Chronotype and subjective sleep quality predict white matter integrity in young people with emerging mental disorders

Protecting brain health is a goal of early intervention. We explored whether sleep quality or chronotype could predict white matter (WM) integrity in emerging mental disorders. Young people (N = 364) accessing early‐intervention clinics underwent assessments for chronotype, subjective sleep quality, and diffusion tensor imaging. Using machine learning, we examined whether chronotype or sleep quality (alongside diagnostic and demographic factors) could predict four measures of WM integrity: fractional anisotropy (FA), and radial, axial, and mean diffusivities (RD, AD and MD). We prioritised tracts that showed a univariate association with sleep quality or chronotype and considered predictors identified by ≥80% of machine learning (ML) models as ‘important’. The most important predictors of WM integrity were demographics (age, sex and education) and diagnosis (depressive and bipolar disorders). Subjective sleep quality only predicted FA in the perihippocampal cingulum tract, whereas chronotype had limited predictive importance for WM integrity. To further examine links with mood disorders, we conducted a subgroup analysis. In youth with depressive and bipolar disorders, chronotype emerged as an important (often top‐ranking) feature, predicting FA in the cingulum (cingulate gyrus), AD in the anterior corona radiata and genu of the corpus callosum, and RD in the corona radiata, anterior corona radiata, and genu of corpus callosum. Subjective quality was not important in this subgroup analysis. In summary, chronotype predicted altered WM integrity in the corona radiata and corpus callosum, whereas subjective sleep quality had a less significant role, suggesting that circadian factors may play a more prominent role in WM integrity in emerging mood disorders.

this subgroup analysis.In summary, chronotype predicted altered WM integrity in the corona radiata and corpus callosum, whereas subjective sleep quality had a less significant role, suggesting that circadian factors may play a more prominent role in WM integrity in emerging mood disorders.
K E Y W O R D S bipolar disorder, depression, endophenotype, imaging, prediction, youth mental health

| INTRODUCTION
Protecting young people's brain health is a goal of the early intervention model (Hickie et al., 2019).Prospective studies demonstrate that exposure to psychopathology in early life predicts later alterations in brain health (e.g.sensory, motor and cognitive impairments) (Wertz et al., 2021).An essential aspect of brain health is white matter (WM), which refers to the myelinated bundles of axons connecting brain regions.Efficient communication of information across the brain relies on the integrity of WM tracts (Wandell, 2016).Studies have found abnormalities in WM integrity in people with depressive, bipolar, psychotic and other mental disorders (Barth et al., 2022;Favre et al., 2019;Kelly et al., 2018;Kochunov, Fan, et al., 2022;van Velzen et al., 2020), which suggests that part of the pathophysiology of mental disorders, in some cases, is perturbed WM integrity.Moreover, some, but not all, studies have linked WM abnormalities to illness course (e.g.age of onset, illness duration, stage), potentially indicating a worsening of brain health with more exposure to psychopathology or specific correlates (Kelly et al., 2018;Kochunov, Fan, et al., 2022;Lagopoulos, Hermens, Hatton, Battisti, et al., 2013;Sacks et al., 2021;van Velzen et al., 2020).If this is true, there is a need to intervene early to try to offset these effects (Moffitt & Caspi, 2019).
Two factors that are implicated in mental disorders (Carpenter et al., 2021) and are associated with neurobiological alterations in other populations are chronotype and sleep quality.These factors are of interest to WM integrity for several reasons.First, sleep loss causes acute changes in brain morphology (e.g.reduced WM integrity after sleep deprivation) (Elvsåshagen et al., 2015).Second, during sleep, genes related to WM integrity (e.g.myelin maintenance and turnover) are upregulated, and relatedly, sleep loss disrupts WM-associated gene transcription (Bellesi et al., 2013;Cirelli et al., 2004;de Vivo & Bellesi, 2019).Third, evening chronotype might be marker of chronic sleep loss and might therefore predict perturbed WM integrity (Merikanto et al., 2012;Roepke & Duffy, 2010).Fourth, it has been hypothesised that a major function of sleep is to maintain healthy WM (de Vivo & Bellesi, 2019).
At the time of our analysis, only two studies have investigated associations between chronotype and WM integrity.While chronotype is operationalised differently depending on the measure, a popular definition is an individual's bio-behavioural preference for the daily timing of sleep, activity, and feeding, and it relates to daily variation of cognitive function, mood, and energy, among other phenomena (Horne & Östberg, 1976).This dimension captures individuals that are referred to as 'night owls' (evening types) and 'morning larks' (morning types), as well as individuals in between.The first relevant study categorised healthy males into chronotypes (early, late and intermediate) and tested differences in five common measures of WM integrity: fractional anisotropy (FA), fibre count, and mean, axial, and radial diffusivities (MD, AD and RD, respectively) (Rosenberg et al., 2014).Compared to early or intermediate chronotypes, late chronotypes had altered WM integrity in regions of the frontal lobe, temporal lobe, and corpus callosum, suggesting evening types, on average, have poorer WM health (Rosenberg et al., 2014).The second relevant study investigated genetic correlations between chronotype and WM integrity.Twenty-nine genetic correlations among evening chronotype and WM integrity were identified.Mendelian randomisation identified 13 one-way causal associations implying that evening chronotype influences increases in MD, AD, and RD (suggesting lower WM integrity) in the superior fronto-occipital fasciculus, internal capsule (posterior limb and reticular region), corpus callosum (splenium), and corona radiata (anterior, posterior and superior); there were no genetic correlations between FA and chronotype (García-Marín et al., 2021).
The relationship between sleep quality and WM has received more attention, but the literature is mixed.Studies in young and older healthy adults have found that poorer subjective sleep is associated with lower FA and/or higher MD, AD, and RD values (suggesting lower WM integrity), typically in frontal and temporal lobes (Khalsa et al., 2017;Sexton et al., 2017;Takeuchi et al., 2018).A study of healthy adolescents found that subjective sleep quality was associated with higher FA in the posterior limb of the internal capsule, but not a range of other tracts, suggesting poor sleep quality may not exert notable effects on WM in early adolescence (Jamieson et al., 2020).These findings contrast with studies of nonclinical samples that have not observed relationships between subjective sleep quality and WM integrity in adults (Toschi et al., 2021) and older adults (Altendahl et al., 2020;Li et al., 2020).Notably, studies in patients with insomnia have identified WM alterations including lower FA in the internal capsule, right anterior and superior corona radiata, right superior longitudinal fasciculus, body of the corpus callosum, and right thalamus (Bresser et al., 2020;Li et al., 2016;Spiegelhalder et al., 2014).Whether these WM alterations are a cause or consequence of sleep disturbance is not known.
It is notable that WM alterations linked to subjective sleep quality and chronotype are in regions that regulate sleep/wakefulness and emotional, cognitive, and sensorimotor functions, which are often altered in mental disorders.The goal of this study was to examine the relationships among WM integrity, subjective sleep quality and chronotype in youth with emerging mental disorders.Based on the results from related literatures (e.g.insomnia), we hypothesised that worse subjective sleep quality and higher eveningness would be associated with alterations in WM integrity (lower FA and higher RD, MD, and AD), particularly in frontal, temporal and callosal tracts.We focus on these metrics given their potential relevance to axonal and myelin integrity specifically, but we note that these analyses are associational and not causal.

| Ethical approval and informed consent
All procedures were approved by the University of Sydney Human Research Ethics Committee (2012/1626; 2012/1631).Written informed consent was obtained from participants aged 16 and older, and parental/guardian consent was obtained for participants aged under 16.

| Study participants
Participants were referred from one of two primary carebased early intervention mental health services (www.headspace.org.au) in Sydney, Australia (Scott, Hermens, Glozier, et al., 2012).The participants were recruited to a transdiagnostic research study at the University of Sydney's Brain and Mind Centre, examining the neurobiology of emerging mental disorders in young people.The wider programme has been described elsewhere (Carpenter et al., 2017;Crouse et al., 2020;Hermens et al., 2015) but briefly, involved administration of a protocol of self-report questionnaires, interview-based clinician-rated scales, a neuropsychological test battery, and neuroimaging.While studies of this sample have examined relationships among WM and neurocognition, clinical stage, and alcohol/substance use (Hatton et al., 2014;Hermens et al., 2018Hermens et al., , 2019Hermens et al., , 2022;;Lagopoulos, Hermens, Hatton, Battisti, et al., 2013;Lagopoulos, Hermens, Hatton, Tobias-Webb, et al., 2013;Sacks et al., 2021), this is the first to examine WM integrity in a predictive framework and the first to examine sleep quality and chronotype.The participants were excluded if they met any of the following: (a) history of neurological disease, (b) medical illness known to affect cognitive/ brain function, (c) received electroconvulsive therapy in the 3 months before assessment, (d) clinically evident intellectual disability, or (e) insufficient understanding of the English language to participate in testing.

| Assessments
Demographic characteristics (e.g.age, biological sex, educational attainment) were collected using standard questionnaire items.As reported elsewhere (Scott, Hermens, Naismith, et al., 2012), referring clinicians provided 'primary' psychiatric diagnoses based on Diagnostic and Statistical Manual of Mental Disorders, fourth edition (DSM-IV) criteria (American Psychiatric Association, 1994), which were grouped into four broad categories: depression, bipolar disorder, psychosis, and other (e.g.anxiety disorder, substance use disorder, learning disorder, personality disorder and autism spectrum disorder).Healthy controls were recruited from the same local community and were excluded if they had any personal history of a mental disorder or any other medical or neurological conditions.

| Pittsburgh sleep quality index (PSQI)
PSQI is a 19-item, self-report rating scale designed to measure sleep quality in clinical populations (Buysse et al., 1989).Seven component scores are derived (scores 0-3) and then summed to produce a total score (higher scores indicating poorer sleep quality).The PSQI has good psychometric properties (Carpenter & Andrykowski, 1998;Dietch et al., 2016).We used the total score as a dimensional index of sleep quality.

| Note
We have also collected data using objective sleep/circadian measures (e.g.dim-light melatonin onset and polysomnography) (Porteous et al., 2021;Robillard et al., 2018); the subset with objective data is only $10% as large as those with self-ratings, which was our focus here.

| Image acquisition, pre-processing and ENIGMA-DTI pipeline
The participants underwent magnetic resonance imaging at the Brain and Mind Centre using a 3-Tesla GE MR750 Discovery scanner (GE Medical Systems, Milwaukee).To enable neuroanatomical analysis at high resolution (0.9-mm isotropic resolution), we used a custom MP-RAGE 3D T1-weighted sequence: repetition time = 7265 ms, echo time = 2784 ms, flip angle = 15 , coronal acquisition, field of view = 230 mm, acquisition matrix = 256 Â 256 and total slices = 196.Whole brain diffusion-weighted images were acquired using an echo planar imaging sequence: repetition time = 7000 ms, echo time = 68 ms, slice thickness = 2.0 mm, field of view = 230 mm, acquisition matrix = 256 Â 256, axial orientation and 69 gradient directions.Eight images without gradient loading (b = 0 s/mm 2 ) were acquired before the acquisition of 69 images (each containing 55 slices) with uniform gradient loading (b = 1159 s/mm 2 ).
As performed elsewhere (Hermens et al., 2019), using the FMRIB Software Library (FSL, v 5.0) (Jenkinson et al., 2012), diffusion-weighted volumes were eddy current corrected and brain tissue was extracted.A tensor model was fitted at each voxel using DTIFIT, and each resulting voxel-wise tensor map of FA was inspected for the appropriate reconstruction.Tract-based spatial statistics (TBSS) within FSL was used to nonlinearly register each participant's tensor map to a 1-mm isotropic target FA image (Andersson, Jenskinson, & Smith, 2007).All transformations were visually inspected before affine registration of aligned FA images into the 1-mm isotropic MNI152 template (Andersson, Jenkinson, & Smith, 2007).
The Enhancing NeuroImaging Genetics through Meta-Analysis diffusion tensor imaging (ENIGMA-DTI) pipeline (Jahanshad et al., 2013;Kochunov et al., 2014) was then used to perform atlas-based voxel-wise segmentation of the major WM pathways, as previously reported (Jahanshad et al., 2013).All FA maps were linearly, elastically registered (Leow et al., 2007) to the ENIGMA-DTI template resulting in three-dimensional deformation fields that were applied to the three diffusivity maps and projected onto the skeletonised ENIGMA-DTI template.Finally, the individual images of AD, RD and MD were transferred to the FA template space, and the ENIGMA skeleton was projected onto the registered images.

| Eligibility criteria
The participants were eligible if they met the following criteria: 1. Participant aged 12-30.2. Completed self-ratings of chronotype (MEQ) and sleep quality (PSQI).3. Underwent DTI that passed quality control checks.4. The MEQ, PSQI, and DTI assessments were within a 3-month window.

| Statistical analysis
Analyses were performed using R (version 4.2.2) (R Core Team, 2022) in RStudio (RStudioTeam, 2020).Given the inconsistencies in the studies linking WM to sleep quality and chronotype, the absence of published studies linking DTI to sleep quality and chronotype in early phase mental disorders, and the risk of loss of information by focussing on a subset of WM tracts, we examined all available WM tracts.Our analyses were conducted in two steps.In the first step, given that many DTI parameters were available, we reduced the parameter space by taking the average across the hemispheres (where possible) and using this derived variable.Using Spearman's rank correlation coefficient, we discarded tracts that were not associated (p < .05)with either the PSQI or MEQ.In the second step, for the tracts that were associated with the PSQI or MEQ, we used ML to examine whether sleep quality and chronotype were robust predictors of WM integrity, and whether they were more important predictors than clinical and demographic variables.The DTI parameters examined at each stage is shown in Table S2.
We intentionally chose ML models from different families (ensemble, regularization and kernel-based) to capture diverse relationships within the data.This is because there is no single ML algorithm which can globally perform better than others for all possible problems, implied by the 'no free lunch theorem' (Wolpert & Macready, 1997).Accordingly, we used four ML algorithms (random forest (Liaw & Wiener, 2002), lasso (Friedman et al., 2010), elastic-net (Friedman et al., 2010), support vector machine) and used recursive feature elimination (RFE)) to identify the most informative features associated with each WM measure (including age, sex, diagnostic group, education level, sleep quality and chronotype).The RFE algorithm removes the lowest ranked features from the model and then builds a model with the reduced feature set, repeating the process until the 'best' set of features is determined.
Each algorithm provides advantages and disadvantages and works well under different assumptions.By comparing the models, we can assess their relative performance, identify potential biases, and choose the appropriate model (or models).The performance of each model was evaluated using the normalised root-mean-squared error (nRMSE), normalised mean absolute error (nMAE) and coefficient of determination (R 2 ).We used 10-fold cross validation to reduce bias and model variance.In each fold, 90% of the sample was used as the training set, and the remaining 10% was used as a test set.In the following, we refer to 'important' predictors as those that were selected by at least 80% of the models (i.e. using four ML algorithms Â 100 iterations = selected in 320/400 models).

| Sample characteristics
The sample included 364 young people aged 12-30 years (M = 20.8;SD = 4.1), with 222 females (61.0%).Most had a primary diagnosis of depression (n = 147; 40.4%), followed by bipolar disorder (n = 78; 21.4%), psychotic disorder (n = 40; 11.0%), and 'other' disorders (n = 48; 13.2%).We included 51 unaffected controls (14.0% of the sample).The clinical and control groups had similar estimated premorbid IQ (103.9 vs 103.4).The clinical group had a lower level of educational attainment (12.7 years vs 14.2 years), which was likely driven by them being younger (20.84 years vs 23.75 years).In the clinical group, the mean total score for the MEQ was 46.6 (SD = 10.1),indicating that, on average, the samples were 'intermediate' types; the mean total score for the PSQI was 8.5 (SD = 4.2), meaning that, on average, the sample experience poor subjective sleep quality.Sample characteristics by sex and diagnostic groups are presented in Table S1.

| Predicting FA
Figure 1 presents the ranked feature selection for models predicting FA across the eligible WM tracts: ALIC, CGC, EC and SFO (nRMSE = 21.13-22.26%)(Table S3).Sex was the most commonly identified important feature (≥80% of models for ALIC, CGC, EC, and SFO), followed by depression (CGC, EC and SFO) and age (CGC).
Chronotype or sleep quality were not important predictors of FA in any tract.

| Predicting AD
Figure 2 shows the ranked feature selection for models predicting AD in only one tract: CGH.Sleep quality was a top-ranked predictor of AD in the CGH (identified in $90% of models), and alongside other features including education level and sex, contributed to a fair prediction of CGH (nRMSE = 24.53-24.70%)(Table S4).S5).

| Subgroup analysis: focus on mood disorders
Depressive and bipolar disorder were the only diagnostic groups that were important predictors of WM integrity.Given this, and that evening chronotype and poor sleep quality are commonly associated with mood disorders, and proposed to play a causative role in key features of mood disorders (Crouse et al., 2021;Freeman et al., 2020), we decided to conduct a subgroup analysis focussing on cases with a mood disorder (n = 78 bipolar disorder; n = 147 depressive disorder) and excluding the healthy controls and other diagnoses.
This hypothesis was partly supported (Figures 4-6).In ML models predicting FA (nRMSE = 24.29-26.74%),chronotype was important for predicting CGC, whereas sleep quality was not important for predicting integrity of any included tract (Table S6).In models predicting AD (nRMSE = 25.26-26.89%),chronotype was the topranked predictor in two of three eligible tracts, including (ACR and GCC), whereas sleep quality was not important for predicting any included tract (Table S7).Finally, in models predicting RD (nRMSE = 23.28-27.61%),chronotype was important for predicting ACR , CR and GCC, however, sleep quality was not important for predicting any included tract (Table S8).

| DISCUSSION
Pooling results from four ML algorithms, we found in a transdiagnostic sample of youth with mood, psychotic and other mental disorders, that chronotype and subjective sleep quality were less important predictors of WM integrity compared to demographic and diagnostic factors.By contrast, when focussing on a mood disorder subgroup, chronotype, but not subjective sleep quality, was a topranked predictor of WM in multiple regions, and notably of metrics thought to be markers of axonal and myelin integrity.In the following, we will discuss our findings with a primary focus on chronotype and sleep quality (given the study hypotheses) and provide a brief discussion of findings related to diagnosis, age and sex.
In the analysis of the transdiagnostic sample, chronotype was not an important predictor of WM in any parameter.By contrast, chronotype emerged as a topranked predictor across several tracts in the mood disorder subgroup, including FA in the CGH; AD in the ACR and GCC; and RD in the ACR, CR and GCC.Based on the correlational analyses (Figures S3 and S4), the clearest pattern of association was that for individuals with a mood disorder, those with higher morningness differ in WM integrity compared to their peers with higher eveningness, particularly in tracts related to the corpus callosum (i.e.lower FA in the body and whole of CC and higher AD and RD in the genu of CC) and the corona radiata (i.e.lower FA in the posterior CR, increased AD and RD in the anterior CR, and increased RD in the CR).Among the limited studies of chronotype and WM, there are findings of lower FA in people with an evening (late) chronotype (Rosenberg et al., 2014), typically in frontal or temporal lobes or corpus callosum, which shares commonalities with our findings.One genome-wide association study observed negative genetic correlations among F I G U R E 3 Machine learning prediction of radial diffusivity (RD) as a measure of white matter integrity in young people with emerging mental disorders.(a and b) Prediction of RD in two white matter tracts.The predictors of RD for these tracts are depicted along the y-axis and ranked in order of their importance to the model.The x-axis illustrates how many of the models each predictor variable was featured in (range: 0-100%).The four machine learning models used (random forest [RF], Lasso, elastic-net and support vector machine [SVM]) are represented in different colours.A bar crossing the dotted line indicates that a given predictor was an important feature for the model (i.e.selected by ≥80% of models).Bipolar disorder, sex, age and education were important predictors of RD in these tracts.
evening chronotype and MD, AD, and RD, but not FA (García-Marín et al., 2021).A UK Biobank study (published after our analysis) found that a single-item selfrating of late chronotype was not associated with FA or MD in any tract (AD/RD were not examined); however, late chronotype was associated with lower intracellular volume fraction (a different measure of WM integrity) in superior thalamic radiation, uncinate fasciculus, corticospinal tract, cingulate gyrus and superior longitudinal fasciculus (but not 10 other tracts) (Stolicyn et al., 2023).In the context of the wider literature, our findings suggest that chronotype might be more relevant to the neurobiology of mood disorders specifically.
Subjective sleep quality was an important predictor for AD in the CGH (in the overall sample) but for no WM tracts in the mood disorder subgroup.It is noteworthy that none of the previous studies examining WM integrity and subjective sleep quality observed a significant relationship with AD in the CGH (Altendahl et al., 2020;Bresser et al., 2020;Jamieson et al., 2020;Khalsa et al., 2017;Li et al., 2020Li et al., , 2016;;Sexton et al., 2017;Spiegelhalder et al., 2014;Takeuchi et al., 2018;Toschi et al., 2021), whereas only one study observed a significant relationship between MD in the left CGH and subjective sleep quality in a small sample of healthy adults (Khalsa et al., 2017).One possible explanation for the link in our study is that AD has been reported to be associated with the slope of the slow wave during Stage 3 sleep (deep sleep), an indirect proxy of sleep quality (Piantoni et al., 2013).Nonetheless, the F I G U R E 4 Machine learning prediction of fractional anisotropy (FA) as a measure of white matter integrity in the mood disorder subgroup.(a-e) Prediction of FA in five white matter tracts.The predictors of FA for these tracts are depicted along the y-axis and ranked in order of their importance to the model.The x-axis illustrates how many of the models each predictor variable was featured in (range: 0-100%).The four machine learning models used (random forest [RF], Lasso, elastic-net and support vector machine [SVM]) are represented in different colours.A bar crossing the dotted line indicates that a given predictor was an important feature for the model (i.e.selected by ≥80% of models).Depression, age, sex and circadian preference were important for predicting FA in one or more tracts.results from the correlational and ML models suggest that subjective sleep quality and WM integrity are not strongly linked in youth with emerging mental disorders (viewed transdiagnostically).
In the transdiagnostic analysis, a diagnosis of depressive disorder, and to a lesser extent bipolar disorder, were important for predicting WM integrity in several tracts, including FA in the ALIC, CGC, and EC (for depression but not bipolar disorder), and RD in the ACR (for bipolar disorder but not depression).These associations are consistent with previous studies of these diagnostic groups (Ching et al., 2022;Kochunov, Hong, et al., 2022;van Velzen et al., 2020) and make sense in the context of the phenomenology common to these disorders.For example, the corona radiata is composed of projection fibres that relay information about the sleep-wake cycle (e.g.slow EEG activity during slow wave sleep (Jones, 2011)) and is involved in emotional processing and attention (Stave et al., 2017).
Age was an important predictor for various indices of WM integrity, including FA in the CGC, and RD in the PLIC and EC.These findings are consistent with observations of age-related microstructural WM changes (Sullivan & Pfefferbaum, 2006;Tsuchida et al., 2021;Voineskos et al., 2012), especially in adolescence during which WM continues to mature with global increases in FA and decreases in RD (Asato et al., 2010;Simmonds et al., 2014;Tsuchida et al., 2021).Overall, sex was the top-ranked predictor of WM (across a variety of tracts).While we did not examine age by sex interactions, some studies have reported that females appear to reach mature WM levels earlier than males (Asato et al., 2010;Wang et al., 2012).
Our study has several important limitations.First, our analyses were cross-sectional, and we cannot know whether associations between WM integrity and sleep quality or chronotype are causal.Relatedly, we cannot know the degree to which WM measures are stable or reflect developmental or pathological change.Third, we used subjective measures of sleep quality and chronotype rather than objective measures (e.g.polysomnography and sleep-wake midpoint).This distinction is relevant given the phenomenon of subjective-objective sleep discrepancy (Herbert et al., 2017;Rezaie et al., 2018), which may be especially salient in people with depressive symptoms, depressive disorders or a ruminative cognitive-emotional style (Krishnamurthy et al., 2018;O'Callaghan et al., 2021;Tsuchiyama et al., 2003).Fourth, we did not investigate the directional changes in WM integrity but focussed on prediction using ML.Fifth, there were inconsistencies in F I G U R E 5 Machine learning prediction of axial diffusivity (AD) as a measure of white matter integrity in the mood disorder subgroup.(a-c) Prediction of AD in three white matter tracts.The predictors of AD for these tracts are depicted along the y-axis and ranked in order of their importance to the model.The x-axis illustrates how many of the models each predictor variable was featured in (range: 0-100%).The four machine learning models used (random forest [RF], Lasso, elastic-net and support vector machine [SVM]) are represented in different colours.A bar crossing the dotted line indicates that a given predictor was an important feature for the model (i.e.selected by ≥80% of models).Sex, circadian preference and education were important predictors of AD in at least one of the tracts.
the identified WM regions in our study and other studies, which may be because of differences in sample characteristics, sample size, MRI sequences and/or statistical methods (voxel-wise or region-of-interest methods); relatedly, there were differences in how often the four ML algorithms identified specific predictors (indicated by the varying widths of the bars in Figures 1-6), which may be because of the different assumptions of the ML algorithms (e.g.linearity vs nonlinearity).Fifth, time-of-day effects on WM have been reported, revealing widespread increases in FA after day of wakefulness (T.Elvsåshagen et al., 2015;Voldsbekk et al., 2020) and decreases in RD, AD and FA from morning to evening (Jiang et al., 2014), which could affect our findings.Sixth, the DTI scalars we examined (FA, RD, AD, MD) are correlated, and there may be value in future studies using multivariate methods to predict WM integrity in these scalars simultaneously.Finally, we used an arguably conservative threshold that required predictors to be identified in at least 80% of ML models before we deemed them 'important'.While we used this threshold to try to obtain more robust associations, we probably discounted the importance of some factors for predicting WM integrity.

| CONCLUSION
In young people in the early phases of mood disorders, chronotype appears to be an important predictor of WM integrity, particularly in tracts within the corona radiata F I G U R E 6 Machine learning prediction of radial diffusivity (RD) as a measure of white matter integrity in the mood disorder subgroup.(a-e) Prediction of RD in three white matter tracts.The predictors of RD for these tracts are depicted along the y-axis and ranked in order of their importance to the model.The x-axis illustrates how many of the models each predictor variable was featured in (range: 0-100%).The four machine learning models used (random forest [RF], Lasso, elastic-net and support vector machine [SVM]) are represented in different colours.A bar crossing the dotted line indicates that a given predictor was an important feature for the model (i.e.selected by ≥80% of models).Circadian preference, sex, depression and education were important predictors of RD in at least one of the tracts.and corpus callosum, whereas subjective sleep quality was less important.Chronotype and subjective sleep quality were much less predictive of WM integrity in the transdiagnostic sample that also included unaffected controls, suggesting chronotype might have specific relevance to the neurobiology of mood disorders.

Figure 3
Figure3shows the ranked feature selections for models predicting RD in two tracts: ACR and PLIC.Neither chronotype nor sleep quality were important predictors of RD in any tract.Sex, age and level of education were important predictors of RD in ACR (nRMSE = 22.76-23.12%),whereas bipolar disorder was the only important predictor of RD in PLIC (nRMSE = 22.34-22.43%)(TableS5).

F
I G U R E 1 Machine learning prediction of fractional anisotropy (FA) as a measure of white matter in young people with emerging mental disorders.(a-d) Prediction of FA in four different white matter tracts.The predictors of FA for each tract are depicted along the yaxis and ranked in order of their importance to the model.The x-axis illustrates how many of the models each predictor variable was featured in (range: 0-100%).The four machine learning models used (random forest [RF], Lasso, elastic-net and support vector machine [SVM]) are represented in different colours.A bar crossing the dotted line indicates that a given predictor was an important feature for the model (i.e.selected by ≥80% of models).Sex, depression and age were important predictors for the four tracts.F I G U R E 2 Machine learning prediction of axial diffusivity (AD) as a measure of white matter integrity in young people with emerging mental disorders.Prediction of AD in one white matter tract.The predictors of RD for this tract are depicted along the yaxis and ranked in order of their importance to the model.The xaxis illustrates how many of the models each predictor variable was featured in (range: 0-100%).The four machine learning models used (random forest [RF], Lasso, elastic-net and support vector machine[SVM]) are represented in different colours.A bar crossing the dotted line indicates that a given predictor was an important feature for the model (i.e.selected by ≥80% of models).Education, sleep quality and sex were important predictors of AD in this tract.RD, radial diffusivity.