Extracting functional networks with spatial independent component analysis: the role of dimensionality, reliability and aggregation scheme

Purpose of review Clinical studies have differentiated functional brain networks in neurological patient and control populations using independent component analysis (ICA) applied to functional MRI (fMRI). Principal component analysis (PCA) is used to reduce the data dimensionality to make this feasible. The role of this choice is reviewed in connection with the accuracy and the reliability of the ICA results and the schemes of data aggregation in population studies. Recent findings It has been pointed out recently that it is important to critically explore the ICA model orders without relying on strictly predetermined PCA cutoffs for the number of components. We further illustrate this aspect empirically by showing that a large enough range of dimensions may exist where ICA components remain accurate but also that the minimum PCA dimension required to reliably extract the best ICA maps may vary substantially across subjects. Moreover, with the aid of a simple simulation, we show that reliable independent components can still be recovered beyond a theoretical PCA cutoff. Summary The role of the PCA cutoff and its impact on the accuracy and reliability of the ICA results should be carefully considered in future clinical fMRI studies.