WeiRD - a fast and performant multivoxel pattern classifier

摘要

Emerging from their initial application to functional imaging data, multivoxel pattern classifiers are increasingly applied to structural data to predict, among many others, disease status and psychological traits. Naturally, the choice of a suitable classifier in such analyses is a critical step. In the present paper, we assessed the performance of two frequently used classifiers (Support Vector Machine, SVM, and Random Forest, RF) and a novel voxel-wise voting classifier (Weighted Robust Distance, WeiRD) against different simulated scenarios and against a realworld structural dataset of alcoholic patients and controls (LeADstudy, N=217). The simulated scenarios comprised a dislocation scenario, in which a certain proportion of functionally corresponding voxels are located at different positions between samples; a noise scenario, in which the proportion of voxels containing informative signal was systematically varied; and a phenotype scenario, in which a given class consists of multiple subclasses or "phenotypes". Our simulation results show that either RF or WeiRD outperformed SVM across large parameter ranges in these scenarios. While WeiRD performed best among the classifiers in the dislocation and noise scenario, RF showed superior classification in the phenotype scenario. Applied to the structural dataset of alcoholic patients and controls, RF and WeiRD outperformed SVM across all tested smoothing kernels (0, 4, 8, 12, 16, 20, 24 mm isotropic) and resampling resolutions (1.5, 3, 6, 9, 12 mm isotropic). While classification scores generally increased with the amount of smoothing, RF performed best among classifiers for smaller smoothing kernels (0-8 mm) and WeiRD best for larger smoothing kernels (12-24 mm). In sum, our results (1) show that the classifier performance critically depends on the structure of the data and (2) provide guidance in cases where prior information about this structure is available. In addition, we introduce WeiRD, a simple and fast distance-tocentroid classifier, which performed well for simulated and realworld structural data.