In summary, FC was altered in adults with FAS compared to controls not exposed to alcohol prenatally both within a majority of cognition-related networks (including the salience / ventral attention networks and dorsal attention networks, as well as, to a lesser degree, the fronto-parietal control network, and the default mode network) and between these networks. These results on the global and network level are based on an HC approach, indicating that at least rare and weak group effects seem to be present [
60]. HC-based findings do, however, not necessarily mean that FC is changed in a majority of connections. Group effects could not be further resolved to connections between individual regions using conventional mass-univariate testing with multiple-comparison correction. In the additional exploratory time-resolved FC analysis, altered FC dynamics in the FAS group could not be observed.
Group differences in attention-related networks
The observation of more obvious effects in attention- and salience-related systems compared with networks underlying other cognitive functions highlights the importance of attentional deficits in FASD [
1,
73‐
77], including adults [
5,
78]. However, this interpretation might be limited by the infeasibility of a direct quantitative comparison between networks as well as reverse inference [
79,
80]. Our rs-fMRI findings are generally in line with previous studies suggesting a particular involvement of attentional functions and underlying neural systems in FAS when compared to other cognitive deficits: Response and activation patterns in a Go/NoGo task in a sample of young female adults overlapping with this study also provide indirect evidence of a particular importance of attentional deficits compared with inhibitory control deficits in this age group [
36]. Attention-related networks were also altered in other alcohol-exposed samples studied with fMRI: Attention networks were among those altered in studies by Fan et al. [
25] and Ware et al. [
27]. In the latter study, FC alterations were associated with differences in attentional performance measures. The authors consequently conclude that the patterns observed (lower within-network, higher between-network FC) provide support for reduced attention network specialization and inefficiency [
27]. Reduced FC between key regions of the salience network with other cognition-related networks were among the key findings by Little et al. [
26]. There is further evidence of altered attention systems from magnetoencephalography [
81] and diffusion tensor imaging [
82].
Group differences in the default mode network
The default mode network (DMN), though classically reported as anti-correlated with task-positive cognitive networks [
86], is considered to be involved in cognitive functions including task-switching and integration of information [
87,
88]. There is evidence of regional differentiation within the DMN, with subdivisions subserving different cognitive functions [
89]. Fan et al. observed altered FC in the anterior part of the DMN, discussed as subserving social perception, judgment, and self-referential processing in individuals with FASD; however, they found no changes within the posterior DMN [
25]. Different from the preprint version of this analysis (i.e., not controlling for covariates in the primary models) [
59], we did not find evidence of a strong predominance of group effects in either the anterior or posterior part of the DMN. Beyond that, there is further evidence of less regionally specific DMN dysfunction in FASD [
32]. One single connection within the DMN was significant when correcting for multiple comparisons within the sub-network only, but not with global correction. Considering the distribution of effects within this network in comparison with the other networks, we consider this a potential outlier.
Specific aspects of the statistical approach underlying main results
The main static FC analysis applied in this study follows a hierarchical statistical approach, partially based on HC statistics. This approach addresses the general limitations of functional neuroimaging analyses in relatively infrequent disorders such as FAS: Conventional high-dimensional FC analysis methods, such as frequently used mass-univariate statistical testing, carry the risk to report only a “tip of the iceberg” of true underlying alterations due to lower than optimal statistical power. There is an increased risk of both published findings being false-positive [
90,
91] or false-negative findings [
90]. Further in-depth discussions on this issue have been published [
91‐
93]. Consequently, statistical thresholding of mass-univariate analyses (mainly for multiple-comparison correction) might in part explain ambiguous FC results in previous studies in children with FASD [
25‐
27]. Such a conventional analysis approach as applied at the most detailed analysis level in this study did not yield significant group effects regarding single connections between regions of interest in this study when correcting for multiple comparisons across all connections. There are increasing efforts to report subthreshold effects in fMRI studies in order to facilitate better interpretation of underlying patterns [
94‐
96]. An example is the additional presentation of unthresholded activation or connectivity maps [
95,
97]. Our hierarchical approach with HC-based joint hypothesis tests [
60,
61] at the network level might help avoid these shortcomings without sacrificing information from individual connections. Compared with conventional mass-univariate fMRI analyses, it avoids selectively reporting and interpreting few selected results that would pass a multiple-comparison threshold but might not well represent the true underlying effect in a medium-power setting. The finding of group effects in HC statistics not being observed in the conventional mass-univariate analysis might indicate that substantially larger samples are desirable in future rs-fMRI studies in neurodevelopmental disorders such as FAS. This is, however, limited by the rarity of these disorders [
98]. Recent work in a related field further highlights the potential specific relevance of a network-perspective and limitations of group mean comparisons at the regional perspective: Segal et al. observed that structural brain alterations in individuals with mental disorders mainly converge at the network level while effects on the regional level are sparse and heterogeneous [
99]. Please see the “
Methods” section for a more detailed description of the HC-based multi-level approach and underlying rationale, as well as the “
Potential limitations” section for further methodical aspects.
Potential limitations
Findings of this first FC analysis in adults with PAE are restricted to young adults (ages 18–32 years) with FAS. They do not necessarily translate to other age groups. Furthermore, findings do also not necessarily translate to other gradations across the FASD spectrum. Despite greater psychopathology, attention deficits, and impulsiveness compared with controls, a recent study did not find network-based FC alterations in a population of adolescents with a wide range of PAE, i.e., less severely exposed individuals [
102]. Adult participants with FAS in this sample had been diagnosed during their childhood using the Majewski criteria [
37] then widely used in German-speaking countries. These criteria did not gain widespread use in other countries [
38]. Although they generally reflect the full clinical picture of FAS (see also Additional_file_
1.pdf: supplementary methods), disease severity cannot be exactly mapped onto newer diagnostic criteria [
2]. Explicit information about the presence and quantity of maternal alcohol exposure is not available in this sample. This represents a more general limitation of FAS diagnosis both, in research and clinical care: As reflected by current clinical guidelines [
8,
103‐
105], FAS can be diagnosed based on the typical clinical picture and does not require explicit knowledge about prenatal alcohol exposure. Similarly, there is a possibility of unknown sub-clinical prenatal alcohol exposure in the control group, potentially reducing the sensitivity for FC group differences. While occasional alcohol consumption has been reported in 14%, regular alcohol consumption during pregnancy has been relatively rarely reported (around 1%) in Germany [
106]. Social and psychopathological characterization of the participants and their families is limited: no information about socioeconomic status is available. Socioeconomic inequalities such as parental income, educational status, or neighborhood context have been related to differences in structural brain development and to disrupted development of cognitive abilities [
107]. Early SES disadvantage in childhood has been associated with altered resting-state functional connectivity of brain networks involved in cognition [
108]. Children with FASD are frequently exposed to such adverse experiences when growing up [
109‐
111]. In addition to the influence of prenatal alcohol exposure itself, SES might therefore pose an independent factor that could have interfered with brain development in our sample [
110]. Future research is needed to highlight the effects of the additional burden imposed upon FASD subjects by socioeconomic disadvantage and the associated implications on brain structure and activity in affected individuals. Information on psychiatric comorbidity for the exclusion criteria is based on anamnestic information and a screening tool (SSQ), but not on detailed assessment within this study or on previously assigned ICD or DSM diagnostic codes. We did not include total brain volume and IQ in the statistical models of FC alterations, since both reduced IQs [
1] and reduced total brain volumes [
112,
113] are considered disease features of FAS with the potential for overcorrection if included as covariates [
114].
Taking differences in head motion into account in rs-fMRI studies in clinical populations is a matter of ongoing critical debate [
115]. Although we have taken precautions to minimize head motion, excluded participants with excessive head motion, include head motion in the analyses and even though measures did not differ significantly between groups, it cannot be excluded that parts of the results are movement-related [
53,
115]. The state-of-the-art motion correction methods during data pre-processing are very similar to those showing a particularly good performance in an additional large-scale benchmarking study of rs-fMRI motion correction strategies [
115]. We have also included head motion as a covariate in the main FC group analyses at the level of individual connections to further take a potential movement bias into account. This also means that connections with different distances between brain regions were modelled separately at this stage. We refrained from including global signal regression since it may confound the directionality of FC estimates and aggravate potential distance-related effects of residual head motion [
115]. In this, context, it should however also be considered that motor restlessness itself is a disease feature in FAS [
116].
Statistical power of the final step of the hierarchical analysis approach (resolving FC alterations to single connections between pairs of regions) is potentially limited by the high dimensionality of the underlying atlas. This notion is also supported by the distribution of effect size estimates for group differences of individual connections. This atlas resolution was chosen because it is extensively validated [
56] and aims to optimally reflect the brain’s functional architecture [
56,
117]. The HC-based approach was chosen here instead of averaging FC estimates to allow global and network-wise inference while maintaining the advantages of the high-resolution atlas. Findings are limited to the cortex and do not include subcortical gray matter nuclei within these networks [
118,
119]. It has to be noted that there is an ongoing debate about functional network nomenclature, so that the networks described here [
56] may deviate from studies using other atlases [
120].
Though widely used in other research areas with high-dimensional data [
60,
63], the HC statistic has only recently been introduced to fMRI [
64,
65,
96]. The HC statistic primarily assumes statistically independent features, since correlations among features can lead to unbalanced
p-value histograms, however without expecting peaks in the first histogram bin (low
p-values). It has thus been argued that the influence of correlations among features is negligible when the underlying histograms show typical behavior [
62]. In addition to the HC statistic, we therefore visually interpreted the underlying
p-value histograms as a plausibility control and observed well-behaved
p-value histograms in the global and within-network analysis and to a slightly lesser degree in the between-network analysis. The HC-based global hypothesis test does not directly result in
p-values but in a primary decision to reject/retain the corresponding global null hypothesis. However,
p-values for typical HC values have been approximated [
121]. Even when using these approximated
p-values, thus departing from the notion of the original HC statistic, main results would be statistically significant at the
p < 0.05 level or stricter
p-values.
Dynamic or time-resolved FC analysis is a promising, already widely used, yet still evolving rs-fMRI analysis method [
15‐
17]. Thus, there is currently a relatively high methodological variability [
15,
17]. Here, we adopted a widely used sliding window approach [
17] and refrained from extensively exploring analysis settings in order to avoid false-positive findings [
122]. Hence, there is a risk to miss group differences of FC dynamics which might have been uncovered with other, less well-established dynamic FC analysis approaches [
15,
17,
123]. Furthermore, FC states themselves, which were estimated in the entire sample here, might be intrinsically different in both groups. This might be addressed by separate FC state estimation, however will need in the future more advanced methodology for subsequent actual group comparison. Similar to methodological heterogeneity and open methodological questions, there is still no consensus of connectivity states to be expected in a normal population. However, a relatively small number of FC states partially reflecting changing interactions of parts of the DMN, similar to those observed here (Additional_file_
1.pdf: Supplementary Fig. 2) has been repeatedly reported [
17,
124]. Silhouette values (Additional_file_
1.pdf: Supplementary Fig. 1) indicate that clusters observed in our analysis may not be well separated. Thus, these two clusters capture dynamic FC changes as a model but might not represent truly discrete FC states in a neurobiological sense. Cluster frequencies suggest a high inter-subject variability. Despite these general limitations of this evolving methodology, we believe that our exploratory approach can be a starting point for further investigations on dynamic FC in FAS and other neurodevelopmental disorders.