Article Figures & Data
Figures
- Fig 1.
Illustration of the diagnostic workflow and data-processing pipeline proposed by most of the studies reviewed here.
- Fig 2.
Overview of MVPA-based diagnostic fMRI studies with an above-average statistical power to detect successful models (n ≥ 25 in every group of the training set; group size ratio ≤ 1.2 as a prerequisite for comparing the overall classification accuracies reported).34,43,45,55,61,72,74,85,95 Symbol sizes represent average group sizes in the respective study. a, Only minimum and maximum classification accuracies are shown here. b, Study of cross-validation findings also reports results in a smaller independent validation set. c, Only Gaussian process classifier overall accuracy is shown here. d, High overlap with another study is not shown here. FS indicates feature selection. For a complete list of studies, see On-line Tables 1 and 2.
Tables
Proposed reporting standards for future studies applying multivariate classification techniques to fMRI data with a diagnostic intention in addition to general reporting standards in fMRI research
Proposed Reporting Standards Desirable Aspects of Study Design Report of overall classification accuracies/generalization rate (in addition to sensitivity and specificity if applicable) Sufficiently large, equally sized groups (25 subjects per group desirable, but depending on specific methodologies used and power analyses) Rigorous statistical tests Sufficiently long scan duration for resting state fMRI (at least 5 minutes, evaluation of a potential benefit of longer acquisitions desirable) Report of all tested models (not only optimized/best models) Systematic model optimization Clearly identify inferential tests and explorative analyses Well-chosen and well-defined clinical problem Clearly identify feature extraction and feature selection/weighting strategies Multisite/multiscanner studies Denominate origin of classification algorithms if an independent toolbox is used for pipeline building Evaluation of functionally/anatomically interpretable classifiers and methods with individual reliability measures Good clinical characterization of the sample including demographic data (psychometric data if applicable) Independent validation set
