A Novel Feature Selection Approach for Analyzing High dimensional Functional MRI Data

Feature selection based on traditional multivariate methods is likely to obtain unstable and unreliable results in case of an extremely high dimensional space and very limited training samples. In order to overcome this difficulty, we introduced a novel feature selection method which combines the idea of stability selection approach and the elastic net approach to detect discriminative features in a stable and robust way. This new method is applied to functional magnetic resonance imaging (fMRI) data, whose discriminative features are often correlated or redundant. Compared with the original stability selection approach with the pure l_1 -norm regularized model serving as the baseline model, the proposed method achieves a better sensitivity empirically, because elastic net encourages a grouping effect besides sparsity. Compared with the feature selection method based on the plain Elastic Net, our method achieves the finite sample control for certain error rates of false discoveries, transparent principle for choosing a proper amount of regularization and the robustness of the feature selection results, due to the incorporation of the stability selection idea. A simulation study showed that our approach are less influenced than other methods by label noise. In addition, the advantage in terms of better control of false discoveries and missed discoveries of our approach was verified in a real fMRI experiment. Finally, a multi-center resting-state fMRI data about Attention-deficit/ hyperactivity disorder (ADHD) suggested that the resulted classifier based on our feature selection method achieves the best and most robust prediction accuracy.