Background Although a majority of studies in cancer biomarker discovery claim to use proportional hazards regression (PHREG) to the study the ability of a biomarker to predict survival, few studies use the predicted probabilities obtained from the model to test the quality of the model. In this paper, we compared the quality of predictions by a PHREG model to that of a linear discriminant analysis (LDA) in both training and test set settings. Methods The PHREG and LDA models were built on a 491 colorectal cancer (CRC) patient dataset comprised of demographic and clinicopathologic variables, and phenotypic expression of p53 and Bcl-2. Two variable selection methods, stepwise discriminant analysis and the backward selection, were used to identify the final models. The endpoint of prediction in these models was five-year post-surgery survival. We also used linear regression model to examine the effect of bin size in the training set on the accuracy of prediction in the test set. Results The two variable selection techniques resulted in different models when stage was included in the list of variables available for selection. However, the proportion of survivors and non-survivors correctly identified was identical in both of these models. When stage was excluded from the variable list, the error rate for the LDA model was 42% as compared to an error rate of 34% for the PHREG model. Conclusions This study suggests that a PHREG model can perform as well or better than a traditional classifier such as LDA to classify patients into prognostic classes. Also, this study suggests that in the absence of the tumor stage as a variable, Bcl-2 expression is a strong prognostic molecular marker of CRC.
An adapted linear discriminant analysis with variable selection for the classification in high-dimension, and an application to medical data