Altered resting-state functional connectivity in newborns with hypoxic ischemic encephalopathy assessed using high-density functional near-infrared spectroscopy

Hypoxic-ischemic encephalopathy (HIE) results from a lack of oxygen to the brain during the perinatal period. HIE can lead to mortality and various acute and long-term morbidities. Improved bedside monitoring methods are needed to identify biomarkers of brain health. Functional near-infrared spectroscopy (fNIRS) can assess resting-state functional connectivity (RSFC) at the bedside. We acquired resting-state fNIRS data from 21 neonates with HIE (postmenstrual age [PMA] = 39.96), in 19 neonates the scans were acquired post-therapeutic hypothermia (TH), and from 20 term-born healthy newborns (PMA = 39.93). Twelve HIE neonates also underwent resting-state functional magnetic resonance imaging (fMRI) post-TH. RSFC was calculated as correlation coefficients amongst the time courses for fNIRS and fMRI data, respectively. The fNIRS and fMRI RSFC maps were comparable. RSFC patterns were then measured with graph theory metrics and compared between HIE infants and healthy controls. HIE newborns showed significantly increased clustering coefficients, network efficiency and modularity compared to controls. Using a support vector machine algorithm, RSFC features demonstrated good performance in classifying the HIE and healthy newborns in separate groups. Our results indicate the utility of fNIRS-connectivity patterns as potential biomarkers for HIE and fNIRS as a new bedside tool for newborns with HIE.

moderate or severe encephalopathy based on a standardized neurological examination 31 .We excluded neonates if they had evidence of major anomalies of the brain or other organs, congenital infections (e.g., TORCH), intrauterine growth restriction (IUGR), identifiable metabolic disorder or ultrasound evidence of a large parenchymal hemorrhagic infarction.
Term-born healthy newborns with no reported brain injury were also recruited to the study from the Mother Baby Care Unit (MBCU).Inclusion criteria were admission to the MBCU, birth > 36 weeks' GA, and inborn.Exclusion criteria were evidence of congenital malformation or syndrome, antenatal infections, antenatal exposure to illicit drugs, small for gestational age (SGA) and IUGR.
The study was approved by the Health Sciences Research Ethics Board at Western University.Informed consent was provided by the parents/caregivers of the newborns.The study was conducted in accordance with the Declaration of Helsinki.

Clinical and demographic variables
Maternal and newborn health data were extracted from electronic medical records by a Paediatric Nurse, Paediatric Resident or NICU Fellow.For participants with HIE, the demographic data extracted included gestational age, birth weight, biological sex, HIE stage (based on Sarnat staging 32 ), resuscitation details, Apgar scores and cord pH.We also collected the following data: 72 h treatment with TH, the presence of brain injury on MRI, and postnatal infections (clinical sepsis or positive culture infection, confirmed necrotizing enterocolitis).

fMRI preprocessing and analysis
The fMRI data were preprocessed using FMRIB Software Library 33 (https:// fsl.fmrib.ox.ac.uk/ fsl/ fslwi ki/ FSL).The preprocessing pipeline included brain extraction, motion correction, spatial smoothing (full width at half maximum = 5 mm), band-pass filtering (0.01-0.2 Hz) and registration to a neonatal atlas 34 .From the frontal, parietal, temporal and occipital lobes of both hemispheres, we extracted the average BOLD sequences.Then, we calculated the Pearson correlation among them to build an 8-by-8 RSFC map for each neonate (Fig. 1B).

T1-weighted image processing and segmentation
The T1-weighted scans were automatically segmented using infant FreeSurfer 35 .Automatic processing steps included intensity normalization, skull stripping, and segmentation of the cortex, white matter and subcortical structures 35 .A multi-atlas approach was employed for the segmentation.Multiple brain atlases of newborns were first registered to native space, and structure labels were transferred.The atlases were developed from neonatal MRI scans 36 .To initially create the atlases, manually segmented labels were developed using MRI scans from a representative sample of neonates (0-2 years of age).In the current study, developmentally appropriate atlases for newborns were employed.Labels were fused into a single segmentation result, providing higher accuracy than single-atlas approaches 37 .Volumetric measurements for anatomical features could then be extracted 35 .Brain subcortical brain structure volumes were extracted, as well as cerebral white matter and cerebral cortex volumes, to compute total cerebral volumes (TCV).Each segmented T1-weighted image was visually inspected using Freeview software, available within FreeSurfer.Manual segmentation was employed to correct any segmentation errors (i.e., partial volume effects) using ITK-SNAP (http:// www.itksn ap.org/).

fNIRS acquisition and analysis
fNIRS was acquired in newborns with HIE (within 3 days post-TH after rewarming) and healthy newborns using the same experimental setup.Participants were scanned using a NIRSport2 (NIRx, Berlin, Germany) system with 8 light sources and 8 detectors covering the whole brain.The system operates at two wavelengths of light (760 nm and 850 nm), and the sampling rate was 10.1725 Hz.For either hemisphere, 4 channels were located on the temporal, 2 on the parietal, frontal and occipital lobes (Fig. 1A).fNIRS scans lasted for a minimum of 6 min during rest or natural sleep with the neonates lying in the incubator, cot, or caregiver's arms.
After data acquisition, Homer3 software 40 was used for preprocessing.The pipeline included spline interpolation for motion correction 41 42 .For each neonate, two connectivity maps, sized differently, were calculated using Pearson correlation.Channels corresponding to one lobe were averaged and then correlated with other cortices to build lobe-wise connectivity maps (8 by 8).20-by-20 channel-wise RSFC maps were also calculated with individual channels corresponding with each other.HbO and Hbr data were both used for building connectivity maps separately (Fig. 1A).

Statistical analysis
Statistical analyses were performed using Matlab (R2020b, Natick, Massachusetts: The MathWorks Inc).For the first aim, which was comparing the RSFC maps yielded from the fNIRS and fMRI data, Euclidean distances and Jaccard distances were calculated between the two 8-by-8 maps at various levels of sparsity 43 .It is considered common practice to filter out negative or low-weighted connections when analyzing RSFC maps, where the percentage of connections remaining is the level of sparsity of a map.For weighted maps, we kept the original weights of connections, while for binarized maps, the remaining connections after filtering were rounded to 1.Given adjacency matrices of two maps A 1 = a 1 ij and A 2 = a 2 ij , where a ij was the weight of connection between node i and j , Euclidean distance between the two was defined as while Jaccard distance was defined as We also calculated similarity maps for weighted and binarized RSFC maps between fNIRS and fMRI.Note that similarity maps had the same dimension as connectivity maps.For weighted cases, entries of similarity maps were the average Euclidean distance among subjects for the corresponding connections.In contrast, for binarized cases, they were the percentage of subjects sharing the connection.Volume counts of cortical grey matter obtained from T1-weighted images were associated with overall fNIRS and fMRI connectivity strength, respectively, using linear regression, adjusted for GA at birth, postmenstrual age (PMA) at scan and sex.
(1) To address the second aim, which was identifying altered connectivity patterns between HIE neonates and healthy controls, we used graph theory-based measurements to quantify the patterns of 20-by-20 RSFC maps.On HbO and Hbr maps, we calculated the clustering coefficient of a map as where N was the number of nodes in the map.Network efficiency was also calculated locally as where d ij was the shortest distance N i between i and j , G i the node set containing i and its direct neighbors only, and |G i | the number of nodes of G i .Modularity was defined as where C was the number of communities identified by a Louvain-like algorithm 44 and e pq denoted the fraction of connections between community p and q .Based on clustering, the between-subject variation was calculated using mutual information 45 .Given a confusion matrix C = [c wv ] , where c wv was the number of nodes in both community w and community v of two separate clusterings, respectively, mutual information was defined as T-tests were used to determine if the two groups were significantly different between a measurement.To further demonstrate the distinctive power of connectivity measurements separating HIE neonates from healthy controls, we trained machine learning models of support vector machine (SVM) with several local metrics, including connectivity, clustering coefficient, nodal efficiency, degree centrality and closeness centrality, with fourfold cross validation.The accuracies of classification were calculated, and receiver operating characteristic (ROC) curves were plotted.Areas under ROC curves (AUC) were also obtained.

Consent statement
Consent was obtained from the guardians of the study participants.

Participants
Twenty-one term-born neonates with HIE were enrolled (GA at birth = 38.92,SD = 1.55 weeks, Table 1).Nineteen (90.4%) of the HIE neonates underwent TH before they were scanned with MRI and fNIRS.Two newborns who did not undergo TH were diagnosed with mild or moderate HIE.All newborns with HIE underwent an fNIRS scan at a PMA of 39.96, SD = 1.43 weeks after rewarming.Twelve (57%) of the HIE neonates had MRI scans (PMA = 39.94,SD = 1.15 weeks).
Twenty healthy term-born infants were recruited from the MBCU of LHSC with a birth GA of 39.08, SD = 1.21 weeks, and fNIRS scans were performed within 48 h of life (Table 2).None of the healthy newborns underwent an MRI.

Brain injury characterization
Most newborns with HIE (n = 20, 95%) had white matter injury identified using MRI (Table 3).Mild white matter injury was commonly seen in the neonates, regardless of HIE severity (n = 19, 90%).Injury to the basal ganglia

Comparing fNIRS and fMRI
Lobe-wise 8-by-8 connectivity maps were obtained from fNIRS for HbO and Hbr for neonates with HIE.Euclidean and Jaccard distances were used to compare fNIRS and fMRI connectivity maps at levels of sparsity (Fig. 2A).Means, standard deviations (SD) and 95% confidence intervals (CI) of the two metrics were calculated based on subject-by-subject comparisons.Sparsity ranged from 0.2 to 0.46.We also calculated the similarity map (Fig. 2B) between the two modalities.
We also examined the association between brain volume and FC (Fig. 3).Correlations were calculated with whole-brain grey matter or white matter volume against HbO, Hbr or fMRI connectivity.Connectivities were summed over all connections, which were weighted and not filtered.We showed that FC increases with either grey or white matter volume.HbO connectivity showed a significant positive correlation with grey and white matter volumes (p-values were 0.015 and 0.011, respectively), adjusted for birth GA, PMA at scan and sex.While Hbr and fMRI connectivities were not significantly correlated with volumes, they demonstrated slight positive increases in relation to the brain volumes.3. Brain injury patterns in HIE participants.White matter injury was scored according to de Vries 38 .Intraventricular haemorrhage grade was using scored using the method of Papile 39 ; WMI using the method of de Vries; Positive = any morphological abnormality.

Altered RSFC patterns of neonates with HIE
Based on the connectivity maps obtained from fNIRS data, graph theory-based metrics were calculated to characterize patterns of RSFC and identify differences between HIE and control groups (Fig. 4).It was shown that the HIE group had increased clustering coefficient, network efficiency and modularity compared to controls, but had lower consistency within the group.
Local measurements, including connectivity, degree centrality, closeness centrality, clustering coefficient and node efficiency, were calculated and used as features for training SVM classifiers to distinguish the HIE group from the control.Note that only features showing significance in t-tests were selected to train SVM models.We achieved a maximal accuracy of 75.61% and a maximal AUC of 0.75 with connectivity features (Fig. 5A) using fourfold cross validation.

Discussion
Despite the wide adoption of TH along with other interventions, HIE remains a major cause of mortality as well as later physical and cognitive difficulties.In this study, we examined functional connectivity patterns in a heterogeneous cohort of newborns with HIE using multi-modal imaging methods.RSFC maps yielded from fNIRS and fMRI were comparable within the cohort of HIE newborns.Further, positive associations between grey and white matter volumes, a marker of brain health, and fNIRS and fMRI connectivity strength were evident.Our findings provide support for fNIRS to be a promising bedside monitoring method to be used alongside other imaging technologies, especially for newborns receiving care in the NICU.We also demonstrated altered fNIRS-based connectivity patterns in HIE neonates compared to healthy newborns and the distinctive power of these connectivity features with machine learning models.Results indicated that fNIRS-based RSFC may be a novel biomarker of brain health in HIE.
The RSFC maps derived from fNIRS (HbO or Hbr) were comparable to fMRI in newborns with HIE who underwent TH.Our findings align with previous studies demonstrating good correspondence between fNIRS and fMRI.In adults, Sasai et al. 46 reported good spatial correspondence between fNIRS and fMRI on several resting-state networks using seed-based analyses.Further, Duan et al. 23 in an adult population reported positive correlations between fNIRS and fMRI connectivity strength as well as similar topological patterns at different levels of sparsity.Anwar et al. 47 recorded simultaneous fNIRS, fMRI and EEG in adults to show bi-directional connectivity that occurs within sensorimotor networks, using Granger causality.In infants, less work has focused on comparing resting-state networks using fNIRS and fMRI.However, Bulgarelli et al. 24 demonstrated a strong resemblance between fNIRS and fMRI connectivity using dynamic causal modeling in a 6-month-old infant.
In the current work, we further examined fNIRS-and fMRI-based connectivity in relation to grey and white matter volumes as separate markers of brain health.Previous MRI studies conducted during the early stages of brain development demonstrated positive associations between FC strength and brain volumes [48][49][50] .fNIRSbased connectivity was significantly associated with both grey and white matter volumes.Our findings indicate that in the hypoxic newborn, FC can be maintained while adapting to changes in brain volume.However, this  www.nature.com/scientificreports/finding was primarily driven by HbO-based connectivity.In contrast, fMRI-and Hbr-based connectivity only showed positive trends towards an association with brain volumes.The physiological basis of the BOLD and Hbr signals are comparable in that they are both sensitive to changes in Hbr and are coupled to underlying changes in neuronal activity.In turn, connectivity based on deoxygenation changes may show less association with grey and white matter development in the rapidly changing newborn brain, as these processes are supported by synaptogenesis and myelination and are highly dependent on oxygen and nutrient delivery.An additional consideration is that the temporal dynamics of RSFC have been widely documented 51,52 , and methodological differences in the fNIRS and fMRI acquisition and analyses could further explain the inconsistency in the results.When comparing fNIRS-based RSFC between neonates with HIE and healthy controls, the HIE group showed increased clustering coefficients, network efficiency and modularity.These metrics are commonly considered to be indicative of the small-worldness of networks.Small-worldness is characterized by high clustering and short path length and is commonly seen in brain networks 53,54 .Findings suggest that brain regions with similar functions demonstrated more synchronized activity and had more highly connected hubs that support efficient global communication.Similar findings were documented by Gozdas et al. 55 , who reported increased modularity in preterm-born neonates, but using fMRI-based connectivity metrics.Additionally, using fNIRS, Zhang et al. 30 also reported increases in several small-world metrics in newborns with HIE.An increase in small-worldness in the HIE population may reflect disruptions in the underlying white matter fibre pathways that may be reflected in severed long-range connectivity but strengthened local connectivity through neural compensation and reorganization.These compensatory mechanisms may be a common feature in neonates born critically ill 56 that can even be seen later in adulthood 52 or neurological disease states 53 .Increased functional connectivity within local networks has been reported previously.For example, newborns scanned with fNIRS demonstrated positive associations between connectivity strength of homologous-interhemispheric networks as well as behavioural measures of negative emotionality 57 .In relation to the current study, our findings of increased functional connectivity could also partially be influenced by exposure to invasive NICU procedures (TH, tape removal, skin breaks, etc.) experienced by the newborns.Further, increased grey matter cerebral blood flow within sensorimotor networks has been reported in preterm born infants 58 , which may have also been a contributing factor to the increased functional connectivity within the networks seen in our study.Future studies could address whether these changes are related to underlying brain injuries associated with HI and predict local changes in connectivity.Further, an increased variation of clustering among neonates with HIE was evident and could also be a result of variations in the severity of HIE, IVH, or be related to white matter injury or a combination of these factors.Given our sample size, we could not further assess whether the variations in clustering were related to clinical factors or variations in care practices, but this could be explored in future work.
Machine learning models were implemented in the current study to determine whether RSFC features can be potential biomarkers of HIE.Optimal performance was realized when using connectivity features that were a product of significant connections that were either interhemispheric or within the right hemisphere.Altered interhemispheric or long-range connections have been reported in multiple studies 19,20 , which may be related to HIE-induced injury or inflammation localized to the white matter or even subcortical areas, which are key to long-range neuronal signal transmission.Connections within hemispheres were significantly different in the right but not the left hemispheres.Further examination as to whether there may be hemispheric vulnerability to HIE is needed.Massaro et al. 54 demonstrated that the magnitude of cerebral pressure in the right hemisphere was predictive of adverse outcomes detectable by MRI.Whether a similar lateralization effect due to changes in intracranial pressure occurred in the current sample remains unknown but could have been an underlying factor.
The majority of the HIE participants were scanned after TH and rewarming.The two participants who had mild HIE were not candidates for TH.Given that these newborns represented a small fraction of the total dataset, their data were unlikely to have unduly influenced the results.While TH is the standard of care for newborns with severe or moderate HIE, our findings demonstrating altered RSFC maps suggest that disrupted functional organization of the brains persists after TH.The persisting disrupted RSFC patterns could also contribute to adverse developmental outcomes in the first few years of life 55,56 .The present study focused on RSFC patterns in newborns with HIE, and future studies should incorporate long-term developmental outcomes.This would be the next step to determine how fNIRS markers of brain health at the bedside can predict developmental abilities in newborns with HIE.
The strengths of our study are grounded in several areas.Firstly, this was a single-centre study with a wellcharacterized cohort of newborns with HIE and healthy newborns.Also, a high-density fNIRS system offered coverage of the newborn brain's frontal, parietal, temporal and occipital areas.We also had the opportunity to characterize functional networks that were related to brain structure in the HIE population; however, not all participants had the same measures available.The MRI scans were conducted as part of clinical care and after TH and rewarming.Given the timing of the MRI scans, some families could only consent to the fNIRS component of the study.While the goal of the study was to compare RSFC maps obtained using fNIRS and fMRI, these two modalities could not be obtained at the same time in our sample due to a lack of resources in a clinical imaging centre.The duration of the fMRI scanning sequence was relatively short.However, the neonates underwent clinical imaging before the research sequences and could only be away from the NICU for a short period of time.The fMRI protocol was tested in an adult prior to scanning the newborns, further the networks were assessed in individual newborns to determine their robustness.Additionally, the fNIRS probes were not precisely located in the MRI space, and the fMRI data were analyzed using an atlas approach.In turn, the spatial correspondence between the fNIRS and fMRI data could be improved in future work.Another consideration is that the fNIRS data may have been influenced by physiological noise.In adults, short-distance channels are frequently implemented in fNIRS studies as a means to assess surface-artefacts, signal quality, source localization, and improve signal decomposition 59 .However, in neonates placing short distance channels on channels that already very close together is ineffective due to the smaller surface area of the head in neonates compared to adults 60 .Another

Figure 1 .
Figure 1.(A) Newborn wearing the fNIRS cap (family provided informed consent for publication of the image).Based on fNIRS recordings of oxygenated hemoglobin (HbO), 20-by-20 connectivity maps were created.(B) Atlas (lobe)-based analysis of the fMRI data.A hemisphere is sectioned into temporal, parietal, frontal and occipital lobes.8-by-8 connectivity maps were built upon fMRI data.
MI = −2 w,v (c wv ln c wv N w c wv v c wv ) w v c wv ln v c wv N + v w c wv ln w c wv N .
up to 0 days 50th percentile: up to 1.0 days 75th percentile: up to 5.3 days Table

Figure 2 .
Figure 2. (A) Similarity of RSFC maps between fMRI with respect to HbO and Hbr, measured by Euclidean distance (right panel) and Jaccard distance (left panel).Both weighted and binarized maps were implemented.(B) Lobe-wise similarity maps at a sparsity of 45%.Color bar denotes the value of similarity.Higher values reflect greater similarity between the two modalities.

Figure 3 .
Figure 3. Brain volumes and fNIRS-and fMRI-based functional connectivity strength.Functional connectivity strength (y axis) of oxygenated (HbO, blue lines and plus [+] signs), deoxygenated (HBr, black lines and open circles [o]) hemoglobin and fMRI-based functional connectivity (red lines and stars [*]) in relation to grey matter volumes (A, x axis) and white matter volumes (B, x axis).

Figure 5 .
Figure 5. (A) ROC curves of SVM classifiers trained with connectivity features.In the order shown in the legend, maximal accuracies of individual feature sets were 68.29%, 68.29%, 68.29%, 51.22% and 75.61%, respectively.AUCs were 0.71, 0.71, 0.69, 0.36 and 0.75, respectively.(B) The connections marked red were those showing significance based on t-tests and were selected to train the SVM model.
, band-pass filtering of 0.01-0.1 Hz and conversion to HbO and Hbr fluctuations with modified Beer-Lambert law

Table 1 .
Participant characteristics.The median values and interquartile ranges.Probability values provide results using the t-test for continuous measures and Chi-square tests for categorical measures.PMA: postmenstrual age.wasidentified in one (5%) neonate with severe HIE.The presence of IVH was only seen in one (4%) participant with severe HIE.No watershed injury was evident in any MRI scans of the newborns with HIE.

Table 2 .
Characteristics for HIE participants.