Investigating the effect of channel pruning on functional near-infrared spectroscopy data collected from children aged 5-24 months

39 Significance 40 Infant functional near-infrared spectroscopy (fNIRS) data are particularly vulnerable 41 to noise; participant behaviour can result in motion artefacts and reduced set-up 42 times can cause poor optode coupling. Accurate channel pruning is therefore essential 43 but approaches vary and often use adult-derived thresholds, risking unnecessary data 44 loss. 45 Aim 46 This work systematically compared pruning approaches and parameter choices to 47 evaluate their effects on data quality and retention in infant fNIRS. 48 Approach 49 Data from 5–24 month-old infants were collected across two cohorts, using two 50 paradigms. Channel pruning was performed using the coefficient of variation (CV) and 51 the Quality Testing of Near Infrared Scans (QT-NIRS) tool, varying key thresholds. 52 Multilevel models assessed effects of pruning method, parameter choice, age, motion, 53 and testing site on signal-to-noise ratio (SNR) and channels retained. 54

55 QT-NIRS produced significantly higher SNR than CV pruning across nearly all age, 56 task, and cohort combinations, when matched for data retention. Higher QT-NIRS 57 thresholds improved quality but reduced retention. Motion prevalence strongly 58 reduced both SNR and retention; testing site and age had smaller but notable effects. 59

60 QT-NIRS offers a better balance of data quality and retention than CV pruning. Lower 61 QT-NIRS thresholds than adult defaults are recommended for infant data. These 62 findings provide practical guidance for preprocessing pipelines in developmental 63 fNIRS research. 64 65

infant fNIRS, channel pruning, processing66 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 3 1 Introduction 67 1.1 Scalp-optode coupling 68 In a typical functional near-infrared spectroscopy (fNIRS) experiment, participants 69 wear a headband or cap embedded with optodes to monitor brain activity by emitting 70 and detecting near-infrared light at two separate wavelengths. When fitted securely, 71 the cap ensures a motion-robust signal, enabling the study of a wide range of cognitive 72 tasks and abilities with less stringent demands for stillness than other neuroimaging 73 modalities [1]. For this reason, fNIRS is widely used with developmental populations 74 [2]. However, signal processing is usually still required to remove artefacts arising 75 from motion, poor scalp-optode coupling, and physiological signal confounds [3], [4], 76 [5]. Infant data is particularly susceptible to motion and poor scalp-optode coupling, 77 as infants typically exhibit increased fussiness, limited compliance with instructions, 78 and shorter attention spans, which increase the likelihood of participant motion, 79 reduced capping time and difficulties in handling the imaging headgear [6]. In fact, 80 poor coupling allow s light from the source optode(s) to escape , or ambient light to 81 flood detector optode(s) [7]. Affected channels can exhibit signal saturation (easily 82 detectable by unrealistically high raw intensity values caused by excessive light 83 reaching the detector) and greater variability, which impacts the estimation of the 84 haemodynamic response [8] whose amplitude is already lower and more variable in 85 infants than in older participants [9], [10]. 86 1.2 Strategies to mitigate poor optode coupling 87 Fitting the fNIRS headgear securely aids scalp-optode coupling [7] but is time-88 consuming and assumes stable coupling throughout recording, which is challenging 89 when working with infant participants. To reduce the impact of poor coupling on data 90 quality, post-hoc channel pruning – the exclusion of data from an entire channel – is 91 therefore often required. An important pruning consideration is the trade-off 92 between data retention and quality: removing poorly-coupled channels improves 93 overall signal quality and mitigates the impact of superfluous signals, but reduces the 94 number of remaining channels and participants available for analysis [6]. This is 95 particularly important in infant research given the already high attrition rates due to 96 low attention span and susceptibility to fussiness [11]. Pruning method selection is 97 important, and manually pruning channels is subjective and time-consuming [8], 98 especially for high-density imaging arrays. Two methods are often used to prune 99 channels: the coefficient of variation (CV) or the ‘Quality Testing of Near-InfraRed 100 Spectroscopy’ (QT-NIRS) tool. 101 CV pruning quantifies the relative signal variability. Channels are pruned if the 102 CV for either wavelength falls below a particular threshold [12], [13], if the CV 103 difference between wavelengths exceeds a threshold [14], or both [15]. While this 104

is faster and less subjective than manual pruning, a parameter choice is still 105 required and some levels of variability - which is expected in the task -based fNIRS 106 signal due to the evoked haemodynamic response [16] - may be interpreted as signal 107 noise. Additionally, CV pruning only examines the signal in the time domain, 108 potentially overlooking important signal properties. 109 QT-NIRS utilizes objective signal measures from both the time- and frequency 110 domains [7], [16], providing a more comprehensive consideration of signal quality. 111 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 4 Channels are pruned via signal characteristic assessments in two domains using the 112 scalp coupling index (SCI) and peak spectral power (PSP) . The SCI is a time -domain 113 approach, which assesses correlation between wavelengths within the cardiac 114 frequency band after bandpass filtering the signal, with high values indicative of a 115 strong cardiac component [17]. This incorporates a time-domain signal characteristic, 116 but also risks retaining channels with high correlation due to motion-induced 117 artefacts. To address this, QT-NIRS incorporates a second measure based on the 118 frequency domain, PSP, which detects strong, recurrent oscillations in the signal. High 119 PSP values in the cardiac frequency band likely correspond to cardiac pulsations, 120 whereas components with inconsistent or varying frequencies usually result in lower 121 PSP values. QT -NIRS utili zes strengths from manual pruning (physiological 122 grounding, consideration of both time- and frequency domains) and CV pruning 123 (objectivity; efficiency) making it a robust tool for assessing data quality at the 124 channel level. 125 Recently, QT-NIRS has been increasingly adopted in infant fNIRS research [18], 126 [19] yet independent comparisons with other pruning approaches are yet to be 127 established, and empirical estimations of SCI and PSP parameters are available only 128 for adult participants [7], [16]. This highlights the need to refine its implementation 129 with infant data. In addition to behaviour, both physiological and anatomical factors 130 may affect channel pruning in infants: they have thinner scalp tissue and higher 131 cardiac signal frequencies (~1.3–3.2 Hz at rest, compared to ~1–1.7 Hz in adults)[20], 132 [21]. The former may result in a weak superficial cardiac signal, whereas the latter 133 may result in coarse representations of the infant cardiac signal by fNIRS 134 instrumentation sampling rates optimized for adult participants [22]. Further, signal 135 quality can be detrimentally affected by skin and hair colour, hair type, age and even 136 head size with darker skin pigmentation and thicker hair corresponding to poorer 137 signal quality when compared to other skin and hair types [23], illustrating the need 138 for inclusion of less-frequently sampled populations in fNIRS studies. 139 Against this backdrop, the objectives of this work are to: 140 (a) Compare QT-NIRS as a pruning method against pruning using CV, which is 141 used frequently by fNIRS users – and provides a baseline for channel pruning 142 via a previously employed method and parameters 143 (b) Investigate contextual and data-derived measures which affect data quality 144 and channel retention (with a particular focus on QT-NIRS parameter choices, 145 age and motion incidence) 146 (c) Provide guidance on channel pruning and QT-NIRS use for infant participants 147 To achieve this, fNIRS data from the Brain Imaging for Global HealTh (BRIGHT) 148 Project [24], a longitudinal study of infant development in Kiang West (The Gambia) 149 and Cambridge (UK), were analyzed. The analyses in this work incorporates data from 150 two different experimental paradigms, collected from both the Gambian and UK sites 151 and pertaining to participants with physical and behavioural characteristics from 152 both a commonly -sampled and a more underrepresented population in fNIRS 153 research [23]. The longitudinal nature of the data (with five time points over the first 154 two years of life) further enables the investigation of the effects of age across early 155 childhood while accounting for variability in cohort and task. 156 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 5 Based on prior literature and exploratory findings, the following hypotheses 157 were formulated: 158 (1) QT-NIRS based pruning will result in a greater balance of data quality and 159 retention compared to CV, due to its multidomain examination of signal 160 characteristics 161 (2) Motion will be negatively associated with signal quality and retention, 162 because of (i) increased likelihood of latent, undetected artefacts in data 163 where more motion is detected, (ii) displacement of optodes affecting 164 signal quality after motion artefacts, or both. 165 (3) Data quality and retention will be diminished in the Gambian cohort, given 166 the increased probability of darker, coarser hair types interfering with 167 optode-scalp coupling 168 2 Methods 169 2.1 Data 170 2.1.1 Participants 171 Participants were recruited into the BRIGHT project, from early 2016 to February 172 2018, and fNIRS data were collected when infants were 1-, 5-, 8-, 12-, 18- and 24 173 months (hereafter 𝑥mo for 𝑥 months of age), plus a follow-up between 3 and 5 years 174 of age [25], [26]. The 1mo fNIRS protocol was limited to auditory stimuli with sleeping 175 participants [27], likely inducing infant motion with a different noise profile; at 3 -5 176 years, a different fNIRS cap for data collection was used and data at this age were not 177 collected in the UK site. To enable matched dataset comparisons, data from the other 178 5 time points (5-, 8-, 12-, 18- and 24mo) was therefore used in this work. Participants 179 met the inclusion criteria if infants were born at 37 –42 weeks’ gestation (both 180 cohorts) and had a minimum birth weight of 2.5 kg (UK only). 181 After applying exclusion criteria, a total of 204 mother-infant dyads were 182 included in the Gambian cohort; of these, 185 remained at the 24mo timepoint. 183 Pregnant, Mandinka -speaking women were recruited during routine antenatal 184 clinical assessments at MRCG@LSHTM Keneba Field Station by fieldworkers in the 185 Gambian BRIGHT Project team. An information sheet and consent form written in 186 English were provided to potential recruits then explained fully in Mandinka by a 187 study staff member. fNIRS data collection took place at MRC Unit The Gambia at the 188 London School of Hygiene and Tropical Medicine (‘MRCG@LSHTM’) Keneba Field 189 Station. Ethical approval was granted by the joint Gambia Government/MRC Ethics 190 Committee under the title: ‘Developing brain function for-age curves from birth using 191 novel biomarkers of neurocognitive function’, SCC number 1451v2. 192 61 mother-infant dyads were enrolled in the UK from the Rosie Hospital, 193 Cambridge University Hospitals NHS Foundation Trust. Information about BRIGHT 194 was provided during antenatal appointments, with families expressing an interest 195 contacted and recruited subsequently via email or phone call. Data collection 196 primarily took place at Evelyn Perinatal Imaging Centre at Rosie Hospital, 197 Addenbrooke’s Hospital, Cambridge, and to a lesser extent at the Centre for Brain and 198 Cognitive Development in Cambridge [24], [28]. Ethical approval was given by the 199 National Research Ethics Service Committee East of England (REC reference 200 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 6 13/EE/02000); informed written consent was obtained from all parents/carers prior 201 to participation. 202 2.1.2 fNIRS Paradigms 203 Social/Non-Social paradigm 204 Full details of the Social/Non-Social (SNS) task paradigm can be found in [15] and 205 [29]. Briefly, the paradigm consisted of alternating visual social (silent), auditory 206 social and auditory non -social stimuli. Stimuli repeated until the participant became 207 bored or fussy or the end of the task was reached; inter-stimulus baselines varied 208 between 9 to 12 seconds. 209 Visual social stimuli consisted of full colour, life-sized videos of adults from the 210 same population as the participant, on a 24 -inch screen ~100cm away. Throughout, 211 adult actors in the video either moved their eyes or played ‘hand-games’ for 9-12 212 seconds. Actors, their actions, and concurrent spoken auditory stimuli were varied to 213 prevent anticipatory brain activity. Auditory stimuli were non-synchronized to the 214 video in terms of both duration (8 seconds) and content, with environmental sounds 215 for non-social stimuli and non-vocal speech sounds for social stimuli. Sounds in each 216 condition (social and non-social) were matched for duration and sound intensity. 217 Habituation and novelty detection (HaND) paradigm 218 The experimental paradigm included 25 trials, each consisting of a spoken 8sec 219 sentence in the family’s first language (i.e. English or Mandinka) followed by 10sec of 220 silence. The first trial was preceded by at least 10sec of silence, acting as a baseline. 221 The same recording, with a female voice, was used for trials 1 -15; a different, 222 male voice was used for trials 16-20; finally, the original, female recording was again 223 used for trials 21 -25. The stimulus sentence: “Hi baby! How are you? Are you having 224 fun? Thank you for coming to see us today. We're very happy to see you” was translated 225 to Mandinka to maintain the same semantic meaning. 226 Technical detail on the recording, processing and playback of the auditory 227 stimulus can be found in previous work [14], [26]. 228 2.1.3 Data collection 229 Custom-made headgear was fitted after head measurements (head circumference, 230 and ear-to-ear both around the forehead and over the top of the head) had been taken 231 to aid with the alignment of fNIRS headgear with the 10/20 system anatomical 232 landmarks. Headgear consisted of custom -built stretchy silicone headbands to 233 increase friction and prevent slippage, with attached probes into which optodes were 234 clipped. Optodes were designed to accommodate glass optic fibres at right-angles to 235 allow them to sit flush on the scalp. The headband was fastened around the head to 236 provide even pressure over the base of the probes [30], [31]. In the left hemisphere, 237 headgear was placed such that source 4 in Figure 1 was centered abo ve the 238 preauricular point, so that the channel it formed with the detector located directly 239 behind it sat above T3 in the 10-20 system; the equivalent right hemisphere channel 240 was above T4. The array angle was guided by the headband, which was placed on the 241 head so that it touched the join between the ear and head and, frontally, lay over the 242 infant brow line (through Fp1 and Fp2 in the 10-20 system) [32]. 243 The headgear was designed to record responses bilaterally from auditory-244 associative brain regions, including the inferior frontal gyrus (IFG), middle and 245 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 7 superior temporal regions, and the temporo-parietal junction [14], [33], [34]. The 246 fNIRS array comprised 17 channels in each hemisphere with a source -detector (S-D) 247 distance of 2cm, corresponding to a penetration depth of ~1cm from the skin surface 248 and thus permitting measurement of the gyri and superficial sulci [14]. fNIRS data 249 were collected using this array design with the NTS optical topography system 250 (Gowerlabs Ltd., UK) with a sampling frequency of 10Hz and source wavelengths of 251 850 and 780nm.π 252 Infants sat on a carer’s lap during data acquisition. Carers were discouraged 253 from interacting with the infant to attempt to minimize confounding stimuli however 254 infants’ attention was engaged , if necessary, with (non-social, non-auditory) bubble-255 blowing and silent demonstration of soft toys, which also minimized infant head 256 movement. The HaND task was part of a larger battery of fNIRS assessments with a 257 total recording time of ~ 21min 30 s (6min social task; 4min functional connectivity 258 data acquisition; 7min 30 s HaND; 4min further functional connectivity). Where 259 possible, paradigms were completed uninterrupted; sessions were paused and 260 subsequently resumed in the event of infant discomfort. [35] 261 2.2 Channel Pruning 262 2.2.1 Pre-pruning steps 263 First, channels were excluded from analyses if their minimum light intensity value 264 was less than 3e-4, based on previous experience with the NTS system [14]; these 265 were labelled ‘channels with signal extrema’ (CSE). Analysis of the pruning methods 266 was conducted on motion -free segments. To find motion -free data, motion artefacts 267 were detected using hmrMotionArtifactByChannel function from Homer2 [36] with 268 established infant fNIRS preprocessing parameters: tMotion = 1, tMask = 1, 269 STDEVthresh = 15 and AMPthresh = 0.4 [37]. Data for each channel was split into 3 270 second windows per channel, as this aligned with QT-NIRS temporal segmentation 271 and thus avoided additional processing complexities. Windows were excluded from 272 pruning analyses if they contained artefacts at any point. If the data for a particular 273 channel at one wavelength was excluded, data for both wavelengths was removed. 274 Figure 1: Array layout during data collection. Red dots indicate position of optodes centred above the preauricular point during headgear fitting. .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 8 The two channel pruning methods assessed in this work were implemented 275 using custom-written scripts in MATLAB [26]. Both pruning methods described use 276 raw light intensity data as input. 277 2.2.2 CV Pruning 278 CV pruning was conducted using an in -house script, and CV itself was calculated on 279 motion-free data for each wavelength and channel using the equation: 280 𝐶𝑉 = 𝜎 |𝜇| 281 where 𝜎 and 𝜇 are the standard deviation and mean of the light intensity signal in a 282 channel, respectively [38]. Channels were pruned if the difference between CV values 283 for each wavelength exceeded 0.2 based on previous experience with the same data 284 and system [14]. 285 2.2.3 QT-NIRS pruning 286 QT-NIRS was implemented using the function qtnirs (available at 287 https://github.com/lpollonini/qt-nirs at the time of writing) for precise control over 288 the pruning and quick, repeated processing of the large volume of data. 289 More detail on QT -NIRS can be found in publications describing the methods 290 [7], [17] but an outline is provided here. First, bandpass filtering is conducted to retain 291 only those frequencies in the cardiac band , ~1.3 − 3.2𝐻𝑧 in the case of infants [20]. 292 The cross -correlation of contemporaneous (zero -lag) wavelength signals in the 293 cardiac frequency band, (i.e., SCI), is then calculated: 294 𝑆𝐶𝐼 = 𝑥̅)! ⊗ 𝑥̅)" (0) 295 where 𝑥̅)* represents the light intensity signal for wavelengths 𝑖 = 1, 2 in motion-free 296 signal. PSP is the maximum signal value in the frequency domain, representing the 297 dominant oscillation in the bandpass -filtered signal and presumed to correspond to 298 the cardiac frequency in well-coupled data. The recorded signal is divided into equal-299 length windows, and both SCI and PSP are calculated for each window. A window's 300 signal is considered of sufficient quality if both the calculated SCI and PSP exceed the 301 user-defined thresholds, sci_threshold and psp_threshold, respectively. 302 The focus was on the alteration of sci_threshold and psp_threshold during 303 analysis, since each provides a threshold for one of the key measures of optode 304 coupling quality used to assess data quality with QT-NIRS, and adult reference values 305 of sci_threshold = 0.8 and psp_threshold = 0.1 are available for these two parameters 306 [16]. Default parameters were used for window size (3 seconds) and quality 307 threshold, or q_threshold (0.75), which prunes channels with less than 75% of 308 windows meeting both SCI and PSP threshold values. 309 2.3 Statistical analyses 310 Multi-level models (MLMs), a form of linear regression that estimates variance at 311 multiple levels, were used for statistical analyses, to effectively account for repeated 312 measures, hierarchical data structures (including participant- and channel-level 313 measures), and missing data [39]. All models were fitted in R 4.4.1 [40] using the 314 lme4 package [41]. All models included random intercepts for each participant to 315 account for individual variation. Final models used for analysis are described in 316 ‘Models and Analyses’. 317 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 9 2.3.1 Measures, Outcomes and Effects 318 Considering the anticipated data quality/inclusion trade-off, which is central to 319 decision making around preprocessing of fNIRS data [6], the performance of pruning 320

and parameters were assessed using two metrics: (i) signal -to-noise ratio 321 (SNR) and (ii) channel inclusion/exclusion percentage. 322 The effects of other factors such as age, motion, and signal extrema, were 323 included as predictors. Other measures were more particular to this dataset - such 324 as task, cohort and optode position. They were included as covariates to account for 325 the variability they may cause. 326 The variables used are listed below, with letters contained in brackets 327 indicating whether they were used as outcome variables (O), predictors (P), or 328 covariates (C). Full rationale can be found in Supplementary Materials 1. 329 2.3.1.1 Task-relevant channel signal-to-noise ratio (O) 330 Signal quality after pruning was measured on the included channels using the signal-331 to-noise ratio (SNR): 332 𝑆𝑁𝑅 = 20 log!+ 𝜇 𝜎 333 with 𝜇 and 𝜎 the mean and standard deviation of the signal, respectively [5]. SNR was 334 calculated in ‘task-relevant channels’ (TRCs) – channels where a true haemodynamic 335 signal was observed. Age-specific TRCs were determined using prior analyses: the 336 SNS TRCs were taken from work by Benerradi and colleagues [25] whereas the HaND 337 TRCs were taken from work by Blasi and colleagues [26]. TRCs for each age and task 338 were those which exhibited a haemodynamic response in both chromophores for 339 either cohort, except the SNS task at 18mo: only 1 channel met this criterion, so 340 channels were added for this age/task combination if they were in the set of TRCs for 341 at least two other ages for the SNS task. This outcome was named the task -relevant 342 channel signal-to-noise ratio (TRC SNR). 343 2.3.1.2 Channels retained (O) 344 The number of channels per participant included after pruning using the described 345

and parameter(s) was summed – this was labelled ‘Channels Retained’ (CR). 346 2.3.1.3 SCI Threshold (P) 347 sci_threshold values ranging from 0.05 to 0.9 were used, with increments of 0.05. The 348 upper threshold of 0.9 sits between the recommended value for adult participants 349 (0.8) and the theoretical maximum of SCI = 1. Initial exploratory analyses (see 350 Supplementary Materials 2) for each age/task/cohort combination indicated that 351 even very low SCI values continued to alter signal quality, so the entire range of lower 352 values was used. 353 2.3.1.4 PSP Threshold (P) 354 The psp_threshold value was varied, ranging from 0.005 to 0.1, with increments of 355 0.005. The upper value of 0.1 is the recommended threshold for adult participants. 356 Values which spanned the entire possible range lower than this – based on initial 357

from simpler MLM analyses – were used; this is supported by the considerable 358 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 10 proportion of infant PSP values lower than the adult threshold apparent during post-359 analysis data examination in the UK cohort (see Figure 2). 360 2.3.1.5 Percentage of motion (P) 361 For each participant/age/task combination, the cross -channel mean of the 362 percentage of windows per channel containing motion as identified by 363 hmrMotionArtifactByChannel and subsequently excluded from analyses (see 364 ‘Pruning’) was labelled the ‘Percentage of Motion’ (PoM) providing a measure of the 365 prevalence of motion. 366 2.3.1.6 Age (P) 367 Age was included as a five-level predictor (5-, 8-, 12-, 18-, 24mo) to assess change 368 with age. 369 2.3.1.7 Cohort (C) 370 Cohort was included as a two-level covariate (Gambia and UK). 371 2.3.1.8 Task (C) 372 Task was included as a two -level covariate (HaND and SNS) to account for the 373 contribution to variance of the different paradigms. 374 2.3.1.9 Channels pruned due to signal extrema (C) 375 The number of CSE per participant, for each age and task, was used as a participant -376 level covariate and proxy measure of poor optode coupling. 377 378 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 11 2.3.2 Models and analyses 379 2.3.2.1 Comparison of QT-NIRS and CV pruning 380 In line with objective (a), each fNIRS recording underwent channel pruning using: 381 1. CV, by pruning channels where the CV values for different wavelength signals 382 differed by 20% or more (‘CV’) 383 2. sci_threshold only at every parameter value, by setting psp_threshold to 0, (‘SCI 384 Only’), and 385 3. every combination of sci_threshold and psp_threshold parameters listed in 386 ‘Measures, Outcomes and Effects’ (‘Full QT-NIRS’) 387 Top: bar charts showing the proportion low channel Average SCI and Average PSP measures in relation to the adult recommended parameters of 0.8 (SCI) and 0.1 (PSP), and half of these threshold values (0.4 and 0.05, respectively). Top -left: Gambia cohort. Top-right: UK cohort. Bottom: number of participants, by age and Task, with less than 60% of acceptable channels exhibiting mean SCI and PSP values compared to adult threshold values of 0.8 and 0.1, respectively. Highest exclusion rate was ~33% for UK infants at 18mo during the SNS task. Figure 2: Data characteristics in relation to adult QT-NIRS thresholds. SNS HaND (5) SNS (5) HaND ( 8) SNS ( 8) HaND (12) SNS (12) HaND (18) SNS (18) HaND ( 24) SNS (24) HaND (5) SNS (5) HaND (8) SNS (8) HaND (12) SNS (12) HaND (18) SNS (18) HaND (24) SNS (24) .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 12 Pruning method 1 provides a baseline for the channel pruning, using a method and 388 pruning criteria which has previously been used for the HaND task [14]. Pruning 389 approach 2 permits comparison of two time-domain methods (methods 1 and 2) to 390 assess the impact of incorporating temporally specific measures. Pruning approach 3 391 provides insight into the benefit of additionally using a PSP threshold and pruning 392 using frequency characteristics of the signal. 393 For every age/task/cohort combination, each sci_threshold parameter choice 394 (SCI) or combination of sci- and psp_threshold values (QT-NIRS) were given two 395 separate rankings according to their similarity to CV pruning, in terms of the mean 396 number of channels retained across all participants and the total number of 397 participants excluded. These two rankings were then combined to find the parameter 398 (SCI) or parameters (QT-NIRS) producing the closest approximation to data retention 399 provided by CV pruning. 400 For each of the 20 age (5 levels)/task (2 levels)/cohort (2 levels) 401 combinations, the following model was then fitted to participant-level data: 402 𝑇𝑅𝐶 𝑆𝑁𝑅 ~ 𝑃𝑟𝑢𝑛𝑖𝑛𝑔 𝑀𝑒𝑡ℎ𝑜𝑑 + (1 | 𝐼𝐷) 403 Equation 1 404 to compare the effect of pruning methods (1)-(3) on signal quality whilst using a 405 random intercept for each infant to account for inter -participant variability. To 406 correct for multiple comparisons, p-values were Bonferroni-corrected. 407 2.3.2.2 Effect of SCI and PSP threshold choice on signal quality and retention 408 Models were designed to investigate the effect of sci_threshold, psp_threshold, age, and 409 motion on TRC SNR and Channels Retained (Objectives (b) and (c)). 410 To examine potential combinations of theoretically viable predictors, 411 covariates, and their interactions used for each outcome, a systematic approach to 412 model building was used as a first step, combining predictors in various model 413 formulae using combinatorial logic before assessing model fit. Model fit was assessed 414 using the Akaike Information Criterion (AIC) [42], a model selection criterion that 415 balances goodness of fit with model complexity [43, p. 824]. Model variables were also 416 included based on the outcomes of the subsidiary investigations described in 417 Supplementary Materials 3 which investigated the factors affecting motion incidence 418 and average SCI and PSP measures at the channel level. These additional variables are 419 included in the bottom line of Equation 2. 420 Mindful of model convergence issues and overfitting, the priority when 421 constructing models was to include predictors of interest, plus interaction terms, 422 random slopes and random intercepts which incorporated them. This resulted in the 423 following model for both outcomes TRC SNR and Channels Retained: 424 𝑂𝑢𝑡𝑐𝑜𝑚𝑒 ~ 𝑆𝐶𝐼 ∗ 𝑃𝑆𝑃 + 𝐴𝑔𝑒 ∗ 𝑆𝐶𝐼 + 𝐴𝑔𝑒 ∗ 𝑃𝑆𝑃 + 425 𝑆𝐶𝐼 ∗ 𝑃𝑜𝑀 + 𝑃𝑆𝑃 ∗ 𝑃𝑜𝑀 + 𝐴𝑔𝑒 ∗ 𝑃𝑜𝑀 + 426 𝑇𝑎𝑠𝑘 + 𝐶𝑜ℎ𝑜𝑟𝑡 + 𝐶𝑆𝐸 + (1 | 𝐼𝐷) + 427 (1 | 𝑆𝐶𝐼: 𝐶𝑜ℎ𝑜𝑟𝑡) + (1 | 𝑃𝑆𝑃: 𝐶𝑜ℎ𝑜𝑟𝑡) 428 Equation 2 429 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 13 where SCI = sci_threshold, PSP = psp_threshold, and symbol ‘∗’ denotes an interaction 430 term (itself denoted using ‘:’) plus all individual terms used in the interaction, as is 431 consistent with notation used in the lme4 package. Random intercept terms (bottom 432 row) were included due to quantitative support based on subsidiary MLM 433 investigations into average SCI and PSP measures. 434 Assessments of the model residuals showed that they were non-normally 435 distributed, so to calculate effects a bootstrapping approach was used. Bootstrap 436 datasets were generated by sampling rows from the original dataset with 437 replacement, using a fixed random seed for reproducibility. The relevant model for 438 each bootstrap sample was fitted using the lmer function from the lme4 package and 439 extracted fixed effect estimates. To mitigate potential biases caused by fitting models 440 to data with different scales, variables 𝑥* were scaled and centred: 441 𝑥*′ = 𝑥* − 𝑥̅ 𝑠, 442 Equation 3 443 where 𝑥*′ is the scaled value, and 𝑥̅ and 𝑠, are the sample mean and standard 444 deviation, respectively, of variable 𝑥. 445 3 Results 446 CV pruning yielded an average TRC SNR of 23.7 ± 1.66 across all 20 age/task/cohort 447 combinations. Pruning with QT-NIRS using just the SCI threshold resulted in a mean 448 increase of 2.13 ± 0.644 in TRC SNR beyond that obtained using CV pruning. Similarly, 449 using both SCI- and PSP thresholding in combination resulted in a mean TRC SNR 450 increase of 2.16 ± 0.627 in comparison to CV pruning (see Figure 3, and 451 Supplementary Materials 4 for the full comparison of TRC SNR values). 452 Figure 3: Example violin plot demonstrating typical differences between obtained mean TRC SNR values using CV pruning, QT-NIRS using SCI threshold only, and QT-NIRS utilising both parameters for Gambian participants at 12mo during the SNS task. TRC SNR for 12 month infants. Task: SNS; Cohort: Gambia .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 14 3.1 Comparison of QT-NIRS and CV pruning 453 3.1.1 Pruning using CV and QT-NIRS 454 The effect size and Bonferroni -corrected significance values of the 3 contrast 455 conditions were calculated using Equation 1; results are displayed in 456 Table 1. For 19 of the 20 age/task/cohort combinations, significant (p < 0.01) positive 457 effects were found on the TRC SNR when controlling for data rejection when using 458 QT-NIRS, using either the SCI Only or Full QT -NIRS approach. In 17 of 20 459 combinations, the differences were found to be statistically significant at the p < 460 0.0001 threshold. Though the effect was positive for the single remaining 461 combination of 20, it did not reach statistical significance. 462 Table 1: Condition contrasts between each of the three pruning methods 463 CV vs SCI Only CV vs Both Parameters SCI Only vs Both Parameters Age (months) Task Cohort Significance Effect Size Significance Effect Size Significance Effect Size 5 HaND Gambia <0.001 20.85 <0.001 20.83 1.0 0.05 8 HaND Gambia <0.001 12.71 <0.001 12.75 1.0 <0.01 12 HaND Gambia <0.001 19.32 <0.001 19.70 1.0 0.27 18 HaND Gambia <0.001 17.04 <0.001 17.10 1.0 0.04 24 HaND Gambia <0.001 12.36 <0.001 12.12 1.0 0.17 5 SNS Gambia <0.001 16.97 <0.001 16.93 1.0 0.02 8 SNS Gambia <0.001 14.69 <0.001 14.70 1.0 0 12 SNS Gambia <0.001 16.95 <0.001 17.65 1.0 0.49 18 SNS Gambia <0.001 15.10 <0.001 15.09 1.0 <0.01 24 SNS Gambia <0.001 13.38 <0.001 13.38 1.0 0 5 HaND UK <0.001 9.40 <0.001 9.40 1.0 <0.01 8 HaND UK <0.001 12.90 <0.001 12.90 1.0 <0.01 12 HaND UK <0.001 8.70 <0.001 8.56 1.0 0.10 18 HaND UK <0.001 10.92 <0.001 10.92 1.0 <0.01 24 HaND UK 0.012 4.68 0.012 4.68 1.0 0 5 SNS UK <0.001 6.93 <0.001 7.80 1.0 0.62 8 SNS UK <0.001 7.18 <0.001 7.28 1.0 0.07 12 SNS UK 0.924 3.23 0.288 3.87 1.0 0.45 18 SNS UK 1.0 2.84 1.0 2.84 1.0 0 24 SNS UK 1.0 0.76 1.0 0.76 1.0 <0.01

are: CV, QT -NIRS using the sci_threshold parameter only, and QT -NIRS using 464 both the sci_threshold and psp_threshold parameters. Significance values calculated 465 during bootstrapping and reported after correction. Effect sizes rounded to 2d.p. 466 3.1.2 Pruning using SCI Only compared with both parameters 467 No significant statistical differences were found between TRC SNR values when 468 pruning using SCI only and Full QT-NIRS approaches. In every case, however, the 469 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 15 mean TRC SNR across participants was higher when using both parameters than 470 when using sci_threshold alone. 471 A representative comparison of the three different pruning methods is given 472 in Figure 3, with higher average TRC SNR values obtained using both QT -NIRS 473 approaches when compared to CV pruning. A higher mean TRC SNR is obtained using 474 both parameters when pruning with Full QT -NIRS compared to using SCI Only, but 475 with a more erratic distribution of values. 476 3.2 Predictors of TRC SNR and Channels Retained 477 The focus here is primarily on reporting results for the outcomes of interest and fixed 478 effects with wider generalisability but full results are included in Figure 4. Effect sizes 479 were classified as small, medium or large if the absolute value of the Estimate was < 480 0.15, < 0.35, or ≥ 0.35, respectively, using Cohen’s 𝑓" thresholds (Cohen, 2013, 481 chap.9). 482 3.2.1 Predictors for TRC SNR 483 Both SCI Threshold (𝛽 = 0.0107, 95% CI [0.0084, 0.0131], 𝑆𝐸 = 0.0012) and PSP 484 Threshold (𝛽 = 0.0530, 95% CI [0.0505, 0.0556], 𝑆𝐸 = 0.0013) had small, positive 485 effects representing an increase in signal quality for higher threshold values. The 486 interaction effect between PSP Threshold and SCI Threshold had a small, negative 487 effect (𝛽 = −0.0019, 95% CI [−0.0037, −0.0001], 𝑆𝐸 = 0.0009). 488 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 16 489 Figure 4: Forest plots showing the results for the effect of each fixed effect in Equation 2 on TRC SNR and Channels Retained, obtained via bootstrapping scaled values. Top: Forest plot showing the type and relative effect size of each fixed effect on TRC SNR. Bottom: Forest plot showing the type and relative effect size of each fixed effect on Channels Retained. .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 17 PoM had the largest impact on TRC SNR, with a large negative effect (𝛽 =490 −0.5004, 95% CI [−0.5028, −0.4979], 𝑆𝐸 = 0.0012) corresponding to a decrease in 491 signal quality in data with more frequent incidences of motion. This decrease in signal 492 quality with motion was offset slightly as participants aged, driven mostly by a more 493 moderate decrease in the 18mo data, illustrated by a small negative interaction effect 494 between and PoM (𝛽 = −0.0485, 95% CI [−0.0506, −0.4062], 𝑆𝐸 = 0.0011). Small 495 interaction effects between PoM and both SCI Threshold ( 𝛽 = 0.0027, 95% CI 496 [0.0008, 0.0047], 𝑆𝐸 = 0.0010) and PSP Threshold ( 𝛽 = 0.0253, 95% CI 497 [0.0234, 0.0272], 𝑆𝐸 = 0.0010) exhibited trends which saw high threshold values 498 mitigate the detrimental effect on TRC SNR; the slightly larger main (PSP) and 499 interaction (PSP:PoM) effect in the case of PSP Threshold led to greater mitigation of 500 the TRC SNR decline due to PoM than in the case of SCI (see Figure 5a). 501 The small, negative effect of Cohort ( 𝛽 = −0.8783, 95% CI 502 [−0.8832, −0.8733], 𝑆𝐸 = 0.0025) is notable since this fixed effect had larger effects 503 on other outcomes, including Channels Retained. The non-significant effect of PSP:Age 504 (Figure 6b) is notable given the size of the dataset and rarity of non-significant 505 predictors throughout this work. 506 Figure 5: The relationship between motion, and data quality and retention. Blue, italicised axes values represent the approximate original values before scaling during analysis. (a) Predicted TRC SNR trend by Percentage of Motion, grouped by PSP Threshold. (b) Predicted Channels Retained trend by Percentage of Motion, grouped by PSP Threshold. (c) Predicted TRC SNR trend by Percentage of Motion, grouped by Age. (d) Predicted Channels Retained trend by Percentage of Motion, grouped by Age. 28 24 20 0% 4% 8% 12% 70 60 50 40 30 20 0% 4% 8% 12% 30 25 20 15 0.0% 2.5% 5.0% 7.5% 10.0% 12.5% 65 60 55 50 45 40 0.0% 2.5% 5.0% 7.5% 10.0% 12.5% .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 18 507 (a) The interaction between and SCI Threshold, showing the mitigating impact of high SCI values for younger participants. (b) The interaction between and PSP Threshold, in which a pattern with age is harder to ascertain. Figure 6: Effect of the interaction between QT-NIRS thresholds and age on channel retention. .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 19 3.2.2 Predictors for Channels Retained 508 Both SCI Threshold (𝛽 = −0.1381, 95% CI [−0.1401, −0.1361], 𝑆𝐸 = 0.0010) and 509 PSP Threshold ( 𝛽 = −0.4067, 95% CI [−0.4087, −0.4048], 𝑆𝐸 = 0.0009) had small 510 and large negative effects on Channels Retained, respectively (see Figure 7). This 511 indicates an increase in the number of channels pruned when using higher threshold 512 values, particularly for the PSP Threshold. The interaction effect between SCI 513 Threshold and PSP Threshold was small and negative ( 𝛽 = −0.8783, 95% CI 514 [−0.8832, −0.8733], 𝑆𝐸 = 0.0025), with each threshold alleviating the negative effect 515 of the other at high parameter values. 516 PoM had a significant, medium negative impact ( 𝛽 = −0.2013, 95% CI 517 [−0.2030, −0.1995], 𝑆𝐸 = 0.0009), with higher amounts of motion associated with 518 decreased channel retention. At the two oldest ages (i.e. 18- and 24mo), greater 519 proportions of motion in the data had a less drastic negative impact on the number of 520 channels pruned (see Figure 5d) which was captured by the small positive interaction 521 effect Age:PoM (𝛽 = 0.0610, 95% CI [0.0593, 0.6264], 𝑆𝐸 = 0.0009). As with TRC 522 SNR, interaction effects of PoM with both SCI Threshold ( 𝛽 = −0.0034, 95% CI 523 [−0.0050, −0.0019], 𝑆𝐸 = 0.0008) and PSP Threshold ( 𝛽 = −0.1962, 95% CI 524 [−0.1977, −0.1945], 𝑆𝐸 = 0.0008) moderated this channel reduction for higher 525 percentages of motion. Though both interactions were significant, PSP:PoM was of 526 medium effect size and also acted on two (negative) medium main effects, leading to 527 a near negation of the detrimental impact of motion on predicted Channels Retained 528 for low PSP Threshold values (see Figure 5b). 529 Task (𝛽 = −0.2038, 95% CI [−0.2069, −0.2007], 𝑆𝐸 = 0.0015) and Cohort 530 (𝛽 = −0.8783, 95% CI [−0.8832, −0.8733], 𝑆𝐸 = 0.0025) had medium and large 531 effects on Channels Retained, respectively, resulting in lower channel retention for 532 the SNS task and UK cohort. 533 534 535 536 537 538 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 20 539 (a) The interaction between thresholds acting on TRC SNR, showing an inconsistent, slight decrease in TRC SNR when compared to the overall trend for high SCI- and PSP Threshold values. (b) The interaction between thresholds acting on Channel Retained, showing a more consistent trend across threshold levels and the dampening of the reduction in channels when both SCI- and PSP Threshold values are high. Figure 7: The effect of the interaction between SCI Threshold and PSP Threshold on signal quality and retention. .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 21 4 Discussion 540 Channel pruning in infant fNIRS research is a crucial step in data -analysis, yet often 541 performed with subjective parameters or adult derived threshold s. Moreover, 542 approaches to channel pruning, including QT-NIRS and CV pruning, in infant fNIRS 543 data have often relied on adult -derived thresholds or single -domain measures, 544 without systematic evaluation of their suitability for infant data. The work described 545 here took a quantitative approach to directly compare pruning strategies and 546 parameter settings for infant sparse array data. It was found that QT-NIRS, in its 547 consideration of both time- and frequency-domain signal quality measures, achieved 548 a better balance of data quality and retention than CV pruning . Additionally, 549 parameter values for QT-NIRS were investigated alongside contextual factors, with 550 lower values than those used for adults likely being preferable. These findings 551 supplement work in the wider literature aimed at improving infant NIR imaging data 552 pipelines more broadly [6], [37], [44]. 553 4.1 QT-NIRS as a channel pruning method 554 QT-NIRS produced infant fNIRS data of significantly greater SNR than pruning in all 555 but one age/task/cohort-specific comparison (SNS task in UK infants at 24mo), whilst 556 controlling for data retention. Higher signal quality remained statistically significant 557 when channels were pruned using SCI only, suggesting that evaluating the signal 558 using a method which accounts for a temporal data characteristic (correlation) is 559 more effective than assessment using time -independent measures. It may also 560 indicate that evaluation of the signal in subsampled windows provides a more 561 accurate reflection of signal quality than holistic signal assessment, as was the case 562 with CV pruning. This difference in signal quality, even without using PSP is observed 563 despite evidence suggesting that SCI may be biased by latent, undetected motion in 564 the signal in at least some participants (see Effects of Motion). 565 Using both SCI- and PSP thresholds further improved signal quality, resulting 566 in higher mean TRC SNR values in all cases. Though non -significant, the higher TRC 567 SNR values reinforce the advantage of incorporating both frequency - and time-568 domain metrics and provide further justification for the use of QT-NIRS when channel 569 pruning of infant fNIRS data. 570 4.2 Effects of QT-NIRS parameter choice 571 Higher SCI- and PSP thresholds reduced data retention by pruning more channels, but 572 improved signal quality in TRCs. The effect of parameter changes was smaller than 573 anticipated, particularly on signal quality. PSP Threshold had more influence than SCI 574 Threshold on both outcomes but particularly on Channels Retained, illustrating that 575 caution is needed when altering this threshold to avoid unnecessary data loss. 576 A positive SCI:PSP interaction reduced the channel pruning rate relative to 577 that expected when considering both thresholds in isolation (see Figure 7b), 578 suggesting that the potential cost to data retention from increasing one threshold 579 value is mitigated when the other value is also high. This aligns with the use of 580 complementary signal metrics in QT-NIRS which must both be of sufficient quality to 581 retain channels: each threshold will exclude channels which may be included by the 582 other, with the overlap in excluded channels increasing with parameter values. 583 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 22 The slight negative SCI:PSP interaction effect on TRC SNR may be driven by the 584 highest values for both parameters (Figure 7a). Higher thresholds increased the 585 likelihood of pruning channels with neuronally -evoked cortical haemodynamic 586 responses and, consequently, high SNR; their removal is likely to reduce the mean TRC 587 SNR more than would be expected for lower threshold values. Change in TRC SNR is 588 also less uniform across SCI - and PSP Threshold values than for Channels Retained. 589 Unlike channel retention, which reflects the whole array, TRC SNR is localized to 590 select channels. Single channels will therefore likely contribute a much greater 591 proportional value to the average TRC SNR and the consequence of pruning these 592 channels is likely to have a larger relative impact on the TRC SNR. Additionally, 593 pruning a channel consistently reduces Channels Retained whereas its impact on TRC 594 SNR depends on the pruned channel’s SNR, further contributing to the inconsistency 595 in change. 596 4.3 Effects of Motion 597 Substantial negative effects of PoM on both signal quality and channel retention 598 suggest that motion reduces signal quality in at least some channels even when 599 artefacts are not considered in the pruning analysis process, as was the case in this 600 work. Motion artefacts may cause optodes to move, dislodge, or uncouple completely, 601 causing a poorer quality signal reflected in decreased TRC SNR. In turn, channels 602 affected by motion may be pruned to a greater extent during QT-NIRS processing, 603 leading to lower Channels Retained values. 604 Interaction effects show PSP Threshold had the greater impact on motion-605 affected data of the two thresholds, improving signal quality but reducing channel 606 retention, especially at high threshold values. In contrast, a comparatively modest 607 effect of SCI Threshold on signal quality and retention was found. Motion exhibited a 608 significant adverse effect on Average PSP, likely due to optode displacement severe 609 enough to disrupt coupling, and a positive effect of motion incidences on Average SCI, 610 suggesting that SCI measures may be capturing correlation induced by latent, 611 undetected motion artefacts in the data. Future work may examine the effect of 612 different motion detection parameters, or alternative motion correction methods 613 altogether such as the Sobel filter [45], acceptance rate adaptive algorithm [46], global 614 variance of temporal derivatives [47], [48] or entropy-based methods [49]. 615 The potential (interaction) effect of high PSP Threshold values on channel 616 retention, especially for channels with a greater proportion of motion, warrants 617 caution for users when looking to employ high PSP Threshold values. This is 618 particularly true given the relatively small beneficial impact on signal quality of 619 increasing the PSP threshold, indicated by its main effect on TRC SNR. This is 620 consistent with the argument for using lower values discussed in the 621 Recommendations section of the Discussion. Conversely, motion incidences have the 622 largest negative impact on TRC SNR; low PSP thresholds may exacerbate this effect if 623 motion is not appropriately addressed. 624 4.4 Effects of age 625 Age had small effects on TRC SNR and Channels Retained, suggesting limited changes 626 in signal quality and retention in infants between the age of 5 and 24 mo. Reflecting 627 this, associated trends were less commonly observed with age than for other 628 predictors; however, it was notable that channel loss was less severe when increasing 629 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 23 the SCI Threshold in younger participants (Figure 6). Additionally, while higher signal 630 PoM is associated with reduced channel retention, older infants retained more 631 channels than younger ones for data with high motion prevalence ( Figure 5d). This 632 may be due to younger infants manually touching or grabbing the cap more, causing 633 more severe artefacts that permanently displace optodes and affect subsequent 634 coupling, an assertion supported by post-hoc data examination which suggested a 635 decrease in motion severity with age (see Supplementary Materials, Figure S 7) 636 Interactions with age for TRC SNR showed no meaningful patterns. This may 637 reflect a lack of consistent changes within the age range sampled here. It may also 638 reflect the multifaceted mix of concepts which ‘age’ represents: the interplay of the 639 physiological and behavioural changes with age may be too complex for a simple fixed 640 effect to capture. Interactions between age and other model terms (e.g. CSE, CL) 641 caused convergence issues and were omitted. Future research may priorit ize 642 investigating age-related change and associated factors affecting signal quality – such 643 as hair characteristics [23], hair style [31], or hair type changes with age [50]. 644 4.5 Other predictors 645 4.5.1 Cohort 646 Cohort had substantial effects on signal quality and retention, likely due to factors 647 such as testing environment, tester experience, parent and infant behaviours, and 648 sample size. Since cohort was not a primary predictor, its main effect was not 649 investigated in depth and interaction terms were not included. Nevertheless, data 650 from the Gambian cohort – whose skin and hair characteristics have been found to 651 pose challenges for fNIRS signal quality [23] – exhibited better signal quality and 652 retention, to the extent that data exclusion was far lower for Gambian infants in 653 general and almost non-existent at the infant level (see Figure 2). There may be 654 several reasons for this. Firstly, UK infants generally had finer hair which was longer 655 at later assessment ages, possibly causing cap slippage and poorer subsequent signal 656 quality, as these values appear to decrease with age. Lower motion incidence in 657 Gambian infants may also play a role, possibly due to a relative lack of familiarity with 658 digital screens in daily life leading to an increased focus on a novel, unfamiliar object. 659 While physical characteristics of participants undoubtedly affect the fNIRS signal, 660 cohort differences in this study emphas ize the need to consider other factors which 661 affect signal quality during data collection. 662 4.5.2 Weak or saturated channel signals 663 A small and significant negative effect of CSE on channel retention was anticipated, 664 considering it is itself a measure of channel removal. CSE was also significantly 665 negatively associated with data quality, however, suggesting poor coupling in 666 channels with extremely low quality signal could be affected by – or causing – signal 667 quality issues elsewhere in the array. It may be of interest to investigate whether 668 signal quality was worse in channels located closely to the CSE channels in future 669 work, as has been the focus of prior work into motion artefact detection [49]. 670 4.6 Strengths and limitations 671 A key strength of this work is the dataset used. Data encompassed five testing time 672 points across the age range 5- to 24mo, allowing assessment of QT -NIRS and its key 673 parameters for infants whilst accounting for age -related structural changes in skull 674 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 24 thickness, cardiac signal properties, and surface vasculature [51]. Data from two sites 675 were also assessed, one of which was a rural setting in a sub-Saharan African country, 676 addressing a common bias which often exists in infant neuroimaging studies when 677 participant recruitment is limited to predominantly White infants from high -income 678 contexts [52]. As such, the findings are likely to be more robust and generalisable than 679 those derived from smaller or single-site samples. The inclusion of longitudinal data 680 allows for the assessment of within- and between-participant changes over time, 681 providing richer insights into the effect of processing methods than would be possible 682 from cross-sectional analyses alone. 683 Another strength is the MLM approach which accounted for the hierarchical 684 variance structure of longitudinal data grouped by task and cohort, reducing bias and 685 enabling random intercepts and slopes during regression analysis to capture 686 individual differences [39]. This approach also handled missing data introduced 687 through prior channel removal (CSE channels), missed visits or incomplete testing, 688 which would not have been possible with many other common analysis procedures 689 [53]. Domain knowledge and data -driven insight were combined to priorit ize model 690 predictors and validate model assumptions and additionally fitted and assessed 691 subsidiary MLMs to ensure the final model (Equation 2) was as comprehensive as 692 possible. 693 QT-NIRS was compared with CV pruning, using a maximum wavelength CV 694 difference of 0.2, as previously applied with data from this study [14]. Significant 695 differences in signal quality between QT -NIRS and CV pruning were found in all but 696 one age/task/cohort combination. However, another common CV pruning approach 697 uses a single-channel threshold instead, pruning both channel wavelength signals at 698 least one of them exceeds it [12], [32], [54] . Future work could investigate whether 699 the significant differences found here persist when using this alternative thresholding 700 method. 701 The focus of this work was on optimizing thresholds of the two QT-NIRS 702 parameters which are most pivotal to performance, for which only reported 703 recommended values for adult participants were found in the literature. Future work 704 may focus on other parameters, such as changing the quality threshold (q_threshold), 705 balancing channel quality discernment with the risk of losing nuance in signal 706 characteristics. Future studies could also investigate the impact of altering window 707 size or overlap: the former will likely balance measurement accuracy within windows 708 against overall temporal sensitivity; the latter may provide more window temporal 709 sensitivity at the expense of computational efficiency. Given the dominance of PSP 710 Threshold change on outcomes reported here, it may also be of interest to explore 711 pruning using only the PSP threshold [55]. 712 This work must also be placed in the context of the increasing impact of 713 machine learning on infant NIR imaging data processing, with future work in the field 714 likely to assess the efficacy and generalisability of such methods. Deep learning 715 approaches are frequently being added to the literature: a machine learning based 716 detector has been developed to identify ‘bad’ channels to be pruned, for example [55]. 717 This approach was shown to be more adaptive, interpretable, and effective across 718 diverse noise types than QT-NIRS and other methods, so future work may seek to 719 assess the efficacy of this independently. In mitigation, care should be taken to ensure 720

common to many deep learning approaches (e.g. overfitting) are 721 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 25 addressed through appropriate validation strategies, including the use of 722 independent test sets and cross-validation. 723 Model design was limited by convergence issues, particularly when including 724 random slopes and intercepts. The most likely and important sources of variation 725 were prioritized, however it was not possible to fully include all relevant interaction 726 terms or random effects. Additionally, linear terms were used in the models to 727 simplify interpretability. Future work may incorporate non -linear terms, as an 728 alternative to the bootstrapping approach for dealing with non-normal residuals. 729 Other alternative approaches could include sensitivity analysis of the influential 730 points and outliers; variance modelling for specific predictors; and utilisation of 731 robust standard errors. 732 While age is likely to reflect more than behaviour and motion, the longitudinal 733 study design may also result in participant familiarity with the testing procedures and 734 the paradigms, in turn possibly affecting behaviour, attention, stress levels, 735 engagement, and – consequently – the recorded data. Thus, caution should be taken 736 when examining the effects of the age model term and its interactions, recognising 737 this as a potential confounding factor in this work. 738 Further evaluation of QT-NIRS as a channel pruning method for infant fNIRS 739 data is still necessary to address the limitations of this work, extend it to high-density 740 systems, and compare it to alternative approaches including those incorporating 741 machine learning [55]. 742 4.7 Recommendations 743 Based on this work, the following guidelines for channel pruning infant fNIRS data are 744 recommended: 745 (1) QT-NIRS as a pruning method is preferable to pruning using CV (when using a 746 minimum threshold difference between wavelengths) 747 (2) Users should conduct channel pruning on motion-free data, with considerable 748 emphasis during initial preprocessing placed on adequately identifying motion 749 artefacts 750 (3) Tuning of the PSP Threshold should be priorit ized over the SCI threshold in 751 data 752 (4) To provide a good trade -off between data quality and retention, lower 753 thresholds (with minima of psp_threshold ≈ 0.04–0.05 and sci_threshold ≈ 0.6) 754 can be used for infant data than th ose recommended for adults, especially in 755 older infants with fine/slippery hair, since (amongst other factors): 756 (i) There are a large proportion of data showing SCI- and PSP values lower 757 than the adult thresholds of 0.8 and 0.1, respectively (Figure 2) 758 (ii) The risk of removing data with higher thresholds is likely greater than 759 the potential gain in improved signal quality 760 (iii) There is a plateau in both average TRC SNR and Channels Retained 761 values when using lower threshold values than the advised minima (see 762 Supplementary Materials, Figure S 6) 763 A pictorial guide to the effect of the most common predictors on data quality and 764 retention is also included in Figure 8. 765 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 26 It is still strongly recommended that fNIRS users take the appropriate time to 766 understand the dataset being analysed since no approach can possibly be universal. 767 To aid final parameter selection, a tool which other users may find useful to guide 768 parameter selection was developed . Alongside code used during processing and 769 analysis reported in this study can be found at: https://github.com/sam-770 beaton/pruningComparisons/. The tool is designed to examine the effects of 771 parameter threshold changes within a group (e.g. age), by establishing a trend 772 capturing the trade-off between data quality and retention, then assessing which 773 specific SCI- and PSP Threshold parameter combinations perform best in relation to 774 this trend. 775 5 Conclusion 776 Recent advances in fNIRS channel pruning approaches show promise for improving 777 the accuracy of preprocessing by evaluating both the time and frequency domains. 778 The work described here compared QT-NIRS with an established pruning method 779 utilising CV, across 5 infant ages, two paradigms, and two sites. It was found that QT-780 NIRS provides data with greater signal quality when controlling for data retention. 781 Figure 8: Guide to the effects of the parameters and data characteristics on data quality and retention. First five rows represent positive (green) or negative(red) associations with increasing numerical parameters (SCI Threshold and PSP Threshold) or data characteristics (Age, PoM, CSE). All positive effects are small in size. Bottom two rows represent categ orical variables, with greyscale shading indicative of the impact on outcome changing the categorical variable may have. .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 27 The consequences of different parameter choices for QT -NIRS were also 782 demonstrated, highlighting the importance of the PSP Threshold, plus the influence of 783 motion. Evidence-based recommendations for QT -NIRS pruning, and parameter 784 choice for infant data with different characteristics, are provided. 785 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 28 Disclosures 786 The authors have no conflict of interest to declare. 787 788 Code and Data Availability 789 All data used in the analyses presented can be made available following relevant 790 approvals. The code used to conduct the analyses and generate the figures presented 791 in this paper are available at https://github.com/sam-beaton/pruningComparisons. 792 Code relies on the following open source R libraries: arules [56], broom [57], car [58], 793 data [59], doParallel [60], dplyr [61], effectsize [62], effects [58], e1071 [63], foreach 794 [64], ggplot2 [65], ggpubr [66], gridExtra [67], lme4 [41], lmerTest [68], MASS [69], 795 matrix [70], MuMIn [71], Scales [72], tidyr [73], viridis [74]. 796 797 Acknowledgments 798 We would like to place on record our thanks to the children, mothers and wider 799 families who took part in this study. In addition, we would like to thank the data 800 collection teams in both The Gambia and the UK . Finally, we thank Luca Pollonini for 801 his contribution to early discussions concerning the framing of the work, and Johann 802 Benerradi, for providing task-relevant channels for the social/non-social task prior to 803 their publication. 804 805 Funding 806 Sam Beaton, Ebrima Mbye, Samantha McCann (to July 2024) and Sophie Moore are 807 supported by a Wellcome Trust Senior Research Fellowship (220225/Z/20/Z) held 808 by Sophie Moore. The BRIGHT Study was funded by the Gates Foundation 809 (OPP1127625) and core funding MC -A760-5QX00 to the International Nutrition 810 Group by the Medical Research Council UK and the UK Department for International 811 Development (DfID) under the MRC/DfID Concordat agreement. Further support was 812 provided through a UKRI Future Leaders Fellowship (MR/S018425/1) held by Sarah 813 Lloyd-Fox. Borja Blanco was supported by a Medical Research Council Programme 814 Grant (MR/T003057/1) and a UKRI Future Leaders fellowship (MR/S018425/1). 815 816 Author Contributions (CRediT taxonomy) 817 SB: conceptualisation, formal analysis, methodology, software, validation, 818 visualisation, writing – original draft, writing – review and editing; BB: 819 conceptualisation, methodology, supervision, writing – review and editing; CB: 820 conceptualisation, methodology, writing – review and editing; CE: funding 821 acquisition, project administration, resources, writing – review and editing; SLF: 822 funding acquisition, project administration, resources, writing – review and editing; 823 EM: data curation, investigation, project administration; SMc: data curation, 824 investigation, project administration, supervision; ABR: conceptualisation, data 825 curation, supervision, visualisation, writing – review and editing; SM: funding 826 acquisition, project administration, resources, supervision, writing – review and 827 editing. 828 829 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 29 Biographies 830 Samuel Beaton is a doctoral researcher and research assistant at King’s College 831 London, developing analytical pipelines for functional near-infrared spectroscopy 832 (fNIRS) and high-density diffuse optical tomography (HD-DOT) data. 833 Chiara Bulgarelli is a Senior Lecturer at the Centre for Brain and Cognitive 834 Development at Birkbeck, University of London. She has over a decade of experience 835 using fNIRS with infants and toddlers to study the neural mechanisms underlying 836 social interactions and the development of brain connectivity. Recently, she 837 pioneered the use of fNIRS in non-traditional lab settings, such as with toddlers 838 interacting in a virtual reality environment. 839 Borja Blanco is a postdoctoral research associate in the Department of Psychology at 840 the University of Cambridge. His work focuses on developing data processing and 841 analysis methods for optical neuroimaging in developmental populations. He applies 842 these methods to investigate infant functional brain development and the contextual 843 factors that influence it. 844 Clare Elwell is a professor of medical physics leading projects to investigate brain 845 function using functional near infrared spectroscopy in high and low resource 846 settings in adult and infants. 847 Sarah Lloyd-Fox is a Principal Research Associate in the Department of Psychology, 848 University of Cambridge. She leads several multi-disciplinary projects focusing on 849 developmental trajectories of early cognitive and brain development during 850 pregnancy, infancy and early childhood. Her research focuses on understanding how 851 family and environmental context - i.e. contextual factors such as poverty associated 852 challenges and enriched multigenerational family support - shape early life. 853 Ebrima Mbye is a Field Coordinator at MRCG@LSHTM, formerly employed on the 854 BRIGHT Project and currently engaged on the INDiGO Trial working under Professor 855 Sophie Moore. 856 Samantha McCann is a Public Health Registrar and formerly Postdoctoral Research 857 Associate in the Department of Women and Children’s Health at King’s College 858 London. Her main research interest is supporting early child development, with a 859 strong focus on the impact of undernutrition in infancy on long -term 860 neurodevelopmental outcomes. 861 Anna Blasi is a Postdoctoral Research Fellow at UCL. Her research interests are 862 centered on functional aspects of human physiology. Her research career started with 863 models of the cardiovascular system and the effects of disease. Through her work at 864 UCL, KCL, and Birkbeck, her research interests have shifted toward the use of 865 functional imaging (fNIRS, fMRI) to study brain function and neurocognitive 866 development in early infancy. 867 Sophie Moore is Professor of Global Women and Children's Health in the Department 868 of Women & Children's Health at King’s College London and an Honorary Associate 869 Professor at the London School of Hygiene and Tropical Medicine (LSHTM). Her 870 research focuses on the nutritional regulation of ‘healthy’ fetal and infant growth, 871 incorporating infant immune and brain development as outcomes, and on the 872 mechanisms through which maternal, infant and childhood nutrition may influence 873 development and later health. 874 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 30

875 [1] L. L. Emberson, B. D. Zinszer, R. D. S. Raizada, and R. N. Aslin, ‘Decoding the infant 876 mind: Multivariate pattern analysis (MVPA) using fNIRS’, PLOS ONE, vol. 12, no. 877 4, p. e0172500, Apr. 2017, doi: 10.1371/journal.pone.0172500. 878 [2] J. Gervain, Y. Minagawa, L. Emberson, and S. Lloyd -Fox, ‘Using functional near -879 infrared spectroscopy to study the early developing brain: future directions and 880 new challenges’, Neurophotonics, vol. 10, no. 2, p. 023519, Apr. 2023, doi: 881 10.1117/1.NPh.10.2.023519. 882 [3] D. A. Boas, K. Chen, D. Grebert, and M. A. Franceschini, ‘Improving the diffuse 883 optical imaging spatial resolution of the cerebral hemodynamic response to 884 brain activation in humans’, Opt. Lett., vol. 29, no. 13, p. 1506, Jul. 2004, doi: 885 10.1364/OL.29.001506. 886 [4] H. Santosa, X. Zhai, F. Fishburn, P. J. Sparto, and T. J. Huppert, ‘Quantitative 887 comparison of correction techniques for removing systemic physiological signal 888 in functional near-infrared spectroscopy studies’, Neurophotonics, vol. 7, no. 3, p. 889 035009, Sep. 2020, doi: 10.1117/1.NPh.7.3.035009. 890 [5] M. A. Yücel et al., ‘Best practices for fNIRS publications’, Neurophotonics, vol. 8, 891 no. 01, Jan. 2021, doi: 10.1117/1.NPh.8.1.012101. 892 [6] J. Gemignani and J. Gervain, ‘Comparing different pre-processing routines for 893 infant fNIRS data’, Dev. Cogn. Neurosci. , vol. 48, p. 100943, Apr. 2021, doi: 894 10.1016/j.dcn.2021.100943. 895 [7] L. Pollonini, H. Bortfeld, and J. S. Oghalai, ‘PHOEBE: a method for real time 896 mapping of optodes-scalp coupling in functional near-infrared spectroscopy’, 897 Biomed. Opt. Express , vol. 7, no. 12, pp. 5104 –5119, Dec. 2016, doi: 898 10.1364/BOE.7.005104. 899 [8] L. M. Hocke, I. K. Oni, C. C. Duszynski, A. V. Corrigan, B. D. Frederick, and J. F. Dunn, 900 ‘Automated Processing of fNIRS Data —A Visual Guide to the Pitfalls and 901 Consequences’, Algorithms, vol. 11, no. 5, Art. no. 5, May 2018, doi: 902 10.3390/a11050067. 903 [9] T. Arichi et al., ‘Development of BOLD signal hemodynamic responses in the 904 human brain’, NeuroImage, vol. 63, no. 2, pp. 663 –673, Nov. 2012, doi: 905 10.1016/j.neuroimage.2012.06.054. 906 [10] C. Issard and J. Gervain, ‘Variability of the hemodynamic response in infants: 907 Influence of experimental design and stimulus complexity’, Dev. Cogn. Neurosci., 908 vol. 33, pp. 182–193, Feb. 2018, doi: 10.1016/j.dcn.2018.01.009. 909 [11] S. Baek, S. Marques, K. Casey, M. Testerman, F. McGill, and L. Emberson, ‘Attrition 910 rate in infant fNIRS research: A meta -analysis’, Infancy, vol. 28, no. 3, pp. 507 –911 531, 2023, doi: 10.1111/infa.12521. 912 [12] E. M. Frijia et al., ‘Functional imaging of the developing brain with wearable high-913 density diffuse optical tomography: A new benchmark for infant neuroimaging 914 outside the scanner environment’, NeuroImage, vol. 225, p. 117490, Jan. 2021, 915 doi: 10.1016/j.neuroimage.2020.117490. 916 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 31 [13] L. H. Collins-Jones et al., ‘Whole-head high-density diffuse optical tomography to 917 map infant audio-visual responses to social and non-social stimuli’, Imaging 918 Neurosci., vol. 2, pp. 1–19, Sep. 2024, doi: 10.1162/imag_a_00244. 919 [14] S. Lloyd-Fox et al., ‘Habituation and novelty detection fNIRS brain responses in 920 5- and 8-month-old infants: The Gambia and UK’, Dev. Sci., vol. 22, no. 5, Sep. 921 2019, doi: 10.1111/desc.12817. 922 [15] S. Lloyd-Fox, A. Blasi, A. Volein, N. Everdell, C. E. Elwell, and M. H. Johnson, ‘Social 923 Perception in Infancy: A Near Infrared Spectroscopy Study’, Child Dev., vol. 80, 924 no. 4, pp. 986–999, Jul. 2009, doi: 10.1111/j.1467-8624.2009.01312.x. 925 [16] S. M. Hernandez and L. Pollonini, ‘NIRSplot: A Tool for Quality Assessment of 926 fNIRS Scans’, in Biophotonics Congress: Biomedical Optics 2020 (Translational, 927 Microscopy, OCT, OTS, BRAIN) , Washington, DC: Optica Publishing Group, 2020, 928 p. BM2C.5. doi: 10.1364/BRAIN.2020.BM2C.5. 929 [17] L. Pollonini, C. Olds, H. Abaya, H. Bortfeld, M. S. Beauchamp, and J. S. Oghalai, 930 ‘Auditory cortex activation to natural speech and simulated cochlear implant 931 speech measured with functional near-infrared spectroscopy’, Hear. Res., vol. 932 309, pp. 84–93, Mar. 2014, doi: 10.1016/j.heares.2013.11.007. 933 [18] B. Blanco et al., ‘Cortical Responses to Social Stimuli in Infants at Elevated 934 Likelihood of ASD and/or ADHD: a Prospective Cross-Condition fNIRS Study’, 935 Cortex, Sep. 2023, doi: 10.1016/j.cortex.2023.07.010. 936 [19] L. K. Gossé, P. Pinti, F. Wiesemann, C. E. E. Elwell, and E. J. H. Jones, ‘Developing 937 customized NIRS-EEG for infant sleep research: methodological considerations’, 938 Neurophotonics, vol. 10, no. 3, p. 035010, Sep. 2023, doi: 939 10.1117/1.NPh.10.3.035010. 940 [20] S. Fleming et al., ‘Normal ranges of heart rate and respiratory rate in children 941 from birth to 18 years: a systematic review of observational studies’, Lancet, vol. 942 377, no. 9770, pp. 1011–1018, Mar. 2011, doi: 10.1016/S0140-6736(10)62226-943 X. 944 [21] MedlinePlus Medical Encyclopedia, National Library of Medicine, ‘Pulse’. 2025. 945 [Online]. Available: https://medlineplus.gov/ency/article/003399.htm 946 [22] K. L. Perdue, A. Westerlund, S. A. McCormick, and C. A. N. Iii, ‘Extraction of heart 947 rate from functional near -infrared spectroscopy in infants’, J. Biomed. Opt. , vol. 948 19, no. 6, p. 067010, Jun. 2014, doi: 10.1117/1.JBO.19.6.067010. 949 [23] M. A. Yücel et al., ‘Inclusivity in fNIRS Studies: Quantifying the Impact of Hair and 950 Skin Characteristics on Signal Quality with Practical Recommendations for 951 Improvement’, Oct. 28, 2024, bioRxiv. doi: 10.1101/2024.10.28.620644. 952 [24] S. Lloyd-Fox et al., ‘The Brain Imaging for Global Health (BRIGHT) Project: 953 Longitudinal cohort study protocol’, Gates Open Res., vol. 7, p. 126, Oct. 2023, doi: 954 10.12688/gatesopenres.14795.1. 955 [25] J. Benerradi et al., ‘Functional specialisation across the first five years of life: a 956 longitudinal characterisation of social perception with fNIRS’, Jul. 29, 2025, 957 bioRxiv. doi: 10.1101/2025.07.28.667189. 958 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 32 [26] A. Blasi Ribera et al., ‘Longitudinal habituation and novelty detection neural 959 responses from infancy to early childhood in The Gambia and UK’, Dev. Cogn. 960 Neurosci., vol. 76, p. 101619, Dec. 2025, doi: 10.1016/j.dcn.2025.101619. 961 [27] M. Rozhko et al., ‘Effect of sleep stages on patterns of fNIRS hemodynamic 962 response to auditory paradigms in one -month-old Gambian and UK infants’, Jul. 963 31, 2025, bioRxiv. doi: 10.1101/2025.07.28.667238. 964 [28] L. Kischkel, ‘Assessing infant neurocognitive development in resource -poor 965 settings: the example of memory development in the UK and The Gambia’, 966 Doctoral thesis, University College London, 2019. [Online]. Available: 967 https://discovery.ucl.ac.uk/id/eprint/10073018/ 968 [29] S. Lloyd-Fox et al., ‘Cortical specialisation to social stimuli from the first days to 969 the second year of life: A rural Gambian cohort’, Dev. Cogn. Neurosci., vol. 25, pp. 970 92–104, Jun. 2017, doi: 10.1016/j.dcn.2016.11.005. 971 [30] S. Lloyd-Fox, A. Blasi, and C. E. Elwell, ‘Illuminating the developing brain: the past, 972 present and future of functional near infrared spectroscopy’, Neurosci. Biobehav. 973 Rev., vol. 34, no. 3, pp. 269 –284, Mar. 2010, doi: 974 10.1016/j.neubiorev.2009.07.008. 975 [31] L. Katus et al., ‘Implementing neuroimaging and eye tracking methods to assess 976 neurocognitive development of young infants in low - and middle -income 977 countries’, Gates Open Res. , vol. 3, p. 1113, Apr. 2019, doi: 978 10.12688/gatesopenres.12951.1. 979 [32] A. Blasi, S. Lloyd -Fox, Mark. H. Johnson, and C. Elwell, ‘Test –retest reliability of 980 functional near infrared spectroscopy in infants’, Neurophotonics, vol. 1, no. 2, 981 Oct. 2014, doi: 10.1117/1.NPh.1.2.025005. 982 [33] S. Lloyd-Fox, J. E. Richards, A. Blasi, D. G. M. Murphy, C. E. Elwell, and M. H. 983 Johnson, ‘Coregistering functional near-infrared spectroscopy with underlying 984 cortical areas in infants’, Neurophotonics, vol. 1, no. 2, p. 025006, Oct. 2014, doi: 985 10.1117/1.NPh.1.2.025006. 986 [34] L. H. Collins -Jones et al., ‘Longitudinal infant fNIRS channel -space analyses are 987 robust to variability parameters at the group-level: An image reconstruction 988 investigation’, NeuroImage, vol. 237, p. 118068, Aug. 2021, doi: 989 10.1016/j.neuroimage.2021.118068. 990 [35] S. McCann, ‘Nutrition, Growth and Neurodevelopment Across Infancy and Early 991 Childhood in Rural Gambia.’, PhD Thesis, King’s College London, London, UK, 992 2020. 993 [36] T. J. Huppert, S. G. Diamond, M. A. Franceschini, and D. A. Boas, ‘HomER: a review 994 of time-series analysis methods for near-infrared spectroscopy of the brain’, 995 Appl. Opt. , vol. 48, no. 10, pp. D280 –D298, Apr. 2009, doi: 996 10.1364/AO.48.00D280. 997 [37] R. Di Lorenzo et al., ‘Recommendations for motion correction of infant fNIRS data 998 applicable to multiple data sets and acquisition systems’, NeuroImage, vol. 200, 999 pp. 511–527, Oct. 2019, doi: 1000 https://doi.org/10.1016/j.neuroimage.2019.06.056. 1001 [38] G. A. Zimeo Morais et al., ‘Non-neuronal evoked and spontaneous hemodynamic 1002 changes in the anterior temporal region of the human head may lead to 1003 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 33 misinterpretations of functional near -infrared spectroscopy signals’, 1004 Neurophotonics, vol. 5, no. 1, p. 011002, Aug. 2017, doi: 1005 10.1117/1.NPh.5.1.011002. 1006 [39] A. Gelman and J. Hill, Data analysis using regression and multilevel/hierarchical 1007 models. Cambridge university press, 2006. 1008 [40] R Core Team, R: A Language and Environment for Statistical Computing . Vienna, 1009 Austria: R Foundation for Statistical Computing, 2024. [Online]. Available: 1010 https://www.R-project.org/ 1011 [41] D. Bates, ‘Fitting linear mixed -effects models using lme4’, ArXiv Prepr. 1012 ArXiv14065823, 2014. 1013 [42] H. Akaike, ‘A new look at the statistical model identification’, IEEE Trans. Autom. 1014 Control, vol. 19, no. 6, pp. 716–723, 1974. 1015 [43] J. W. Harris and H. Stöcker, Handbook of mathematics and computational science. 1016 Springer Science & Business Media, 1998. 1017 [44] H. F. Behrendt, C. Firk, C. A. N. Iii, and K. L. Perdue, ‘Motion correction for infant 1018 functional near-infrared spectroscopy with an application to live interaction 1019 data’, Neurophotonics, vol. 5, no. 1, p. 015004, Feb. 2018, doi: 1020 10.1117/1.NPh.5.1.015004. 1021 [45] S. Jahani, S. K. Setarehdan, D. A. Boas, and M. A. Yücel, ‘Motion artifact detection 1022 and correction in functional near -infrared spectroscopy: a new hybrid method 1023 based on spline interpolation method and Savitzky –Golay filtering’, 1024 Neurophotonics, vol. 5, no. 1, p. 015003, Feb. 2018, doi: 1025 10.1117/1.NPh.5.1.015003. 1026 [46] S. Sutoko et al., ‘Adaptive algorithm utilizing acceptance rate for eliminating 1027 noisy epochs in block -design functional near -infrared spectroscopy data: 1028 application to study in attention deficit/hyperactivity disorder children’, 1029 Neurophotonics, vol. 5, no. 4, p. 045001, Oct. 2018, doi: 1030 10.1117/1.NPh.5.4.045001. 1031 [47] A. Sherafati, A. T. Eggebrecht, K. M. Bergonzi, T. M. Burns-Yocum, and J. P. Culver, 1032 ‘Improvements in functional diffuse optical tomography maps by global motion 1033 censoring techniques’, in Microscopy Histopathology and Analytics , Optica 1034 Publishing Group, 2018, pp. JW3A-51. 1035 [48] A. Sherafati et al., ‘Global motion detection and censoring in high-density diffuse 1036 optical tomography’, Hum. Brain Mapp., vol. 41, no. 14, pp. 4093–4112, 2020, doi: 1037 10.1002/hbm.25111. 1038 [49] G. Lee, S. H. Jin, S. H. Lee, B. Abibullaev, and J. An, ‘fNIRS motion artifact correction 1039 for overground walking using entropy based unbalanced optode decision and 1040 wavelet regression neural network’, in 2017 IEEE International Conference on 1041 Multisensor Fusion and Integration for Intelligent Systems (MFI) , Nov. 2017, pp. 1042 186–193. doi: 10.1109/MFI.2017.8170427. 1043 [50] S. A. Furdon and D. A. Clark, ‘SCALP HAIR CHARACTERISTICS IN THE NEWBORN 1044 INFANT’, Adv. Neonatal Care , vol. 3, no. 6, p. 286, Dec. 2003, doi: 1045 10.1016/j.adnc.2003.09.005. 1046 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 34 [51] L. L. Emberson, S. L. Crosswhite, J. R. Goodwin, A. J. Berger, and R. N. Aslin, 1047 ‘Isolating the effects of surface vasculature in infant neuroimaging using short -1048 distance optical channels: a combination of local and global effects’, 1049 Neurophotonics, vol. 3, no. 3, p. 031406, Apr. 2016, doi: 1050 10.1117/1.NPh.3.3.031406. 1051 [52] E. T. Margolis et al. , ‘Modality -level obstacles and initiatives to improve 1052 representation in fetal, infant, and toddler neuroimaging research samples’, Dev. 1053 Cogn. Neurosci., vol. 72, p. 101505, Apr. 2025, doi: 10.1016/j.dcn.2024.101505. 1054 [53] A. P. Field and D. B. Wright, ‘A Primer on Using Multilevel Models in Clinical and 1055 Experimental Psychopathology Research’, J. Exp. Psychopathol. , vol. 2, no. 2, pp. 1056 271–293, May 2011, doi: 10.5127/jep.013711. 1057 [54] E. Mercure et al., ‘Language Experience Impacts Brain Activation for Spoken and 1058 Signed Language in Infancy: Insights From Unimodal and Bimodal Bilinguals’, 1059 Neurobiol. Lang., vol. 1, no. 1, pp. 9–32, Mar. 2020, doi: 10.1162/nol_a_00001. 1060 [55] C. Gerloff, M. A. Yücel, L. Mehlem, K. Konrad, and V. Reindl, ‘NiReject: toward 1061 automated bad channel detection in functional near -infrared spectroscopy’, 1062 Neurophotonics, vol. 11, no. 4, p. 045008, Nov. 2024, doi: 1063 10.1117/1.NPh.11.4.045008. 1064 [56] M. Hahsler, C. Buchta, B. Gruen, and K. Hornik, arules: Mining Association Rules 1065 and Frequent Itemsets . 2025. [Online]. Available: https://CRAN.R -1066 project.org/package=arules 1067 [57] D. Robinson, A. Hayes, and S. Couch, broom: Convert Statistical Objects into Tidy 1068 Tibbles. 2025. [Online]. Available: https://CRAN.R-project.org/package=broom 1069 [58] J. Fox and S. Weisberg, An R Companion to Applied Regression , 3rd ed. Thousand 1070 Oaks, CA: Sage, 2019. 1071 [59] T. Barrett et al., data.table: Extension of `data.frame` . 2024. [Online]. Available: 1072 https://CRAN.R-project.org/package=data.table 1073 [60] M. Corporation and S. Weston, doParallel: Foreach Parallel Adaptor for the 1074 ‘parallel’ Package . 2022. [Online]. Available: https://CRAN.R -1075 project.org/package=doParallel 1076 [61] H. Wickham, R. François, L. Henry, K. Müller, and D. Vaughan, dplyr: A Grammar 1077 of Data Manipulation . 2025. [Online]. Available: https://CRAN.R -1078 project.org/package=dplyr 1079 [62] D. Lüdecke, M. S. Ben-Shachar, D. Makowski, and I. Patil, effectsize: Indices of 1080 Effect Size and Standardized Parameters . 2024. [Online]. Available: 1081 https://CRAN.R-project.org/package=effectsize 1082 [63] D. Meyer, E. Dimitriadou, K. Hornik, A. Weingessel, and F. Leisch, e1071: Misc 1083 Functions of the Department of Statistics, Probability Theory Group (Formerly: 1084 E1071), TU Wien . 2024. [Online]. Available: https://CRAN.R -1085 project.org/package=e1071 1086 [64] Microsoft and S. Weston, foreach: Provides Foreach Looping Construct. 2022. 1087 [Online]. Available: https://CRAN.R-project.org/package=foreach 1088 [65] H. Wickham, ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New 1089 York, 2016. [Online]. Available: https://ggplot2.tidyverse.org 1090 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint 35 [66] A. Kassambara, ggpubr: ‘ggplot2’ Based Publication Ready Plots . 2023. [Online]. 1091 Available: https://CRAN.R-project.org/package=ggpubr 1092 [67] B. Auguie, gridExtra: Miscellaneous Functions for ‘Grid’ Graphics . 2017. [Online]. 1093 Available: https://CRAN.R-project.org/package=gridExtra 1094 [68] A. Kuznetsova, P. B. Brockhoff, and R. H. B. Christensen, ‘lmerTest Package: Tests 1095 in Linear Mixed Effects Models’, J. Stat. Softw., vol. 82, no. 13, pp. 1–26, 2017, doi: 1096 10.18637/jss.v082.i13. 1097 [69] W. N. Venables and B. D. Ripley, Modern Applied Statistics with S, 4th ed. New 1098 York: Springer, 2002. 1099 [70] D. Bates and M. Maechler, Matrix: Sparse and Dense Matrix Classes and Methods . 1100 2024. [Online]. Available: https://CRAN.R-project.org/package=Matrix 1101 [71] K. Bartoń, MuMIn: Multi -Model Inference . 2024. [Online]. Available: 1102 https://CRAN.R-project.org/package=MuMIn 1103 [72] H. Wickham and D. Seidel, scales: Scale Functions for Visualization. 2023. [Online]. 1104 Available: https://CRAN.R-project.org/package=scales 1105 [73] H. Wickham, D. Vaughan, and M. Girlich, tidyr: Tidy Messy Data . 2024. [Online]. 1106 Available: https://CRAN.R-project.org/package=tidyr 1107 [74] S. Garnier, N. Ross, R. Rudis, A. Filippi, A. Kaplan, and A. Hackett, viridis: 1108 Colorblind-Friendly Color Maps for R . 2024. [Online]. Available: https://CRAN.R -1109 project.org/package=viridis 1110 1111 .CC-BY 4.0 International licensemade available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is The copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.06.680288doi: bioRxiv preprint