Introduction
In recent years, psychiatry researchers have shifted to a greater use of dimensional symptom-based methods to identify traits across the full phenotypic range of populations (Ball et al., Reference Ball, Malpas, Genc, Efron, Sciberras, Anderson and Silk2018; Insel, Reference Insel2014; Woo, Chang, Lindquist, & Wager, Reference Woo, Chang, Lindquist and Wager2017), in accordance with the Research Domain Criteria framework (Shephard et al., Reference Shephard, Stern, van den Heuvel, Costa, Batistuzzo, Godoy and Miguel2021). These approaches are conducive to the discovery of dimensional biomarkers for psychiatric disorders, and beneficial for translating neuroimaging into clinical and precision medicine (Gillan, Fineberg, & Robbins, Reference Gillan, Fineberg and Robbins2017; Woo et al., Reference Woo, Chang, Lindquist and Wager2017). This approach has also entered the field of obsessive–compulsive disorder (OCD) neuroimaging research (Gillan et al., Reference Gillan, Kalanthroff, Evans, Weingarden, Jacoby, Gershkovich and Simpson2020; Hettwer et al., Reference Hettwer, Larivière, Park, van den Heuvel, Schmaal, Andreassen and Valk2022; Jacobs et al., Reference Jacobs, Voineskos, Hawco, Stefanik, Forde, Dickie and Ameis2021; Robbins, Gillan, Smith, de Wit, & Ersche, Reference Robbins, Gillan, Smith, de Wit and Ersche2012).
OCD is a common psychiatric disorder in the population with a high lifetime prevalence (2–3%; Stein et al., Reference Stein, Costa, Lochner, Miguel, Reddy, Shavitt and Simpson2019). Previous research has linked specific underlying neural features with different core symptoms of OCD, especially compulsion (Gillan et al., Reference Gillan, Kalanthroff, Evans, Weingarden, Jacoby, Gershkovich and Simpson2020; Voon et al., Reference Voon, Derbyshire, Rück, Irvine, Worbe, Enander and Bullmore2015; Zhu et al., Reference Zhu, Fu, Chen, Yu, Zhang, Zhang and Wang2022). However, OCD patients are often comorbid with other psychiatric disorders (Kaur, Garg, & Arora, Reference Kaur, Garg and Arora2018; Sharma et al., Reference Sharma, Sharma, Balachander, Lin, Manohar, Khanna and Stewart2021), and high levels of comorbidity are nearly unavoidable even after careful diagnosis (Ruscio, Stein, Chiu, & Kessler, Reference Ruscio, Stein, Chiu and Kessler2010). As a result, OCD patient groups often include significant levels of anxiety, depression, and other psychiatric symptoms. In addition, the presence of compulsive thoughts and behaviors falls along a continuum from the general population to OCD patients and is not exclusive to patients themselves. In order to account for this variability in symptoms and diagnostic categories, it is helpful to take a cross-symptom dimensional approach utilizing methods such as principal components analysis (PCA) to identify key dimensions across all subjects (Gillan et al., Reference Gillan, Kalanthroff, Evans, Weingarden, Jacoby, Gershkovich and Simpson2020; Sharma et al., Reference Sharma, Sharma, Balachander, Lin, Manohar, Khanna and Stewart2021; Voon et al., Reference Voon, Derbyshire, Rück, Irvine, Worbe, Enander and Bullmore2015). Furthermore, such a dimensional approach can be helpful in identifying the relationships between obsessive–compulsive symptoms, OCD diagnosis, and the neuropathological features of OCD.
Many previous studies have found that OCD is subserved by changes within the corticostriatal system (Saxena, Brody, Schwartz, & Baxter, Reference Saxena, Brody, Schwartz and Baxter1998), and thus, many neuroimaging studies have focused on functional and structural connections with the striatum (Park et al., Reference Park, Kim, Kwak, Cho, Lee, Moon and Kwon2022; Xu et al., Reference Xu, Hou, He, Ruan, Chen, Wei and Peng2022). The striatum has been shown to be a hub region that underlies abnormal reward processing and instrumental learning in OCD (Robbins, Vaghi, & Banca, Reference Robbins, Vaghi and Banca2019). OCD has also been associated with neurotransmitter abnormalities (glutamate, N-acetylaspartate) in the striatum (Brennan, Rauch, Jensen, & Pope, Reference Brennan, Rauch, Jensen and Pope2013). Recent research has benefitted from taking a neurocircuit-based approach relating specific corticostriatal circuits to specific symptoms of OCD (Shephard et al., Reference Shephard, Stern, van den Heuvel, Costa, Batistuzzo, Godoy and Miguel2021; van den Heuvel et al., Reference van den Heuvel, van Wingen, Soriano-Mas, Alonso, Chamberlain, Nakamae and Veltman2016). Typically in these studies, the striatum was first selected as a seed region, then functional or structural connectivity was calculated between the striatum and other brain regions, and then finally these connectivity patterns were related to core symptoms of OCD (Park et al., Reference Park, Kim, Kwak, Cho, Lee, Moon and Kwon2022) and biomarkers of OCD (Vaghi et al., Reference Vaghi, Vértes, Kitzbichler, Apergis-Schoute, van der Flier, Fineberg and Robbins2017). This network approach was also made use of by researchers who found that OCD is characterized by an imbalance between corticostriatal systems involved in goal-directed instrumental behavior and habitual behavior (Zhang et al., Reference Zhang, Fan, Zhu, Tan, Chen, Gao and Xiao2017). Even in healthy participants, compulsivity is linked to reduced development of goal-directed control and frontostriatal functional connectivity (Vaghi et al., Reference Vaghi, Moutoussis, Váša, Kievit, Hauser, Vértes and Vanes2020). However, examining corticostriatal systems via single imaging modalities is insufficient for a full understanding of OCD. Research has identified the potential involvement of other systems in OCD, including the inferior parietal cortex (Boedhoe et al., Reference Boedhoe, Schmaal, Abe, Alonso, Ameis, Anticevic and van den Heuvel2018), posterior cingulate cortex, and cerebellum (Sha et al., Reference Sha, Edmiston, Versace, Fournier, Graur, Greenberg and Phillips2020; Zhou et al., Reference Zhou, Xu, Ping, Zhang, Chen, Shen and Cheng2018). Previous OCD studies have not utilized multiple modalities and have not included a broad range of demographic variables (Picó-Pérez et al., Reference Picó-Pérez, Moreira, de Melo Ferreira, Radua, Mataix-Cols, Sousa and Morgado2020). Comprehensive research that is not limited to case–control designs and single neuroimage modalities is needed to identify associations between neuroimaging features and clinical phenotypes and clarify the heterogeneity of the results.
Gray matter volume, fiber tract integrity, and spontaneous functional features are all local brain indicators that are commonly examined in neuroimaging research, with each method having the potential to reveal different pathological features (Picó-Pérez et al., Reference Picó-Pérez, Moreira, de Melo Ferreira, Radua, Mataix-Cols, Sousa and Morgado2020). Gray matter volume can be measured using voxel-based morphometry (VBM) to analyze high-resolution anatomical scans. Gray matter volume is thought to reflect gray matter features including neuronal and dendritic spine density (Ashburner & Friston, Reference Ashburner and Friston2000). Fiber tract integrity is most frequently measured by calculating fractional anisotropy (FA) and mean diffusivity (MD) using diffusion-weighted magnetic resonance imaging (MRI) scans (Soares, Marques, Alves, & Sousa, Reference Soares, Marques, Alves and Sousa2013). Spontaneous functional features can be measured using resting-state functional MRI (rs-fMRI) scans. Two of these features, amplitude of low-frequency fluctuation (ALFF) and fractional ALFF (fALFF) (Zang et al., Reference Zang, He, Zhu, Cao, Sui, Liang and Wang2007; Zou et al., Reference Zou, Zhu, Yang, Zuo, Long, Cao and Zang2008), are based on the fluctuations within the magnitude of the regional BOLD signal amplitude and reflect the intensity of spontaneous neural activity. Previous studies have found that fALFF is less sensitive to physiological noise than ALFF and is helpful as an additional measure (Zou et al., Reference Zou, Zhu, Yang, Zuo, Long, Cao and Zang2008). A third spontaneous functional feature is regional homogeneity (ReHo) (Zang, Jiang, Lu, He, & Tian, Reference Zang, Jiang, Lu, He and Tian2004). ReHo measures localized similarity and is thought to reflect local alterations in the brain function.
OCD has been characterized using local measures of brain structure and function in many studies, which have revealed several biomarkers that have the potential to be used for understanding, diagnosis, and prediction of symptom severity (Bruin et al., Reference Bruin, Abe, Alonso, Anticevic, Backhausen and Balachander2023; Bu et al., Reference Bu, Hu, Zhang, Li, Zhou, Lu and Huang2019; Xu et al., Reference Xu, Hou, He, Ruan, Chen, Wei and Peng2022). VBM studies have found that patients with OCD have lower gray matter volume in the prefrontal cortex, striatum, and limbic brain regions (de Wit et al., Reference de Wit, Alonso, Schweren, Mataix-Cols, Lochner, Menchón and van den Heuvel2014). FA and MD differences in OCD have been reported in white matter tracts connecting the striatum with cortex and limbic regions (Hu et al., Reference Hu, Zhang, Bu, Li, Gao, Lu and Gong2020; Koch, Reeß, Rus, Zimmer, & Zaudig, Reference Koch, Reeß, Rus, Zimmer and Zaudig2014). ALFF, fALFF, and ReHo have each been used to effectively predict the diagnosis of OCD (Bu et al., Reference Bu, Hu, Zhang, Li, Zhou, Lu and Huang2019) and symptom severity (Zhang et al., Reference Zhang, Wang, Li, Wang, Li, Zhu and Zhang2019a). Furthermore, by using family-based designs, some of these indicators have been identified as potential vulnerability indicators for OCD (Menzies et al., Reference Menzies, Achard, Chamberlain, Fineberg, Chen, del Campo and Bullmore2007; Yang et al., Reference Yang, Luo, Zhong, Yang, Yao, Wang and Li2019). However, each of these different local neuroimaging modalities has usually been considered in isolation, and these unimodal images can only explain some aspects of phenotypes. A full understanding of how brain mechanisms relate to psychiatric disorders requires that we move beyond unimodal neuroimaging.
Combining multimodal neuroimage data sources (Mišić & Sporns, Reference Mišić and Sporns2016) is helpful to provide a more comprehensive view of the association between neuroimaging features and clinical phenotypes (Calhoun & Sui, Reference Calhoun and Sui2016). Although gray matter volume, white matter connectivity, and spontaneous brain activation fluctuations have different biological bases, integrating these features using multimodal fusion methods is helpful to uncover the underlying biology of psychiatric disorders, which can be related to symptoms and behaviors (Llera, Wolfers, Mulders, & Beckmann, Reference Llera, Wolfers, Mulders and Beckmann2019). Our goal in the present study was to combine structural and functional local neuroimaging methodologies using data fusion techniques and to relate these local features to OCD symptoms. To achieve this goal, we first performed data reduction using PCA on the behavioral measures. We took a dimensional approach (across OCD and healthy control (HC) subjects) and used PCA to extract components reflecting demographic and clinical variables across all subjects. Second, we used linked independent component analysis (LICA) to fuse functional and structural image modalities to identify independent multimodal patterns that were present in the subjects. We focused on local measures of brain structure and function that could be defined at the voxel level. Third, we related the behavioral components we extracted in step 1 to the multimodal neuroimaging patterns we identified in step 2. We used correlation analysis to identify multimodal imaging components that were positively or negatively related to the OCD symptoms, diagnosis, and demographic variables.
Materials and methods
Sample
Fifty-one patients with OCD from Shenzhen Kangning Hospital (China) and 59 HC subjects from the local community participated (from August 2018 to June 2021) in this study. All subjects were from the Chinese Han population and were assessed by experienced clinical psychiatrists. Patients met the Diagnostic and Statistical Manual of Mental Disorders, 5th edition (DSM-V) (American Psychiatric Association, 2013) criteria for OCD and were screened using the Chinese version of 6th Mini-International Neuropsychiatric Interview (M.I.N.I.) (Sheehan et al., Reference Sheehan, Lecrubier, Sheehan, Amorim, Janavs, Weiller and Dunbar1998). As long as OCD was the primary clinical diagnosis, patients were not excluded for comorbid anxious and depressive symptoms. A total of 19 of the patients with OCD were diagnosed with a comorbid condition: seven with anxiety, six with depression, and six with both anxiety and depression. Thirty-five of the patients with OCD were receiving pharmaceutical treatment. All potential subjects were excluded if they had current or past: (1) brain trauma or neurological disease, (2) contraindications to MRI scans, and (3) alcohol or substance abuse. HCs were also excluded if they had a personal or family history of mental illness.
An additional five subjects (two HC and three OCD patients) participated in the experiment but were excluded due to imaging artifacts identified during quality assurance checks. Before data preprocessing, a trained technician checked for artifacts in T1, diffusion, and functional images, such as venetian blind, gradient-wise motion, or ghosting, slice by slice.
All subjects signed an informed consent form after receiving a detailed description of the study from research staff and having their questions about the study answered. The study was approved by the Institutional Research and Ethics Committee of Shenzhen Kangning Hospital. All research procedures complied with the ethical standards of the relevant national and institutional committees on human experimentation.
Clinical measures
We used the Yale-Brown Obsessive-Compulsive Scale (Y-BOCS) to assess OCD symptom severity, from which we calculated separate obsessions and compulsions subscores. Categories of OCD symptoms were measured by the Obsessive-Compulsive Inventory-Revised (OCI-R) (Foa et al., Reference Foa, Huppert, Leiberg, Langner, Kichic, Hajcak and Salkovskis2002; Peng, Yang, Miao, Jing, & Chan, Reference Peng, Yang, Miao, Jing and Chan2011), which assesses six different types of symptoms: cleaning, obsession, hoarding, checking, neutralizing, and ordering. Obsessive beliefs were assessed using a 44-item Obsessive Belief Questionnaire (OBQ-44) (Obsessive Compulsive Cognitions Working Group, 2005; Wang, Wei, Wang, Jiang, & Peng, Reference Wang, Wei, Wang, Jiang and Peng2015), which includes three subscales: responsibility (duty), perfectionism, and control of thoughts. We used the Beck Depression Inventory (BDI) to assess depressive symptoms (Beck, Reference Beck1961) and the State-Trait Anxiety Inventory (STAI) to assess separate scores for state and trait anxiety symptoms (Spielberger, Gorsuch, Lushene, Vagg, & Jacobs, Reference Spielberger, Gorsuch, Lushene, Vagg and Jacobs1983). The Temporal Experiences of Pleasure Scale (TEPS) was used to measure individual trait dispositions in both anticipatory and consummatory experiences of pleasure (Gard, Gard, Kring, & John, Reference Gard, Gard, Kring and John2006). The Emotional Expressivity Scale (EES) was used to measure the extent to which people outwardly display their emotions, including two subscales: emotional constraint (inhibition) and emotional expression (Kring, Smith, & Neale, Reference Kring, Smith and Neale1994).
Data reduction for clinical/demographic variables
To reduce the dimensionality of the demographic and clinical variables and to achieve a better separation of factors related to OCD, principal component analysis (PCA) was conducted on the primary demographic variables and the clinical measures described earlier (Fig. 1c). We retained PCs with eigenvalues ⩾1 (maximum iterations = 25). The loading of each variable within each PC across subjects was calculated from the rotated component matrix using Varimax with Kaiser normalization. We then calculated individual subject loadings on each behavior PC (PC scores) using regression for use in the analyses relating these demographic/clinical PCs to the image components.
Figure 1. Flowchart for multimodal neuroimage analysis and main results. (a) Features were extracted from diffusion (FA and MD), functional (ALFF, fALFF, and ReHo), and anatomical (VBM) image data within the defined group masks (see text for more detail). (b) Left: The neuroimaging features were analyzed using LICA, from which we extracted 27 independent components. Middle: for each component, we identified the contribution of each subject. Right: the proportional weighting of each of the six neuroimaging features, see Figure S2 for full graphs. (c) Data reduction on demographic and clinical variables using PCA identified six different PCs. The figure illustrates the weight of each demographic and clinical variable within each of the PCs. Three of the PCs were related to clinical measurements. PC1 was sensitive to diagnostic variables: it combines a high weight for the subject's diagnosis, along with measures from the YBOCs and OCI inventories. PC2 and PC3 both weight OCD symptoms, with PC2 more strongly weighting compulsive symptoms (ordering, checking, cleaning, neutralizing, hoarding), and PC3 more strongly weighting obsessive symptoms (duty, perfection, control). (d) Finally, the loading of each image component and loading of each behavior component were correlated across subjects. We identified a total of eight significant correlations (p < 0.05) after false discovery rate (FDR) correction in seven different image components, indicated by *. Red colors indicate positive correlation, and blue indicates negative correlation. FA, fractional anisotropy; MD, mean diffusivity; ALFF, amplitude of low-frequency fluctuations; fALFF, fractional ALFF; ReHo, regional homogeneity; VBM, voxel-based morphometry; LICA, linked independent component analysis; PC, principal component; YBOCS, Yale-Brown Obsessive-Compulsive Scale; OCI, Obsessive–Compulsive Inventory; OBQ, Obsessive Belief Questionnaire; BDI, Beck Depression Inventory; STAI, State-Trait Anxiety Inventory; C1-C27, image components 1–27.
Neuroimaging measures
All neuroimaging measures were acquired on a 3.0-Tesla Discovery MR750 system (General Electrical Healthcare, USA) equipped with an eight-channel phased-array head coil. Details about the parameters for T1, resting-state functional MRI (rs-fMRI), and diffusion tensor imaging (DTI) data acquisition can be found in the online Supplementary Materials.
Preprocessing and formation of feature maps for individual modalities
Before the multimodal neuroimaging analysis, data from all three of the scans were individually preprocessed, and then relevant local brain feature image maps were formed. The overall analysis pipeline is illustrated in Fig. 1. FA and MD maps were formed using FSL v6.0 (Smith et al., Reference Smith, Jenkinson, Woolrich, Beckmann, Behrens, Johansen-Berg and Matthews2004) and combined with tract-based space statistics (Smith et al., Reference Smith, Jenkinson, Johansen-Berg, Rueckert, Nichols, Mackay and Behrens2006), VBM was processed using CAT 12 (Gaser & Dahnke, Reference Gaser and Dahnke2022), and ALFF, fALFF, and ReHo feature maps were formed using SPM12 and DPARSF (Yan, Wang, Zuo, & Zang, Reference Yan, Wang, Zuo and Zang2016) (further details are given in online Supplementary Material).
Multimodal neuroimaging analysis
We performed multimodal neuroimaging analyses for all the local structural indicators and functional activity indicators within appropriate masks based on cortico-striatal connectivity. Two separate masks were formed using seed-based methods with the striatum as the seed: a primarily functional mask (for the VBM and rs-fMRI image modalities) and a primarily white matter mask (for the DTI modalities). The functional and structural masks are illustrated in online Supplementary Fig. S1 (bottom) and online Supplementary Fig. S1 (top), respectively. It should be noted that previous research has reported impaired connectivity in OCD between the striatum and many neural regions, including the frontal, parietal, and temporal cortex, and cerebellum (Eng, Sim, & Chen, Reference Eng, Sim and Chen2015), and that the striatum has been shown to be structurally and functionally connected to most of the cortex and the cerebellum. Therefore, our functional mask was large and encompassed all areas with known striatum-based functional connectivity (Marquand, Haak, & Beckmann, Reference Marquand, Haak and Beckmann2017). Details for constructing the functional and structural masks can be found in the online Supplementary Materials.
We then used FMRIB-LICA (Groves, Beckmann, Smith, & Woolrich, Reference Groves, Beckmann, Smith and Woolrich2011) to find neuroimaging patterns that spanned multiple neuroimage modalities (FA, MD, ALFF, fALFF, ReHo, and VBM) across subjects. LICA is a data-driven method that uses a Bayesian extension of the ICA algorithm that allows one to directly compare image modalities even with different dimensionalities by applying separate ICAs for each modality. LICA has advantages over other methods in that it can balance the information across different modalities by taking into account the spatial correlation of each modality, which results in better performance than other concatenated fusion methods for multimodal neuroimaging data (Francx et al., Reference Francx, Llera, Mennes, Zwiers, Faraone, Oosterlaan and Beckmann2016; Groves et al., Reference Groves, Beckmann, Smith and Woolrich2011, Reference Groves, Smith, Fjell, Tamnes, Walhovd, Douaud and Westlye2012). We linked all individual ICA factorizations through a shared common mixing matrix that reflects the subject's contribution to each component (Groves et al., Reference Groves, Smith, Fjell, Tamnes, Walhovd, Douaud and Westlye2012). This analysis resulted in a set of independent multimodal neuroimage patterns, along with the loadings that describe the degree to which the patterns were ‘driven’ by the specific modalities (online Supplementary Figs 1 and S2). We chose to extract 27 components following the formula proposed by Groves et al. (Reference Groves, Beckmann, Smith and Woolrich2011, Reference Groves, Smith, Fjell, Tamnes, Walhovd, Douaud and Westlye2012) who argued that the number of components should be less than the number of subjects divided by 4 (in our study, we had 110 subjects).
The individual subject loadings defined within the cross-subject variation of multimodal effects were also calculated. Previous research has established that these loadings can be related to behavioral or clinical variables such as age (Douaud et al., Reference Douaud, Groves, Tamnes, Westlye, Duff, Engvig and Johansen-Berg2014) and clinical symptoms (Ball et al., Reference Ball, Malpas, Genc, Efron, Sciberras, Anderson and Silk2018) by using simple correlation. We examined the relationship between the demographic/clinical PC measures and the imaging ICs using Pearson correlations in order to be consistent with methods used in the previous work and in order to simplify interpretation. We chose not to use canonical correlation analysis (Llera et al., Reference Llera, Wolfers, Mulders and Beckmann2019) because the resulting complex relationships can be hard to interpret. Permutation testing (10 000 times) was used to assess significance (p < 0.05), and this procedure has been shown to be appropriate for small sample sizes. Correction for multiple comparisons was performed using the false discovery rate (FDR) method (Benjamini & Hochberg, Reference Benjamini and Hochberg1995). To further confirm the stability of our results, we controlled for total intracranial volume (TIV) and head motion in a supplementary partial correlation analysis (online Supplementary Fig. S4).
Results
Descriptive statistics for demographic and clinical variables
A summary and comparison of the demographic and clinical measurements between the OCD and HC groups are presented in online Supplementary Table S1. Overall, the two groups did not differ significantly for most demographic variables (age, gender, weight, marital status). However, the HC group had significantly greater years of education than the OCD group. The OCD group had significantly higher scores on all the YBOCs and OCI subscales, along with higher scores on the clinical measures of anxiety and depression. The two groups did not differ significantly in the clinical measures of emotional expression and inhibition.
Dimensionality reduction for demographic and clinical variables
We first performed the Kaiser–Meyer–Olkin test (0.84) and Bartlett's test of sphericity (1965.52, df = 276, p < 0.001); both tests indicated that our variable sets and our sample were suitable for PCA. We then performed the PCA on the behavioral variables; this analysis revealed six principal components with eigenvalues ⩾1, which together explained 74% of the total variance. As shown in Fig. 1c and online Supplementary Table S2, the first component (PC1) included a number of factors that support the interpretation that this component reflects the overall presence of OCD and OCD diagnosis. These factors include high weighting on the diagnosis label, high weighting for the subscales from the Y-BOCS and OCI, and high weighting for psychiatric symptoms often comorbid in OCD including anxiety, anhedonia, and depression. In addition, this component included weighting of years of education; this is likely due to the significant difference in years of education between the OCD and HC groups rather than indicating an independent role for years of education in OCD symptomology.
Components 2 and 3 each showed high weighting for different symptoms of OCD, but not in conjunction with a high weighting on the diagnosis label. Thus, these components may reflect obsessive and compulsive symptoms across the whole population studied rather than within OCD specifically. Component 2 showed high weighting of several of the OCD subscales of the OCI, largely related to compulsive behavior. Component 3 showed high weighting of the three subscales of the OBQ, obsession subscales within the OCI, and emotion inhibition subscales. The variables that had the highest loadings in PC2 are symptoms of specific subtypes of OCD but were not variables that had a high weighting in PC1, which was associated with OCD diagnosis. One reason that these variables may have loaded on separate components rather than a single OCD component may be that diagnosis of OCD by clinicians relies mainly on the Y-BOCS and clinical observations and does not include as important criterion information about the specific symptom types in OCD that are measured by the OCI-R. Furthermore, compulsion (PC2) is not uniquely a characteristic of OCD but exists in the larger population along a spectrum.
Components 4 and 5 were related to demographic variables. Component 4 had a high negative weighting of age along with a high positive weighting of marital status (within which marriage was coded with a higher value than being single). Because marriage and age are positively correlated (older subjects are more likely to be married), we interpret this component as primarily reflecting age. Component 5 had a high negative weighting of weight along with a high positive weighting of gender (within which female was coded with a higher value than male). Because gender and weight are positively correlated (males are more likely to weigh more), we interpret this component as primarily reflecting gender. Finally, Component 6 reflected the two emotion expression subscales from EES, with a high weighting of emotional expressiveness and a low weighting of emotional suppression.
Multimodal neuroimaging patterns for diagnosis, symptoms, and demographic characteristics
The LICA resulted in 27 component maps, each of which included weightings of the contributions of the six different neuroimaging feature maps, including the diffusion features (FA and MD), functional activation patterns (ALFF, fALFF, and ReHo), and gray matter volume (VBM). Each component was associated with a single vector of contributions that described the degree to which that component was ‘driven’ by the different modalities (feature loadings) (online Supplementary Fig. S2). Each component was also associated with a single vector for each subject that described how that individual subject contributed to the component (subject loadings). We performed a correlation analysis between these individual subject loading vectors and the individual's PC loadings on the clinical and demographic characteristics. The results were corrected for multiple comparisons using the FDR method. We identified a total of eight significant correlations between the image components and the clinical components, which are indicated in Fig. 1 via an asterisk. Detailed coefficients and p values for the analyses are provided in online Supplementary Table S3.
We found that PC1, the behavioral component that included high weighting of OCD diagnosis and symptoms, showed a negative correlation with IC9 (r = −0.29, p = 0.0018, FDR corrected p = 0.041). Within IC9, three of the six imaging modalities were significantly weighted: ALFF, FA, and MD. The contributing areas within each imaging modality are illustrated in Fig. 2. The relative contributions from each of these different modalities were 1% for FA, 5% for MD, and 91% for ALFF (see online Supplementary Fig. S2). As shown in Fig. 2, positive weightings for ALFF were located in the bilateral prefrontal cortex, superior temporal gyrus, insula, and superior parietal lobule. Negative weightings for ALFF were in the brainstem and insular cortex. The structural connectivity measures made smaller but significant contributions to the IC. These measures (both FA and MD) were mainly concentrated within the orbitofrontal tract, the anterior limb of internal capsule, and the inferior fronto-occipital fasciculus, especially the parts of these tracts connecting striatum with thalamus, lateral prefrontal, and temporal cortex (Fig. 2).
Figure 2. Multimodal neuroimaging pattern IC9, which was negatively correlated with PC1. PC1 reflected OCD diagnosis and symptoms (refer to Fig. 1c). Spatial maps are illustrated for the three significantly weighted modalities within this IC: FA, MD, and ALFF. For better visualization, the value of voxels of each modality has been normalized, and thresholded at a z-value of 1.6. This value indicates how much a voxel in the spatial map exceeds the estimated noise floor in either a positive or negative direction. ALFF, amplitude of low-frequency fluctuation; FA, fractional anisotropy; MD, mean diffusivity; OCD, obsessive–compulsive disorder; PC, principal component.
Two of the ICs were correlated with PC2. PC2 included high weightings of compulsive symptoms across measures and across both subject groups (Fig. 1c). One IC was negatively correlated with PC2 [IC10 (r = −0.34, p = 0.0008, FDR corrected p = 0.021)] and another was positively correlated with PC2 [IC27 (r = 0.30, p = 0.0006, FDR corrected p = 0.019)]. These two ICs are illustrated in Fig. 3. IC10 (Fig. 3, top row) included primarily anatomical measures, with significant weighting from the FA (19%), MD (77%), and VBM (2%) images. In IC10, structural connections measured by both FA and MD were heavily weighted in the orbitofrontal tract, fornix, and anterior part of internal capsule, especially the parts of these tracts close to striatum, thalamus, and orbitofrontal cortex. Gray matter volume (VBM) was heavily weighted in caudate, putamen, thalamus, and superior temporal cortex. The negative correlation with the PC2 behavioral measure indicates that highly weighted areas (red-orange color scheme) were negatively correlated with compulsion, and highly negatively weighted areas (blue color scheme) were positively correlated with compulsion. IC27 in contrast included strongest weightings from the three imaging measures based on the rs-fMRI scan, including ALFF (19%), fALFF (33%), and ReHo (44%). The positive correlation between PC3 and IC27 can be interpreted as a greater positive correlation between compulsion and highly weighted areas in the images (red-orange color scheme). Within component 27, ALFF showed high regional weights in medial prefrontal, temporal, and occipital cortex; fALFF in lateral prefrontal, temporal, and occipital cortex; and ReHo in temporal and occipital cortex (Fig. 3).
Figure 3. Multimodal neuroimaging patterns (ICs) correlated with PC2, which reflected compulsive symptoms across all subjects (Fig. 1c). Note that IC10 was negatively correlated with PC2, whereas IC27 was positively correlated with PC2. Spatial maps are illustrated for the significantly weighted modalities within each IC: for IC10, FA, MD, and VBM; for IC27, ALFF, fALFF, and ReHo. For better visualization, the value of voxels of each modality has been normalized, and thresholded at a z-value of 1.6. ALFF, amplitude of low-frequency fluctuations; fALFF, fractional ALFF; FA, fractional anisotropy; MD, mean diffusivity; ReHo, regional homogeneity; VBM, voxel-based morphometry.
Age as identified in principal component 4 was correlated with three ICs: positively with IC1 (r = 0.55, p = 0.0002, FDR corrected p = 0.011) and IC15 (r = 0.30, p = 0.0018, FDR corrected p = 0.041) and negatively with IC23 (r = −0.42, p = 0.0002, FDR corrected p = 0.011). Gender as identified in principal component 5 was also correlated with IC1, though negatively (r = −0.42, p = 0.0002, FDR corrected p = 0.011). These three ICs are illustrated in Fig. 4. IC15 and IC23 included significant weightings from across all six structural modalities (no single modality contributed >50% to the total variance of component). In contrast, IC1 was dominated by VBM (94%). Across these three image components, the multimodal neuroimage weightings were distributed across many brain regions. Finally, principal component 6, associated with the ESS measures of emotional expression, negatively correlated with IC8 (r = −0.31, p = 0.0008, FDR corrected p = 0.019), and IC8 included primary weighting of FA (81%) and is illustrated in online Supplementary Figs S2 and S3. We also performed an analysis to control for TIV and head motion by applying partial correlation analysis and permutation test. The overall pattern of results remained the same in this analysis with the exception that the correlation between IC1 and PC5 no longer met the criteria for statistical significance under FDR correction (online Supplementary Fig. S4).
Figure 4. Multimodal neuroimaging patterns (ICs) linked to demographic variables. IC1 was positively correlated with age (PC4) and negatively correlated with gender (PC5). IC15 was positively correlated with age (PC4). IC23 was negatively correlated with age (PC4). Spatial maps are illustrated for the significantly weighted modalities within each IC: IC1 included five of the six modalities (FA, ALFF, fALFF, ReHo, and VBM), whereas ICs 15 and 23 included all six modalities (FA, MD, ALFF, fALFF, ReHo, and VBM). For better visualization, the value of voxels of each modality has been normalized, and thresholded at a z-value of 1.6. ALFF, amplitude of low-frequency fluctuations; fALFF, fractional ALFF; FA, fractional anisotropy; MD, mean diffusivity; PC, principal component; ReHo, regional homogeneity; VBM, voxel-based morphometry.
Discussion
We identified a set of multimodal local brain imaging patterns linked to phenotypic variation, especially compulsive behaviors and OCD diagnosis factors, across patients with OCD and HCs. One of the imaging patterns (IC9) was associated with PC1, which had high weights for the OCD diagnosis label along with the most common OCD measure (YBOCS) and comorbid depression, anxiety, and anhedonia symptoms (Fig. 1). The symptoms and measures included in PC1 were those most often relied upon in clinical practice for diagnosis; experienced clinicians not only consider OCD symptoms but also consider comorbidities during diagnosis. IC9 was dominated by ALFF, which is thought to measure the intensity of the regional brain activity. Many psychiatric disorders including OCD have been associated with ALFF differences (Hou et al., Reference Hou, Wu, Lin, Wang, Zhou, Guo and Li2012). One study found that ALFF in the dorsolateral prefrontal cortex and insula could be used to classify whether subjects have OCD with higher accuracy than other neuroimaging features (Bu et al., Reference Bu, Hu, Zhang, Li, Zhou, Lu and Huang2019). Another research study found that ALFF in medial prefrontal cortex and superior temporal gyrus was also correlated with the severity of social anhedonia in OCD patients (Xia et al., Reference Xia, Fan, Du, Liu, Li, Zhu and Zhu2019). It should be noted that our results were based on a simple Pearson correlation test. Furthermore, the correlation was negative, so that lower weights of this component were associated with a higher likelihood of OCD. The spatial distribution of ALFF included especially high weights for the prefrontal cortex, which is consistent with previously reported image patterns for OCD diagnosis (Bu et al., Reference Bu, Hu, Zhang, Li, Zhou, Lu and Huang2019). Anatomically, the FA and MD features that also contributed to IC9 mainly weighted white matter tracts that are known to pass between the striatum and lateral prefrontal and temporal cortex. Previous research found that FA across a distributed network, including white matter tracts near bilateral prefrontal and temporal regions, and the inferior fronto-occipital fasciculus, could be used to correctly identify patients with OCD v. controls with high accuracy (Li et al., Reference Li, Huang, Tang, Yang, Li, Kemp and Gong2014). Lochner et al. (Reference Lochner, Fouché, du Plessis, Spottiswoode, Seedat, Fineberg and Stein2012) found that FA in bilateral regions of the anterior limb of the internal capsule was correlated with depression and anxiety symptoms in patients with OCD. In summary, we provided a more comprehensive view of local neuroimaging features linked to OCD diagnosis and comorbid symptoms across both functional and structural levels.
Two of the complementary neuroimaging patterns, IC10 and IC27, were associated primarily with compulsion, a core symptom of OCD (identified in PC2; see Fig. 1). IC10 included weightings of anatomical features, whereas IC27 included weightings of the functional features. Within IC10, fiber tract integrity in the orbitofrontal tract, fornix, and anterior part of internal capsule, along with gray matter density in the striatum, thalamus, superior temporal cortex, and cerebellum was correlated negatively with compulsion. The previous research has also emphasized the importance of these regions to OCD. The disruption of orbitofrontal white matter has been found in OCD (Hu et al., Reference Hu, Zhang, Bu, Li, Gao, Lu and Gong2020), and structural connectivity in the salience network can be used to predict compulsion (Zhu et al., Reference Zhu, Fu, Chen, Yu, Zhang, Zhang and Wang2022). Wu et al. (Reference Wu, Yang, Xu, Huo, Seger, Peng and Chen2022a, Reference Wu, Yu, Zhang, Feng, Zhang, Sahakian and Robbins2022b) found that the volume of the bilateral precentral gyrus, inferior parietal locus, and right precuneus correlated negatively with externalized problems including compulsion. In addition, a positive relationship also has been identified by our results. Within IC27, ALFF, fALFF, and ReHo functional neuroimaging features correlated positively with compulsion. Across the three measures, differences were localized within medial and lateral prefrontal, temporal, and occipital cortices, and cerebellum. The previous research reported that ALFF in the left superior temporal gyrus of subjects with OCD was positively correlated with compulsion score and that fALFF in the cerebellum was positively correlated with the total score of both compulsion and obsession (Zhang et al., Reference Zhang, Wang, Li, Wang, Li, Zhu and Zhang2019a, Reference Zhang, Hu, Li, Lu, Li, Hu and Huang2019b). Another study found that ReHo in the temporal, inferior orbitofrontal, and precentral cortices was positively correlated with compulsive behavior scores in OCD patients (Niu et al., Reference Niu, Cheng, Song, Yang, Chu, Liu and Li2017). A third study found that a change in the ReHo value in the left cerebellum could predict the reduction of compulsion scores in OCD after therapy (Yang et al., Reference Yang, Sun, Luo, Zhong, Li, Yao and Li2015). Integrating across the negative and positive correlations of IC10 and IC27 with compulsion, we propose that this pattern reflects a resilience (IC10) vulnerability (IC27) factor for developing compulsive symptoms. This is similar to the finding of a positive-negative behavior spectrum (Llera et al., Reference Llera, Wolfers, Mulders and Beckmann2019). Connectivity between frontal cortical, middle occipital cortical, and cerebellar regions was identified as a vulnerability factor for OCD in a previous study (Hampshire et al., Reference Hampshire, Zadel, Sandrone, Soreq, Fineberg, Bullmore and Chamberlain2020), and the regions identified in this study overlapped with IC27. However, due to the differences in analysis methods and study design, this inference of resilience-vulnerability patterns requires more evidence.
Although we did not perform connectivity analyses on the rs-fMRI data, the areas implicated in our results are consistent with those identified in rs-fMRI functional connectivity studies. Functional connectivity among the salience network, the frontal parietal network, and the default mode network has been associated with compulsion in OCD (Zhu et al., Reference Zhu, Fu, Chen, Yu, Zhang, Zhang and Wang2022). Interindividual variations in connectivity within cerebellar-visual, striato-limbic, and frontal networks were associated with OCI-R scores (Kashyap et al., Reference Kashyap, Eng, Bhattacharjee, Gupta, Ho, Ho and Chen2021). Connectivity between dorsal caudate and superior frontal gyrus was positively correlated with hoarding severity (Kashyap et al., Reference Kashyap, Eng, Bhattacharjee, Gupta, Ho, Ho and Chen2021). The previous research largely focused on the connectivity between the striatum and prefrontal and parietal cortex in OCD. However, OCD is also characterized by abnormal connectivity between the striatum and cerebellum during rest (Sha et al., Reference Sha, Edmiston, Versace, Fournier, Graur, Greenberg and Phillips2020) and response inhibition (Eng et al., Reference Eng, Sim and Chen2015). Our analyses were restricted to the striatum-based circuit, providing further information about the links between compulsion and multiple spontaneous functional activity and anatomical measures. These features had high weights not only on the cerebral cortex but also on several parts of the cerebellum (IC10 and IC27). The previous research has highlighted the role of cerebellar functional connectivity in OCD (Anticevic et al., Reference Anticevic, Hu, Zhang, Savic, Billingslea, Wasylink and Pittenger2014). Our results provide further evidence that the cerebellum plays an important role in compulsion and diagnosis of OCD.
We did find a behavioral component associated with obsessive symptoms (PC3); however, it did not significantly correlate with any of the neuroimaging patterns. Obsession may be better considered as secondary factor in OCD rather than the core symptom (Gillan & Sahakian, Reference Gillan and Sahakian2015). This is consistent with recent arguments by researchers that compulsion is the main cause of OCD, and obsession is an accessory symptom caused by compulsion (Gillan & Robbins, Reference Gillan and Robbins2014; Gillan & Sahakian, Reference Gillan and Sahakian2015). The previous work from our laboratory examining structural–functional coupling found a direct relationship between instrumental learning systems and compulsion, but only an indirect relationship with obsession (Wu et al., Reference Wu, Yang, Xu, Huo, Seger, Peng and Chen2022a, Reference Wu, Yu, Zhang, Feng, Zhang, Sahakian and Robbins2022b; Xu et al., Reference Xu, Hou, He, Ruan, Chen, Wei and Peng2022).
We also found three multimodal neuroimaging patterns (IC1, IC15, and IC23) that were correlated with demographic components including age (PC4) and gender (PC5). Although most previous case–control studies matched or controlled the demographic variables within the study sample, they did not examine their effects on the neural correlates of OCD. Identifying the effects of demographic variables on neuroimaging features may be helpful in identifying how these variables interact with the effects of clinical symptoms. The previous research has identified morphometric alterations in the prefrontal, parietal, and temporo-occipital regions that are related to age (Piras et al., Reference Piras, Piras, Chiapponi, Girardi, Caltagirone and Spalletta2015), including volume loss in temporal cortex specifically in OCD patients with the increasing age (de Wit et al., Reference de Wit, Alonso, Schweren, Mataix-Cols, Lochner, Menchón and van den Heuvel2014). VBM studies have also reported sex differences in the brain (Zhang et al., Reference Zhang, Wang, Li, Wang, Li, Zhu and Zhang2019a, Reference Zhang, Hu, Li, Lu, Li, Hu and Huang2019b). Our results extend beyond these unimodal studies and provide a more comprehensive view, finding the effects of these demographic variables in almost all neuroimaging features and across extensive brain regions. Finally, we found one neuroimaging pattern, IC8, that correlated mainly with emotional expression across all subjects. Lower weights on FA, MD, and VBM were related to the more intense expression of emotion (OCD > HC, online Supplementary Table S1). OCD patients have shown reduced functional activation in the thalamus to facial expressions of disgust (de Wit et al., Reference de Wit, Alonso, Schweren, Mataix-Cols, Lochner, Menchón and van den Heuvel2014), and greater activation in posterior thalamus and parahippocampal cortex to fearful faces (Cardoner et al., Reference Cardoner, Harrison, Pujol, Soriano-Mas, Hernández-Ribas, López-Solá and Menchón2011). The present study further supports the finding that these areas are related to emotional expression in OCD.
Limitations
The current study has several limitations that warrant consideration. For computational reasons, we focused on the VBM measurement of the gray matter volume. Other structural neuroimaging features such as cortical thickness and area were not included in the analysis and should be examined in future research. We used Pearson correlation coefficients to assess the relationship between behavioral measure PCs and neuroimaging features; correlation has the advantage of simplicity, but is subject to interpretational limitations (e.g. possible third variable contributions). The OCD patients varied in the medications they were receiving; in future research, effects of medication should be examined by enrolling a larger sample size with sufficient power to test for any medication effects. The current study used a single-session cross-sectional design. Future research including a large longitudinal cohort with unified therapeutic design will be needed to identify the relationships between multimodal neuroimaging patterns and clinical symptom changes over time and treatment. In order to confirm the pattern of results and test the generalizability of the methods used, it will be important to replicate the results in an independent dataset in the future.
Conclusion
Our results demonstrate how local structural and functional neuroimaging data and dimensionality reduction methods can be used to identify the neurobiological correlates of core symptoms, diagnosis factors, and demographic variables in OCD. To our knowledge, this is the first study to discriminate patterns across these local structural and functional imaging modalities associated with multiple symptoms and diagnosis of OCD. We further provide evidence that compulsive behavior regardless of OCD diagnosis is an independent phenotype with its own unique multimodal neuroimaging patterns.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291723002386.
Data
The code for FMRIB LICA can be found at https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FLICA, and the code for visualization was downloaded from https://github.com/allera/Llera_elife_2019_1.
Financial support
This work was supported by the National Natural Science Foundation of China [Grant Nos. 31871113, 32171081, and 31920103009 (to Z. P.) and 32071049 (to Q. C.)], the National Science and Technology Innovation 2030 Major Program [Grant No. 2021ZD0203800 (to Q. C.)], the Natural Science Foundation of Guangdong Province [2023A1515011782 (to Z. P.)], Guangdong Basic and Applied Basic Research Foundation [2022A1515012185 (to Q. C.) and 2023A1515011782 (to Z.P.)], the Neuroeconomics Laboratory of Guangzhou Huashang College [2021WSYS002 (to Q. C.)], the General Program of Shenzhen Science and Technology Innovation Commission [JCYJ20220530155204009 (to J. R. C.)], Shenzhen Key Medical Discipline Construction Fund (Grant No. SZXK071), and Key-Area Research and Development Program of Guangdong Province (2019B030335001).
Author contributions
C. X. and G. H. collected the data. C. X., Q. C., and Z. P. developed the idea for the study and conducted part of the analyses. C. X., Z. P., Q. C., and C. A. S. contributed to substantial revisions and drafted the manuscript. X. G., J. C., T. H., Z. R., and Z. W. conducted part of the analyses. Z. P., Q. C., C. A. S., and G. H. provided supervision for the project and revisions. All authors have approved the final article.
Competing interest
None.
