Reproducibility of graph measures at the subject level using resting‐state fMRI

INTRODUCTION: Graph metrics have been proposed as potential biomarkers for diagnosis in clinical work. However, before it can be applied in a clinical setting, their reproducibility should be evaluated. METHODS: This study systematically investigated the effect of two denoising pipelines and different whole‐brain network constructions on reproducibility of subject‐specific graph measures. We used the multi‐session fMRI dataset from the Brain Genomics Superstruct Project consisting of 69 healthy young adults. RESULTS: In binary networks, the test–retest variability for global measures was large at low density irrespective of the denoising strategy or the type of correlation. Weighted networks showed very low test–retest values (and thus a good reproducibility) for global graph measures irrespective of the strategy used. Comparing the test–retest values for different strategies, there were significant main effects of the type of correlation (Pearson correlation vs. partial correlation), the (partial) correlation value (absolute vs. positive vs. negative), and weight calculation (based on the raw (partial) correlation values vs. based on transformed Z‐values). There was also a significant interaction effect between type of correlation and weight calculation. Similarly as for the binary networks, there was no main effect of the denoising pipeline. CONCLUSION: Our results demonstrated that normalized global graph measures based on a weighted network using the absolute (partial) correlation as weight were reproducible. The denoising pipeline and the granularity of the whole‐brain parcellation used to define the nodes were not critical for the reproducibility of normalized graph measures.