Characterization of Hemodynamic Alterations in Schizophrenia and Bipolar Disorder and Their Effect on Resting-State fMRI Functional Connectivity

Abstract

Common and distinct neural bases of Schizophrenia (SZ) and bipolar disorder (BP) have been explored using resting-state fMRI (rs-fMRI) functional connectivity (FC). However, fMRI is an indirect measure of neural activity, which is a convolution of the hemodynamic response function (HRF) and latent neural activity. The HRF, which models neurovascular coupling, varies across the brain within and across individuals, and is altered in many psychiatric disorders. Given this background, this study had three aims: quantifying HRF aberrations in SZ and BP, measuring the impact of such HRF aberrations on FC group differences, and exploring the genetic basis of HRF aberrations. We estimated voxel-level HRFs by deconvolving rs-fMRI data obtained from SZ (N = 38), BP (N = 19), and matched healthy controls (N = 35). We identified HRF group differences (P < .05, FDR corrected) in many regions previously implicated in SZ/BP, with mediodorsal, habenular, and central lateral nuclei of the thalamus exhibiting HRF differences in all pairwise group comparisons. Thalamus seed-based FC analysis revealed that ignoring HRF variability results in false-positive and false-negative FC group differences, especially in insula, superior frontal, and lingual gyri. HRF was associated with DRD2 gene expression (P < .05, 1.62 < |Z| < 2.0), as well as with medication dose (P < .05, 1.75 < |Z| < 3.25). In this first study to report HRF aberrations in SZ and BP, we report the possible modulatory effect of dopaminergic signalling on HRF, and the impact that HRF variability can have on FC studies in clinical samples. To mitigate the impact of HRF variability on FC group differences, we suggest deconvolution during data preprocessing.

Introduction

Although schizophrenia (SZ) and bipolar disorder (BP) have long been viewed as distinct disorders based on differing clinical presentations; ^1–3 there is substantial evidence suggesting that these disorders share neuropsychological features⁴ and cognitive dysfunction.^5,6 SZ is a psychotic disorder characterized by altered perception, thought processes, and behaviors⁷; while BP is a mood disorder involving prolonged states of depression and mania.⁸ A popular noninvasive technique for characterizing the neural bases of SZ and BP has been resting-state functional magnetic resonance imaging (rs-fMRI) based on the blood-oxygen-level dependent (BOLD) contrast. A primary reason for the popularity of rs-fMRI is that it does not require the subjects to perform an explicit task, which can be difficult in clinical populations. Rs-fMRI also allows us to investigate interactions between neurophysiological events in spatially remote brain regions by assessing the temporal correlation of their respective BOLD signals (i.e. functional connectivity [FC]),⁹ which can characterize large-scale brain networks both in healthy and clinical populations. To date, several FC studies have attempted to distinguish the neural aberrations in SZ from BP, but the results have been inconsistent. While most studies have identified more extensive disconnection in SZ than BP^10,11,12 regional diagnostic specificity has not been evident on a consistent basis For example, cingulate and thalamic FC has been reported to be abnormal in both patients groups^13–15 in some but not other studies^16–18 Nevertheless, functional connectivity patterns have been considered to hold great promise in differentiating the various disorders in the psychosis continuum^19,20 and resting-state FC studies are commonplace due to the ease of pooling large datasets, as well as the lack of task performance, effort, or sensory processing confounds when investigating psychosis.

Interpreting findings from rs-fMRI FC is not straightforward given that fMRI BOLD contrast is an indirect measure of neural activity.^21–23 The measured fMRI signal is considered as a convolution of latent neural activity with a transfer function – the hemodynamic response function (HRF).²⁴ As such, the fMRI signal contains variance arising from the variability of the HRF as well. The issue, however, is that HRF shape has been shown to vary not only across subjects and populations, but also across different brain regions within the same subject.^25,26 Different factors contribute to this variability. First, it is partly driven by factors affecting neurovascular coupling including neurotransmitters (e.g. γ-aminobutyric acid (GABA), glutamate, dopamine, serotonin, and nitric oxide)²⁷ and components of the neurovascular unit (glial cells, endothelia), both of which are affected by genetic factors.²⁸ Second, nonneural factors such as baseline cerebral blood flow, vasculature differences, caffeine/alcohol/lipid ingestion, partial volume imaging of veins, hematocrit, pulse or respiration differences, global magnetic susceptibilities, and slice timing differences^29–31 can also affect the HRF.

The vascular hypothesis in psychosis suggests disruption of appropriate vascular response to changes in cerebral metabolic activity,^32,33 possibly due to a dysfunction of the neurovascular unit or neurotransmitters that are vasoactive (e.g. dopamine^34–36,37). Neurovascular uncoupling may result in insufficient support for synaptic activity, which in turn may trigger synaptic loss or promote an inflammatory response,³⁸ induce mitochondrial dysfunction and oxidative stress leading to neuronal atrophy.³⁹ Characterizing HRF variability provides an opportunity to systematically study the putative neurovascular uncoupling. Furthermore, HRF variability could confound FC estimates obtained from fMRI data, as a time series pair may appear to be synchronized when the underlying latent neural signals are not and vice versa^21,40,41 (figure 1). Thus systematic FC aberrations differentiating clinical groups (SZ and BP) could be driven by factors affecting neurovascular coupling and hence HRF shape.^22,42–44

Fig. 1.

The effect of HRF variability on Pearson’s correlation between two-time series obtained from experimental fMRI data. The measured fMRI time series, the estimated HRF and the deconvolved fMRI (or latent neural) time series are shown in the top, middle, and bottom rows respectively. We specifically illustrate two scenarios: (A) The acquired fMRI time series are highly correlated while the correlation between underlying latent neural signals is low (after minimizing HRF variability), and (B) The acquired fMRI time series are uncorrelated while the correlation between underlying latent neural signals is high (after minimizing HRF variability).

Deconvolution methods are capable of estimating voxel-specific HRFs from the measured fMRI signal, and subsequently recovering the latent neural time series.^24,45 The procedure is relatively straightforward for fMRI task paradigms because the timing of external events driving neural activity is known.^46,47 For rs-fMRI, which lacks explicit timing of neural events,^48,49 Wu et al.⁴⁸ proposed a deconvolution method to identify pseudo-events from rs-fMRI time series, and estimate HRF and latent neural time series using it. We have used this method in the current work. Voxel-specific HRFs were estimated from 3D+time fMRI data (after completing all the preprocessing steps), which were characterized by three shape parameters³⁸: time-to-peak (TTP), response height (RH), and full-width at half-max (FWHM) (see figure 2).

Fig. 2.

Typical HRF with its three characteristic shape parameters.

Response Height; Time-To-Peak; Full-Width At Half Max

This work aimed to address three important questions. (1) The first aim was to identify brain regions with systematic HRF differences between SZ, BP, and matched healthy controls, wherein the neurovascular coupling mechanisms are likely to differ. (2) The second aim was to identify possible confounds introduced by HRF variability in identified functional networks, for which we compared FC estimated from deconvolved (HRF-variability-minimized) and nondeconvolved (HRF-variability ignored) data. (3) The third aim was to understand the relationship between dopamine system (especially D2 receptor) that is therapeutically highly relevant for SZ and BP, by studying the DRD2 gene expression as well as the effect of antipsychotic dose on the HRF aberrations. Probing these questions has been undertaken for the first time in the literature, with potential impact on better mechanistic understanding of shared and distinguishing neurovascular coupling mechanisms of SZ and BP and reduced confounds during studying functional brain networks in these disorders.

Materials and Methods

Participants

The subject sample consisted of 38 patients satisfying DSM-IV criteria for SZ, 19 patients having BP with psychosis, and 35 matched healthy controls. Details about the sample and recruitment process are provided in supplementary sections S1 and section S11 and supplementary table S9. 54 out of 57 patients were receiving psychotropic medications at the time of scan. We calculated the median Defined Daily Dose (DDD) for antipsychotics.⁵⁰ A detailed breakdown of the prescribed medications is presented in supplementary section S11 and supplementary table S8.

fMRI Data Acquisition

Subjects were scanned in a 3T Philips Achieva MRI scanner (Philips, the Netherlands), using an 8-channel SENSE head coil with SENSE factor 2 in anterior–posterior direction. To enhance sensitivity, dual-echo gradient-echo echo-planar images (GE-EPI) were acquired.⁵¹ The sequence parameters are provided in supplementary section S2.

fMRI Data Preprocessing

After following a standard resting state fMRI preprocessing procedure (details in Supplementary Section S3), the pipeline was split in two. In the first pipeline, no further processing was pursued and the preprocessed data was designated as the nondeconvolved (NDC) dataset. In the second pipeline, voxel-level HRF deconvolution⁴⁸ was performed and the resulting data (voxel-level latent neural time series) was designated as the deconvolved (DC) dataset. Finally, both DC and NDC datasets were temporally bandpass filtered (0.01–0.1 Hz).

Deconvolution and HRF Estimation

We applied a blind deconvolution technique developed for rs-fMRI by Wu et al.⁴⁸ to retrieve latent neural signals as well as estimate voxel-specific HRFs. This method is based on the increasing evidence that the resting-state BOLD spikes can be seen as the response to spontaneous neuronal events and the dynamics of the brain at resting state is governed by nonrandom patterns. Spontaneous neural events were detected by identifying BOLD fluctuations of relatively large amplitude (with a threshold of mean + 1 SD, after despiking and removing sources of noise that might lead to spurious spikes) and modeling spontaneous neuronal events as pseudo-events in the rs-fMRI time series (using a point process model).^52,53 A mask with relative head motion <0.3 mm was imposed to further avoid motion-driven events.⁴⁸ The estimated neural events were convolved with the canonical HRF with different values of RH, TTP, and FWHM varying within their biologically plausible range, and the HRF resulting in the best fit (in the least squares sense) between the resulting time series and experimental data was chosen as the final estimated HRF for each voxel. Finally, latent neural time series were estimated from the HRF and raw BOLD signal using Wiener deconvolution.²⁴ As shown in figure 2, the HRF shape is characterized by RH, TTP, and FWHM, representing the amplitude, timing, and spread of the response, respectively. More details are provided in Supplementary Section S4.

HRF Differences Between Schizophrenia, Bipolar Disorder, and Healthy Controls

We applied a blind deconvolution technique developed for rs-fMRI by Wu et al.^48,52,53 to retrieve latent neural signals as well as estimate voxel-specific HRFs. Further details about deconvolution and HRF estimation are provided in Supplementary Section S5. To address the first aim of identifying impaired neurovascular coupling in SZ and BP, we compared HRF parameters between the groups. For each of the three HRF parameters, two sample two-tailed t-tests were conducted for three comparisons separately: SZ vs control, BP vs control, and SZ vs BP. With each comparison, whole-brain maps that indicate voxels showing significant (FDR corrected P-value < .05, cluster-size > 50) group differences in three HRF parameters were obtained. This resulted in 6 maps (SZ < control, SZ > control, BP < control, BP > control, SZ < BP, and SZ > BP) per HRF parameter and 18 maps overall. Finally, we sought to identify voxels that exhibited significantly different HRFs across all three pairwise group comparisons (with at least one HRF parameter). This relied on a conjunction approach that is more stringent than the group comparison maps, as the threshold required passing 3 levels of FDR correction, one for each comparison.

We acknowledge that FDR is an adaptive approach that is seen as more liberal than family-wise error correction methods for type 1 errors. We chose FDR correction over other familywise error correction methods as the problem of interest (differences in HRF parameters) demands avoidance of false negatives over the acceptance of false positives.⁵⁴ Furthermore, this approach avoided the issue of arbitrary cluster inclusion⁵⁵ when studying the three HRF properties whose spatial properties (“ground truth”) are not known a priori. Nevertheless, to ensure the inadvertent false positives are kept minimal, we used a cluster extent threshold as well (50 voxels) that takes into account the possible dependencies among adjacent voxels.

Impact of HRF Variability on Seed-Based Functional Connectivity Group Differences

To address the second aim of identifying the effect of HRF variability on group differences in FC, we compared seed-based FC across the groups using DC and NDC datasets. Specifically, we identified the seed region(s) from our earlier analysis as that whose HRF shape (i.e. at least one of the three HRF parameters) was different across all the three groups (only the thalamus was found from that analysis, which was defined as the seed region of interest [ROI]). For each participant, the mean time series within the thalamus seed ROI was obtained. FC between the thalamus seed ROI time series and the remaining gray matter voxels in the brain were evaluated using Pearson’s correlation coefficient. The correlation coefficients were transformed by Fisher’s z-transform to ensure Gaussianity for statistical comparisons.^56,57 This procedure for obtaining FC maps was implemented separately for the two datasets: (1) NDC: data processed without deconvolution, and (2) DC: data processed with deconvolution.

For each group, the FC maps from individual subjects were fed to a random effects one-sample t-test (FDR corrected P-value < .05, cluster-size > 50) to determine voxels whose connectivity with the thalamus seed were significantly different from zero. Subsequently, the FC maps from individual subjects were also fed to a random effects two-sample t-test (FDR corrected P-value < 0.05, cluster-size > 50) to identify brain regions showing significantly different FC with the thalamus seed ROI in three comparisons separately (SZ vs control, BP vs control, and SZ vs BP). These procedures were also implemented separately for both DC and NDC datasets.

With the aim of investigating the effect of HRF variability on between-group differences in FC, we performed a two-way repeated-measures ANOVA using SPSS software (version 20, IBM Inc., USA). Here the groups (SZ, BP, and Control) were considered as one factor and data with/without deconvolution as the other factor. We then identified the voxels showing a significant interaction between the two factors (FDR corrected P-value < 0.05, cluster size > 50 voxels).

Association Between HRF Parameters and Gene Expression

The Allen Human Brain Atlas of gene expression has been mapped to⁵⁸ Freesurfer’s Desikan–Killianey (DK) whole-brain parcellation with 62 ROIs,⁵⁹ allowing for the exploration of associations between gene expression and imaging measures. To address our third aim of understanding the relationship between gene expression and HRF aberrations in these disorders, we probed the associations between DK parcellation gene expression map and the maps of HRF parameters in the DK parcellation space. For this, we found mean HRF parameters among all the voxels within each DK parcellation ROI, giving us one RH value (and one TTP and one FWHM value) for each DK-ROI in each subject. We subsequently obtained group difference maps for the HRF parameters for three group comparisons: BP vs Control, SZ vs Control, and BP vs SZ, in DK-space.

We then examined the association between the HRF difference values from each NDC-DC data pair and gene expressions of Dopamine Receptor D2 (DRD2) gene, which encodes the D2 subtype of the dopamine receptor, highly relevant to the pathophysiology as well as treatment response in schizophrenia^60–63 and BP.⁶⁴ As a control for the gene expression – HRF relationship, we studied the expression of the Polybromo 1 (PBRM1) gene which encodes a subunit of ATP-dependent chromatin-remodeling complexes and implicated in both BP and SZ^65,66 but no known direct effects on the vasoactive neuromodulators.

We specifically chose DRD2 expression as regional cerebral blood flow changes that occur on administration of antipsychotics have been shown to be directly proportional to brain D2R densities and DRD2 mRNA expression measures (Selvaggi et al.³⁴).

With respect to the gene expression analysis, our aim was to demonstrate that HRF changes may not be on the primary mechanistic pathways that lead to the clinical expression of schizophrenia or bipolar disorder; but neuromodulators such as dopamine, which are directly relevant to the treatments used in practice, can affect the HRF. We did not have any specific hypothesis on DRD2’s effects being higher than the other genetic loci; thus, we do not study or report the interaction effects (i.e. difference in strength of slopes between 2 genetic loci). To this end, we selected a control locus that has susceptibility signal for multiple psychiatric disorders (see Cross-Disorder Group of the PGC, 2014⁶⁷; Liu et al.⁶⁸ and Yang et al⁶⁹; also see Golovino⁷⁰), but not directly linked to dopamine or other neuromodulators, exhibit low endothelial profile in line with DRD2 (based on cerebral cortex tissue expression summary from proteinatlas.org) despite having established cortical and thalamic expression profile (based on Allen Human Brain Tissue Gene Expression Profiles).

Relationship Between HRF Parameters and DDD of Medications

Finally, we probed the association between DDD of medications with HRF parameters. Specifically, we probed DDD values for antipsychotics in SZ. The association between DDD measures and the three HRF parameters at each voxel were assessed. Significant clusters are reported (P < 0.05 FDR corrected, cluster size > 50).

Results

Between-Group HRF Differences

In the SZ vs control comparison, we found that bilateral thalamus, midbrain, and precuneus had significantly wider FWHM (figure 3A) in the SZ group, while bilateral cerebellum posterior lobe, left middle frontal gyrus (MFG), and bilateral precentral gyrus had significantly wider FWHM (figure 3A) in the control group. Right ACC and right parahippocampal gyrus had significantly higher RH (figure 3A) in the SZ group, while bilateral cerebellum posterior lobe, posterior cingulate cortex (PCC), precuneus, and lingual gyrus had significantly higher RH (figure 3A) in the control group. Bilateral midbrain had significantly longer TTP (figure 3A) in the SZ group, while bilateral cerebellum posterior lobe had significantly longer TTP (figure 3A) in the control group.

Fig. 3.

Spatial maps showing regions with significantly different HRF parameters in the three comparisons: (A) schizophrenia (SZ) vs control, (B) bipolar disorder (BP) vs control, and (C) BP vs SZ.

In the BP vs control comparison, we found that bilateral ACC and precuneus had significantly wider FWHM (figure 3B) in the BP group, while bilateral cerebellum posterior lobe, left middle temporal gyrus (MTG) and bilateral precentral gyrus had significantly wider FWHM (figure 3B) in the control group. Bilateral ACC, bilateral cingulate gyrus, bilateral mPFC, and right supramarginal gyrus had significantly higher RH (figure 3B) in the BP group, while bilateral cerebellum posterior lobe, bilateral cerebellum anterior lobe, bilateral thalamus and bilateral mPFC had significantly higher RH (figure 3B) in the control group. Bilateral cingulate gyrus had significantly longer TTP (figure 3B) in the BP group, while bilateral cerebellum posterior lobe, left MTG, and bilateral MFG had significantly longer TTP (figure 3A) in control group.

In the BP vs SZ comparison, we found that bilateral cerebellum posterior lobe had significantly wider FWHM (figure 3C) in the BP group, while bilateral thalamus, cerebellum anterior lobe, and midbrain had significantly wider FWHM (figure 3C) in the SZ group. Bilateral mPFC, left lingual gyrus, bilateral PCC, and bilateral inferior parietal lobule had significantly higher RH (figure 3C) in the BP group, while bilateral parahippocampal gyrus and bilateral superior frontal gyrus (SFG) had significantly higher RH (figure 3C) in the SZ group. Bilateral cerebellum posterior lobe had significantly longer TTP (figure 3C) in the BP group, while bilateral pons and thalamus had significantly shorter TTP (figure 3C) in the BP group.

These results support the notion of HRF aberrations (and thus neurovascular coupling impairments) in SZ and BP, and provide a mechanistic map of such aberrations. These results also illustrate certain distinct HRF aberrations between SZ and BP populations. Detailed information such as cluster sizes, cluster centroids etc. are shown in table 1.

Table 1.

Voxel Clusters that had Significantly Different (FDR Corrected P-Value < .05, Cluster Size > 50 Voxels) HRF Parameters in Three Comparisons: BP vs Control, SZ vs Control, and BP vs SZ. The Anatomical Labels, MNI Coordinates of Cluster Centroids, Cluster Size in Terms of the Number of Voxels, T Score, and Corresponding Brodmann Areas (Where Appropriate) Are Listed

		Coordinates
Comparison	Cluster Anatomical Location	x	y	Z	Cluster Size	Hemisphere	BA	T Score
Full-width at half max SZ > Control	Thalamus	−20	−16	0	4678	Bilateral		4.19
	Midbrain	0	−32	−14				3.91
	Precuneus	14	−56	32			31	3.52
Full-width at half max SZ < Control	Cerebellum Posterior Lobe	−62	−40	−12	4352	Bilateral	21	−4.47
	Middle Frontal Gyrus	−36	44	26	208	Left	10	−4.84
	Precentral Gyrus	18	−28	68	320	Right	4	−3.91
	Precentral Gyrus	−20	−28	66	715	Left	4	−4.48
Response height SZ > Control	Anterior Cingulate Cortex	18	36	18	104	Right	32	4.15
	Parahippocampal Gyrus	20	−4	−28	326	Right	28	4.71
Response height SZ < Control	Cerebellum Posterior Lobe	−30	−64	−50	2378	Bilateral		−3.96
	Posterior Cingulate	18	−60	4	527	Bilateral	30	−4.44
	Precuneus	−6	−52	30	1236	Bilateral	31	−3.85
	Lingual Gyrus	−4	−76	0			18	−3.55
Time-to-peak SZ > Control	Midbrain	2	−30	−14	540	Bilateral		3.70
Time-to-peak SZ < Control	Cerebellum Posterior Lobe	−26	−74	−50	1850	Bilateral		−5.08
Full-width at half max BP > Control	Anterior Cingulate Cortex	4	30	−8	1752	Bilateral	32	3.91
	Precuneus	−18	−46	30	329	Left	31	4.12
	Precuneus	14	−54	32	999	Right	31	4.57
Full-width at half max BP < Control	Cerebellum Posterior Lobe	−40	−54	−50	1229	Bilateral		−3.28
	Middle Temporal Gyrus	−62	−40	−12	86	Left	21	−3.58
	Precentral Gyrus	−18	−32	70	4352	Right	4	−4.69
	Precentral Gyrus	14	−38	72		Left	3	−3.93
Response height BP > Control	Cingulate Gyrus	−16	4	30	167	Left		3.66
	Cingulate Gyrus	−16	−26	40	228	Left	31	3.92
	Cingulate Gyrus	16	−58	28	96	Right	31	4.27
	Medial Prefrontal Cortex	4	−10	66	624	Bilateral	6	4.01
	Supramarginal Gyrus	42	−48	34	438	Right	40	3.92
Response height BP < Control	Cerebellum Posterior Lobe	2	−58	−48	524	Bilateral		−4.03
	Cerebellum Anterior Lobe	20	−42	−26				−3.59
	Cerebellum Anterior Lobe	−20	−40	−26				−3.39
	Thalamus	4	−20	4	80	Bilateral		−4.65
	Medial Prefrontal Cortex	−2	−10	66	481	Bilateral	6	−3.22
Time-to-peak BP > Control	Cingulate Gyrus	−10	12	34	262	Left	24	3.72
	Cingulate Gyrus	10	−6	38	150	Right	24	4.29
Time-to-peak BP < Control	Cerebellum Posterior Lobe	−10	−66	−50	909	Bilateral		−3.29
	Middle Temporal Gyrus	−36	2	−50	59	Left	20	−3.25
	Middle Frontal Gyrus	−26	−6	60	1157	Left	6	−4.12
	Middle Frontal Gyrus	32	−8	60	231	Right	6	−3.88
Full-width at half max BP > SZ	Cerebellum Posterior Lobe	−22	−80	−44	267	Left		4.36
	Cerebellum Posterior Lobe	16	−74	−50	81	Right		4.02
Full-width at half max BP < SZ	Thalamus	−6	−24	0	1126	Bilateral		−4.63
	Cerebellum Anterior Lobe	−16	−32	−18				−3.91
	Midbrain	−6	−32	−20				−3.52
Response height BP > SZ	Medial Prefrontal Cortex	10	36	−20	196	Bilateral	11	4.26
	Lingual Gyrus	−12	−90	−2	477	Left	17	4.21
	Posterior Cingulate	−4	−60	18			23	3.84
	Posterior Cingulate	18	−66	8	280	Right	30	3.76
	Inferior Parietal Lobule	40	−40	26	358	Right		4.13
	Inferior Parietal Lobule	−42	−40	24	527	Left		4.24
Response height BP < SZ	Parahippocampal Gyrus	34	−6	−26	231	Right		−3.81
	Parahippocampal Gyrus	−16	−14	−26	103	Left	28	−4.06
	Medial Prefrontal Cortex	−8	−12	68	218	Bilateral	6	−4.26
Time-to-peak BP > SZ	Cerebellum Posterior Lobe	−32	−78	−42	1193	Bilateral		3.56
Time-to-peak BP < SZ	Pons	−16	−26	−30	997			−3.96
	Thalamus	−8	−24	0				−3.71
Intersection	Thalamus	2	−22	4	101	Bilateral

Coordinates indicate location of maximum Z-scores for clusters or location of local maxima. For clusters with more than one peak, local maxima are listed. Cluster size in voxels. Coordinates are in standard MNI space. BA, Brodmann Area.

We found that a thalamic cluster showed significant differences in at least one HRF parameter in all 3 comparisons: BP vs control, SZ vs control, and BP vs SZ. (figure 4A). Note that in the identified thalamus seed ROI, the differences in HRF parameters exhibited the following pattern – FWHM: SZ > Control, RH: BP < Control, and FWHM: SZ > BP. The overlap of our thalamic seed (shown in white, figure 4B) with thalamic zones as described in the Oxford thalamic atlas based on diffusion tractography⁷¹ revealed that our seed had the highest probability of structural connectivity with the temporal and prefrontal cortex. By registering our seed ROI to the digital 3D Morel atlas,^72,73 we found that the thalamic seed ROI encompassed the mediodorsal (MD) nucleus, habenular (Hb) nucleus, and the central lateral (CL) nucleus (figure 4C).

Fig. 4.

Thalamus seed. (A) Voxels within the thalamus that showed significant differences in all three HRF parameters in the following three comparisons: BP vs control, SZ vs control, and BP vs SZ. Sagittal (x = 2), and axial (z = 4) views are shown in standard MNI space. In the thalamus seed ROI, alterations in HRF parameters exhibited the following pattern: FWHM: SZ > control; RH: BP < control, and FWHM: SZ > BP. (B) The overlap of our thalamic seed (shown in white) with thalamic zones as described in the Oxford thalamic atlas.⁷¹ (C) The overlap of our thalamic seed (shown in white) with thalamic nuclei described in the Morel thalamic atlas.^72,73 The seed encompassed the mediodorsal nucleus, habenular nucleus, and the central lateral nucleus.

Group Differences in Seed-Based Thalamic Functional Connectivity

Within group FC patterns using the thalamic seed as well as group differences in thalamic FC patterns were assessed (SZ vs controls, BP vs controls, and SZ vs BP) separately using DC and NDC datasets. The corresponding maps and regional statistics have been provided in Supplementary Section S6. A summary of those observations has been depicted in figure 5.

Fig. 5.

Summary of between-group differences in seed-based thalamic functional connectivity for (A) deconvolved, and (B) nondeconvolved datasets. The regions in red boxes displayed functional connectivity differences between groups that varied by whether the data was deconvolved or not. (C) The 3D positions of involved regions. Note that the Insula is shown as an interior region within a semitransparent rendering of the brain.

Impact of HRF Variability on Group Differences in Functional Connectivity

Voxel clusters in bilateral lingual gyrus, bilateral insula, and bilateral SFG showed a significant interaction effect between the group factor (control vs BP vs SZ) and the deconvolution factor (data with and without deconvolution) (figure 6, table 2). This implied that seed-based FC differences between the three groups (SZ, BP, and controls) in these regions were influenced by HRF variability, and thus resulted in differing group differences with DC and NDC datasets.

Fig. 6.

Brain regions showing significant interaction between the group factor (control vs BP vs SZ) and the deconvolution factor (data with and without deconvolution).

Table 2.

Anatomical Labels, Cluster Size, and P-Value of Local Maxima in Brain Areas Showing Significant Interaction Between the Group Factor (Control vs BP vs SZ) and the Deconvolution Factor (Data With and Without Deconvolution)

	Peak Coordinates
Cluster Anatomical Location	x	y	z	Cluster Size	Hemisphere
Lingual Gyrus	−8	84	−4	728	Bilateral
Insula	−30	−28	12	42	Left
Insula	30	−18	22	190	Right
Superior Frontal Gyrus	−24	20	54	45	Left
Superior Frontal Gyrus	28	44	10	41	Right

Coordinates referring to the peak of the cluster in MNI Space.

Anatomical labels based on Automated Anatomical Labeling (AAL).

Relationship Between HRF Parameters and Gene Expression

In order to understand the basis of HRF group differences, our third aim was to assess the association between gene expression (DRD2 and PBRM1 genes) and the significant differences observed in HRF parameters. Figure 7 shows the difference value of HRF parameters between the groups (BP-minus-control, SZ-minus-control, and BP-minus-SZ) in DK-space, which were used to test associations with gene expressions. Figure 8 shows the gene expression maps for DRD2 and PBRM1 genes. The association of gene expression values with HRF difference values is presented in table 3. We found that DRD2 gene expression was significantly positively associated with RH and TTP difference values for SZ-minus-control, and with RH difference value for BP-minus-SZ (table 3). No significant associations with the PBRM1 gene expression were observed.

Fig. 7.

The difference value in HRF parameters (RH, FWHM, and TTP) between the groups. The difference values have been normalized across the whole brain (range 0–1). Figure shows HRF difference value for: (A) BP-minus-control, (B) SZ-minus-control, (C) BP-minus-SZ.

Fig. 8.

Gene expression map of the two genes under consideration (normalized to the range 0–1). (A) DRD2 gene expression. (B) PBRM1 gene expression.

Table 3.

The Association (R-Value and Corresponding P-Value) of Regional HRF Parameters with DRD2 Gene and PBRM1 Gene expressions. Before calculating the association, mean group difference maps for the HRF parameters for three group comparisons: BP vs Control, SZ vs Control, and BP vs SZ, were obtained in DK-space. We then examined the association between the mean HRF difference values from each NDC-DC data pair and gene expressions of DRD2 and PBRM1 genes by using corresponding values across all ROIs as the sample

	DRD2(R-Value/P-Value)			PBRM1 (R-Value/P-Value)
	RH	TTP	FWHM	RH	TTP	FWHM
BP-Control	−0.02/.87	0.15/.24	0.04/.74	−0.16/.21	−0.19/.14	−0.17/.20
SZ-Control	0.35/.005*	0.31/.015*	0.22/.09	−0.15/.25	0.03/.80	−0.01/.91
BP-SZ	−0.29/.02*	0.14/.28	0.17/.19	0.005/.99	0.20/.12	0.17/.20

*Significant associations are labeled in bold letters.

Relationship Between HRF Parameters and DDD of Medications

In the SZ group, DDD of antipsychotics was significantly associated with RH in left mPFC (figure 9A, supplementary table S7 in supplementary section S10), with FWHM in right cingulate gyrus and right MFG (figure 9B, supplementary table S7), and with TTP in cerebellum posterior lobe and occipital lobe (figure 9C, supplementary table S7).

Fig. 9.

Brain regions showing significant association between DDDantipsych and HRF parameters (FWHM, RH, and TTP) in the SZ group.

Discussion

This study had three aims: (1) quantifying neurovascular coupling impairments (using the HRF) in SZ and BP, (2) measuring the impact of such neurovascular coupling impairments (i.e. HRF variability as a confound) on the commonly used measure of functional connectivity, and (3) understanding the role of dopamine receptor (DRD2) gene expression and D2-blocking agents on the neurovascular coupling parameters. To this effect, we identified HRF differences between SZ, BP, and control groups, and identified the significant impact of HRF variability on FC group differences hitherto attributed to disease related disruptions of neural level statistical dependencies. In particular, both patient groups had aberrant HRF parameters affecting the thalamus, but differing in degree between the 2 disorders. We also uncovered the likely influence of DRD2 gene expression as well as psychotropic medication use on HRF parameters, highlighting the importance of accounting for HRF variations while estimating functional connectivity in clinical samples.

HRF Results in Relation to Neurochemicals

HRF variability across brain regions and subjects due to nonneural factors tend to be more randomly distributed^{25,26,74–76} and are expected to cancel out when comparing different populations. On the other hand, pathological neurochemical alterations are expected to introduce more systematic differences in the HRF between groups. In the current study, we show that D2 receptor expression, but not PBRM1 expression, relates to HRF parameters as predicted. Unlike DRD2, the PBRM1 gene has no known effect on neurotransmitters that modulate the HRF. Furthermore, higher D2 blockade in patients as indexed by higher DDD of antipsychotics related to reduced HRF response height in cingulate cortex, in line with prior reports of D2 receptors being responsible for a decrease of hemodynamic response.³⁵ While we selected a single gene expression profile with therapeutic relevance to both SZ and BP, multiple other neurochemicals such as GABA, serotonin, glutamate, and nitric oxide (NO) also control neurovascular coupling.²⁷Figure 10 shows a schematic of these putative neurovascular coupling mechanisms.

Fig. 10.

A schematic illustrating neurovascular coupling (the hemodynamic coupling between neural activity and blood flow). The HRF is a mathematical function which represents this neurovascular coupling mechanism. Aberrations in neurochemicals in SZ and BP subjects, which may impact neurovascular coupling and hence HRF shape, are shown. Abnormalities in GABA are shown as 1, dopamine as 2, nitric oxide as 3, and glutamate as 4.

HRF Differences Across All Three Groups in the Thalamus

We detected a cluster of voxels within the thalamus (figure 4A) that showed significant HRF differences across all three groups. According to the Morel thalamic atlas,^72,73 this thalamic sub-region encompasses MD, Hb, and CL nuclei (figure 4C). Kumar et al⁷⁷ proposed a thalamic parcellation based on both rs-fMRI FC and structural connectivity of the thalamus. They compared their parcellations with the Morel atlas as well.^72,73 They reported that their parcellated thalamic clusters that overlapped with MD and CL thalamic nuclei were functionally connected at rest with the occipital and parietal lobes, the temporoparietal junction, pre- and post-central gyrus, cuneus, inferior, and medial temporal lobes, the ACC, and parts of the medial and lateral prefrontal cortices, in line with our reported observations. Previous studies have also reported aberrations in thalamic connectivity in SZ^{78,12,79–81} and BP^{12, 82} similar to those reported here. Taken together, these patterns of connectivity between the MD, CL, and Hb thalamic nuclei and the rest of the brain, as well as the observed aberrations involving the thalamus in SZ and BP disorder are in line with prior reports.

Impact of HRF Variability on Group Differences in Functional Connectivity

It is noteworthy that group differences in FC were different between deconvolved and nondeconvolved data (see figure 6 in main text as well as supplementary figures S1–S3). Although mostly similar overarching clusters were identified with both DC and NDC data, the size, peak coordinates, and corresponding statistics were different. On some instances, certain between-group FC differences were observed only in one of the datasets (DC or NDC). We found significant interaction effects in the FC paths between thalamus seed and the following regions: bilateral lingual gyrus, bilateral insula, and bilateral SFG (figure 6, table 2). This implies that FC differences across groups between the thalamic seed and these regions would be inferred differently in the NDC and DC datasets. This finding suggests that if FC results are to be interpreted purely in terms of underlying neural processing rather than in terms of confounding nonneural and neural (viz. neurovascular signaling mechanisms) components affecting the HRF, then it would be preferable to characterize FC using the latent neural time series (i.e. deconvolved data) rather than nondeconvolved fMRI data. In regions where the interaction effect was not significant, the alteration in numerical FC values due to HRF variability may impact analyses at the single-subject level, including (but not limited) to machine learning analyses used in precision psychiatry.

This result must be viewed in light of other reports²¹ suggesting that ignoring HRF variability can cause a mean error of 14.7% in connectivity estimates, with a tenth of the connections confounded by over 33%. Also, consistent with Rangaprakash et al's study,²¹ we found more false positives than false negatives if HRF variability were ignored (i.e. comparing NDC and DC datasets, more clusters and larger cluster sizes in within-group comparisons were found with the former; supplementary figures S1–S3). NDC dataset (that ignored HRF variability) was also generally less sensitive to group differences (i.e., fewer clusters and smaller cluster sizes were seen in most cases in between-group comparisons with NDC dataset, as compared to DC dataset; supplementary figures S4–S6). Given the evidence from previous studies^{21,22,42–44} and evidence presented in the current study, we suggest that: (1) HRF variability cannot be ignored while performing fMRI connectivity modeling as it confounds within-group FC estimates as well as between-group FC differences; (2) HRF variability can be minimized by performing a deconvolution step after completing all other preprocessing steps and prior to postprocessing steps such as connectivity modeling.

Among the limitations of this study were the fact that we investigated the effect of HRF variability on FC only with the thalamus seed; the magnitude and characteristics of HRF confound on FC elsewhere in the brain could be different. A similar analysis at the whole brain level, though computationally expensive, may provide a more complete picture. We selected this seed based on a stringent selection approach that required at least three significant “hits” in the group comparison maps of the 3 parameters. We used a multiple testing correction approach aimed at improving sensitivity over specificity⁸³ in this exploratory work; we acknowledge that this is not the gold standard correction for mass univariate search approaches. The regions highlighted here could inform further motivated studies of HRF in schizophrenia in the future. Second, the discussion points we provided regarding how neurochemical factors might affect the HRF is still speculative, given that there is yet very limited literature directly investigating the relationship between HRF and these factors. In the absence of direct measurement of neurochemicals (using in vivo MR spectroscopy, positron emission tomography or single-photon emission computed tomography) in regions showing alterations in the shape of the HRF, it is not possible to make definitive claims regarding how these neurochemicals might control different aspects of the HRF shape. Third, in both deconvolved and nondeconvolved data, we estimated the correlation between HRF parameters and FC as well as clinical variables at the individual subject level. None of the correlations were significant after correcting for multiple comparisons, while a couple of associations approached trend level. While we agree that this type of analysis would provide interesting insights, we feel that the variability in the data as well as the sample size must be larger for this type of individual difference analysis. This might be an interesting research direction for future studies. Third, the sample size used in this study is modest, especially for bipolar disorder. Therefore, the conclusions obtained from this study must be replicated in relatively larger samples in order to inform best practices in the use of functional MRI in studies involving subjects with schizophrenia or bipolar disorder. Finally, we used gene expression maps obtained from the Allen Human Brain Atlas. However, more definitive conclusions can be reached if one had gene expression and neurochemical data from the same subjects (SZ, BP and Control) undergoing the MRI.

To conclude, neurovascular coupling is highly variable among brain regions within an individual, as well as between different individuals. These variations relate to psychosis, dopamine receptor distribution, and psychotropic dose exposure. In psychotic disorders, functional disconnection among brain regions may result from factors differentially affecting regional neurovascular coupling. Parsing this uncoupling could potentially improve the consistency of mechanistic studies based on functional connectivity.

Author Contributions

G.D., L.P., and P.F.L. conceived and designed the study. L.P. and P.F.L. acquired the data. W.Y. performed the bulk of the analysis with some of the analyses performed by G.D. and D.R. All authors wrote and reviewed the manuscript.

Data Availability

Raw and pre-processed data as well as results can be made available upon reasonable request to the corresponding author.