Optimal choice of parameters in functional connectome-based predictive modelling might be biased by motion: comment on Dadi et al

In a recent study, Dadi and colleagues make recommendations on optimal parameters for functional connectome-based predictive models. While the authors acknowledge that “optimal choices of parameters will differ on datasets with very different properties”, some questions regarding the universality of the recommended “default values” remain unanswered. Namely, as already briefly discussed by Dadi et al., the datasets used in the target study might not be representative regarding the sparsity of the (hidden) ground truth (i.e. the number of non-informative connections), which might affect the performance of L1- and L2-regularization approaches and feature selection. Here we exemplify that, at least in one of the investigated datasets systematic motion artefacts might bias the discriminative signal towards “non-sparsity”, which might lead to underestimating the performance of L1-regularized models and feature selection. We conclude that the expected sparsity of the discriminative signal should be carefully considered when planning predictive modelling workflows and the neuroscientific validity of predictive models should be investigated to account for non-neural confounds.