Introduction
Contemporary psychiatry has been rooted in the conceptualization that schizophrenia (SCZ), bipolar disorder (BPD), and major depressive disorder (MDD) are three distinct major psychiatric disorders (MPDs). However, mounting evidence indicates significant commonalities in genetic (Lee et al., Reference Lee, Ripke, Neale, Faraone, Purcell, Perlis and Wray2013), neural (Wei et al., Reference Wei, Chang, Womer, Zhou, Yin, Wei and Wang2018), neuroinflammatory (Pape, Tamouza, Leboyer, & Zipp, Reference Pape, Tamouza, Leboyer and Zipp2019), and clinical features (Barch & Sheffield, Reference Barch and Sheffield2014). Recent studies have increasingly focused on direct comparisons across SCZ, BPD, and MDD. Their subsequent findings have converged on the reconceptualization of MPDs as a transdiagnostic continuum across SCZ, BPD, and MDD, rather than distinct diagnostic categories (Goodkind et al., Reference Goodkind, Eickhoff, Oathes, Jiang, Chang, Jones-Hagata and Etkin2015). A dimensional reconceptualization of these disorders implies that common manifestations should be associated with common substrates or pathophysiology (Buckholtz & Meyer-Lindenberg, Reference Buckholtz and Meyer-Lindenberg2012).
Notably, executive deficits, one subset of cognition, have been considered a core feature across MPDs (Barch & Sheffield, Reference Barch and Sheffield2014; Goodkind et al., Reference Goodkind, Eickhoff, Oathes, Jiang, Chang, Jones-Hagata and Etkin2015; Sheffield et al., Reference Sheffield, Kandala, Tamminga, Pearlson, Keshavan, Sweeney and Barch2017). Patients with SCZ have demonstrated relatively generalized executive deficits (Bora, Yucel, & Pantelis, Reference Bora, Yucel and Pantelis2009). Evidence has suggested the persistence of cognitive dysfunction, particularly in executive functions, across all phases of BPD, including euthymic, manic, and depressive states (Latalova, Prasko, Diveky, & Velartova, Reference Latalova, Prasko, Diveky and Velartova2011). Preliminary results have shown cognitive impairments in executive function, working memory, and attention may appear early in the course of MDD and may persist even during remission (Trivedi & Greer, Reference Trivedi and Greer2014). The Wisconsin Card Sorting Test (WCST) is a widely used neuropsychological assessment of executive function (Gardizi, King, McNeely, & Vaz, Reference Gardizi, King, McNeely and Vaz2019). Compared to healthy controls (HC), patients with SCZ, BPD, and MDD show significantly greater deficits in multiple WCST measures (Kosger, Essizoglu, Baltacioglu, Ulkgun, & Yenilmez, Reference Kosger, Essizoglu, Baltacioglu, Ulkgun and Yenilmez2015; Rady, Elsheshai, Abou El Wafa, & Elkholy, Reference Rady, Elsheshai, Abou El Wafa and Elkholy2012). The biological mechanisms for shared cognitive manifestations are unclear but may relate to common biological substrates in MPDs (Barch & Sheffield, Reference Barch and Sheffield2014). A dimensional approach to examining cognitive deficits across MPDs using multi-level analyses could lead to valuable integration of biological understanding and clinical practice (Krystal & State, Reference Krystal and State2014).
Decades of research have revealed that brain connectivity is critical in determining a neurobiological mechanism related to executive deficits in MPDs (Krystal & State, Reference Krystal and State2014). Diffusion tensor imaging (DTI) provides an opportunity for evaluating the organization and coherence of white matter (WM) fiber tracts. The WM microstructure plays a pivotal role in maintaining processing speed and executive measures, and WM microstructural abnormalities have been linked to executive deficits in MPDs (Kochunov et al., Reference Kochunov, Coyle, Rowland, Jahanshad, Thompson, Kelly and Hong2017; Perez-Iglesias et al., Reference Perez-Iglesias, Tordesillas-Gutierrez, McGuire, Barker, Roiz-Santianez, Mata and Crespo-Facorro2010; Wise et al., Reference Wise, Radua, Nortje, Cleare, Young and Arnone2016). Fractional anisotropy (FA) has been among the most commonly used measures for determining structural connectivity and the extent of myelination in axonal bundles (Kochunov et al., Reference Kochunov, Coyle, Rowland, Jahanshad, Thompson, Kelly and Hong2017). DTI studies have consistently shown that lower FA values are associated with executive deficits among MPDs (Kochunov et al., Reference Kochunov, Coyle, Rowland, Jahanshad, Thompson, Kelly and Hong2017; Perez-Iglesias et al., Reference Perez-Iglesias, Tordesillas-Gutierrez, McGuire, Barker, Roiz-Santianez, Mata and Crespo-Facorro2010). Additionally, intrinsic functional connectivity, including regional and remote functional connectivity, has become a transdiagnostic substrate across MPDs (Sheffield et al., Reference Sheffield, Kandala, Tamminga, Pearlson, Keshavan, Sweeney and Barch2017). Intrinsic functional connectivity plays a central role in maintaining the balance of neurotransmitters and the number of excitatory v. inhibitory connections in the brain (Jiang & Zuo, Reference Jiang and Zuo2016; Power, Schlaggar, & Petersen, Reference Power, Schlaggar and Petersen2014). The connectivity between neuron density and cell type may contribute to functional connectivity within a small region, and regional alterations in micro-level homogeneity most likely contribute to regional functional connectivity alterations (Jiang et al., Reference Jiang, Xu, He, Hou, Wang, Cao and Zuo2015). Regional homogeneity (ReHo) is a highly reproducible and reliable index of regional functional connectivity (Jiang & Zuo, Reference Jiang and Zuo2016), which has been considered to reflect anatomical, morphological, and intrinsically geometric features in a local brain structure as well as a topology-functionality interplay (Wei et al., Reference Wei, Duan, Womer, Zhu, Yin, Cui and Wang2020). Accumulating evidence implicates ReHo alterations in the pathophysiology of MPDs (Ji et al., Reference Ji, Meda, Tamminga, Clementz, Keshavan, Sweeney and Pearlson2020; Jiang et al., Reference Jiang, Xu, He, Hou, Wang, Cao and Zuo2015; Zuo et al., Reference Zuo, Xu, Jiang, Yang, Cao, He and Milham2013). Moreover, an association between abnormalities in ReHo and executive deficits was also reported in MPDs (Ji et al., Reference Ji, Meda, Tamminga, Clementz, Keshavan, Sweeney and Pearlson2020). Based on these findings, identifying structural and functional dysconnectivity could help illuminate the transdiagnostic mechanism underlying executive deficits across MPDs.
Inflammation has been hypothesized to be a possible neurobiological mechanism linking executive deficits in MPDs (Fineberg & Ellman, Reference Fineberg and Ellman2013; Miller, Maletic, & Raison, Reference Miller, Maletic and Raison2009; Pape et al., Reference Pape, Tamouza, Leboyer and Zipp2019). Genome-wide association studies have found shared variants in inflammatory pathways across psychiatric disorders (Network & Pathway Analysis Subgroup of Psychiatric Genomics, 2015). Most critically, there is interplay between the brain and peripheral systems via the glymphatic and meningeal lymphatic system (Deverman & Patterson, Reference Deverman and Patterson2009; Pape et al., Reference Pape, Tamouza, Leboyer and Zipp2019). Pro-inflammatory cytokines can mediate neurodevelopmental processes, including myelination, axonal growth, and synaptogenesis (Bartzokis, Reference Bartzokis2012; Deverman & Patterson, Reference Deverman and Patterson2009). Interestingly, inflammatory processes, characterized by elevated interleukin (IL)-1β, IL-6 and tumor necrosis factor-α (TNF-α), have a significant effect on cognitive function in patients with MPDs (Frodl & Amico, Reference Frodl and Amico2014). Furthermore, structural changes associated with inflammation have been observed in the prefrontal cortex of individuals with SCZ, BPD, and MDD (Frodl & Amico, Reference Frodl and Amico2014; Lin et al., Reference Lin, Shao, Wang, Lu, Zou, Chen and So2019). Recent research also implicates the relationship between inflammation and functional connectivity in the prefrontal cortex (Felger et al., Reference Felger, Li, Haroon, Woolwine, Jung, Hu and Miller2016). Although preclinical and clinical studies (Fineberg & Ellman, Reference Fineberg and Ellman2013; Kiehl, Reference Kiehl2006; Lin et al., Reference Lin, Shao, Wang, Lu, Zou, Chen and So2019; Sheffield et al., Reference Sheffield, Kandala, Tamminga, Pearlson, Keshavan, Sweeney and Barch2017; Wise et al., Reference Wise, Radua, Nortje, Cleare, Young and Arnone2016) have demonstrated that structural or functional dysconnectivity and an increase of pro-inflammatory cytokine play a key role in cognitive dysfunctions, the underlying mechanisms whereby such multi-level manifestations contribute to cognitive impairments remain unclear across MPDs.
To address this issue, the first aim of the study was to explore the relationship between structural and functional connectivity (FA and ReHo), pro-inflammatory cytokines (IL-1β, IL-6, and TNF-α), and executive cognition in patients with MPDs. We test the hypothesis that structural and functional dysconnectivity, as well as the identified alterations in pro-inflammatory cytokine levels are associated with executive deficits. Based on these possible associations, our second aim was to employ a mediation analysis to identify whether brain connectivity and inflammation are neurobiological mechanisms underlying executive deficits. The third aim was to test whether the brain connectivity and inflammation could precisely predict executive cognition. If supported, the findings would provide strong evidence of common neuroanatomical substrates and neurobiological mechanisms associated with executive deficits across MPDs.
Methods
Participants
The study enrolled 1543 individuals from a single site: 396 with SCZ, 286 with BPD, 376 with MDD, and 485 HC. All patients were recruited from the inpatient and outpatient units of the Department of Psychiatry, the First Affiliated Hospital of China Medical University, and Shenyang Mental Health Centre, Shenyang, China. Subjects were independently diagnosed with SCZ, BPD, or MDD by two trained psychiatrists according to the Structured Clinical Interview for Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) Axis I Disorders in participants 18 years and older and the Schedule for Affective Disorders and Schizophrenia for School-Age Children-present and Lifetime Version (K-SADS-PL) in those younger than 18 years. The two psychiatrists had a high inter-rater reliability (Cohen's κ score > 0.80). Age- and gender-matched HC were recruited from the local area via advertisements. They were confirmed to have no current or lifetime Axis I disorder using the SCID Non-Patient Edition, as well as no history of Axis I disorders among first-degree relatives (determined by obtaining a detailed family history). The exclusion criteria for all participants included the following: (i) history of DSM-IV defined alcohol or substance abuse in the last 3 months or substance abuse/dependence within the preceding 6 months; (ii) history of severe medical or neurological disorders, such as type 2 diabetes, hypertension, cardiovascular disease, stroke, or transient ischemic attack; (iii) clinical indication of inflammation; (iv) infectious diseases such as HIV/AIDS and SARS; and (v) magnetic resonance imaging (MRI) contraindications. Participant characteristics are provided in the online Supplementary Method section and Fig. S1. The study was approved by the Institutional Review Board of China Medical University and was performed in accordance with the Declaration of Helsinki. All experiments and methods were performed in accordance with approved guidelines and regulations. All subjects provided written informed consent.
Clinical assessment
Clinical assessments were administered by two well-trained assessors with highly reliable inter-rater evaluation results (k > 0.80). All participants completed the Hamilton Depression Rating Scale (HAMD), Young Mania Rating Scale (YMRS), and Brief Psychiatric Rating Scale (BPRS). We assessed executive cognition using the Wisconsin Sorting Card Test (WCST), which is a repeatable computerized cognitive test with good validity and reliability.
MRI parameters and data processing
DTI and functional MRI were acquired in the same scanning session for each subject. MRI scans were performed using a 3.0 T GE Sigma system (Sigma EXCITE HDx; GE Healthcare, Milwaukee, MI, USA) with a standard eight-channel head coil at the First Affiliated Hospital of the China Medical University, Shenyang, China. DTI was carried out with a single-shot spin-echo echo-planar imaging (EPI) sequence. The parameters were as follows: 25 non-collinear directions (b = 1000 s/mm2), together with an axial acquisition without diffusion weighting (b = 0), repetition time (TR) = 17 000 ms; echo time (TE) = 85.4 ms; matrix = 120 × 120; field of view (FOV) = 240 × 240 mm2, voxel size = 2.0 mm3 and 65 contiguous slices of 2 mm without gap. The scan lasted for 7 min and 39 s. Functional images were acquired with a gradient EPI sequence for ReHo measures. The resting-state functional parameters were as follows: TR = 2000 ms; TE = 30 ms; matrix = 64 × 64; FOV = 240 × 240 mm2 and 35 contiguous slices of 3 mm without gaps. The scan lasted for 6 min and 40 s. Data processing details are described in online Supplementary Method section.
Measurement of pro-inflammatory cytokines
Only 325 of the 1543 subjects had whole blood sample according to their willingness to participate. Three pro-inflammatory cytokine levels were measured with the Human Magnetic Luminex Assay (IL-6: BR13; IL-1β: BR28; TNF-α: BR12) using a Human Premixed Multi-Analyte Kit (R&D Systems, Inc., Minneapolis, MN, USA). A detailed description can be found in the online Supplementary Method section.
Statistical analysis
Demographic and clinical characteristics
The patient group comprised SCZ, BPD, and MDD cases, and analyses were conducted to compare patients and HC groups. One-sample Kolmogorov–Smirnov (K-S) test was used to test the normality of plasma cytokine levels. Two-tailed t tests or Mann–Whitney U tests were used for continuous variables, and χ2 tests for categorical variables. Statistical significance was set at p < 0.05. The effects of medication, medication classes, illness duration, the first episode, and age were also examined to confirm data reliability (see online Supplementary Materials).
Comparison between MPDs and HC according to FA, ReHo, and pro-inflammatory cytokines
Between-group analyses of FA values were performed in SPM12 using voxel-based, two-sample t tests with age and gender as covariates. DPABI was used to perform voxel-based, two-sample t tests comparing ReHo values between patients and HC, with age, gender, and mean framewise displacement (FD) as covariates. A voxel-wised threshold was set at p < 0.005 with a false discovery rate (FDR) correction. For each cluster with significant group differences, mean FA and ReHo values were extracted for subsequent analyses.
Correlation analyses
Partial correlation analyses (two-tailed) were performed in the patient group to determine variables significantly related to altered FA, ReHo, plasma cytokine levels, WCST scores, and HAMD total scores. Statistical significance was set at p < 0.05 (FDR corrected for multiple comparisons). We treated age and gender as covariates in the structural connectivity dataset, and we considered age, gender, and mean FD as covariates in the functional connectivity dataset.
Mediation analyses
To determine the role of brain dysconnectivity and inflammation in executive deficits, mediation analyses were performed using Mplus (Version 7.0) statistical software with 1000 bias-corrected bootstrap samples for significance testing. Causal variables included FA/ReHo values; outcome latent variables were constructed for executive cognition measures including correct responses (CR), categories completed (CC), total errors (TE), perseverative errors (PE), and non-perseverative errors (NPE), and the proposed mediator was inflammatory cytokine levels. Age and gender were identified as covariates in the structural connectivity dataset, while age, gender, and mean FD were treated as covariates in the functional connectivity dataset. In the study, we test the effect of these covariates on mediator variable and the latent dependent variable. Bootstrapped bias-corrected 95% confidence interval of mediation effect not covering zero indicated statistical significance. Mediation proportion was calculated as the ratio of the indirect effect to the total effect.
Predictive models for executive cognition
To further validate the transdiagnostic associations between brain connectivity and pro-inflammatory cytokines, we design a series of machine learning models to predict executive cognition. Brain connectivity, pro-inflammatory cytokines, and their combination were used to predict five executive cognition scores. Specifically, we used six different data to predict executive cognition, such as (i) FA values in right uncinate fasciculus, right external capsule, and right anterior limb of internal capsule; (ii) ReHo values in left orbital frontal cortex, left putamen, and left insula only; (iii) IL-1β; (iv) IL-6; (v) FA and IL-1β combined; and (vi) ReHo and IL-6 combined. Therefore, a total of 30 predictive models were trained and evaluated.
We used Artificial Neural Network (ANN) to predict executive cognition scores. The details of the network architecture and training parameters are given in the online Supplementary Methods section and Tables S1 and S2. We used fivefold cross-validation to evaluate the performance of the prediction models. The ratio of train set and test set was 4:1. Specifically, the structural connectivity dataset (N = 155) was divided into a training (N = 124) and a test set (N = 31), and the functional connectivity dataset (N = 151) was split into two sets: training (N = 120) and test set (N = 31). Subsequently, the prediction performance of executive cognition was evaluated with mean absolute error (MAE), root mean square error, the correlations between the predicted value and true value (R), and coefficient of determination (R 2). The model with the highest R 2 and the lowest MAE were chosen as the best model.
Results
Demographic and clinical characteristics
No significant differences in age, gender, handedness, BMI, smoking, and mean FD were observed between patients and HC. The K-S test showed that IL-1β, IL-6, and TNF-α in each group were normally distributed (p > 0.05). Tables 1 and 2 present detailed demographic and clinical data of the MPDs and HC, whereas online Supplementary Tables S3 and S4 list detailed characteristics of the SCZ, BPD, and MDD groups.
Table 1. Demographic, clinical characteristics, and pro-inflammatory cytokines of participants in the structural connectivity dataset
BMI, body mass index; BPRS, Brief Psychiatric Rating Scale; CC, categories completed; CR, correct responses; HAMD, Hamilton Depression Scale; IL-1β, interleukin-1β; IL-6, interleukin-6; NA, not available; NPE, non-perseverative errors; PE, perseverative errors; TE, total errors; TNF-α, tumor necrosis factor-α; WCST, Wisconsin Card Sorting Test; YMRS, Young Mania Rating Scale.
Descriptive statistics table for continuous numeric variables are reported as mean (s.d.) for each group. Descriptive statistics for categorical variables are reported as frequencies and percentages (%).
Table 2. Demographic, clinical characteristics, and pro-inflammatory cytokines of participants in the functional connectivity dataset
BMI, body mass index; BPRS, Brief Psychiatric Rating Scale; CC, categories completed; CR, correct responses; HAMD, Hamilton Depression Scale; IL-1β, interleukin-1β; IL-6, interleukin-6; NA, not available; NPE, non-perseverative errors; PE, perseverative errors; TE, total errors; TNF-α, tumor necrosis factor-α; WCST, Wisconsin Card Sorting Test; YMRS, Young Mania Rating Scale.
Descriptive statistics table for continuous numeric variables are reported as mean (s.d.) for each group. Descriptive statistics for categorical variables are reported as frequencies and percentages (%).
Comparison between MPDs and healthy controls according to FA, ReHo, and pro-inflammatory cytokines
MPDs had lower FA in the right uncinate fasciculus; right external capsule; bilateral anterior and posterior limbs of the internal capsule; genu, body, and splenium of the corpus callosum; bilateral cingulum; bilateral posterior thalamic radiation; bilateral anterior and posterior corona radiata; bilateral superior fronto-occipital fasciculus; bilateral superior corona radiata and left sagittal stratum compared to HC (Fig. 1a and online Supplementary Table S5).
Fig. 1. Significantly altered regions of fractional anisotropy and regional homogeneity values in major psychiatric disorders. Panel (a) shows regions with lower fractional anisotropy values in patients with major psychiatric disorders compared with healthy controls. Panel (b) shows significantly altered regions of regional homogeneity values in patients with major psychiatric disorders compared with healthy controls. Significance level was set at false discovery rate corrected p < 0.05. The color bar indicates the t-values. Blue and red denote decreased and increased fractional anisotropy and regional homogeneity values in the patients with major psychiatric disorders, respectively.
MPDs had significantly higher ReHo in the bilateral orbitofrontal cortices, bilateral insulae, and left putamen and lower ReHo in the right primary and association visual cortices, left visual association cortex, right primary auditory cortex, left primary motor and somatosensory cortices, and right supplementary motor area compared to HC (Fig. 1b and online Supplementary Table S6).
Plasma IL-1β, IL-6, and TNF-α levels in patients with MPDs were significantly elevated compared with HC (Tables 1 and 2). Online Supplementary Tables S3 and S4 list the specific IL-1β, IL-6, and TNF-α levels of the SCZ, BPD, and MDD groups. Specific FA, ReHo, pro-inflammatory cytokines, and WCST scores of the SCZ, BPD, and MDD groups can be found in online Supplementary Figs S2–S5.
The details of effects of medication, medication classes, the first episode illness duration, and age on FA, ReHo, and pro-inflammatory cytokines levels were in the online Supplementary Material (Tables S7–S12).
Correlations analyses
After controlling age and gender, WCST scores (CR, CC, TE, and NPE) were significantly associated decreased FA (in the right uncinate fasciculus, right external capsule, and right anterior limb of internal capsule; left anterior and posterior limb of internal capsule, left superior fronto-occipital fasciculus and left superior corona radiata; right external capsule, right anterior and superior corona radiata, right anterior and posterior limb of internal capsule, and right superior fronto-occipital fasciculus and right superior corona radiata, left sagittal stratum, bilateral cingulum, bilateral posterior thalamic radiation, bilateral anterior and posterior corona radiata, and genu, body, and splenium of corpus callosum) and increased IL-1β and TNF-α levels in structural connectivity dataset (online Supplementary Tables S13–S15).
In the functional connectivity dataset, the correlations among WCST scores (CR, TF, PE, and NPE), increased ReHo in the left orbitofrontal cortex, left putamen, and left insula and increased IL-6 levels were still significant when controlling for age, gender, and mean FD (online Supplementary Tables S16–S18).
No significant correlations between FA or ReHo values – pro-inflammatory cytokines (IL-6, IL-1β, or TNF-α) – and HAMD total score were observed in structural and functional connectivity datasets (online Supplementary Tables S19 and S20).
Mediation analyses
After controlling for age and gender, the relationship between FA values in the right uncinate fasciculus, right external capsule, and right anterior limb of the internal capsule and WCST scores was significantly mediated by IL-1β in structural connectivity dataset. IL-6 partially mediated the relationship between ReHo in the left orbitofrontal cortex, left putamen, left insula, and WCST scores. Analyses controlled for age, gender, and mean FD in the functional connectivity dataset (Fig. 2 and online Supplementary Tables S21 and S22).
Fig. 2. Mediation models. Panel (a) illustrates the mediation of IL-1β on the relationship between FA values within the fronto-limbic-striatal regions and executive deficits, using causal mediation analysis. Path A represents the association between lower FA values within the fronto-limbic-striatal regions and IL-1β. Path B represents the association between IL-1β and executive deficits. Path C represents the association between lower FA and executive deficits. Path C′ is used to assess how IL-1β mediates the effect of FA on executive deficits, whereas path AB represents the indirect effect of FA on executive deficits mediated by IL-1β. Panel (b) illustrates the mediation of IL-6 on the relationship between ReHo values within the fronto-limbic-striatal regions and executive deficits, using causal mediation analysis. Path A represents the association between increased ReHo within the fronto-limbic-striatal regions and IL-6. Path B represents the association between IL-6 and executive deficits, which, in combination with A, is used to assess how IL-6 mediates the effect of increased ReHo on executive deficits. Path C represents the association between increased ReHo and executive deficits. Path C′ is used to assess how IL-6 mediates the effect of ReHo on executive deficits, whereas path AB represents the indirect effect of ReHo on executive deficits mediated by IL-6. FA, fractional anisotropy; ReHo, regional homogeneity; IL-1β, interleukin-1β; IL-6, interleukin-6.
Predictive models for executive cognition
The models based on brain connectivity and pro-inflammatory cytokine (FA-IL-1β and ReHo-IL-6) showed a better performance than the models based on brain connectivity or pro-inflammatory cytokine alone. Specifically, the combined FA in the right uncinate fasciculus, external capsule, and anterior limb of the internal capsule-IL-1β model had the best performance (the average MAE of five cognition: 1.047–3.170), followed by FA-only (the average MAE: 1.410–7.002), and IL-1β-only model (the average MAE: 1.683–9.680). Among all models, the combined ReHo in the left orbitofrontal cortex, left insula, and left putamen-IL-6 model has the lowest MAE value (the average MAE: 1.110–3.529), followed by ReHo-only model (the average MAE: 1.329–5.211), and IL-6-only model (the average MAE: 1.714–9.787) in the functional connectivity dataset. Figures 3a, b present the architecture of ANN models. Distribution of WCST scores can be found in online Supplementary Fig. S6. Online Supplementary Table S23 summarizes the prediction performance of different models. The scatter plots of predicted v. actual WCST scores are presented in Figs 3c, d and online Supplementary Fig. S7.
Fig. 3. Architecture of artificial neural network model and the prediction performance of the best fitting model for the structural and functional connectivity datasets. Panel (a) illustrates the architecture of single input artificial neural network model. For the prediction models based on brain connectivity or pro-inflammatory cytokine alone, we design a single input artificial neural network model, which consists of one input layer, two hidden layers, and one output layer. The hidden layers used 50 and 20 neurons, respectively. Panel (b) shows the architecture of multi-input neural network model. For the prediction models based on combination of brain connectivity and pro-inflammatory cytokine indicators (FA-IL-1β and ReHo-IL-6), we design a multi-input neural network model, which includes two input branches. The first input branch accepts brain connectivity data, including two hidden layers with 50 and 20 neurons. The second input branch accepts pro-inflammatory cytokines and contains a hidden layer with 10 neurons. The 30-dimentional data were concatenated by the output layer of these two branches and followed by two fully connected layers that contain five and one neuron in each layer, respectively. Panel (c) shows scatter plots of the predicted v. actual WCST scores in the structural connectivity datasets. Predicted executive cognition derived used FA values and IL-1β as input data. Panel (d) shows scatter plots of the predicted v. actual WCST scores for the functional connectivity datasets. Predicted executive cognition derived used ReHo values and IL-6 as input data. The values of MAE and coefficient of determination (R 2) are indicated in the plots. CC, categories completed; CR, correct responses; FA, fractional anisotropy; IL-1β, interleukin-1β; IL-6, interleukin-6; MAE, mean absolute error; NPE, non-perseverative errors; PE, perseverative errors; ReHo, regional homogeneity; TE, total errors; WCST, Wisconsin Card Sorting Test.
Discussion
This study has three major findings. Firstly, executive deficits were associated with decreased FA in the right uncinate fasciculus, external capsule, and anterior limb of the internal capsule; increased ReHo in the left orbitofrontal cortex, left insula, and left putamen; elevated IL-1β and IL-6 across MPDs. Interestingly, overlapping structural and functional dysconnectivity within fronto-limbic-striatal regions contributed to executive deficits. Secondly, mediation analyses showed that the association between FA and WCST scores was mediated by IL-1β; while IL-6 mediated the relationship between ReHo and WCST scores. Thirdly, executive cognition was better predicted by both brain connectivity and cytokine measures than either one alone for FA-IL-1β and ReHo-IL-6. Collectively, we found a strong association between brain connectivity in fronto-limbic-striatal regions, pro-inflammatory cytokines (IL-1β and IL-6), and executive cognition, reflecting common dimensional neurobiological substrates across MPDs. Machine learning-based models have indicated that brain connectivity-pro-inflammatory cytokines could be leveraged to produce objective and reliable biomarkers of executive deficits in MPDs.
Consistent with our hypotheses, this study revealed a transdiagnostic association among FA values in the right uncinate fasciculus, external capsule, anterior limb of the internal capsule, IL-1β, and executive deficits in patients with MPDs. Reductions in WM integrity are predominantly located within the fronto-limbic-striatal brain regions (Kiehl, Reference Kiehl2006), which are well known for their association with executive cognition in MPDs. These findings are in line with previous studies, Rocío Pérez-Iglesias et al. reported lower FA in pathways connecting cortical and subcortical regions are associated with executive deficits in patients with first-episode psychosis (Perez-Iglesias et al., Reference Perez-Iglesias, Tordesillas-Gutierrez, McGuire, Barker, Roiz-Santianez, Mata and Crespo-Facorro2010). Lower FA may represent a loss of axonal integrity, indicating myelin damage (Kochunov et al., Reference Kochunov, Coyle, Rowland, Jahanshad, Thompson, Kelly and Hong2017). Interestingly, oligodendrocytes are not only the predominant constituents of WM structures, but also the cells responsible for producing and maintaining myelin in the brain. Oligodendrocyte dysfunction triggers neuroinflammation and promotes the release of pro-inflammatory cytokine IL-1β in the periphery (Deverman & Patterson, Reference Deverman and Patterson2009; Pape et al., Reference Pape, Tamouza, Leboyer and Zipp2019). Therefore, our results suggest that lower FA values within the fronto-limbic-striatal brain regions may contribute to elevated IL-1β through oligodendroglia dysfunction. Moreover, IL-1β was found to be a mediator in the association between FA alterations and executive deficits. IL-1β is released by oligodendrocytes and microglial cells. Mounting evidence suggests that oligodendrocyte damage may cause IL-1β elevation at the level of the periphery that may positively affect executive cognition. Consistent with our data, George Bartzokis demonstrated subcortical myelin and oligodendrocyte damage as a shared mechanism in SCZ, BPD, and MDD (Bartzokis, Reference Bartzokis2012). Therefore, we can conclude that structural dysconnectivity within the fronto-limbic-striatal brain regions as well as elevated IL-1β in the peripheral blood may contribute to more severe executive deficits through oligodendrocytes dysfunction and dysmyelination in MPDs.
In parallel, we found the association among ReHo within fronto-limbic-striatal regions (left orbitofrontal cortex, left insula, and left putamen), IL-6, and executive deficits across MPDs. Orbitofrontal cortex, insula, and putamen are critical for executive cognition (Buckholtz & Meyer-Lindenberg, Reference Buckholtz and Meyer-Lindenberg2012; Kiehl, Reference Kiehl2006; Nusslock et al., Reference Nusslock, Brody, Armstrong, Carroll, Sweet, Yu and Miller2019). ReHo has been proven to be a robust psychosis biomarker (Ji et al., Reference Ji, Meda, Tamminga, Clementz, Keshavan, Sweeney and Pearlson2020). Using the same method, a recent study also found that higher ReHo values in inferior orbital frontal, middle frontal cortices, and insula were positively correlated with cognition across SCZ, BPD, and schizoaffective disorder (Ji et al., Reference Ji, Meda, Tamminga, Clementz, Keshavan, Sweeney and Pearlson2020). Higher ReHo has been found to be related to increased neuronal excitability, neuronal damage, and microglial activation (Wei et al., Reference Wei, Chang, Womer, Zhou, Yin, Wei and Wang2018). Interestingly, microglial activation might trigger inflammation in the central nervous system, leading to increased pro-inflammatory cytokine IL-6 release in the periphery (Fiala, Spacek, & Harris, Reference Fiala, Spacek and Harris2002). From this, it could be speculated that abnormalities within the fronto-limbic-striatal regions may lead to microglial activation, a marked increase in microglia-induced IL-6 expression, and correspondingly neuronal damage. We also found that IL-6 had partial mediation effect for the influence of ReHo and executive cognition. A possible explanation for this is that functional connectivity may directly impact executive cognition. Consistent with this model, previous findings suggested that abnormalities of fronto-limbic-striatal regions might contribute to executive deficits in MPDs (Chai et al., Reference Chai, Whitfield-Gabrieli, Shinn, Gabrieli, Nieto Castañón, McCarthy and Ongür2011; Clark, Chamberlain, & Sahakian, Reference Clark, Chamberlain and Sahakian2009; Edwards, Barch, & Braver, Reference Edwards, Barch and Braver2010; Price & Duman, Reference Price and Duman2020). In addition, functional dysconnectivity might be acting indirectly on executive deficits through an intermediate factor (i.e. IL-6) or other unknown factors (Barch & Sheffield, Reference Barch and Sheffield2014; D'Mello & Swain, Reference D'Mello and Swain2017; Miller et al., Reference Miller, Maletic and Raison2009). Moreover, IL-6 elevation can not only aggravate cognitive impairment directly, but also indirectly through mediating effects, which is characterized by activated microglial cells (Miller et al., Reference Miller, Maletic and Raison2009; Rudolph et al., Reference Rudolph, Graham, Feczko, Miranda-Dominguez, Rasmussen, Nardos and Fair2018). Hence, we can speculate that functional dysconnectivity within the fronto-limbic-striatal regions as well as elevated IL-6 in the peripheral blood, may contribute to more prominent executive deficits via microglial activation and neuronal damage in MPDs. IL-1β and IL-6 could influence cognitive functions via indirect action through different brain-inflammation-cognition pathways.
Intriguingly, executive cognition could be better predicted by both brain connectivity and cytokine measures than either one alone for FA-IL-1β and ReHo-IL-6 by ANN models, which account for the nature of the neurobiological substrates of executive cognition and reveal complex and mostly non-linear relationships. Thus, predictive models were applied to support our mediation results and to investigate the potential mechanism underlying the transdiagnostic associations, due to the complexity and heterogeneity of psychiatry disorders. More specifically, alterations in brain connectivity within fronto-limbic-striatal regions reflect oligodendrocytes dysfunction and microglial activation that may potentially lead to executive deficits in MPDs. These findings add to a growing literature implicating brain dysconnectivity, and inflammation in the psychopathology of executive deficits, and lend further support to the dimensional approach of executive deficits across MPDs. Previous studies have also demonstrated that ANNs might help to elucidate the inherent relationships between input and output, by analyzing the non-linear relationships among multiple variables (Bosia et al., Reference Bosia, Bechi, Bosinelli, Politi, Buonocore, Spangaro and Cavallaro2019). Recent study from Lanillos et al. supports this view. ANN models of SCZ and autism spectrum disorder could provide a novel tool to fill the gap between theoretical and biological evidence (Lanillos et al., Reference Lanillos, Oliva, Philippsen, Yamashita, Nagai and Cheng2020). From a clinical perspective, an advantage of ANN is that it identifies the number of latent variables in an unbiased, data-driven manner (Galletly, Clark, & McFarlane, Reference Galletly, Clark and McFarlane1996). We observed the proposed model of combining brain imaging data and inflammatory biomarkers may precisely capture non-linearity in the feature space and shows better performance than brain imaging or blood alone model. This indicated that models based on ANN can give a stable prediction performance and increase biological interpretability. Therefore, inflammation-related brain dysconnectivity within fronto-limbic-striatal regions may represent common dimensional substrates and neurobiological mechanisms for executive deficits that could be an objective predictor and a potential therapeutic target for executive deficits in patients with MPDs.
This study has several limitations. First, the cross-sectional design limits our interpretation of causal relationships. Future longitudinal research is required to define the causal relationship and neurobiological mechanisms of executive deficits. Second, we used a multivariate database, including clinical characteristics, neuroimaging, and blood data. However, the ideal database should integrate comprehensive multi-level information (detailed inpatient or outpatient information) from each subject. We did not have an executive function composite score in this study. Future studies will focus on this aspect. Another limitation is the presence of missing data, especially during the collection of blood samples. Future works should incorporate blood data and replicate the results presented herein. Third, our study is limited by possible confounding effects from medication and illness duration. Approximately 60% of patients with MPDs were taking psychotropic medications. Although medication did not have an effect on the results, one cannot rule out the confounding effect of medication use. After adding illness duration as an additional covariate, the mediation effects were not significant, indicating that possible variations in illness duration should be longitudinally investigated. The relatively wide age range also limits this study. Future better-designed prospective studies are required to understand the aberrant developmental trajectory of brain connectivity at different developmental stages across MPDs. More comprehensive information should be collected to confirm our findings.
In conclusion, we demonstrated transdiagnostic associations among brain connectivity in fronto-limbic-striatal regions, inflammation, and executive cognition across MPDs. These results may provide novel insights into common neurobiological substrates and mechanisms for executive deficits. Our findings suggest that inflammation-related brain dysconnectivity within fronto-limbic-striatal regions may serve as a potential predictor and a promising avenue for therapeutic interventions in patients with MPDs.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291722000174
Acknowledgements
This work was supported by research grants from the National Science Fund for Distinguished Young Scholars (F.W., grant number 81725005) the National Natural Science Foundation of China (F.W., grant number 81571331; Z.Z., grant number 62176129; Y.T., grant number 81571311, 81071099 and 81271499), National Key Research and Development Program (F.W., grant number 2016YFC0904300), National High Tech Development Plan (863) (F.W., grant number 2015AA020513), and China Postdoctoral Science Foundation (Y. W., grant number 2021M691643).
Conflict of interest
None.
