Mild Traumatic Brain Injury (MTBI) is a growing public health problem with an underestimated incidence of over one million people annually in the U.S. Neuropsychological tests are used to both assess the patient condition and to monitor patient progress. This work aims to use features extracted from MR images taken shortly after injury to predict the performance of MTBI patients on neuropsychological tests one year after injury. Successful prediction can enable early patient stratification and proper treatment planning. The goal of this project is to determine the most effective features and corresponding prediction method for such prediction. The main challenge is that we have only limited training data, from which we need to develop the prediction method that can be expected to provide accurate prediction results for unseen data. In the machine learning literature, feature selection is usually done to minimize the cross validation error, which still has a chance for overfitting if the available data set is small and noisy. We propose a novel criterion for feature selection, which considers both the cross validation error and the prediction model variance, to further reduce the chance for overfitting. The algorithm is applied to a data set of 15 MTBI patients. The proposed method was able to determine a subset of MR image features for predicting each neuropsychological test to yield both small prediction error and prediction variance.