Fronto-parietal white matter microstructure associated with working memory performance in children with ADHD

INTRODUCTION
Attention Deficit/Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder with many functional impairments thought to be underpinned by difficulties in executive function domains such as working memory. The superior longitudinal fasciculus (SLF) plays an integral role in the development of working memory in neurotypical children. Neuroimaging research suggests reduced white matter organization of the SLF may contribute to working memory difficulties commonly seen in ADHD. This study aimed to examine the relationship between white matter organization of the SLF and working memory in children with ADHD.


METHODS
We examined the association of tract volume and apparent fibre density (AFD) of the SLF with working memory in children with ADHD (n = 64) and controls (n = 58) aged 9-11years. Children completed a computerized spatial n-back task and underwent diffusion magnetic resonance imaging (dMRI). Constrained spherical deconvolution-based tractography was used to construct the three branches of the SLF bilaterally and examine volume and AFD of the SLF.


RESULTS
Regression analyses revealed children with ADHD exhibited poorer working memory, and lower volume and AFD of the left SLF-II compared to healthy controls. There was also an association between reaction time and variability (RT and RT-V) and the left SLF-II. Further analyses revealed volume of the left SLF-II mediated the relationship between ADHD and working memory performance (RT and RT-V).


DISCUSSION
These findings add to the current body of ADHD literature, revealing the potential role of frontoparietal white matter in working memory difficulties in ADHD.

Individuals with ADHD can exhibit functional impairment in a number of domains, including academic, cognitive and social functioning (Efron et al., 2020;Holst & Thorell, 2019;Rosell o et al., 2020). A leading theory suggests that many of these impairments are underpinned by difficulties in executive functioning (Welsh & Pennington, 1988;Willcutt, Doyle, Nigg, Faraone, & Pennington, 2005), and in particular working memory (Kofler et al., 2019;Rapport, Chung, Shore, & Isaacs, 2001). Working memory is a limited capacity neural system that facilitates temporary storage of information and simultaneous manipulation in relation to complex tasks. Meta-analytic studies, in both children and adults with ADHD, consistently report that working memory is poorer in those with ADHD compared to those without ADHD (Alderson, Kasper, Hudec, & Patros, 2013;Kasper, Alderson, & Hudec, 2012;Martinussen, Hayden, Hogg-Johnson, & Tannock, 2005;Willcutt et al., 2005). Interestingly, there is evidence that as working memory demands increase, children with ADHD exhibit more inattentive and hyperactive symptoms (Alderson, Rapport, Kasper, Sarver, & Kofler, 2012;Burgess et al., 2010;Kofler, Rapport, Bolden, Sarver, & Raiker, 2010;Rapport et al., 2009). Research has also highlighted the influence of working memory in a range of fundamental skills impacted by ADHD, including math and reading (Friedman, Rapport, Calub, & Eckrich, 2018) and daily executive and social functioning (Brydges, Ozolnieks, & Roberts, 2017;Kofler et al., 2011Kofler et al., , 2018. In examining the neurobiology underpinning working memory, numerous functional magnetic resonance imaging (fMRI) studies have demonstrated consistent neural patterns. Meta-analyses of working memory studies, in both children and adults, show consistent activation in a widespread network of primarily frontal and parietal regions (Yaple & Arsalidou, 2018;Yaple, Stevens, & Arsalidou, 2019). Frontal and parietal cortices are primarily connected by an association white matter fibre bundle known as the superior longitudinal fasciculus (SLF) in each hemisphere. The SLF can be separated into three branches (I, II and III) each forming different connections and subsuming specific functions, such as attention and visuospatial processing (Klarborg et al., 2013;Parlatini et al., 2017). However, the exact role of each branch is not fully understood and has only recently been investigated in human subjects (Makris et al., 2005;Wang et al., 2016). Emerging evidence suggests that the SLF may play a role in the functioning of working memory.
To date, much of the work exploring brain-behaviour links for working memory has focused on clinical studies in adults. Kinoshita et al. (2016) (n ¼ 34) found a significant correlation between spatial working memory difficulties and the region that overlapped the first and second branches of the SLF. In the same study, two patients exhibited difficulties providing correct responses on the spatial 2-back task, a robust measure of spatial working memory, during direct subcortical stimulation of the first branch of the SLF during awake surgery. No difficulties in working memory were exhibited in immediate and long-term postoperative periods. Furthermore, diffusion magnetic resonance imaging (dMRI) studies have associated the SLF in working memory of neurotypical children (Farah, Tzafrir, & Horowitz-Kraus, 2020;Vestergaard et al., 2011), adolescents (Østby, Tamnes, Fjell, & Walhovd, 2011;Peters et al., 2012) and adults (Metzler-Baddeley et al., 2017). Research examining working memory difficulties in clinical populations, such as schizophrenia (Karlsgodt et al., 2008) and systemic lupus erythematosus (Zhao et al., 2018), also suggest significant association between SLF organization and their poor performance on a working memory task. Together, this research highlights the potential importance of the SLF in neurotypical working memory and provides a target tract to examine in association with working memory difficulties.
Although increasing neuroimaging evidence supports the hypothesis that the SLF is associated with working memory difficulties commonly exhibited in ADHD, fundamental limitations of DTI mean that previous work must be interpreted with caution (Tournier, Mori, & Leemans, 2011). Low FA values have often been interpreted as a reduction in white-matter integrity or organization. However, degree of anisotropy can be influenced by several factors such as inter-axonal spacing, axon diameter, membrane permeability, myelination and consistency of axon orientations (Beaulieu, 2002). This indicates that there is great ambiguity surrounding what element of white matter FA is ultimately measuring. A further limitation of the DTI framework is that it only models a single fibre direction within a voxel. Given that up to 90% of voxels throughout the brain contain multiple fibre directions (Jeurissen, Leemans, Tournier, Jones, & Sijbers, 2010), it is difficult for FA to precisely represent underlying fibre populations, both in terms of accurate tractography (defining the anatomy of the tract), as well as inferences about the underlying microstructure (Alexander, 2008;Wheeler-Kingshott & Cercignani, 2009). These are important implications for the SLF as it crosses through many other major white matter tracts.
Application of higher order models, such as constrained spherical deconvolution (CSD), has been shown to produce anatomically superior modelling of white-matter tracts (Farquharson et al., 2013) as the approach is less-susceptible to the issue of crossing fibres. Further, CSD also provides scope to generate more specific metrics of white-matter microstructure, such as apparent fibre density (AFD) (Raffelt et al., 2012). AFD represents the density of fibres measured as a proportion of space occupied within a voxel. Utilizing these more advanced neuroimaging metrics to investigate the SLF microstructure in ADHD has the potential to reveal more about the neurobiology of this disorder.
This study aims to examine the role of the SLF in working memory in children with ADHD. We hypothesized that, compared to neurotypical controls, children with ADHD would 1) perform more poorly on a measure of working memory and 2) exhibit atypical white matter organization of the SLF. We also expected that the relationship between ADHD diagnosis and working memory performance would be mediated by SLF white matter microstructure. This study intends to further understand the neurobiology of working memory difficulties associated with ADHD.

2.
Methods and materials

Participants
Data were collected as part of the Neuroimaging of the Children's Attention Project study. For full details see Sciberras et al. (2013), Silk et al. (2016). We report how we determined our sample size, all data exclusions (if any), all inclusion/ exclusion criteria, whether inclusion/exclusion criteria were established prior to data analysis, all manipulations, and all measures in the study. Briefly, participants were recruited in their second year of schooling (ages 6e8 years) and diagnostically defined at ADHD or control following a two-stage screening and caseconfirmation procedure using the Conners 3 ADHD Index (AI) (Conners & Angeles, 2008) and a parent face-to-face structured diagnostic interview (NIMH Diagnostic Interview Schedule for Children IV e DISC-IV; Shaffer, Fisher, Lucas, Dulcan, & Schwab-Stone, 2000). Diagnostic status was also reassessed at a 36-month follow-up (aged 9e11 years), where a subset of participants was invited to complete a magnetic resonance imaging (MRI) scan and computerized cognitive battery. Clinical, behavioural and neuroimaging assessments were conducted by researchers blinded to the diagnostic status of the participants. This study was approved by The Royal Children's Hospital (RCH) Human Research Ethics Committee (HREC #34071), and parents or guardians gave informed consent while participating children provided assent.
For the current study, the ADHD group required a positive diagnosis of ADHD at either recruitment or the 36-month follow-up. The neurotypical control group were required to screen negative for ADHD diagnosis at recruitment and 36month follow-up.
After MRI quality control the final sample, with complete working memory data, consisted of 64 ADHD individuals and 58 non-ADHD controls. Thirteen (20%) children in the ADHD group reported currently taking medication for their behaviour, at the time of the assessment (n ¼ 12 taking methylphenidate [with one also taking clonidine], n ¼ 1 taking atomoxetine and risperidone).
While recruitment procedure and data collection methods (including inclusion/exclusion criteria) for this study were carried out as described in the published protocol (Sciberras et al., 2013;Silk et al., 2016), no part of the study procedures or analyses were pre-registered prior to the research being conducted. Exclusion/inclusion criteria were implemented prior to data analysis. The sample size for this study was based on the feasibility of recruitment from the CAP study (see Silk et al., 2016 for details).

Demographic information
Socioeconomic status was determined using the Socio-Economic Indexes for Areas Disadvantage Index (SEIFA) for the family's postcode of residence (mean ¼ 1000; standard deviation ¼ 100), with lower scores reflecting greater disadvantage (Australian Bureau of Statistics, 2011). A brief estimate of intellectual function (IQ) was measured at recruitment using the 2-subtest version (vocabulary and matrix reasoning) of the Wechsler Abbreviated Scale of Intelligence e Second Edition (WASI-II) (Wechsler, 1999). Raw scores of the subtests are converted to age-based standard scores and then summed to generate a composite score. Composite scores are then transformed into an estimated full-scale intelligence quotient (FSIQ) score using normative data conversion tables. Child handedness was collected via self-report.

Working memory
Working memory performance was assessed using a computerized spatial n-back task using Psychology Software tool e-Prime. Participants were presented with 10 squares on the screen, which flash one at a time. Participants completed a 2-back version, in which they were required to press the spacebar when the square that was flashing also flashed the time before last (see Fig. 1). Participants received a demonstration of the block progression, followed by a practice block of 26 trials. The experimental task consisted of 78 trials (presented in 3 blocks) with targets on 24 of those trials. The performance measures are detailed in Table 1. Because simple measures of misses and false alarm responses cannot differentiate between ability to detect signals and response criteria, based on signal detection theory perceptual sensitivity (dprime [d 0 ]) and response bias (c), were calculated, as well as reaction time (RT) and reaction time variability (RT-V), as the key measures in each condition (Sorkin, 1999). Each measure listed in Table 1 records various elements of executive function; misses measure inattention, false alarms relate to impulsivity and response inhibition, while RT measures speed of performance and RT-V the ability to sustain attention. Working memory is best captured by RT variables as misses and false alarms relate more so to attention.

Mock scanner
All children were first prepared in a 30-min mock scanner session. The mock scanner replicates the environment of the MRI, familiarizing the child with the MRI environment, trains them in the requirements of laying still, and helps to reduce anxiety.

MRI scan
Neuroimaging data were collected on a 3-T Siemens TIM Trio MRI scanner (Siemens, Erlangen, Germany) at the Murdoch Children's Research Institute, Melbourne. Participants were scanned without sedation, lay supine with their head support in a 32-channel head coil and scans lasted approximately 45 min. For structural images, multi-echo MPRAGE (T1weighted) images were acquired in the sagittal plane, with in-scanner motion correction (TR ¼ 2530 msec, TE ¼ 1.77, 3.51, 5.32, 7.2 msec, flip angle 7 , voxel size ¼ .9 mm 3 ). To probe white matter microstructure, high-angular resolution diffusion imaging (HARDI) data were acquired in the transverse plane with an anterioreposterior phase encoding direction (PE). 60 gradient directions, b-value ¼ 2800 sec/mm 2 and four interleaved b ¼ 0 volumes were acquired (TR ¼ 3200 msec, TE ¼ 110 msec, echo spacing ¼ .69 msec, FOV ¼ 260 mm, multi- The index of the ability to discriminate between targets and non-targets. Higher values of d 0 indicate better ability to discriminate between targets and non-targets. Calculated using the following algorithm d 0 ¼ z (hit rate) e z (false alarm rate).
À4.392 (0 hits, 54 false alarms) to 4.392 (24 hits, 0 false alarms) d 0 ¼ 0 when performance is at chance level c The participants' willingness to respond. c ¼ 0 e zero or neutral bias c > 0 e conservative bias (a bias against responding 'yes') c < 0 e liberal bias (a bias towards responding 'yes') c o r t e x 1 6 6 ( 2 0 2 3 ) 2 4 3 e2 5 7 band factor ¼ 3 and 2.4 mm isotropic voxels). A reverse phase encoded image was acquired to correct for magnetic susceptibility-induced distortions during EPI acquisition.

Image processing
Diffusion data were preprocessed using MRtrix3 (version 0.3.15) software package. Preprocessing steps included denoising, motion, eddy current and susceptibility distortion correction, and bias field correction. For each participant, response functions were estimated, and fibre orientation distribution (FOD) maps were generated. Using anatomicallyconstrained tractography (ACT), whole brain tractography was run with 2 million streamlines seeded from the grey/ white matter boundary (Smith, Tournier, Calamante, & Connelly, 2012).

Method for defining the SLF tracts
The three segments of the SLF (SLF-I, -II, -III) were defined in each hemisphere, using the following process. Firstly, each participant's whole brain tractogram was filtered using a series of include and exclude regions of interest (ROI) previously described in Thiebaut De Schotten et al. (2011). ROIs consisted of on the coronal slice the level of the anterior commissure (AC), three ROIs demark superior, middle, and inferior frontal gyri, and a single large parietal ROI was drawn at the level of the posterior commissure (PC). Streamlines that pass through both superior frontal and parietal ROIs define SLF-I, both the middle frontal and parietal ROIs for SLF-II and both inferior frontal and parietal ROIs of the SLF-III. Exclusion ROIs ensured any streamlines passing through the arcuate fasciculus, midline, subcortical and brainstem, and internal and external capsule were omitted. To ensure standardized ROIs across individuals, the ROIs were specified on a cohort specific T1 template, generated using the Advanced Normalization Tools (stnava.github.io/ANTs). ROIs were subsequently transformed to each individual's native space using inverted non-linear symmetric and diffeomorphic warping. A second filtering step was also applied based on cortical endpoints of the SLF tracts previously described (Kamali, Flanders, Brody, Hunter, & Hasan, 2014;Wang et al., 2016). For this step, tract endpoints were restricted based on cortical parcellations, generated in Freesurfer (v5.3.0) using the Desi-kaneKilliany atlas (See Fig. 2). This was performed on structural T1 images aligned to diffusion space using a linear registration (Jenkinson & Smith, 2001). SLF-I tracts were filtered to streamlines terminating anteriorly in the Superior Frontal gyrus and posteriorly in the Precuneus or Superior Parietal parcellations. SLF-II tracts were filtered to streamlines terminating anteriorly in the Middle Frontal gyrus (combined caudal and rostral middle frontal) and posteriorly in the Superior Parietal or Inferior Parietal parcellations. SLF-III tracts were filtered to streamlines terminating anteriorly in the Inferior Frontal Gyrus (combined pars opercularis, pars orbitalis and pars triangularis) and posteriorly in the Inferior Parietal or Supramarginal parcellations.
Tract apparent fibre density (AFD) was calculated is the 'afdconnectivity' command in MRTrix3 (Raffelt et al., 2012). A probability map of the number of streamlines passing through each voxel was thresholded, removing voxels that contained <5/1000 streamlines to generate tract volume from the included voxels. Tensor-derived fractional anisotropy (FA) scalar maps were also generated, and mean FA extracted.

Head motion
To determine the level of head motion during the diffusion sequence for each participant, the FSL Motion Outlier script was used to extract the mean and the maximum framewise displacement (FWD) ( Table 1.) (Power, Barnes, Snyder, Schlaggar, & Petersen, 2012). Preliminary analyses revealed no significant group differences in either the mean or maximum FWD (p > .05), therefore FWD was not included as a covariate in subsequent analyses.

Statistical analysis
All statistical analyses were conducted in Stata (Version 15.1) and R (2021.09.1) and IBM SPSS (Version 28). Demographic and symptom characteristics were compared between ADHD and neurotypical developing control groups. Between-group mean comparisons tested via independent samples t-tests, and chi-squared goodness of fit test for proportion comparisons.
An overview of the statistical analysis plan is given in Fig. 3. Between-group regression analyses were conducted in Stata (V15.1) for all working memory and tract outcomes (Step 1 & 2,  Fig. 3.), with model fit of outcomes determining regression c o r t e x 1 6 6 ( 2 0 2 3 ) 2 4 3 e2 5 7 model to be employed. To improve model fit, RT and RT-V were log-transformed to improve normality (as is common for RT), negatively skewed data was reversed transformed to employ Poisson/negative binomial models. Additionally, "c" was positively shifted by 2 on the x-axis to remove negative values. Ordinal Least Squares (OLS) regression models were carried out for all tract outcomes. Because of heavy skew, zero-inflated Poisson regression models were utilized for misses, false alarms, and d 0 . For both working memory and tract analyses the alpha was set at p < .05. Results are displayed also noting correction where findings survived FDR correction (Benjamini & Hochberg, 1995). Those returning uncorrected statistically significant differences between ADHD and control groups were retained for further regression modelling. To address the assumptions of mediation (Step 4, Fig. 3.), the culmination of significant models resulted in an OLS regression analyses to examine the association between working memory performance and SLF properties (Step 3, Fig. 3.).
All models included age, sex, and medication as covariates, whereby medication was defined as either the presence or absence of current medication. In accordance with Dennis et al., 2009, FSIQ was not included as a covariate. Across analyses, effect size was estimated by zero-inflated Poisson regression incidence rate ratio (IRRZ) for Poisson modelling and adjusted-R 2 or standardised regression coefficient for OLS regression modelling.
Mediation analysis was conducted using Model 4 in the PROCESS macro for SPSS (Hayes, 2013) (Step 4, Fig. 3.). The model tested whether left SLF-II properties (mediator variable M) mediated the effect of ADHD group (predictor variable X) in predicting working memory performance (outcomes variable Y), with age, sex and medication as covariates. P M , the ratio of the indirect effect to the total effect, was used as the measure of effect size, whereby effects were considered significant if 95% confidence intervals (CIs), based on 5000 bootstrap samples, did not include zero (Hayes, 2013).

Data and code availability
Data from the Children's Attention Project cohort are available via Lifecourse: https://lifecourse.melbournechildrens.com/ cohorts/cap-and-nicap/. Due to ethics restraints, data is available upon request with ethics approval request. Code used for neuroimaging processing and analysis is publicly available and provided on the MRtrix3 (https://www.mrtrix. org/) website. All stimuli and code used for statistical analysis can be found on github (https://github.com/ldipnall/ ADHD-Working-Memory-and-SLF-White-Matter-Microstructure).

3.1.
Step 1: group differences in working memory performance Regression analyses explored the effect of ADHD diagnosis on working memory performance, adjusting for age, sex and medication. Statistically significant between-group differences were found for Misses, False Alarms, d 0 , RT and RT-V (Table 3) with the ADHD group showing poorer performance on each measure compared to healthy controls. No significant differences were found for c (b ¼ .004, p ¼ .965). All variables returning significant group difference at an uncorrected level (p < .05) survived FDR correction for multiple comparisons and were retained for further analyses. c o r t e x 1 6 6 ( 2 0 2 3 ) 2 4 3 e2 5 7

3.2.
Step 2: group differences in SLF microstructure Similar between-group analyses were conducted for tract SLF measures (See Table 4). After adjusting for age, sex and medication, significantly lower tract volume (b ¼ À199.12, p ¼ .018) and AFD (b ¼ À.23, p ¼ .042) were found at un uncorrected level (p < .05) for the ADHD group the left SLF-II only, in comparison to controls. These two tract measures were retained for further analyses. No tracts survived FDR correction for multiple comparisons.

3.3.
Step 3: relationship between working memory performance and left SLF-II white matter properties Findings from the above working memory performance and tract analyses culminated in final regression analyses across the entire sample, addressing whether working memory performance (misses, false alarms, d 0 , RT, RT-V) was associated with left SLF-II volume and AFD (adjusting for age, sex and medication). The only statistically significant working memory effects found were RT (b ¼ À.225, p ¼ .013) and RT-V (b ¼ À.210, p ¼ .019) predicting left SLF-II volume; and RT (b ¼ À.228, p ¼ .012) and RT-V (b ¼ À.181, p ¼ .045) predicting left SLF-II AFD, whereby slower and more variable RT predicted lower tract volume and density in the left SLF II. No other working memory effects were statistically significant (See Table 5). These findings held when controlling for diagnostic group, for RT (Volume: b ¼ À.193, p ¼ .034; AFD: b ¼ À.200, p ¼ .029), but dropped to trend level for RT-V (Volume: b ¼ À.174, p ¼ .054; AFD: b ¼ À.148, p ¼ .103). The relationships are visually represented in Fig. 5. The association was strongest in the control group, due to less variability in the ADHD group (more clustered towards poorer performance and lower tract metric). c o r t e x 1 6 6 ( 2 0 2 3 ) 2 4 3 e2 5 7 Fig. 4 e Visual representations of distributions between groups within variables (note: purple dot signified mean, blue diamond median and IQR boxplot shown for each group respectively. For exact values please see Table 2.) c o r t e x 1 6 6 ( 2 0 2 3 ) 2 4 3 e2 5 7 3.4.
Step 4: mediating effect of left SLF-II white matter properties on between-group differences in working memory Given that RT and RT-V are proposed to significantly predict AFD and volume of the left SLF-II, a mediation analysis was run to investigate the potential role of white matter organization (AFD and volume) of the left SLF-II in mediating the working memory difficulties in children with ADHD (See Table 6). Fig. 6 depicts the model testing the mediating role of left SLF-II volume in the demonstrated relationship between ADHD diagnosis and RT and RT-V. Volume of the left SLF-II mediated the relationship between ADHD diagnosis and RT, indirect effect b ¼ .022, 95% CI [.001, .056], accounting for 22% of the association between ADHD diagnosis and RT (See Fig. 3: Step 1). ADHD diagnosis significantly predicted volume of the left SLF-II (b ¼ À.037, SE ¼ .016) and volume of the left SLF-II significantly predicted RT (b ¼ À.590, SE ¼ .275). Volume of the left SLF-II also mediated the relationship between ADHD and RT-V, indirect effect b ¼ .028, 95% CI [.002, .064], accounting for 18.5% of the association between diagnosis and RT-V (See Fig. 3: Step 3). ADHD significantly predicted volume of the left SLF-II (b ¼ À.037, SE ¼ .016) however left SLF-II volume only predicted RT-V at a trend level (b ¼ À.737, SE ¼ .378, p ¼ .053). For mediating models of left SLF-II AFD, the indirect effect did not significantly improve the explanatory power.

Discussion
The aim of this study was to examine the relationship between working memory and white matter organization of the SLF in children with ADHD. Compared to neurotypical controls, children with ADHD exhibited poorer working memory and reduced white matter properties of the left SLF-II. We identified novel associations between common measures of working memory (RT and RT-V) and fronto-parietal white matter organization (specifically the left SLF-II). Volume, but not AFD of the left SLF-II was also found to mediate the relationships between ADHD and working memory performance.
The potential mediating role of fronto-parietal white matter properties in the relationship between ADHD and working memory was the key finding of this study. Volume and AFD of the left SLF-II was associated with working memory performance (RT and RT-V) across the entire cohort, however, only volume demonstrated a significant mediating relationship between ADHD and working memory (RT and RT-V). This indicates that the potential neural mechanisms underpinning the relationship between ADHD and working memory difficulties may lie within the white matter organization of the left SLF-II. As per the authors' knowledge, to date no dMRI studies have specifically examined the relationship between fronto-parietal white matter organization, specifically the SLF, and working memory in children with ADHD. In neurotypical children, increased white matter organization of   Note: B ¼ standardised regression coefficient (b); SE ¼ standard error; RT ¼ reaction time; RT-V ¼ reaction time variability; AFD ¼ apparent fibre density; bold signifies p < .05.
c o r t e x 1 6 6 ( 2 0 2 3 ) 2 4 3 e2 5 7 the left SLF has been associated with working memory development (Tamnes, Fjell, Westlye, Østby, & Walhovd, 2012), indicating that as working memory skills, including RT-V, develop during childhood and adolescence, the white matter organization of the left SLF strengthens. Here our results indicate that the inverse may be true during development for children with ADHD, with reduced white matter organisation of the left SLF-II mediating the relationship between ADHD and working memory performance. This study adds to the growing body of literature implicating frontoparietal neural involvement in the poor working memory performance of those with ADHD. Consistent with previous literature, our findings revealed children with ADHD performed more poorly on a common   c o r t e x 1 6 6 ( 2 0 2 3 ) 2 4 3 e2 5 7 empirical measure of working memory than neurotypical controls (Alderson et al., 2013;Kasper et al., 2012;Martinussen et al., 2005;Willcutt et al., 2005). On the spatial n-back task, children with ADHD committed significantly more misses and false alarms. Their d 0 values were also significantly lower, highlighting reduced capacity to discriminate between targets and non-targets. Children with ADHD also showed significantly longer reaction time (RT) and greater reaction time variability (RT-V). Of the aforementioned measures, increased RT and RT-V are consistently reported in ADHD literature across several working memory tasks in both children (Willcutt, Sonuga-Barke, Nigg, & Sergeant, 2008;Ilbegi et al., 2021) and adults (Adams, Roberts, Milich, & Fillmore, 2011) with ADHD compared to controls. A meta-analysis of 319 studies investigating RT-V in children, adolescent and adults with ADHD (Kofler et al., 2013) found that individuals with ADHD were more variable in their performance, but not necessarily slower, suggesting that difficulties in working memory, and not processing speed, are associated with ADHD. Although our study found relationships between ADHD and RT and ADHD and RT-V, of the two relationships RT-V was found to be stronger. These results therefore add to the increasing evidence suggesting RT variables as neuropsychological endophenotypes for ADHD (Uebel et al., 2010).
In addition to working memory performance, differences in white matter organization were observed in children with ADHD compared to controls. Using the CSD-based method of diffusion weighted imaging, both macro-and microstructural properties (volume and AFD) of the left SLF-II were found to be significantly lower in children with ADHD. These results are consistent with the current diffusion literature of ADHD, with multiple meta-analyses reporting the white matter organization of the SLF to be significantly reduced in ADHD (Albajara S aenz et al., 2020;Aoki, Cortese, & Castellanos, 2018;Chiang et al., 2019;Chiang et al., 2016;Kobel et al., 2010;Luo et al., 2020;Makris et al., 2008;McAlonan et al., 2007;Nagel et al., 2011;O'Conaill et al., 2015;Timothy J Silk et al., 2009;Vaidya, 2011;van Ewijk et al., 2012;Wu et al., 2017;Wu et al., 2020). This finding also adds to the current body of literature by providing increased specificity. The CSD-based dMRI methodology applied here is more accurate than the DTI framework previously applied (Farquharson et al., 2013) and was the first to apply CSD-dMRI voxel-wise analysis across the three branches of the SLF.
The SLF-II originates in the posteriolateral parietal lobe, runs through the supramarginal gyrus, postcentral, precentral and middle frontal gyri, terminating at the dorsolateral prefrontal cortex (Thiebaut De Schotten et al., 2011). Associated cortical regions have been implicated in many fMRI studies addressing working memory in ADHD (McCarthy et al., 2013). An fMRI study similarly using the 2-back task to assess working memory in youth with ADHD (n Total ¼ 45; n ADHD ¼ 24), found reduced efficiency of the dorsolateral prefrontal cortex (DPFC) in ADHD during performance of the working memory task (B edard et al., 2014). This proves pertinent as the DLPFC is the termination point for the SLF and when taken in conjunction with the current findings, it is speculated that lower volume and AFD of the left SLF-II in children with ADHD may impact capacity for the left SLF-II to participate in the top-down and bottom-up processing involving cortical regions such as the DLPFC. Although literature investigating the neurobiological underpinnings of working memory in ADHD is sparse, the broader construct of executive function within which working memory lies, has however been investigated. Wolfers et al. (2015) reported that higher RT-V on an executive function task was associated with lower white matter organization of the right SLF in adults with ADHD. Although the finding of Wolfers et al. (2015) is contrary to the current results, in that the relationship was lateralized to the right hemisphere, it is worth noting that our results report a trend towards lower volume (p ¼ .078) and AFD (p ¼ .072) for the right SLF-II in children with ADHD compared to neurotypical controls. The incongruence of findings is possibly primarily due to differences in study cohorts. Here children were studied, while adults were the population of interest in the study by Wolfers et al. (2015). This limits the conclusions that can be drawn when comparing both studies as the brain of a child is significantly different to that of an adult. To clarify the relationship between working memory and brain development in ADHD, multi-modal longitudinal imaging studies such NICAP is a necessity.
Apart from including a larger sample size than most previous ADHD dMRI studies, this study is the first to apply novel CSD-dMRI modelling to examine the relationship between ADHD working memory and white matter organization of the SLF. The application of higher-order imaging techniques allows for the acquisition of superior diffusion images when compared to those acquired via the DTI-framework. Ultimately this improves the accuracy of the tractography which in turn improves the validity and translatability of the results. FA was estimated to allow for comparison against the DTI literature, while AFD provided a more specific biological estimate of white matter microstructure. Future work should focus on more advanced measures of fibre direction within a voxel (e.g., fixel-based analysis).
This study, nevertheless, is not without limitations. Firstly, directionality of the results of the mediation analysis cannot be assumed due to the cross-sectional nature of the study. However, as the NICAP cohort is being followed longitudinally, future studies aim to more closely investigate the directionality of the relationship between white matter microstructure and measures of working memory in children with ADHD. Secondly, as the aim of this study was to examine the neurological underpinnings of working memory in children with ADHD, the role of sustained attention was not examined. It is possible that along with working memory, some variability in RT seen in children with ADHD could be attributed to difficulties in sustained attention as the n-back task progresses. Working memory and attentional networks appear integrated in neurotypical individuals, whereby attentional lapses lead to worse working memory (deBettencourt, Keene, Awh, & Vogel, 2019). As, to the authors knowledge, this has not been explored in children with ADHD, it is possible that a similar relationship exists amongst ADHD cohorts. To address this ambiguity, future research is encouraged to apply a whole-report working memory task that accounts for trial-bytrial fluctuations in attention, such as developed by Adam, Mance, Fukuda, and Vogel (2015). Thirdly, only visuospatial working memory was assessed. Future concurrent assessment of visuospatial and auditory working memory in c o r t e x 1 6 6 ( 2 0 2 3 ) 2 4 3 e2 5 7 children with ADHD is suggested, as it will ensure a more comprehensive understanding of working memory in children with ADHD, especially in the context of endophenotypes. A final limitation of this study is the low proportion of girls with ADHD. However, with a ratio of 3:1 for boys to girls, this study does accurately reflect the broader paediatric clinical population (Nøvik et al., 2006).
The findings here add to the growing body of evidence supporting the hypothesis that atypical fronto-parietal white matter underpins the poor working memory performance frequently seen in ADHD. Caution must however be taken when interpreting the results as there is need for replication of the findings in a larger dataset and across a broader age range.
Understanding the neurobiology of ADHD, and specifically the potential endophenotype of working memory deficits, will only further our understanding of ADHD. The identification and development of valid and reliable neural markers can only help improve the specificity, and thus outcomes, of both pharmaceutical and behavioural intervention for ADHD. As this cohort is being followed longitudinally, future studies plan to examine how developmental changes of the SLF are associated with working memory development.
Further to this, due to the heterogeneity of clinical presentation, there is a need for future work to adopt a more dimensional assessment of ADHD rather than the dichotomous categorisation of diagnostic status. This may aid in the elucidation of neural biomarkers that more specifically align with ADHD itself, as well as potential sub-types.
Overall, these results provide evidence supporting the role of the left SLF-II in working memory difficulties commonly seen in ADHD. This suggests that lower volume and fibre density of fronto-parietal white matter may underpin the nature of poor working memory in ADHD. As working memory difficulties appear to impact the lives of many of those with ADHD in a multitude of significant ways, large-scale neuroimaging studies offer future research the opportunity to further understand the neural underpinnings of working memory in ADHD.

Open practices
The study in this article earned Open Material badge for transparent practices. The materials for this study are available at: https://github.com/ldipnall/ADHD-Working-Memoryand-SLF-White-Matter-Microstructure.