Salience detection is a fundamental component of attention(Parr and Friston 2019). Attention is a complex function that requires top-down sensitivity control and a bottom-up mechanism for filtering stimuli(Parr and Friston 2017). The top-down sensitivity control is a process that regulates the relative signal strengths of the different information channels that compete for access to working memory (HE and S 1997). The bottom-up mechanism for filtering stimuli is composed of a “reality filter” played by the SN, a large scale network of the heteromodal cortex involved in detecting and filtering salient stimuli and by a second network involved in the detection of unexpected task-relevant stimuli to trigger attentional shifts in stimulus-driven attention (Corbetta and Shulman 2002; V. S, JJ, and GR 2014): the VAN.
The SN contributes to a variety of brain functions including social behavior and self-awareness through the interaction of sensory, emotional and cognitive information(Menon and Uddin 2010; V 2011). A major function of the anterior insula (AI) node of the SN is the detection of behaviorally relevant stimuli (Crottaz-Herbette and Menon 2006; Eckert et al. 2009; Sridharan, Levitin, and Menon 2008; P and A 2010), while the anterior cingulate cortex (ACC) send strong motor output. The subcortical nodes of the SN provide preferential context-specific access to affective and reward cues(Lindquist et al. 2012).
The VAN is composed of the temporoparietal junction (TPJ) and the ventral frontal cortex (VFC) and is thought to be lateralized to the right hemisphere of the brain (Corbetta and Shulman 2002; Corbetta, Patel, and Shulman 2008). However, several studies have observed bilateral TPJ activation in tasks tapping attentional reorienting and the processing of rare deviant stimuli(Downar et al. 2000; Geng and Mangun 2011; Serences et al. 2005; Vossel et al. 2009). The VAN and the DAN, a bilateral network comprising the intraparietal sulcus (IPS) and the frontal eye fields (FEF) of each hemisphere involved in top-down voluntary allocation of goal-driven attention(V. S, JJ, and GR 2014), form a twofold attentional control system.
The study of salience detection and attention requires to focus on the activity of the DMN, a large-scale brain network primarily composed of the medial prefrontal cortex, posterior cingulate cortex/precuneus and angular gyrus(Raichle 2015; RL, JR, and DL 2008), identified with the stream of self-referential thoughts, the so called “resting state” (JR et al. 2010; Raichle 2015). The mental state of stimulus-independent thoughts counteracts attention. This finding is supported by the evidence that the DMN is anticorrelated with the DAN(Fair et al. 2007). Moreover, in a previous study adopting causality stochastic methods of analysis, the SN has shown to play a crucial role switching between the DMN and the Central Executive Network (CEN) in a previous study adopting causality stochastic methods of analysis(N et al. 2014). The present study tries to focus on the bottom-up and top-down attention systems. Adding the CEN to the data would have largely complicated the interpretation of the results, especially considering the degree of immaturity of the CEN in neonates. Nevertheless, the study of connectivity between the CEN and attention networks shall be the focus of our future studies.
A number of fMRI studies have demonstrated that infants and even preterm newborns possess immature forms of many of the networks described in the adult (Doria et al. 2010; CD et al. 2010; Smyser et al. 2016; Stoecklein et al. 2020; H et al. 2015). These studies show that higher order networks may be present, even if in a fragmented, immature form, even before term, as opposed to the previous concept by which the networks develop in parallel with the cognitive competences associated with stimulus-dependent thought.
The rapid expansion in the use of machine learning techniques to analyze neuroimaging data has led to employ these methods to model functional connectivity in newborn brains. Ball et al.(G et al. 2016) have combined high dimensional independent component analysis (ICA) with multivariate machine learning techniques to test the hypothesis that preterm birth results in specific alterations to functional connectivity by term-equivalent age. Their study demonstrates that functional connectivity of the basal ganglia and higher-level frontal regions are significantly altered in in preterm infants by term-equivalent age.
Smyser et al.(Smyser et al. 2016) have applied a Support Vector Machine (SVM) – multivariate pattern analysis (MVPA) classification method to infants’ resting state fMRI (rs-fMRI) data and developed a model to estimate an infant’s GA at birth based upon rs-fMRI data collected at term equivalent PMA. The SVM identified widespread intra- and interhemispheric connections within and between the rs networks able to categorize term-born infants from preterm infants, thus enabling quantitative prediction of GA at birth in individual subjects. By adopting the same SVM method Pruett et al.(Pruett et al. 2015) were able to classify, above chance, 6 versus 12 months old infants on FC data.
Shang et al.(Shang et al. 2019) used multivariate machine learning methods to classify young adults born prematurely when compared to full-term on the basis of volumetric data and by measuring the amplitude of low frequency fluctuations (ALIFF) within a repeated and nested cross-validation design. The authors compared the structural and functional preterm features, validated them by assessing the clinical history and assessed their contribution to the prediction of IQ. This study shows that volumetric imaging related to subcortical brain damage present in infancy also appears in early adulthood and that these abnormalities are interconnected with a pattern of predominantly decreased AFF in adults born preterm. Additionally the ALFF appeared to be able to predict among other clinical features, performance IQ. Moreover, the prediction of general IQ was improved by the addition of the ALIFF decision score.
In a recent study by Chiarelli et al.(Chiarelli et al. 2021) the authors examined the effect of prematurity on measures of rs-FC, resting state functional connectivity nodal strength (rs-FCNS), fractional amplitude of low frequency fluctuations (fALFF) and regional volume in 90 ROIs covering the whole brain by performing region-based univariate analysis of each metric to explore the association with GA at birth and the spatial consistency across metrics. To this aim they implemented a Machine learning framework using partial least square regression (PLS). The study demonstrated that prematurity is associated with a complex pattern of bidirectional alterations of FC metrics and regional brain volume and, to a lesser extent, with modifications of fALFF.
We analyzed the same data-set of the recent study from Chiarelli et al.(Chiarelli et al. 2021) and adopted the same Machine Learning multi-variate data-driven framework using partial least square regression (PLS), which allows to consider all the connections within the predicting network at once without any a-priori assumptions. However, the goal of our study was fairly different from that of Chiarelli et al.. First, we focused our attention on a group of specific networks known to be involved in the mechanisms of salience detection and attention. Second, we analyzed how the functional connectivity of each network influences each other network.
In the present study the SN connectivity was found unable to infer the GA at birth. It has long been observed that the insula is the first cortex to differentiate, beginning from 6 weeks after conception, providing the structural basis for its hub role even before term(Gao et al. 2011). Subregional segregation of the insula into an anterior insular cortex, mainly connected with the anterior cingular cortex, the frontal cortex and the posterior insular cortex, densely connected with somatosensory, temporal and parietal cortex, is always present at birth(Alcauter et al. 2015). These findings may suggest that because the insula development starts early in fetal life, its connectivity, under the influence of early external stimulation, matures faster and earlier than other networks(Afif A et al. 2007). The connectivity of the SN after birth might therefore be found at a stage that is too late to infer the GA at birth. In other words, the functional connectivity of the insula after birth might be too mature to variate accordingly with the gestational age.
The VAN was found to significantly infer the GA at birth. The strongest effects on the GA predictability were negative effects (anti-correlation) found for the left medial frontal gyrus-inferior parietal lobule. A possible explanation for this finding might be that the VAN goes under early lateralization during infancy, although the process might be still incomplete in early life(V. S, JJ, and GR 2014). Further studies are needed in order to assess the timing of lateralization of the VAN. The DAN is able to infer the GA at birth. This finding supports the evidence by which the VAN operates with the Dorsal Attention Network (DAN) forming a twofold largely interconnected attentional control system(Corbetta and Shulman 2002; Corbetta, Patel, and Shulman 2008; V. S, JJ, and GR 2014).
The DMN was found to significantly infer the GA at birth, with the strongest effects on the GA predictability found for the connectivity of the right medial prefrontal cortex-right pre-cuneus. These findings are in line with previous studies showing that the DMN is present, even if in a fragmented, immature form, even before term(Doria et al. 2010). More specifically, the cortical hubs in infants seem to be less present in the prefrontal cortex and in the precuneus than in adults. The connectivity of these cortical hubs might therefore reflect the GA.
The SN connections were found to significantly infer the average connectivity of the VAN (p = 0.026). Interestingly, also the VAN was found to strongly infer the mean functional connectivity of the SN (p = 10− 3). These findings seem to be in agreement with previous studies showing that the function of the VAN overlaps with that of the SN in salience detection(Dosenbach et al. 2008; LQ 2015; F. K and LQ 2015). Uddin and Farrant proposed that the VAN and the SN have overlapping nodes in the region surrounding the VFC and anterior insula. They also speculated that in children, these two networks may be less segregated than in adults, and that bottom-up salience processes and attention to environmental stimuli may be over-represented in the child’s brain(F. K and LQ 2015). The strongest positive effects on the predictability of the SN’s mean connectivity by the VAN were found for the connectivity of the medial frontal gyrus with the inferior parietal lobule of the contralateral side, bilaterally. The SN connections also inferred the average connectivity of the DAN (p = 6∙10− 3), supporting the theory by which the SN and VAN have overlapping roles.
All the connections of SN did not correctly infer the average connectivity of the DMN (p > 0.05). The DMN, on the contrary, was found to significantly infer the average connectivity of the SN (p = 7∙10− 3). This finding seems to oppose to the role of the SN in modulating the background activity of the DMN in order to elicit detection of salient stimuli and facilitate goal directed behavior. However, since the function of the CEN is likely reduced in infancy, this finding might also indicate that the SN function in switching attention to goal directed behavior matures with the development of the executive functions(Teffer and Semendeferi 2012).
The VAN was found to significantly infer the average connectivity of the DMN (p = 0.044). We believe this is the first evidence of the influence of the VAN over the DMN in neonates subjects. This finding might be explained by the overlap of function of the SN and VAN and by the over representation of bottom-up saliency detection seen in children (F. K and LQ 2015). The VAN was able to strongly infer the average connectivity of the DAN (p = 4.8∙10− 6). On the contrary, the DAN did not infer correctly average connectivity of any of the other networks, suggesting that it might be too early for the DAN to properly interact with other networks. A number of studies have shown that VAN and DAN are in competition during visual and verbal tasks (Anticevic A et al. 2010; Todd, Thaler, and Dijkstra n.d.; M. D et al. 2010; M. S et al. 2012). Specifically, the higher the load of the task, the higher the activation of the DAN and the higher the deactivation of the VAN. In other words, there seems to be a trade-off between the two attention networks, when the recruitment of task-related attention is high stimulus attention is deactivated allowing for successful task performance. Our finding lead us to hypothesize a conjoint development of the VAN and DAN with a prominent role of the VAN and, therefore, the bottom-up attention system during fetal life and early infancy. The two networks form an immature twofold attention system that matures with the growing ability of the top-down attention system to counteract the bottom-up system. Our hypothesis is further supported by a recent study by Suo et al.(Suo et al. 2021), underpinning the stable and reliable anatomical connections of the DAN and the VAN via functional connectors, demonstrating that the functional interaction of the VAN and the DAN is supported by a solid anatomical structure.
The mediation analysis of the VAN’s inference over the DAN, mediated by the GA found that VAN’s connectivity significantly modify the transmittance of change of the DAN’s connectivity assumed form evidence of cross-validated inference on GA (VAN mediates DAN: Sobel Test t = 2.14, p = 0.032). A specular analysis of the DAN’s inference over the VAN found that DAN’s connectivity significantly modify the transmittance of change on the VAN’s connectivity, assumed from evidence of cross-validated inference on GA (DAN mediates VAN: Sobel Test t = 1.969 p = 0.048 ). This finding further supports the theory of a twofold, largely interconnected attentional control system(Corbetta and Shulman 2002; Corbetta, Patel, and Shulman 2008; V. S, JJ, and GR 2014) and suggests that the interaction between the bottom-up and top-down attention system influence the development of each other. This twofold relationship among VAN and DAN is mediated by the GA; this led us to hypothesize that maturation is crucial for the development of attention networks and prematurity may, therefore, substantially affect their architecture and function.
This is the first study performed on a cohort oh neonates to show that the VAN may influence the connectivity of the SN, the DMN and the DAN. Interestingly the SN influenced the VAN and the DAN, but not the DMN. We believe our findings suggest a prominent role of the VAN in the bottom-up salience detection in early infancy and that the VAN and the SN may overlap in their roles of bottom-up control of attention. Although a small number of studies have adopted machine learning methods to assess if FC metrics were able to predict the GA at birth, this is the first study to focus on a small group of networks of higher order. The evidence that the SN is not able to predict GA at birth, as opposed to the other higher order networks we analyzed, is a novel finding that we thought to be the result of the early development of the insula, which goes under a complex process of maturation between the 13th and the 28th week of gestation(Afif A et al. 2007). Such an early pattern of maturation would be less impaired by the effect of prematurity.
Our studies has some limitations. First, the atlas we used to select the seed regions for each network only identifies coarse regions. However, this limitation is provided by the absence of efficient brain anatomy segmentation tool in newborns. Second the number of subjects was not very large considering that we studied 4 different networks using a multivariate analysis framework. A short BOLD acquisition time (4min), driven by the limited time available for conducting a relatively non-clinical evaluation in a clinical environment. In order to reduce motion artifacts and maximize the efficacy of the standard clinical evaluation, newborns were also mildly sedated using Midazolam. Although Midazolam might have altered brain activity and hemodynamics, an effect of Midazolam should be observed in all subjects and should therefore not affect regression analyses that exploit subject-specific alterations. Another limitation of the present retrospective study is the minimally essential clinical information available. Although infants with evident alterations at standard radiological assessment were excluded from the analysis and no relationship was found between the main available clinical variable, the APGAR score soon after birth, and the extent of prematurity, the presence of subtle clinical confounders cannot be definitely ruled out. Future studies should replicate the present study with a larger cohort of patient with larger sets of clinical data available. Moreover, longer acquisition time should be taken into consideration together with a no sedation approach. Manual segmentations of ROI should be also performed for a more precise delineation of brain regions in neonates.