White Matter Microstructure in Pediatric Bipolar Disorder and Disruptive Mood Dysregulation Disorder

Abstract

Objective:

Disruptive mood dysregulation disorder (DMDD) codifies severe, chronic irritability. Youths with bipolar disorder (BD) also present with irritability, but with an episodic course. To date, it is not clear whether aberrant white matter microstructure–a well-replicated finding in BD–can be observed in DMDD and relates to symptoms of irritability.

Method:

We acquired diffusion tensor imaging data from 118 participants (BD = 36, DMDD = 44, healthy volunteers (HV = 38). Images of fractional anisotropy (FA), axial diffusivity (AD), and radial diffusivity (RD) were processed with tract-based spatial statistics controlling for age and sex. The data were also used to train Gaussian process classifiers to predict diagnostic group.

Results:

In BD vs DMDD, FA in the corticospinal tract was reduced. In DMDD vs HV, reductions in FA and AD were confined to the anterior corpus callosum. In BD vs HV, widespread reductions in FA and increased RD were observed. FA in the anterior corpus callosum and corticospinal tract was negatively associated with irritability. The Gaussian process classifier could not discriminate between BD and DMDD, but achieved 68% accuracy in predicting DMDD vs HV and 75% accuracy in predicting BD vs HV.

Conclusion:

Aberrant white matter microstructure was associated with both categorical diagnosis and the dimension of irritability. Alterations in DMDD were regionally discrete and related to reduced AD. In BD, we observed widespread increases in RD, supporting the hypothesis of altered myelination in BD. These findings will contribute to the pathophysiological understanding of DMDD and its differentiation from BD.

Clinical trial registration information:

Studies of Brain Function and Course of Illness in Pediatric Bipolar Disorder; https://clinicaltrials.gov/; NCT00025935; Child & Adolescent Bipolar Disorder Brain Imaging and Treatment Study; https://clinicaltrials.gov/; NCT00006177.

Researchers have debated whether severe, chronic irritability without episodic mania constitutes a developmental phenotype of bipolar disorder (BD).¹ Longitudinal data show that severe, chronic pediatric irritability does not confer risk for BD,² leading to the creation of the DSM-5 diagnosis of disruptive mood dysregulation disorder (DMDD). Further research on pathophysiological differences between BD and DMDD is important to guide the development of targeted interventions for these different, but equally impairing, pediatric phenotypes.

Neurobiological models of BD emphasize aberrant development of white matter (WM), which can be assessed using diffusion tensor imaging (DTI), as a key mechanism.³ WM microstructure is often described in terms of fractional anisotropy (FA), which is positively correlated with the directional coherence, increased axon packing density, and smaller axon diameter of WM.⁴ The 12 DTI studies conducted in youth with bipolar diosrder (n = 8–26) report findings consistent with those in adults,^5,6 that is, reduced FA in the corpus callosum (CC), superior longitudinal fasciculus (SLF), uncinate fasciculus (UF), and cingulum bundle.^7,8 FA is influenced by various factors including myelination, which is reflected in measures of radial diffusivity (RD),⁹ and axonal organization, which is associated with axial diffusivity (AD).¹⁰ In BD, reduced FA is often accompanied by increased RD, suggesting that altered myelination might be an important mechanism in the disorder.¹¹ Moreover, characteristics of WM microstructure might be used by machine learning algorithms to classify individuals as BD or healthy.¹²

In contrast to the extensive literature in BD, no study to date has investigated WM microstructure in DMDD. However, neurobiological models of both BD¹³ and DMDD¹⁴ emphasize the relevance of the prefrontal cortex (PFC) and amygdala. Thus, in DMDD, one might also expect reduced FA in the anterior CC connecting the prefrontal cortices of both hemispheres and in the UF connecting the ventral prefrontal cortex with the amygdala. Interestingly, in depressed patients, a previous study reported associations between elevated levels of irritability and reduced FA in the inferior longitudinal fasciculus (ILF) and inferior fronto-occipital fasciculus (IFOF). The study also found increased RD in the SLF and increased AD in the corticospinal tract (CST) in depressed patients,¹⁵ findings that did not replicate in a large population sample in which no association between irritability and FA was found.¹⁶ Thus, DMDD might be associated with altered WM microstructure in several tracts (eg, UF, CC SLF, ILF, IFOF, and CST), and alterations in some of these tracts might relate to symptoms of irritability.

Here, our goals were to determine whether: (1) aberrancies in WM microstructure are present in youth characterized by chronic irritability (ie, DMDD), compared to healthy volunteers (HV) and BD youth; and (2) abnormalities in WM microstructure are differentially associated with the symptom dimension of irritability in BD and DMDD. In addition, we sought to answer the question whether differences at the group level might be of practical value at the level of the individual. To address this third goal, we used machine learning algorithms to make individual predictions regarding diagnosis. Given inconsistencies in the literature and the considerable size of our sample, we used a whole-brain approach for this first DTI study comparing WM microstructure and its association with irritability in BD and DMDD.

METHOD

Participants

We collected data on 118 youths (aged 11–21 years) diagnosed with either DMDD (n = 44) or BD (n = 36; n = 22 BD I, n = 14 BD II) and healthy volunteers (HV, n = 38). The three subgroups were comparable regarding sex, age, IQ and socioeconomic status. BD youth were more likely than DMDD youth to receive mood stabilizing medication (antiepileptics, lithium) and benzodiazepines, reflected in a higher medication load score¹⁷ (Table 1), which was covaried in the DTI analyses. Participants over age 18 and parents of minor participants gave written informed consent after receiving a complete description of the study; minors gave written assent. Procedures were approved by the Institutional Review Board of the National Institute of Mental Health.

TABLE 1.

Demographic and Clinical Characterization

	BD (n = 36)	DMDD (n = 44)	HV (n = 38)	Statisticsa	p
Demographics
Sex (female/male)	(17/19)	(15/29)	(21/17)	χ2(2) = 3.74	.15
Age, mean (SD), y	17.2 (3.2)	15.9 (2.6)	16.0 (2.7)	F2,115 = 2.62	.08
Intelligence score, mean (SD)	107.6 (12.4)	109.3 (11.9)	111.6 (11.9)	F2,115 = 0.92	.40
SES Hollingsheadb, mean (SD)	54.5 (32.7)	45.4 (24.9)	41.2 (20.1)	F2,96 = 2.13	.12
Mood Symptoms
YMRS, mean (SD)c	6.1 (2.12)	n/c	n/c	n/a	n/a
CDRSb, mean (SD)	24.6 (11.2)	23.9 (9.0)	n/c	t51 = 0.26	.80
SIGH-SADb, mean (SD)	6.7 (7.0)	4.4 (3.9)	n/c	t21 = 0.87	.39
ARI, mean (SD)d	4.38 (3.07)	4.89 (2.50)	0.82 (1.32)	F2,115 = 33.39	< .001
Euthymice, n (%)	26 (72%)	n/a	n/a	n/a	n/a
Hypomanicf, n (%)	5 (14%)	n/a	n/a	n/a	n/a
Depressedg, n (%)	4 (11%)	3 (7%)	n/a	χ2(1) = 2.89	.24
Mixed stateh, n (%)	1 (3%)	n/a	n/a	n/a	n/a
Current K-SADS diagnoses
ADHD, n (%)	19 (53%)	28 (64%)	n/a	χ2(1) = 0.96	.33
ODD or CD, n (%)	3 (8%)	25 (57%)	n/a	χ2(1) = 20.46	< .001
Anxiety disorders, n (%)	18 (50%)	21 (48%)	n/a	χ2(1) = 0.04	.84
GAD, n (%)	13 (36%)	17 (39%)	n/a	χ2(1) = 0.05	.82
SAD, n (%)	4 (11%)	4 (9%)	n/a	χ2(1) = 0.09	.76
SoPh, n (%)	9 (25%)	9 (20%)	n/a	χ2(1) = 0.24	.63
Psychotropic medication
Medication load, mean (SD)	3.5 (2.84)	2.0 (1.92)	n/a	t79 = 2.77	.007
Total, n taking 0 / 1 / 2 / 3 / ≥ 4i	4 / 7 / 7 / 10 / 9	10 / 15 / 7 / 9 / 3	n/a	χ2(4) = 7.76	.10
Antidepressants, n taking 0 / 1 / 2	27 / 8 / 1	33 / 10 / 1	n/a	χ2(1) = 0.22	.99
Anti-epileptics, n taking 0 / 1 / 2	21 / 12 / 3	37 / 7 / 0	n/a	χ2(1) = 8.01	.02
Lithium, n taking 0 / 1	23 / 13	41 / 3	n/a	χ2(1) = 10.62	.001
Antipsychotics, n taking 0 / 1 / 2	16 / 17 / 3	29 / 14 / 1	n/a	χ2(1) = 4.29	.12
Tranquilizer, n taking 0 / 1 / 2	29 / 5 / 2	43 / 1 / 0	n/a	χ2(1) = 6.66	.04
Stimulants, n taking 0 / 1 / 2	17 / 17 / 2	16 / 26 / 2	n/a	χ2(1) = 1.13	.64

Note: ADHD = attention-deficit/hyperactivity disorder; BD = bipolar disorder; CD = conduct disorder; CDRS = Children’s Depression Rating Scale; DMDD = disruptive mood dysregulation disorder; GAD = generalized anxiety disorder; HV = healthy volunteers; K-SADS = Kiddie Schedule for Affective Disorders and Schizophrenia; n/a = not applicable, n/c = not collected; ODD = oppositional defiant disorder; SAD = separation anxiety disorder; SIGH-SAD = Structured Interview Guide for the Hamilton Depression Rating Scale; SoPh = social phobia; YMRS = Young Mania Rating Scale.

^aDifferent degrees of freedom reflect missing data.

^bData unavailable: CDRS score: 2 DMDD; SES Hollingshead Four-Factor Index of Socioeconomic Status: 5 BD, 8 DMDD, 6 HV; SIGH-SAD: 1 BD, 1 DMDD.

^cThe YMRS was not administered in the DMDD sample because a (hypo)manic episode is exclusionary for this diagnosis. Although DMDD youth might have shown irritable mood or disruptive behavior during the 48 hours prior to the assessment with the YMRS, the symptoms would not have been indicative of an acute (hypo)manic episode, given that the K-SADS assessment had shown that none of these children had a history of such episodes. Thus, the irritability would be a manifestation of the child’s baseline irritable mood, rather than irritable mood occurring in the context of a (hypo)manic episode.

^dPost hoc tests showed significant differences between BD and HV (p < .001), DMDD and HV (p < .001), but not BD and DMDD (p = .612). Despite the fact that the clinical presentation of irritability differs between DMDD and bipolar disorder, the symptom is important in both disorders. Although chronic irritability is characteristic of DMDD, BD youth may also have irritability while euthymic (trait irritability), and irritability may increase during affective episodes (state-related irritability). Thus, what differentiates DMDD from bipolar disorder is not the severity of the irritability, but the fact that only youth with bipolar disorder have a history of manic or hypomanic episodes.

^eEuthymia is defined by CDRS score ≤40 or SIGH-SAD score ≤7 and YMRS score ≤12.

^fHypomania is defined by CDRS score ≤40 or SIGH-SAD score ≤7 and YMRS score between 12 and 18.

^gDepression is defined by CDRS score >40 or SIGH-SAD score >7 and YMRS score ≤12.

^hMixed state is defined by SIGH-SAD score >20 and YMRS score >12.

ⁱThis indicates the number of participants who take 0, 1, 2, 3, 4 or more different psychotropic medications. To illustrate this further, in the BD sample, 4 patients were medication free, 7 were taking 1 psychotropic medication, 7 were taking 2 psychotropic medications, 10 were taking 3 psychotropic medications, and 9 were taking 4 or more psychotropic medications. The same scheme was used to code the consumption of the specific types of medication.

Clinical Assessment

Master’s degree–level and doctoral-level clinicians administered the Kiddie Schedule for Affective Disorders and Schizophrenia (K-SADS) with the DMDD Supplement 1, available online.¹⁸ The primary diagnosis in patients was confirmed by a senior clinical psychologist (M.A.B.) or psychiatrist (E.L., K.T.). Exclusion criteria for all groups were neurological disorders, autism spectrum disorders, substance use within the last 2 months, conditions for which MRI is contraindicated, and full-scale IQ of <70.

The BD participants’ manic symptoms were measured by the Young Mania Rating Scale (YMRS)¹⁹ using information from both child and parent. Depressive symptoms were measured in BD and DMDD participants with the Children’s Depression Rating Scale (CDRS)²⁰ in patients less than 18 years of age (n = 55) and the Structured Interview Guide for the Hamilton Depression Rating Scale, Seasonal Affective Disorders Version (SIGH-SAD)²¹ in patients more than 18 years of age (n = 25). YMRS and CDRS/ SIGH-SAD were administered by master’s degree–leveland doctoral-level clinicians, with excellent interrater reliability (κ > 0.9 for all ratings) within 1 week of scanning. The YMRS was collected only in BD youth and the CDRS/SIGH collected in the two patient groups. Irritability was assessed in the clinical groups with the Affective Reactivity Index (ARI), a 7-item measure completed by both parent and child²²; we used the mean of the parent and child ratings, which were moderately correlated (r_S = 0.64, p < .001), for analyses.

Acquisition and Processing of the DTI Data

We acquired 75 images with diffusion gradients (6: b = 300 seconds/mm², 69: b = 1,000 seconds/mm²) and 5 unweighted images along with a T2-weighted image with fat suppression on a 3.0 Tesla GE scanner. Pre-processing of the DTI data was done with TORTOISE version 1.4.0²³ and comprised motion, eddy current distortion, and echo-planar image distortion corrections. Next, a nonlinear diffusion tensor model was fitted to the corrected data using the Robust Estimation of Tensors by Outlier Rejection (RESTORE) algorithm,²⁴ thereby obtaining FA and eigenvalue maps from which RD and AD maps were calculated. Details on the acquisition protocol and the preprocessing are provided in Supplement 1, available online.

The FA data were processed using the tract-based spatial statistics (TBSS) pipeline,²⁵ which includes the registration of individual FA volumes to the FMRIB58 template and the creation of a mean FA image. The mean FA image was thinned to the mean FA skeleton, which represents the centers of all tracts common to the group. The FA skeleton was thresholded at 0.2 to exclude areas in the periphery of the tracts known for high intersubject variability. The transformation matrix obtained by registering individual FA images to the template was applied to individual RD and AD images.

Goal 1: WM Microstructure in DMDD vs BD vs HV

We compared FA across the three groups by calculating an analysis of covariance (ANCOVA) with age and sex as nuisance variables. To determine the groups that were driving the effects in this omnibus test, we calculated 6 post hoc tests with age and sex as nuisance variables (BD vs DMDD, BD vs HV, DMDD vs HV), using the results of the ANCOVA (thresholded at p_{FWE-corrected} < .05) as mask. Furthermore, to determine whether FA differences were differentially driven by RD and LD, we repeated the post hoc tests on the RD and LD maps. For each statistical test, we performed 5,000 permutations within Randomise²⁶ and used threshold-free cluster enhancement to correct for multiple comparisons.²⁷ For all tests, we applied a familywise error correction of p < .05. In the case of the initial omnibus test, this threshold was applied for all voxels in the FA-skeleton. For the 18 post hoc tests, this threshold was applied for the 5,997 significant voxels identified in the omnibus F test. We report coordinates, p value (p_min) of the peak voxel and cluster size for significant clusters. Results were back-projected into native space to ensure that significant voxels fell within WM pathways. We confirmed that results remained when medication load, sex, and oppositional defiant disorder (ODD) were added as additional covariates. Furthermore, we calculated three additional models determining the main and interaction effects of age, sex, and attention-deficit/hyperactivity disorder (ADHD).

Goal 2: Association Between WM Microstructure and Irritability

We determined whether findings were related to the symptom dimension of irritability by extracting FA values from the peak voxels of the significant clusters from the post hoc tests and mean ARI for each participant and calculating Spearman correlation coefficients. We used a false discovery rate of q = 0.05 to correct for multiple comparisons.

Goal 3: Prediction of Diagnosis Based on WM Microstructure

To complement the univariate analysis, we conducted multivariate analyses using the Pattern Recognition for Neuroimaging Toolbox (PRoNTo).²⁸ For each diffusion measure (FA, RD, AD) and all three features concatenated, we trained a Gaussian Process Classifier (GPC) on the kernel matrix, which uses probabilistic modelling to estimate the likelihood that a test sample belongs to a particular class.²⁹ We used the mean FA-skeleton mask at the first-level for all classifications. Performance of the GPC was assessed using the leave-one-out cross-validation method to estimate the balanced accuracy, sensitivity, and specificity. Mean-centering was applied during all classifications. In the combination model, we also normalized the data vectors by dividing them by their Euclidian norm. Finally, we retrained the model with randomized labels and 1000 permutations. To correct for multiple comparisons (4 models [FA, RD, AD, combination] for 3 pairings [BD vs HV, DMDD vs HV, BD vs DMDD]), we used a false discovery rate of q = 0.05 and consider accuracies only with an uncorrected significance level of p ≤ 0.0166 to be positive results.

Statistical Analysis of Demographic Data

All statistical analyses of clinical data were performed using SPSS version 24. For demographic data conforming to the assumptions of parametric analysis, we used either analyses of variance or Student t tests. For nominal demographic data, χ² tests were computed. Results were considered significant at p < 0.05.

RESULTS

Goal 1: WM Microstructure in DMDD vs BD vs HV

Fractional Anisotropy.

The omnibus ANCOVA showed a main effect of diagnosis in a large cluster centered at the anterior CC (x = −10, y = 25, z = 15, 5997 voxels, p_min = 0.0008) that extended along the CC and included fibers of the ATR, CST, IFOF, SLF, and UF. Post hoc comparison showed lower FA values in BD vs DMDD in the superior corona radiata/ CST (x = −28, y = −17, z = 19, 719 voxels, p_min = 0.004), splenium of the CC (x = −2, y = −36, z = 20, 352 voxels, p_min = 0.013), and SLF (x = −33, y = −43, z = 22, 128 voxels, p_min = 0.030). Post hoc t tests specifically indicated decreased FA in the genu of the CC in DMDD vs HV (x = −14, y = 36, z = 5, 2,177 voxels, p_min = 0.0004). Post hoc tests also revealed reduced FA in BD vs HV in most voxels of the cluster that was identified in the omnibus test, but with the peak voxel being located in the CST (x = −13, y = −12, z = −12, 5,649 voxels, p_min = 0.0002) (Figure 1, Table S1, available online).

FIGURE 1. Visualization of Tract-Based Spatial Statistics (TBSS) Analysis.

Note: Mean maps of fractional anisotropy (grayscale) and mean fractional anisotropy skeleton (grayscale). Yellow colored voxels represent results of the omnibus F test. (A) Red-colored voxels signify decreased fractional anisotropy (p_{FWE-corrected} < 0.05). (B) Green-colored voxels show areas of reduced axial diffusivity. (C) Blue-colored voxels signify increased radial diffusivity (p_{FWE-corrected} < 0.05). We used the tbss-fill script to improve visualization of the results. NS = not significant. Please note color figures are available online.

Radial Diffusivity.

To explore factors underlying microstructural alterations, we repeated the post hoc tests on the RD skeleton, as RD has been linked to the degree of myelination.⁹ This analysis showed increased RD in the CST in BD vs DMDD (x = −27, y = −18, z = 20, 915 voxels, p_min = 0.004). In BD vs HV, RD was increased in a large cluster, the peak voxel of which was located in the anterior CC (x = −14, y = −37, z = 2, 5,503 voxels, p_min = 0.002). No differences in RD were observed between DMDD and HV (Figure 1, Table S1, available online).

Axial Diffusivity.

Post hoc tests showed reduced AD in BD vs DMDD in the posterior limb of the internal capsule/ CST (x = −25, y = −15, z = 17, 330 voxels, p_min = 0.026). AD was also lower in the genu of the CC in DMDD vs HV (x = −14, y = −25, z = 20, 1,532 voxels, p_min = 0.002). Finally, a post hoc t test indicated reduced AD in BD vs HV in the superior corona radiata/ CST (x = −28, y = −19, z = 20, 2,596 voxels, p_min = 0.0008) (Figure 1, Table S1, available online).

Results of post hoc t tests for the entire FA skeleton are provided in Figure S1, available online. When medication load, ODD and socioeconomic status were added as nuisance variables, all reported clusters remained significant except for 2 clusters (1 in the splenium of the CC and 1 in the SLF) where BD had originally shown lower FA than DMDD (Table S2, available online).

Goal 2: Association Between WM Microstructure and Irritability

Irritability was negatively associated with FA values extracted from the genu of the CC (across groups: r_S = −0.35, p = .000013; BD: r_S = −0.39, p = .02; DMDD: r_S = −0.31, p = .04; HV: r_S = 0.02, p = .90) and the CST (across groups: r_S = −0.31, p = .001; BD: r_S = −0.32, p = .06; DMDD: r_S = −0.29, p = .06; HV: r_S = .03, p = .87), but not the other clusters (all p values >.50). Furthermore, FA in the splenium of the CC was positively associated with irritability in DMDD (r_S = 0.55, p = .0001), but not in BD (r_S = 0.08, p = .63) or HV (r_S = −.04, p = .83) (Figure 2).

FIGURE 2. Associations Between Irritability and Fractional Anisotropy (FA).

Note: (A) Association between irritability and FA in the cluster in the anterior corpus callosum, where we observed FA differences between DMDD and HV as well as BD and HV. (B) Association between irritability and FA in the cluster in the corticospinal tract, where BD showed lower FA compared to HV and DMDD. (C) Association between irritability and FA in the cluster in the posterior corpus callosum, where BD showed lower FA compared to HV and DMDD. ARI = Affective Reactivity Index; BD = bipolar disorder participants; DMDD = disruptive mood dysregulation disorder participants; HV = healthy volunteers. Please note color figures are available online.

Goal 3: Prediction of Diagnosis Based on WM Microstructure

Based on FA, the GPC could classify individuals as BD or HV with a sensitivity of 66.67% (p = 0.001), a specificity of 84.21% (p = 0.001), and a balanced accuracy of 75.44% (p = 0.001). Most positive discriminant values were observed in the periphery of the FA skeleton, mainly in the lateral prefrontal and temporal areas. If all other voxels in the brain remained the same, higher FA values in these regions would increase the probability of being classified as BD. Most negative weights were found in the IFOF, SLF, CC, and medial prefrontal WM. If all other voxels in the brain remained the same, individuals with higher FA values in these regions would more likely be classified as HV (Figure 3A). Importantly, however, the combination of all weights defines the model.

FIGURE 3. Depiction of the Predictions Across all Folds, Receiver Operating Characteristic (ROC) Curves, and Model Weights Overlaid on the Mean Fractional Anisotropy (FA) Map.

Note: Depiction of the predictions across all folds, receiver operator curves (ROC) curves, and model weights overlaid on the mean FA map for the classification of the following: (A) BD vs HV based on information about FA; (B) DMDD vs HV based on information about AD; and (C) BD vs DMDD based on information about AD obtained with the Gaussian process classifier algorithm, although this result was not significant when correcting for multiple comparisons. Color scale indicates the importance of the voxel location for the classification. High values in regions with positive weights (red) increase the likelihood of being classified as belonging to the group marked with the red circle in the prediction values graph, whereas high values in regions with negative weights (blue) will more likely result in being classified as belonging to the other group. AD = axial diffusivity; BD = bipolar disorder participants; DMDD = disruptive mood dysregulation disorder participants; FA = fractional anisotropy; HV = healthy volunteers. Please note color figures are available online.

Furthermore, the GPC was able to classify participants as DMDD or HV with a sensitivity of 68.18% (p = 0.016), a specificity of 68.42% (p = 0.009), and a balanced accuracy of 68.30% (p = 0.001) using information about AD. Highest positive discriminant values related to an increased likelihood of being classified as DMDD were observed in the fornix, posterior cingulate bundle, CST, and WM near the hippocampus and amygdala. Highest negative discriminant values were assigned to voxels in the ATR near the thalamus, anterior cingulate bundle, anterior CC, IFOF, and ILF. High AD values in these areas increase the likelihood of being classified as HV given that AD values in all other voxels in the brain remain the same (Figure 3B).

Finally, the GPC was unable to differentiate BD from DMDD. It performed best using information about AD only, achieving a balanced accuracy of 60.10% (p = 0.045), a sensitivity of 61.11% (p = 0.025), and a specificity of 59.09% (p = 0.059). However, these findings are nonsignificant when corrected for multiple comparisons (Figure 3C).

Prediction values, receiver operating characteristic (ROC) curves, and averaged model weights are depicted in Figure 3. Details about the performance of the GPCs for other DTI measures and the combination of different measures are presented in Table 2. Adding irritability as a covariate or using a support vector machine algorithm did not improve accuracies (Tables S3, S4, available online).

TABLE 2.

Model Performance for Different Classifications and Models

Models	BD vs. DMDD (n = 81)	BD vs. HV (n = 75)	DMDD vs. HV (n = 82)
FA	46.23 (30.56, 61.90)	75.44a (66.67a, 84.21a)	59.76 (65.91a, 52.63)
RD	58.33 (50.00, 66.67)	69.31a (67.57a, 71.05a)	68.48a (71.05a, 65.91)
AD	60.10 (61.11, 59.09)	62.28 (66.67a, 57.89)	68.30a (68.18a, 68.42a)
Combination	57.54 (38.89, 76.19a)	60.89 (63.89a, 57.89)	68.29a (79.55a, 55.26)

Note: Boldface type indicates the models the results of which are described in more detail. Model performance for different classifications and models displayed as balanced accuracy values and the sensitivity and specificity (in parentheses). These values represent averages over the 10 folds. AD = axial diffusivity; BD = bipolar disorder; FA = fractional anisotropy; DMDD = disruptive mood dysregulation disorder; HV = healthy volunteers; RD = radial diffusivity.

^aAccuracies that met the threshold of p < .01667 after being tested with 1,000 permutations.

Exploratory Analyses

Both patient groups display a high rate of comorbid ADHD, which has been related to aberrant FA and RD.³⁰ To test whether ADHD symptoms were driving our findings, we added ADHD as an explanatory variable to our univariate analyses. There was no main effect of ADHD and no interactions with BD or DMDD. To test whether findings reflect differential development of WM tracts,³¹ we modeled the diagnosis-by-age interaction in addition to the main effects, which yielded no significant results. A third analysis confirmed the absence of differential sex effects. Main effects of age and sex are reported in Supplement 1, available online. Detailed analyses of medication effects, which yielded no significant findings, are also presented in Supplement 1, available online. Finally, FA, RD, and LD were comparable between individuals with BD I and BD II in clusters identified during the group comparisons reported under Goal 1 (F_12,22 = 0.67, p = .759).

DISCUSSION

The present study aimed (1) to investigate aberrancies in WM microstructure in DMDD compared to BD and HV youth; (2) to determine whether abnormalities in WM microstructure identified in the first analysis relate to symptoms of irritability; and (3) to test whether diagnosis can be predicted based on WM microstructure. Regarding the first goal, we observed lower FA in the CST in BD vs DMDD; lower FA values in the anterior CC that were driven by reduced AD values in DMDD vs HV; and widespread reductions in FA that were driven mainly by increased RD in BD vs HV. Regarding our second goal, we found that FA values in the anterior CC and the CST were negatively associated with irritability in both patient groups and across all groups, whereas we observed a positive association between FA in the posterior CC and irritability in DMDD only. Regarding our third goal, the GPC was unable to differentiate between DMDD and BD groups, but did predict membership of DMDD and HV with an accuracy of 68% based on AD data, and between BD and HV with an accuracy of 75% based on FA data.

Goal 1: WM Microstructure in DMDD vs BD and vs HV

BD vs DMDD.

In our univariate analyses, BD and DMDD differed in their WM microstructure in the CST, a tract previously implicated in BD.⁸ CST microstructure has been associated with processing speed.³² Thus, reduced FA in the CST might contribute to slower reaction times³³ and impaired motor inhibition³⁴ previously reported in BD compared to DMDD. Unfortunately, we did not collect data on processing speed in this sample, but future studies might probe this hypothesis.

Interestingly, BD and DMDD both showed reduced FA values in the anterior CC, which has previously been associated with working memory performance.³⁵ In BD, reduced FA in the anterior CC seems to be driven by increased RD, which is known to be a sensitive marker of myelination⁹; this finding is consistent with the hypothesis that aberrant myelination might be a key mechanism in BD.¹¹ In DMDD, lower FA in the anterior CC was driven by reduced AD, which might be more sensitive to the level of axonal organization. In sum, our findings show that differential mechanisms underlying BD and DMDD include differences in WM microstructure that might relate to abnormalities in the cognitive domain (eg, processing speed, working memory).

DMDD vs HV.

Reduced FA values in DMDD vs HV were confined to the genu of the CC. This finding could not be attributed to effects of psychotropic medication, ODD, or ADHD, which represent common confounding factors in DMDD research. The forward-curving fibers of the anterior CC form the forceps minor and interconnect medial prefrontal cortices and anterior cingulate cortices (ACC), regions that have been implicated in cognitive control³⁶ and emotion regulation.³⁷ In this context, it is of interest that the left hemisphere is primarily associated with the motivation to act and to approach positive/ rewarding stimuli, whereas the right hemisphere may mediate avoiding dangers and inhibiting actions that may lead to painful consequences. Abnormal microstructure of the corpus callosum might disturb the functional balance of both hemispheres and thereby contribute to emotional dysregulation. Of note, aberrant WM microstructure in the genu of the CC has been associated with an early onset of major depression.³⁸ This is especially interesting, as this cluster was positively associated with the level of irritability, and we know from previous studies that irritability confers risk for major depression.² Future studies should investigate whether altered development of the anterior CC is one of the mechanisms that underlie this association.

BD vs HV.

Observed widespread reductions of FA in BD are consistent with previous reports of (1) reduced FA in prefrontal–limbic tracts (ATR, UF) presumably associated with emotional instability³⁹; (2) aberrancies in association tracts (SLF, ILF) related to cognitive impairments⁶; and (3) altered WM microstructure in the corpus callosum, possibly associated with both mood instability and cognitive dysfunction.⁴⁰ Consistent with previous studies, we also found that alterations in FA in BD youth were driven mainly by increased RD, which serves as an indirect estimate of the axonal myelination level.⁹ Thus, our results support the hypothesis of aberrant myelination in BD,¹¹ which proposes that a deficit/dysfunction in oligodendrocytes compromises trophic axonal support, resulting in suboptimal conduction of action potentials. We did not observe an age effect on WM microstructure in BD compared to HV. However, a recent longitudinal study reported the absence of normative age-related increases in the FA of the UF in youth with bipolar disorder.³¹ Future longitudinal studies are needed to document the developmental course of WM microstructure in BD youth.

Goal 2: Association Between WM Microstructure and Irritability

FA values in the CST and the genu of the CC were negatively associated with levels of irritability in DMDD and BD. Particularly, the association between irritability and the CST, a major motor pathway, was a surprising finding that warrants replication. However, the CST has also been implicated in information processing.³² Based on this report, we speculate that reduced information processing speed increases the difficulty of tasks, which might then elicit frustration more easily. The anterior CC has also been associated with cognitive functions such as working memory.³⁵ Working memory deficits are known to have an impact on instrumental learning,⁴¹ which has been proposed to play a crucial role in irritability.¹⁴ To summarize, we hypothesize that impaired cognitive functioning (eg, processing speed, working memory) mediates the association between reduced anisotropy in the CST and the anterior CC and irritability. Future studies should assess WM microstructure and these cognitive functions in BD and DMDD to probe this hypothesis.

Interestingly, higher FA values in the posterior CC were positively related to irritability in DMDD only. However, future studies are necessary to investigate if this differential association relates to specific functions.

Goal 3: Prediction of Diagnosis Based on WM Microstructure

BD vs DMDD.

Despite significant group differences in the CST revealed by the univariate analysis, the Gaussian process classifier was unable to discriminate between BD and DMDD at the individual level. Future machine learning studies might yield improved results using larger patient samples, different classifiers, or restricting analyses to areas showing highest weights in this analysis, for example, anterior thalamic radiation (BD>DMDD), fornix, posterior cingulum bundle, and WM near the amygdala–hippocampus complex (DMDD>BD).

DMDD vs HV.

Notably, the machine learning algorithm was able to discriminate between DMDD and HV with an accuracy of 68%. Furthermore, the multivariate approach indicated that, aside from aberrancies in the anterior CC, which were also identified using a univariate approach, reduced AD in the ATR near the thalamus, and increased AD in the fornix, the posterior cingulum bundle, and WM near amygdala and hippocampus are typical for DMDD compared to HVs. The ATR interconnects the thalamus, striatum, and prefrontal cortex and was previously associated with reversal learning deficits,⁴² and abnormal activity of the amygdala has been reported during aberrant threat processing.^43,44 Thus, aberrant WM microstructure might contribute to altered reward and threat processing in DMDD.¹⁴ Moreover, the ability of machine learning classifiers to make inferences at the individual subject level could be used in future studies to explore whether reduced AD can predict the course of DMDD or treatment outcome.

BD vs HV.

We replicated a previous report that machine learning algorithms can classify BD and HV based on diffusion measures with an accuracy of 75% or more.¹² Clearly, studies of these algorithms are in early phases and will require considerable effort before these methods can be applied clinically. One would hope that such future, improved methods would generate considerably higher levels of accuracy. Future studies also might explore whether machine learning algorithms using FA can predict which first-degree relatives of BD patients will develop the illness.⁴⁵

Because we did not apply a cardiac-gated acquisition protocol, an impact of pulsatile motion artifacts on our results cannot be excluded, although the use of robust tensor fitting renders this possibility very unlikely. Moreover, it is unlikely that our results were caused by differential subject motion,⁴⁶ as volumes with excessive motion were reacquired. In addition, most patients were taking psychotropic medication and reported comorbid ADHD, ODD, or/and anxiety disorders. Additional analyses indicate that effects reported here do not result from these confounding factors, but given the sample size, small effects might have been undetected. However, the use of psychotropic medication and high rates of comorbidities are common in these patient groups^47,48 and thus speak for the generalizability of the results. It will be important to replicate these initial findings, and particularly the ability of the GPC to distinguish the groups (BD vs HV, DMDD vs HV), in an independent cohort ideally also using more robust validation methods such as k-fold cross-validation or bootstrapping. Furthermore, it is possible that a tractographic approach would be more sensitive, particularly to the differences between patient groups. Finally, we would like to point out that manic symptoms were not assessed in all groups, which limited our ability to discern the effets of manic and irritable symptoms on WM microstructure.

Our results support a role of altered WM microstructure in pathophysiological models of both BD and DMDD but, consistent with prior work, indicate different underlying mechanisms, particularly in tracts associated with cognitive functions, such as the anterior CC and CST. Notably, FA in the anterior CC and the CST was negatively associated with irritability. In DMDD, alterations in FA were most pronounced in the anterior CC and were driven by altered AD, presumably related to the degree of axonal organization. In BD, we replicated findings of widespread reductions in FA and increases in RD, supporting the hypothesis of altered myelination as an underlying mechanism of the disorder.