Feature selection and classifier performance in computer-aided diagnosis: the effect of finite sample size.
In computer-aided diagnosis (CAD), a frequently used approach for distinguishing normal and abnormal cases is first to extract potentially useful features for the classification task. Effective features are then selected from this entire pool of available features. Finally, a classifier is designed using the selected features. In this study, we investigated the effect of finite sample size on classification accuracy when classifier design involves stepwise feature selection in linear discriminant analysis, which is the most commonly used feature selection algorithm for linear classifiers. The feature selection and the classifier coefficient estimation steps were considered to be cascading stages in the classifier design process. We compared the performance of the classifier when feature selection was performed on the design samples alone and on the entire set of available samples, which consisted of design and test samples. The area Az under the receiver operating characteristic curve was used as our performance measure. After linear classifier coefficient estimation using the design samples, we studied the hold-out and resubstitution performance estimates. The two classes were assumed to have multidimensional Gaussian distributions, with a large number of features available for feature selection. We investigated the dependence of feature selection performance on the covariance matrices and means for the two classes, and examined the effects of sample size, number of available features, and parameters of stepwise feature selection on classifier bias. Our results indicated that the resubstitution estimate was always optimistically biased, except in cases where the parameters of stepwise feature selection were chosen such that too few features were selected by the stepwise procedure. When feature selection was performed using only the design samples, the hold-out estimate was always pessimistically biased. When feature selection was performed using the entire finite sample space, the hold-out estimates could be pessimistically or optimistically biased, depending on the number of features available for selection, the number of available samples, and their statistical distribution. For our simulation conditions, these estimates were always pessimistically (conservatively) biased if the ratio of the total number of available samples per class to the number of available features was greater than five.