would require postoperative histological examination of the resected tissues or biopsy samples of the patients, unless future radiological examination using dynamic MRI could provide such definitive diagnosis. Nevertheless our primary contribution is the modeling framework which incorporates multiple parameters in prediction prognosis. Its flexible nature allows us to easily remove parameters (e.g. VENINV when either biopsy samples are not available or diagnostic radiography data is not affirmative) or to add new biomarkers (e.g. newly identified gene signatures). Several studies of HCC have reported the ability to use clinicopathologic parameters or biomarkers towards grouping subjects and predicting survival outcomes [2, 3, 14]. However, most studies were small or without an independent validation set. Likewise, there are no models built to systematically incorporate all factors to predict outcome. In addition, the existing prediction rules appear to be ad hoc, involve multiple arbitrary cutoffs and lack statistical rigor. Lastly, the performance of such rules has not been formally assessed by such analyses, e.g., ROC.
The salient contribution of this publication is the identification of certain common clinicopathologic parameters and biomarkers strongly associated with the prognostic outcome that are further confirmed by two different protocols: LOO and independent validation. Previous studies have revealed some of these predictors, but often lack reproducibility which may due to the difference among patient populations. For example, the majority of the Chinese patients carry hepatitis B [15], while hepatitis C virus is prevalent in Western Europe [16]. The relatively small sample sizes of the aforementioned data sets contribute to the inconsistent results. Herein, we provide confirmation to a previous report using a large sample size and validation data. The first group of predictors highlighted in this paper is the tumor stage. There are growing numbers of staging systems available but unfortunately not one is perfect with each having individual strengths and weakness. Conversely, our data sets show staging systems are highly correlated. The following parameters were found to be strongly associated with prognosis and prediction value: AFP, albumin level, venous infiltration and the number of tumor nodules. Among these predictors, the serum albumin level may reflect the conditions of patients liver physiology, while the rest of other factors hint to the biology or characteristics of the HCC tumor per se. This may reflect the importance of both the host (or the microenvironment) and tumor factors contributing to the outcomes. The self-reported variable alcohol consumption was marginally significant in a univariate model but was not used and determined to be non-significant in our multivariate Cox model. This finding is highly consistent with previous report on Chinese HCC patients [15, 17] where dichotomized serum albumin and tumor size, number of tumor nodules, venous infiltration, and tumor stage were significantly associated with cancer prognosis after surgery.
The prediction power of the predicted hazard (h) was demonstrated in this report, which is literally the linear combination of many predictors. A number of studies were conducted quantifying the predictive power of clinicopathological parameters, and our results are consistent with previous reports on vascular invasion [18, 19], AFP level [15, 19] and tumor size [19] in term of hazard ratios. However, this study did not measure some variable, e.g., portal hypertension and bilirubin [20], therefore, we could not directly assess their predictive value. 97.3% of our patients were classified as Child-Pugh grade A, and we found Child-Pugh grade not to be a significant predictor (possibly due to lack of statistical power). Nevertheless, our primary goal is to develop the objective and flexible framework, which can easily accommodate additional biomarkers when becoming available. There are well known scoring systems to classify HCC, including Child-Pugh [18], the Oku [11], Advanced Liver Cancer Prognostic System (ALCPS, applicable to patients with advanced HCC who were not amendable to surgery or locoregional therapy) [21], and Barcelona Clinic Liver Cancer Group [20, 22]. As the drawback, these scores have been devised by a series of ad hoc rules. In the other hand, predicted hazard is objective and can readily incorporate a new patient and biomarkers information. This can be simply performed by updating the linear model and the set of coefficients and will allow us to continuously update the model when new data becomes available. This is particularly important as new biomarkers based on molecular studies of HCC could be incorporated into the model. HCC has a heterogeneous etiology and many factors (e.g. patients' ethnicity and genetic background) may affect prognosis. Therefore, this model may not be directly applicable to a different HCC cohort in the Western countries. This approach should serve as a general framework, where the Cox linear classifier can be trained on a particular cohort and applied to future patients.
A considerable fraction of our HCC patients achieved 5+ years of DFS after surgery, especially for those with favorable clinicopathologic profiles as highlighted in our previous reports [15]. The cutoff of 5 or 10 years is widely used to define the cure of HCC, however, such thresholds are arbitrary. The categorization of a patient with 5 year of DFS as a cure, while classifying another patient with 4.9 year DFS as a failure does not lead to valid conclusions. Therefore, treatment of survival and DFS as quantitative traits was performed without the application of any cutoffs or subjective dichotomization of the clinicopathologic parameters. HCC patient follow-up is a non-trivial task given the ever increasing mobility of an urban population. Due to improved treatment and management, HCC patients live longer and thus make follow-up evaluations a challenge. One strategy is to only analyze the patients with observed events [15, 23], but this approach would greatly reduce the sample size (as well as statistical power), since many of the data point were censored (Figure 1). Instead, we employed of time-to-event methods (e.g. Cox model), where the censored data also contributed to the test and improved statistical power. Moreover, the Cox model also quantifies the strength of the association between clinicopathologic parameters and prognosis (in the form of relative hazard), which is actually the foundation for prediction.
In the past decade we have observed a fast growth of literature using gene expression profiles and DNA abnormality to predict cancer prognosis. However, these approaches have not achieved real clinicopathologic applicability because of the inter-institutional variation caused by numerous factors such as; array platform, statistical algorithm, reagent, laboratory condition/protocol. In contrast, clinicopathologic parameters are more standard, robust, and available worldwide. Models built on clinicopathologic factors can be more readily used in today's hospital practice. Further, the linear model should certainly include new biomarkers (to improve prediction power if there is conclusive evidence that a biomarker gives additional information conditioning on known factors. Recent studies [24, 25] revealed gene expression in tumor and adjacent normal liver tissues (suggesting a so called "field-effect") were predictive for HCC prognosis. It has been suggested that mechanistically these signatures capture tumor status, damage to liver tissue and the state of inflammation which relates to the likelihood of subsequent tumors arising [23, 25]. However, all such studies failed to incorporate clinicopathological and expressional predictors together. In such cases, the identified expressional predictors first capture the same information as clinicopathological parameters (e.g. cancer stage), and the performance for the expressional biomarkers would not necessarily outperform clinicopathological predictors (unpublished results). Herein, we argue that in order to capture information from gene expression that was not redundant to clinicopathologic data, the clinicopathologic parameters would be included during the search for expression signatures, and our hazard score model will serve for this purpose.
In summary, we have retrospectively analyzed the association of common clinicopathologic parameters with clinical outcomes in a large sample cohort (n = 572) of HCC patients after curative surgery. We quantified the strength of the association and then built a classifier based on multi-variates Cox regression. This approach has excellent prediction power. Nonetheless, there is still a large fraction of variance that has not been explained since the area under the ROC is around 70% in validation data. The search for other biomarker to improve the prediction of HCC survival will certainly be an important direction for future research. As a caveat, the models that employ clinicopathologic parameters are often competitive with most of the models incorporating gene expression data. This is due to both set of predictors capturing the tumor stage [8]. Hence, the model described in this paper should serve as a foundation for future work when biomarkers at the protein, mRNA or DNA level are considered. Furthermore, the selection of additional biomarkers should be based on common clinicopathologic parameters; otherwise, they may recapture similar information and provide little extra predictive value.