Title:Multiscale statistical testing for connectome-wide association studies in fMRI

Abstract:Alteration in fMRI brain connectivity has been found to be associated with a variety of clinical disorders. Yet, the systematic identification of connections associated with a disease state is a "needle in a haystack" problem. There are many connections to test and potentially few of them show a strong association with the variable of interest. The number of brain parcels (or scale) used to measure connections is a key parameter, as more parcels means more anatomically accurate measures, but also means many more connections to test and thus more risk to find spurious associations by chance. This work presents a multiscale approach designed to assess the impact of scale on connectome-wide association studies. A mass univariate general linear model is applied independently at each connection, for functional networks generated through clustering at different scales. The multiple comparison problem is addressed at each scale independently, using the false-discovery rate. An omnibus permutation test is implemented to reject the global null hypothesis of no association across all scales. Our results support the fact that, when the global null is rejected, the false-discovery rate is well controlled both within and across scales. On simulations, we showed that multiscale group differences can be detected with excellent sensitivity in large sample size (n=100), and good sensitivity in moderate sample size (n=40). On three real resting-state datasets (schizophrenia, congenital blindness, motor practice), our results presented an excellent face validity. In addition, the multiscale method performed better than current techniques such as network-based statistics or multivariate distance matrix regression, both on simulations and real datasets. Taken together, our findings demonstrate that our approach is a powerful tool to mine for associations across the whole connectome, at multiple scales.