A Paradigm for Measuring Resting State Functional Connectivity in Young Children Using fNIRS and Freeplay

Resting state functional connectivity (RSFC) reflects the organization of functional networks in the brain. Functional networks measured during “resting”, or task-absent, state are correlated with cognitive function, and much development of these networks occurs between infancy and adulthood. However, RSFC research in the intermediate years (especially between ages 3 and 5 years) has been limited, mainly due to a paucity of child-appropriate neural measures and behavioral paradigms. This paper presents a new paradigm to measure RSFC in young children, utilizing functional near-infrared spectroscopy (fNIRS) and Freeplay, a simple behavioral setup designed to approximate resting state in children. In Experiment 1, we recorded fNIRS data from children aged 3-8 years and adults aged 18-21 years and examined feasibility and validity of our measure of RSFC, and compared measures across the two groups. In Experiment 2, we recorded longitudinal data at two points (approximately 3 months apart) from children aged 3-5 years, and examined reliability under a variety of measures. In both experiments, all children were able to complete testing and provide usable data, a significant improvement over fMRI-based RSFC measurement in children. Results suggest this paradigm is practical and has good construct validity and test-retest reliability, and may contribute towards increasing the availability of reliable data on resting state networks in early childhood. In particular, these are some of the first positive results on the feasibility of reliably measuring functional connectivity in children aged 3-5 years.

Discovery of resting state functional connectivity (RSFC) and its link to function [5][6][7][8] has 53 widened the door to investigating mechanisms of cognitive function by providing some advantages over 54 task-based studies: 1) whereas measuring functional connectivity during task provides us with the 55 connectivity of only those regions specifically involved in the task, measurement during resting state can 56 provide connectivity information simultaneously among many regions; 2) certain populations (e.g., 57 clinical populations) or cognitive functions (e.g., motor) that are difficult to engage with specific tasks in 58 the scanner, can be just as well studied in resting state. For these reasons, resting state measurement has 59 become a standard paradigm for studying functional connectivity in human adults. increased connectivity between distant regions [11,12]. This simultaneous increase in both segregation 71 (pruning short-range connections) and integration (strengthening long-range connections) of brain regions 72 over development likely reflects a transition from organization around spatial proximity to organization 73 around higher-order function [12]. Finally, aberrant connectivity in RSNs have been associated with a 74 variety of psychopathologies from affective disorders such as depression, to diseases of cognitive function 75 such as Alzheimer's disease, to neurodevelopmental disorders such as autism or ADHD (for review, see 76 [10]), suggesting that RSN development may be integral to healthy brain and cognitive development. 77 Together, these findings highlight the need for studying RSNs over the time course of development and 78 especially in early childhood, when perhaps the most change is occurring in some brain networks. 79 However, studying RSNs in children has been challenging, mainly for two reasons. First, 80 traditional neuroimaging tools (e.g., functional MRI (fMRI), electroencephalography (EEG), or 81 magnetoencephalography (MEG)) are difficult to utilize with awake children [13,14]. Second, the 82 standard procedure for measuring resting state connectivity --to sit still for a period of time --proves to 83 be a difficult task for children, and some methods previously employed to increase compliance in children another more recent study of fMRI feasibility with children and adolescents showed that clinical groups 109 scanned even less successfully than typically developing controls [14]. 110 fNIRS is a relatively recent light-based neuroimaging method that overcomes many of the 111 challenges with obtaining brain activity measures in child populations. In fNIRS, near-infrared light is 112 used to obtain an estimate of changes in both oxygenated and deoxygenated hemoglobin concentrations 113 in a region of the brain. Thus, like fMRI, fNIRS gives an indirect measure of neural activity based on 114 blood oxygenation levels. Compared to fMRI or EEG, fNIRS is robust to and unrestrictive of motion, 115 comfortable, quick to set up, and cost effective (see [23] for comparison of techniques). These factors 116 make it especially appropriate for use with children. The main limitation of fNIRS in the context of 117 studying RSFC is that measurement is limited to regions near the surface of the brain (<17 mm of brain 118 tissue deep). For researchers interested in studying RSFC among surface regions of the brain, this is a 119 viable neuro-measurement tool. Our study, as discussed below, measures from the surface of the prefrontal 120 cortex.   compared to rest (e.g., early visual network decreased its connectivity with dorsal attention network and 141 increased its connectivity with the default mode network as well as the fronto-parietal network, during 142 movie watching [33]. These findings bring into question the validity of using movies for studying "resting 143 state" -functional connectivities and -networks in particular. However, movie or video-watching still 144 remains the most feasible method by which to record RSFC from children in an fMRI setting (with varying 145 success depending on where the movie lies in the spectrum from too engaging to not engaging enough).

146
Because fNIRS recording set-up is quiet, comfortable to wear, and unrestrictive compared to the 147 fMRI environment, a less engaging set-up for aiding compliance can be sufficient. Our study proposes an 148 experimental paradigm called Freeplay that may closely approximate resting state, and which takes 149 advantage of the spatially unrestrictive nature of an fNIRS recording situation. In this "task", participants 150 are seated at a table, presented with a set of simple toys (e.g., wooden blocks, small plastic animals) and 151 asked to quietly play for a few minutes. The premise is that children can naturally comply much more 152 easily with sitting still and quietly for a period of time when presented with even simple and unengaging 153 toys. In addition, the fNIRS possible sampling rate is much greater than that of fMRI, thus requiring less  of time without any specific task, and in the same way, allows natural individual variation to be present.

158
Allowing natural variability is important as it allows researchers to be more confident that connectivity 159 patterns consistent across many participants are generalizable and not arising from task-specific or stimuli-160 specific states.

162
Since measuring RSFC in children relies on using approximations of the resting state task (e.g., 163 movie watching, silent rest), it is important to also develop multiple such approximations, to combat the 164 task impurity problem. Freeplay contributes as one method for approximating resting state. Moreover, 165 Freeplay avoids several biases that are inherent in some previous methods. For example, relatively 166 engaging or externally-guided stimuli, such as movie clips or screensavers, potentially compromise the 167 unrestrained and non-externally-directed nature of thinking that is characteristic of true resting state. They 168 may also potentially engage functional networks differently from rest (e.g., naturalistic movie viewing has 169 been shown to alter connectivity patterns among certain networks compared to rest) [33]. In contrast,   This study aims to demonstrate feasibility of using fNIRS and Freeplay to measure RSFC in pre-  Participants sat quietly and freely played with a set of toys for about 8 minutes. Toys included: 239 lincoln logs, wooden nuts and bolts, plastic animal figurines, toy cars, and simple coloring pages (a flower, 240 turtle, duck, or fish). Toys were chosen to be simple and minimally engaging, to help induce quiet boredom, 241 a state that we expect may closely approximate resting state (See Figure 1, right panel).    Figure 1), to prevent probe from slipping as well as to cover the probe to prevent ambient 261 light from reaching the sensors. The participant sat in a rigid, stationary chair to reduce movement artifacts.

262
After fitting the fNIRS cap to the participant's head, signal quality was checked for each source-263 detector channel and calibrated if needed to make sure the fNIRS fiber optics made good contact with the 264 scalp of the participant, and that the detector was sensitive to cardiac pulsation as a sign of good signal-265 to-noise (SNR) ratio. Any detector saturation was also adjusted for in this step. fNIRS data was recorded 266 for each participant using custom data collection software that interfaced with the fNIRS system, described   represented graphically in correlation matrices (as in Figure 3) were the main quantities studied in this 290 paper. We used partial correlation as the index of RSFC in this study because it can factor out correlation 291 between fNIRS channels due to shared extracerebral components, and is thus thought to characterize 292 relationships between brain regions more precisely than Pearson's correlation [48,49]. Our first goal was to test for significant homologous connectivity, characteristic of RSFC. To do 297 this, we compared functional connectivity between regions that were homologous (bilaterally symmetric) 298 to that between non-homologous regions.

299
Our second goal was to test the validity of Freeplay as a task for measuring resting state. To do so, we 300 compared the functional connectivity in adults between the two conditions, "at rest" and Freeplay. Since 301 it is difficult to directly measure similarity between two groups of functional connectivity patterns, we did 302 so by estimating the accuracy of classifiers trained to distinguish different conditions (e.g., between "at 303 rest" and Freeplay in adults). Higher accuracies suggest greater distinguishability, and hence greater 304 dissimilarity, between classes. Given the high-dimensionality (45 CPs) of our problem, we used logistic       (2) 404 ICC is defined as the proportion of between subjects variation to total variation. That is, where BSV denotes the between-subjects variance (i.e., the average (over scans) of the variance across 407 subjects) and WSV denotes the within-subjects variance (i.e., the average (over subjects) of the variance 408 across scans). In addition to the partial correlation matrix itself, studies of RSFC are often interested in the 413 network structure of functional connectivity. Studying this requires binarizing the partial correlation 414 matrix (i.e., identifying each CP as either "connected" or "disconnected". Therefore, to study reliability 415 of the functional connectivity network, we binarized each CP by thresholding the absolute value of its 416 partial correlation value at a "connectivity threshold" θ; absolute values below θ were replaced with 0 417 (denoting an unconnected CP), and absolute values above θ replaced by 1 (denoting a connected CP). 418 We then used two indices of inter-scan reliability: the F1 score (a.k.a., Dice coefficient), a general CPs. Accuracy (proportion of agreement) between the two binarized scans was not used as a measure of 423 similarity because it is extremely sensitive to the connectivity threshold; for example, using a threshold 424 of 0 (full connectivity) or 1 (no connectivity) results in a perfect accuracy of 1. The raw (continuous, un-425 thresholded) correlation was also not used, as it is relatively difficult to interpret as a measure of reliability. 426 We chose θ to maximize (over 1000 equally spaced values between 0 and 1) each reliability index (F1 427 score and MCC) and used LOOCV to obtain an unbiased estimate of each reliability index.

Post-hoc tests
As is apparent from the group RSFC matrices shown in Figures 2 and 4 usable data, speaking to the practical utility of the Freeplay paradigm for studying RSFC in children.

523
Further, correlations in adults and children in Freeplay showed different patterns, from which a trained 524 classifier was able to predict "adult" or "child" with high accuracy --this is in line with our expectations 525 given that we know RSNs develop significantly with age. Finally, Experiment 2 demonstrated inter-scan 526 reliability, in that similar connectivity patterns are found between 2 independent scans of the same 527 individual, and that within-subject variance is significantly lower than between-subject variance. This 528 reliability is observed for both the raw RSFC matrices and the resulting connectivity networks after 529 thresholding appropriately, and is observed both with and without adjacent channel pairs. The test-retest reliability value for ICC (0.53) was lower in our paper than in a previous paper (0.7) 534 by Niu et al. looking at RSFC test-retest reliability in adults [19]. Some possible reasons for this may be 535 that: 1) since we are measuring in children, the data may be noisier, and 2) they measured functional 536 connectivity in terms of Pearson (rather than partial) correlations, which can be inflated by other sources 537 of correlation between channels besides functional brain correlations.

538
Although our study compared "Freeplay" to the traditional "at rest" condition measured with adults, 539 it did not compare to either a true task condition, or a "movie-watching" condition. Comparing the 540 functional connectivity patterns in "Freeplay" with not just "at rest" but also these other conditions may 541 help us better characterize RSFC. For example, if we are able to show that a trained classifier can 542 effectively distinguish between adults in the "at rest" or "Freeplay" conditions from those in the movie 543 watching condition, it would provide a comparison condition for "Freeplay" as well as corroborate 544 previous findings that "movie watching" alters resting state.

545
The current study introduced a new paradigm for measuring resting state functional connectivity found that network properties of RSFC in the PFC, as measured by fNIRS, were correlated with varying 553 performance in EF tasks in adults. An important next step will be to extend those investigations to children.

554
Future work can explore the changes in PFC connectivity structures that parallel observed cognitive 555 development in EF or other domains subserved by the PFC. As a practical setup for measuring RSFC in 556 children, the fNIRS-Freeplay paradigm will allow investigation of questions such as these to advance 557 understanding of the neural mechanisms of cognitive development.