Influence of atlas-choice on age and time effects in large-scale brain networks in the context of healthy aging.

There is accumulating cross-sectional evidence of decreased within-network resting-state functional connectivity (RSFC) and increased between-network RSFC when comparing older to younger samples, but results from longitudinal studies with healthy aging samples are sparse and less consistent. Some of the variability might occur due to differences in network definition and the fact that most atlases were trained on young adult samples. Applying these atlases to older cohorts implies the generalizability of network definitions to older individuals. However, because age is linked to a less segregated network architecture, this assumption might not be valid. To account for this, the Atlas55+ (A55) was recently published. The A55 was trained on a sample of people over the age of 55, making the network solutions suitable for studies on the aging process. Here, we want to compare the A55 to the popular Yeo-Krienen atlas to investigate whether and to what extent differences in network definition influence longitudinal changes of RSFC. For this purpose, the following networks were investigated: the occipital network (ON, "visual network"), the pericentral network (PN, "somatomotor network"), the medial frontoparietal network (M-FPN, "default network"), the lateral frontoparietal network (L-FPN, "control network"), and the midcingulo-insular network (M-CIN, "salience network"). Analyses were performed using longitudinal data from cognitively healthy older adults (= 228, mean age at baseline = 70.8 years) with five measurement points over 7 years. To define the five networks, we used different variants of the two atlases. The spatial overlap of the networks was quantified using the dice similarity coefficient (DSC). RSFC trajectories within networks were estimated with latent growth curve models. Models of varying complexity were calculated, ranging from a linear model without interindividual variability in intercept and slope to a quadratic model with variability in intercept and slope. In addition, regressions were calculated in the models to explain the potential variance in the latent factors by baseline age, sex, and education. Finally, the regional homogeneity and the silhouette coefficient were computed, and the spin test and Wilcoxon-Mann-Whitney test were used to evaluate how well the atlases fit the data. Median DSC across all comparisons was 0.67 (range: 0.20-0.93). The spatial overlap was higher for primary processing networks in comparison to higher-order networks and for intra-atlas comparisons versus inter-atlas comparisons. Three networks (ON, PN, M-FPN) showed convergent shapes of trajectories (linear vs. quadratic), whereas the other two networks (L-FPN, M-CIN) showed differences in change over time depending on the atlas used. The 95% confidence intervals of the estimated time and age effects overlapped in most cases, so that differences were mainly evident regarding the-value. The evaluation of the fit of the atlases to the data indicates that the Yeo-Krienen atlas is more suitable for our dataset, although it was not trained on a sample of older individuals. The atlas choice affects the estimated average RSFC in some networks, which highlights the importance of this methodological decision for future studies and calls for careful interpretation of already published results. Ultimately, there is no standard about how to operationalize networks. However, future studies may use and compare multiple atlases to assess the impact of network definition on outcomes. Ideally, the fit of the atlases to the data should be assessed, and heuristics such as "similar age range" or "frequently used" should be avoided when selecting atlases. Further, the validity of the networks should be evaluated by computing their associations with behavioral measures.