Changes in hemodynamic response function resulting from chronic alcohol consumption

Abstract

Background:

Functional MRI (fMRI) task-related analyses rely on an estimate of the brain’s hemodynamic response function (HRF) to model the brain’s response to events. Although changes in the HRF have been found after acute alcohol administration, the effects of heavy chronic alcohol consumption on the HRF have not been explored, and the potential benefits or pitfalls of estimating each individual’s HRF on fMRI analyses of chronic alcohol use disorder are not known.

Methods:

Participants with alcohol use disorder (AUD) and controls (CTL) received structural, functional, and vascular scans. During fMRI, participants were cued to tap their fingers, and averaged responses were extracted from the motor cortex. Curve fitting on these HRFs modeled them as a difference between 2 gamma distributions, and the temporal occurrence of the main peak and undershoot of the HRF was computed from the mean of the first and second gamma distributions, respectively.

Results:

ANOVA and regression analyses found that the timing of the HRF undershoot, increased significantly as a function of total lifetime drinking. Although gray matter volume in the motor cortex decreased with lifetime drinking, this was not sufficient to explain the undershoot timing shifts, and vascular factors measured in the motor cortex did not differ among groups. Comparison of random effects analyses using custom fitted and canonical HRFs for CTL and AUD groups showed better results throughout the brain for custom fitted versus canonical HRF for CTL subjects. For AUD subjects, the same was true except for the basal ganglia.

Conclusions:

These findings suggest that excessive alcohol consumption is associated with changes in the HRF undershoot. HRF changes could provide a possible biomarker for the effects of lifetime drinking on brain function. Changes in HRF topography affect fMRI activation measures, and subject-specific HRFs generally improve fMRI activation results.

Introduction

Heavy alcohol consumption is linked to a number of adverse changes in brain structure and function. Principal brain areas affected are the frontal lobes and cerebellum and their associated functions (Adams et al., 1993; Gilman et al., 1990; Pfefferbaum et al., 1997; Sullivan et al., 2003; Cardenas et al., 2007; Chanraud et al., 2007; Kubota et al., 2001) (for reviews, Oscar-Berman et al. (1993), Oscar-Berman and Marinkovic (Oscar-Berman and Marinković, 2007), Moselhy et al. (2001), and Sullivan (2000)). In vivo MRI evidence for regional tissue shrinkage is consistent with postmortem studies that have revealed neuronal loss in frontal lobes (Harper and Kril, 1990; Kril and Harper, 1989; Kril et al., 1997; Courville, 1955; Brun and Andersson, 2001) and reduction of soma size and processes in cerebellar vermis and hemispheres (Victor et al., 1989; Torvik and Torp, 1986; Phillips et al., 1987; Baker et al., 1999).

Heavy alcohol consumption can also produce cognitive as well as motor changes (Sullivan et al., 2000), and early studies have shown abnormalities in glucose metabolism that correlated with reduced cognitive performance (Adams et al., 1993; Dao-Castellana et al., 1998). More recently, functional MRI (fMRI) has been used in individuals with alcohol use disorder (AUD) to examine alcohol related changes in brain activation in cognitive tasks such as reactivity to drink cues (Braus et al., 2001; Myrick et al., 2004; Tapert et al., 2003; Alba-Ferrara et al., 2016), inhibitory control and conflict processing (Schulte et al., 2012), verbal (Desmond et al., 2003) and spatial (Pfefferbaum et al., 2001) working memory, dual task performance (Chanraud et al., 2010), and auditory language processing (Chanraud-Guillermo et al., 2009). FMRI analysis relies on an estimate of the hemodynamic response function (HRF), an impulse response of the blood oxygen level dependent (BOLD) signal over time to a brief event. Convolution of this HRF with task related square waves - representing the timing of events of interest - results in regressors whose time course represents a prediction of the measured BOLD signal. The temporal accuracy of these regressors is therefore a major factor in the magnitude of observed activations in an fMRI analysis.

The HRF is characterized by an initial peak followed by an undershoot (Buxton et al., 1998; Glover, 1999; Hua et al., 2011; van Zijl et al., 2012). This has been modelled using different approaches (Lindquist et al., 2009), but one approach used in fMRI statistical analysis software packages such as Statistical Parametric Mapping is to describe the HRF as a difference in 2 gamma distributions, the first one modelling the initial peak and the second one modelling the undershoot of the HRF. In normal healthy subjects, a “canonical” HRF is often used, under the assumption that healthy subjects should not be too deviant from the canonical waveform. However, in patient populations, changes in vascular factors such as blood flow and blood volume, or changes in gray matter, could produce alterations in the HRF, leading to inaccuracies in regressors. Such changes might be particularly detrimental to event-related investigations, which involve relatively brief events, and thus, rely on temporal accuracy of the regressor more so than block fMRI designs. For the latter, regardless of the exact form of the HRF, the longer duration event results in the regressor quickly increasing to maximum level and remaining at that level for many seconds, until the event terminates, so variations in the HRF waveform should have less impact.

Changes in HRF properties have been noted in a number of populations, including traumatic brain injury (Mayer et al., 2014), HIV/AIDS (Juengst et al., 2007), schizophrenia (Hanlon et al., 2016), type 2 diabetes (Duarte et al., 2015), cerebrovascular disease (Bonakdarpour et al., 2007), and healthy aging (Morsheddost et al., 2015; D’Esposito et al., 1999). Alterations in HRF have been documented for acute administration of alcohol (Luchtmann et al., 2010), but have not been explored in abstinent AUD patients. Given that global CBF in a sober state has been found to decrease as a function of self-reported weekly alcohol consumption (Christie et al., 2008), one might predict that chronic alcohol consumption could also adversely affect the magnitude or shape of the HRF. Thus, the first goal of the present study was to examine effect of the amount of lifetime drinking on the HRF as measured by a simple task. If found, such changes could potentially be used as a biomarker to measure cumulative effects of alcohol on brain function.

Our procedure involved identifying the peak response to finger tap events by finding the strongest response in the motor cortex to a set of Fourier basis functions. The averaged finger tap response waveform was then extracted from this peak location, and curve fitting was used to find the best set of parameters that fit the waveform as a difference between 2 gamma distributions. The timing of the initial peak and undershoot could then be quantified by computing the mean of the first and second gamma distributions (referred to as mean₁ and mean₂), and variations in these values as a function of alcohol consumption could be assessed. We hypothesized that lifetime alcohol consumption would result in changes in one or both of these HRF timing parameters.

Our second goal was to assess to what extent a custom fitted HRF derived from the motor cortex, versus a canonical HRF, would be beneficial or detrimental in a random effects analysis. We hypothesized that a custom fitted HRF would result in both greater spatial extent of observed activations as well as stronger magnitudes of activation, relative to the canonical HRF, and that these benefits would not be confined to the motor cortex.

Materials and Methods

Participants

Prospective participants were provided with study information and screened for eligibility via email and phone according to standard MRI safety protocols. A total of 89 right-handed native English speaking females and males from the greater Baltimore region, between the ages of 18 and 60 with no history of neurological or psychiatric disorders, were invited for in person assessment and enrolled in the study. Blood Alcohol Content (BAC) was assessed via breathalyzer, and a urine sample was assessed for pregnancy and recent drug use. Participants with a positive pregnancy test, drug screen, and/or a non-zero BAC were determined ineligible to complete the study. 57 participants were excluded for various reasons, including positive drug tests, current drinking, and MRI contraindications. The remaining 32 participants completed finger tapping (the focus of the present study) and cognitive fMRI tasks: 16 of these participants were classified – based on meeting the criteria at any point in their life - as adults with AUD (9 male), and 16 were healthy control participants (9 male). All participants provided written informed consent on the day of their first visit and were compensated for their time. Study procedures were reviewed and approved by the Johns Hopkins Institutional Review Board.

Assessment

Eligible participants completed written assessments including the Alcohol Use Disorders Identification Test (AUDIT) (Saunders et al., 1993), Fagerstrom Test for Nicotine Dependence (Fagerström, 1978), Beck’s Depression Inventory-II (Beck et al., 1996), and Shipley’s Vocabulary and Abstraction Assessments (Zachary, 1986). The AUDIT and Fagerstrom test were used to quantify alcohol and nicotine use, respectively, and the BDI-II was used to quantify depressive symptoms. Participants who scored less than 12 on the Shipley’s Vocabulary Assessment were determined ineligible. These tests were further used in combination with semi-structured interview assessments to determine eligibility.

Eligible participants then completed the Semi-Structured Assessment for the Genetics of Alcoholism (SSAGA) (Bucholz et al., 1994), Lifetime Drinking History (LDH) (Skinner and Sheu, 1982), and Time-line Follow-back (TLFB) (Sobell and Sobell, 1992, 1995; Sobell et al., 1988) with a Masters-level research interviewer to further investigate prior and current alcohol and drug use as well as medical and psychiatric history. The SSAGA yielded diagnoses for alcohol abuse/dependence and other commonly occurring psychiatric disorders under a polydiagnostic system that included Research Diagnostic Criteria (RDC), DSM-III, DSM-IIIR, DSM-IV and Feighner criteria. This interview was used to rule out study participants with current DSM-IV Axis I disorders. The LDH used structured interview prompts to quantify the following: total lifetime alcohol consumption (number of standard drinks in a lifetime), number of years consuming over 80 grams/day, length of heavy drinking (number of years since drinking level reached a mean of 80 grams/day for at least one month), age of heavy drinking onset (age drinking level reached a mean of 80 grams/daily for at least one month), maximum consumption on a single occasion, total quantity of alcohol consumed during the last six months, and date of last drink. At the time of the initial in person assessment, TLFB interview quantified alcohol and drug use on each day during the past 90 days. Interim TLFB interviews were conducted on each day that the participant visited the lab to monitor drinking behaviors and eligibility. All interviews were supervised and evaluated by a neuropsychologist with considerable expertise in psychological assessment and diagnosis to confirm that participants did not have any psychological disorders or substance dependencies that were exclusionary.

Eligible participants underwent physical examination and laboratory testing at the Johns Hopkins General Clinical Research Center (GCRC). The team’s physician or other qualified personnel performed a complete physical exam and a brief neurological exam designed to quantify possible cerebellar damage, including assessments for saccade accuracy, pursuit, and gaze holding, as well as the Brief Ataxia Rating Scale (BARS). A blood count, comprehensive metabolic panel, and syphilis screening were performed to detect any physical disorders that could complicate interpretation of functional brain activation measurements. AUD participants were required to be abstinent for at least 30 days prior to participation, confirmed by carbohydrate deficient transferrin (CDT) or Phosphatidylethanol (PEth) blood testing, so that group differences would reflect chronic rather than acute effects of alcohol or alcohol withdrawal symptoms. Prospective alcohol participants were excluded (n=0) if results were non-zero. For control subjects, TLFB data and subject reports were used to derive months of abstinence estimates, which varied considerably, as some subjects had not consumed alcohol in many months, whereas others were regular social drinkers (mean = 25.4 months, SD = 89.5; median = 0.275, IQR = 2.4). For both control and AUD subjects, breathalyzer testing was used to confirm 0 BAC at the time of scanning.

Eligible participants returned for a second and third visit to complete cognitive testing (outside of the scanner) and additional brain imaging, including cognitive and eyeblink conditioning paradigms that will be described in separate reports. At the beginning of each visit, participants were screened for recent drug use and alcohol consumption via urinalysis and breathalyzer. They were then administered cognitive testing with the team’s neuropsychologist or other qualified personnel to assess IQ, memory (or possible dementia), working memory, executive function, and motor performance. The following tests were conducted: Trail Making, Initial Letter Fluency, Hopkins Adult Reading Test, RBANS, Digit Symbol (WAIS-IV), Visual Puzzles (WAIS-IV), Digit backwards (WAIS-IV), Brief Test of Attention, and Wisconsin Card Sorting Test (WCST) (modified by Nelson, only 48 cards). Participants were then escorted to the F.M. Kirby Research Center for Functional Brain Imaging to receive fMRI scanning.

In two cases (one CTL, one ALC) the participants reported current nicotine use and indicated that they would be unable to complete fMRI scanning without smoking 2 hours beforehand or during the scanning session. To better control nicotine levels during fMRI testing, these participants were offered a 14 mg Nicoderm patch two hours prior to the fMRI scan, which they accepted. These participants were assessed for contraindications, and were given a baseline carbon monoxide (CO) test at the beginning of the visit. At two hours prior to fMRI scanning, a second CO test was administered to insure that covert cigarette smoking had not occurred prior to the nicotine patch administration. The patch was removed at the completion of fMRI scanning.

MRI Scanner

Scanning was performed on a Philips Achieva 3-Tesla scanner with a 32-channel head coil. Subjects received a functional MRI scan during a finger tapping task in order to measure the BOLD hemodynamic response function (HRF) from the motor cortex. To assess whether BOLD affects could be attributable to vascular changes, subjects also received scans for measuring cerebral blood flow, cerebral blood volume, and cerebrovascular reactivity. To examine possible group differences in gray matter, a T1-weighted structural scan was used for segmenting the brain and deriving gray matter volumes for each subject.

Finger tapping fMRI task

For the functional MRI scan, subjects received 19 trials of a finger tapping task in which they were instructed to alternately press two buttons with the right index and middle fingers for a 1 second interval, which was cued by a visual stimulus. The period of tapping was followed by a rest interval of 28–33 s. This task has been shown to reliably elicit activation in the left motor cortex (Chen and Desmond, 2005). E-Prime 2 was used to control stimulus timing and record button responses. The reaction time (RT) of the first button press for each tapping period was used to measure RT for each subject.

BOLD functional images were collected by using a T2*- weighted gradient echo planar imaging (EPI) pulse sequence. Twenty oblique (perpendicular to the axis of the hippocampus) slices (3 mm × 3 mm × 6 mm acquired resolution) were collected (TR: 1000 ms, TE: 30.0 ms; FOV: 22 cm; flip angle: 61°) in a series of 630 sequential images (for a total of 630 s).

Anatomical Scan

Structural images were acquired by using a T1-weighted magnetization prepared rapid acquisition gradient echo (MPRAGE) pulse sequence. Parameters used were: TR = 7.0 ms, TE = 3.2 ms, flip angle = 8°, Inter-shot time = 3000 ms, FOV = 240 mm, 1 mm slice thickness acquired sagittally, 288 × 272 matrix. Total scan duration was 6 min 25 s with SENSE reduction factor of 2.

Cerebral Blood Volume

CBV was measured using dynamic susceptibility contrast (DSC) MRI. DSC MRI images were acquired following the administration of Gd-DTPA contrast agent with gradient-echo (GRE) 2D echo-planar-imaging (EPI) readout, TR/TE = 1500/40 ms, voxel = 3 × 3 × 6 mm³, 20 slices, 40 dynamics, 1 minute.

Cerebral Blood Flow

Regional CBF in the whole brain was measured with a pseudo-continuous arterial spin labeling (pCASL) technique (Dai et al., 2008). Pairs of label and control images were acquired using the following imaging parameters: 2D gradient echo EPI, FOV = 220 × 220 mm, voxel size = 3 × 3 × 6 mm, 19 slices, TR/TE = 4000/14 ms, parallel imaging (SENSE) = 2; label duration = 1650 ms, post-labeling delay (PLD) = 1505 ms, 35 pairs of controls/labels; scan duration = 4 min, 40 s.

Cerebrovascular Reactivity

To measure cerebrovascular reactivity, subjects performed 4 cycles of a breath hold procedure during fMRI (BOLD) imaging. This procedure consisted of normal breathing for 50 s followed by a 5 s period in which subjects were cued to exhale, followed by a 15 s period in which they were cued to take a deep breath and hold it for 15 s. fMRI parameters were similar to those used for the finger tapping task, except that the TR was 2500 ms and a total of 132 dynamic volumes were collected.

Anatomical and Functional Imaging Data Analysis

Structural and functional imaging preprocessing and statistical analyses were performed with Statistical Parametric Mapping (SPM8) software (Wellcome Department of Cognitive Neurology, London, UK). Preprocessing began with slice scan time correction used to correct for differences in image acquisition time between the 20 axial slices that were acquired in an inferior to superior direction. For this step, TR was entered as 1 sec and acquisition time (TA) was calculated as TA = TR-(TR/Nslices) = 0.95 sec. Corrected volumes were referenced to the middle slice (slice 10). The slice timing corrected volumes were then motion corrected using default settings in the SPM8 Realign: Estimate & Reslice module, which included point sampling separation of 4 mm, 5 mm FWHM smoothing before estimating the realignment parameters, and 2^nd degree B-Spline interpolation during transformation estimation. Using a two-pass procedure, volumes were registered to the mean after the first alignment. Once motion correction was achieved, functional and structural (MRPRAGE) scans were then spatially coregistered to each other using the SPM Coregister module. Once coregistered, the structural scan was then spatially normalized using the gray and white matter segmentation-based method implemented in SPM (Ashburner and Friston, 2005). This type of normalization was chosen because it yields cerebellar output that closely corresponds to that of the spatially unbiased infra-tentorial template (SUIT) of the human cerebellum and brainstem (Diedrichsen et al., 2009). A modulated normalized gray matter volume was obtained for each subject from the segmentation process, and these volumes were used for examining group differences in gray matter at the finger tap coordinate described below. Whole brain structural and functional data were transformed into standard stereotaxic space according to the Montreal Neurological Institute (MNI) protocol. Functional images were spatially smoothed with a Gaussian filter (full-width half-maximum 5 mm) and temporally high-pass filtered at 128 sec.

Finger tap analysis

Extraction of Finger tap HRFs for Custom Fitting

To estimate the hemodynamic response function, for each subject, a Fourier set of four (sine/cosine) basis functions was used to model finger tap responses in a 32 s window. The average resulting response from the basis set (averaged over the 19 finger tap trials that the subject performed) was then obtained from the maximum peak found within a left precentral gyrus region of interest from the Hammers adult maximum probability atlas (Gousias et al., 2008; Hammers et al., 2003). The extracted waveforms were converted into percent signal change for analysis of group differences in percent signal change. For curve-fitting, the waveforms were transformed into Z scores by subtracting each value from the mean of the waveform and dividing by the standard deviation.

Curve fitting procedure

The Z-scored waveforms were entered into GraphPad Prism software, where parameters were estimated in order to fit the data to an equation describing the difference of two gamma distributions, ie, $f\left(t\right)=c_1\left(\frac{L_1^{h_1}{t}^{\left(h_1-1\right)}{e}^{-L_{1}t}}{\Gamma \left(h_1\right)}\right)-c_2\left(\frac{L_2^{h_2}{t}^{\left(h_2-1\right)}{e}^{-L_{2}t}}{\Gamma \left(h_2\right)}\right)$, where Γ is the gamma function. The parameters c₁, L₁, h₁ for the first gamma distribution, and c₂, L₂, and h₂ for the second gamma distribution were estimated; from these parameters, the mean of the first and second gamma distribution, describing the time at which the peak and undershoot of the HRF, respectively, occurred were calculated as mean₁ = h₁/L₁ and mean₂ = h₂/L₂ (Evans et al., 1993; Lindquist et al., 2009). Examples of the curve fitting results are depicted in Figure 1.

Figure 1.

Example of curve fitting results for two CTL subjects and two AUD-H subjects. Filled circles represent actual points (averaged over 19 trials) of finger tap HRF waveform data extracted from the X,Y,Z peak activation focus in the precentral gyrus, using the Fourier basis functions analysis. Solid lines represent the curve fitting results from GraphPad Prism software representing the difference of two gamma functions. Largest differences are seen in mean₂ values, which for the CTL subjects are 12.7 s (top) and 13.4 s (bottom), but for the AUD-H subjects are 23.8 s (top) and 25.6 s (bottom).

Cerebral blood volume analysis

DSC MRI images were analyzed using a plugin toolbox DSCoMAN in the ImageJ software (http://imagej.nih.gov/ij/index.html). Maps of CBV were generated for each subject.

Cerebral blood flow analysis

In the ASL scan, a standard surround subtraction method (Lu et al., 2006) was used to compute CBF maps by comparing the labeled and corresponding control images, as described in Hua et al. (Hua et al., 2011). In house Matlab code was programmed for this analysis.

Cerebrovascular reactivity analysis

To analyze the breath hold data, event times for normal breathing, exhale, and hold breath events were convolved with the respiratory response function (RRF) described by Birn et al. (2008) to generate regressors for statistical analysis. Because respiratory related BOLD signal changes tend to be slower than signal changes induced by neural activation, this RRF, rather than the canonical HRF in SPM8, was more appropriate for generating breath hold-induced activation maps. Group comparisons of breath hold responses were therefore based on differences in beta values that characterized how well the breath hold regressor predicted the observed BOLD signal.

Analysis of the effects of alcohol on the HRF peak and undershoot

To examine the effects of alcohol on the magnitude of the HRF over time, we began by testing for group differences in the percent signal change waveforms by conducting a repeated measures ANOVA with one between subjects variable (group: AUD vs CTL) and one repeated measure (time). To focus on the temporal dynamics of the waveforms, we repeated the above analysis using Z score transformed HRF waveforms, and tested for group differences in the values of gamma mean₁ and mean₂ that were obtained from curve fitting for each subject using t-tests.

To further investigate the effects of the total lifetime quantity of alcohol consumed on the HRF, we divided our subjects into 3 groups, control subjects (CTL), AUD subjects who consumed relatively lower amounts of alcohol (< 50,000 total lifetime drinks, AUD-L, N=9), and AUD subjects who consumed relatively higher amounts of alcohol (>= 50,000 total drinks, AUD-H, N=7). Statistical tests on age, education, sex distribution, pack years, and packs smoked last year were calculated to determine if these 3 groups were equated on these variables. We then repeated the above analyses using 3 groups. We also performed correlations to determine if total lifetime drinks was correlated with either the peak or undershoot timing of the HRF.

To test whether group differences found in the temporal dynamics of the HRF (ie., gamma mean₁ or mean₂) might reflect vascular or structural differences occurring at the region where the HRF was measured, separate between subject ANOVAs were performed for CBF, CBV, breath hold, and gray matter measures taken at the HRF region. The values for these analyses were obtained by defining a sphere of 10 mm radius centered on the X,Y,Z coordinate of the HRF of each subject, and extracting the mean value of each measure within the sphere. For any of these variables that showed significant group differences, an ANCOVA was performed to determine if group differences in HRF timing continued to achieve significance when the vascular or structural measure was entered as a covariate.

Comparison of custom fitted HRF versus canonical HRF for fMRI analysis

To determine if the custom fitted HRF and the SPM canonical HRF would result in differences in the extent of fMRI activations, and/or the magnitude of activations, found when convolutions with those HRFs were used to generate regressors for statistical analysis, two analyses were conducted. For the first (spatial extent) analysis, SPM was used to generate two activation maps for each subject, one based on the canonical HRF and the other based on the custom fitted HRF. A random effects analysis from the contrast files was then computed separately for the CTL and AUD groups, and for the canonical and custom fitted HRFs. Then, for each subject group, a p < .0001 significance threshold was set for the canonical and custom fitted spmT maps, and a conjunction map was created to visually show voxels that were: (a) significant using both canonical and custom fitted HRFs, (b) significant for the custom fitted HRF but not the canonical HRF, and (c) significant for the canonical HRF but not the custom fitted HRF (Figure 5).

Figure 5.

Conjunction map of (A) healthy control subjects (N=16) and (B) AUD subjects (N=16) depicting comparison of random effects analysis results using the SPM canonical HRF and the custom fitted HRF. Using a p value significance threshold of 0.0001, voxels in red indicate regions that were significant for both HRFs. Yellow voxels depict regions that reached significance for the custom HRF, but not the canonical HRF. Blue voxels depict regions that reached significance for the canonical HRF, but not the custom fitted HRF.

For the second (activation magnitude) analysis, we investigated if the mean SPM-derived t statistical values generated for each subject using the canonical and custom fitted HRFs were different. To focus on relevant regions for finger tapping, we created 10 spherical regions of interest of 10 mm radius that were centered on the coordinates reported in an ALE meta analysis of finger tapping functional activation (Witt et al., 2008). The regions used were those identified in the main effects of all finger tapping task variations, and included bilateral sensorimotor cortices (LSMC, RSMC), supplementary motor area (SMA), left ventral premotor cortex (LVPMC), bilateral inferior parietal cortices (LIP, RIP), bilateral basal ganglia (LBG, RBG), and bilateral anterior cerebellum (LCBL, RCBL). Because the Witt et al. meta-analysis expressed coordinates for the latter regions in Talairach space, Talairach coordinates were transformed into MNI space using the SPM transformation described by Lancaster et al (Lancaster et al., 2007). The resulting MNI coordinates were used as the center of the spherical ROIs: LSMC: (−39.2, −21.1, 54.2); RSMC: (40.8, −16.2, 56.8); SMA: (−2.5, −1.7, 53.9); LVPMC: (−56.8, 2.4, 32.0); LIP: (−52.6, −24.3, 21.1); RIP: (45.0, −38.5, 47.8); LBG: (−22.6, −6.7, 0.9); RBG: (24.9, −8.5, 2.6); LCBL: (−22.8, −56.2, −23.3); RCBL: (18.3, −53.7, −22.0). For each ROI, and for each subject, the mean t value, computed from all the voxels in the ROI, was obtained for the canonical HRF and for the custom fitted HRF. A repeated measures ANOVA was then performed for each ROI with group (CTL, AUD-L, AUD-H) as a between-subject factor and HRF type (canonical, or custom fitted) as a within-subject factor. If no significant effects of group (group main effect, or group x HRF type interaction) were found, then the groups were pooled and the difference between canonical and custom fitted HRFs was computed from a paired t-test. The results of these tests are illustrated in Figure 6.

Figure 6.

Comparison of average t values obtained in subjects’ finger tap activation maps derived from canonical versus custom fitted HRFs. For all regions except basal ganglia, CTL and AUD subjects were pooled because an ANOVA did not find any group differences or group x HRF type interaction. Abbreviations: PMC = Premotor Cortex, SMA = Supplementary Motor Area, BG = Basal Ganglia, ROI = Region of Interest. Significant difference key: *** p < .001; ** p < .01; * p < .05.

Results

Sample Characteristics

Demographic, drinking and smoking history, and screening assessment data are summarized in Table 1. No significant group differences in age, education, sex distribution, or smoking history were observed (p’s > 0.05), and no differences were found for the BDI-II or Fagerstrom tests. However, AUD participants drank significantly more than healthy participants as indicated by the Total Number of Drinks Consumed over the Lifespan (t(30) = 5.22, p < 0.0001), Total Number of Years of Frequent Drinking (t(30) = 3.31, p = 0.0024), and the AUDIT screening test (t(30) = 3.95, p = 0.0004). AUD participants were abstinent for a mean of 93.3 months (SD = 115.9, range 1.5 – 312).

Table 1.

Sample characteristics

	Healthy (n=16)	AUD (n=16)	t or χ2
Demographics
Age at scan (years)	48.9 (6.6)	51.1 (6.1)	1.01
Sex (% male)	75.0	75.0	0
Education (years)	14.0 (2.4)	14.4 (2.7)	0.45
Race (% caucasian)	31.3	75.0	6.15*
Drinking History
Number of Years Frequent Drinkinga	4.2 (8.2)	14.4 (9.2)	3.31**
Total Number of Drinks Consumed (lifespan)	6000 (6560)	54796 (36836)	5.22 ***
Total Number of Months Abstinent	n/a	93.3 (115.9)	n/a
Smoking History
Number of Pack Years	5.4 (9.1)	11.2 (14.0)	1.41
Total Packs Last Year	70.8 (107.7)	66.1 (117.7)	0.12
Screening Assessments
AUDITb	2.6 (2.3)	15.7 (13.1)	3.95***
BDI-IIc	2.4 (4.0)	8.8 (13.0)	1.86
Fagerstromd	1.31 (1.96)	2.50 (2.42)	1.53

^*p < .05,

^**p < .01,

^***p < .001. Values are Mean (SD) or Percent

^aFrequent Drinking is defined by the Lifetime Drinking History Assessment as 15+ days of drinking per month

^bAlcohol Use Disorders Identification Test (Saunders et al, 1993)

^cBeck Depression Inventory-II (Beck et al, 1996)

^dFagerstrom Test for Nicotine Dependence (Fagerstrom, 1978)

Three group sample characteristics.

Because we wanted to examine the progressive effect of alcohol consumption on the HRF, we subdivided our AUD subjects based on total lifetime drinks consumed, and re-tested for possible group differences in age, education, and smoking on the 3-group sample. These analyses revealed no group differences in age (CTL = 48.9 +/− 6.6, mean +/− SD; AUD-L = 50.3 +/− 6.6; AUD-H = 52.1 +/− 5.6; F(2,29) = 0.65, NS), education (CTL = 14.0 +/− 2.4; AUD-L = 15.3 +/− 2.8; AUD-H = 13.1 +/− 2.3; F(2,29) = 1.66, NS), Number of Pack Years (CTL = 5.4 +/− 9.1; AUD-L = 14.3 +/− 16.0; AUD-H = 7.3 +/− 10.9; F(2,29) = 1.70, NS), Number of Packs Last Year (CTL = 70.8 +/− 107.7; AUD-L = 66.8 +/− 127.1; AUD-H = 65.2 +/− 114.4; F(2,29) = 0.01, NS), or gender distribution (CTL = 4F, 12M; AUD-L = 3F, 6M; AUD-H = 1F, 6M; χ² = 0.76, NS). The 3 groups, however, were significantly different in Number of Drinks Consumed (CTL = 5,999.7 +/− 6,559.5, AUD-L = 28,162.7 +/− 16,407.8; AUD-H = 89,039.0 +/− 24,519.3; F(2,29) = 16.1 p < 0.0001), and each pairwise contrast of the groups was significantly different from 0 (CTL, AUD-L: p = 0.0012; CTL, AUD-H: p < 0.0001; AUD-L, AUD-H: p < 0.0001). Groups also differed in AUDIT scores (CTL = 2.56 +/− 2.3; AUD-L = 12.3 +/− 10.7; AUD-H = 20.0 +/− 15.4; F(2,29) = 9.64, p < 0.001), with pairwise comparisons of CTL/AUD-L and CTL/AUD-H reaching significance (p = 0.016, and p= 0.0002, respectively), and AUD-L/AUD-H approaching significance (p = 0.11).

Finger tap reaction time

For each of the 19 “tap” cues presented to the subject, E-prime recorded the reaction time of the first button press that occurred during the 1 second tap interval. The median RT for each subject was computed, and a one way ANOVA was used to test for group differences in mean median RT. This analysis revealed no group differences (CTL = 332.8 ms +/− 107; AUD-L = 364.6 +/− 82; AUD-H = 321 +/−58; F(2,29) = 0.52, NS).

Effects of alcohol consumption on the HRF peak and undershoot

A group (CTL, AUD) x time (32 points at 1 sec intervals) repeated measures ANOVA analysis of HRF percent signal change revealed a significant main effect of time (F(30,900)=56.2, p < 0.0001) but no main effect of group (F(1,30) = 0.02, NS), or group x time interaction (F(30,900) = 1.02, NS). Similar results were obtained for the Z score transformed HRFs (time main effect: F(31,930) = 103.1, p < 0.0001; group main effect: F(1,30) = 1.0, NS; group x time interaction: F(31,930) = 1.26, NS) and for percent signal change for the 3 group repeated measures ANOVA (time main effect: F(30,870) = 51.4, p < 0.0001; group main effect: F(2,29) = 0.86, NS; group x time interaction: F(60,870) = 1.22, NS). However, the 3 group repeated measures ANOVA on the Z score-transformed HRFs did reveal a significant group x time interaction (F(62,899) = 1.58, p = 0.004), indicating that, while overall percent signal change was comparable across the groups, the topography of the waveforms was different (Figure 2A).

Figure 2.

A. Average Z-transformed HRF for the 3 groups of subjects. B. Average curve-fitted waveforms illustrating a shift in the undershoot time of the HRF in AUD subjects.

Further information concerning the nature of the temporal differences in the HRFs of the groups was obtained from the analysis of the mean₁ and mean₂ times in the fitted curves. For mean₁, a one way ANOVA revealed no significant differences among the groups (CTL = 6.63s +/− 1.36; AUD-L = 7.17 s +/− 1.23; AUD-H = 6.87s +/− 0.69; F(2,29) = 0.57, NS), indicating that the time of initial peak of the HRF was not different across the groups. However, a one way ANOVA of mean₂ times revealed a significant difference across the groups (CTL = 15.50s +/− 3.56; AUD-L = 16.38s +/− 4.45; AUD-H = 29.71s +/− 16.65; F(2,29) = 7.64, p = 0.0022). Analysis of simple main effects indicated that there were significant differences between CTL and AUD-H (p = 0.0008) and between AUD-L and AUD-H (p = 0.0035). For the latter ANOVA, the relatively higher SD of the AUD-H group was driven by one AUD-H subject whose mean₂ time was more than 4 SD deviant from other subjects. Because of concern that the homogeneity of variance assumption of the test was violated, we removed this outlier, tested for homogeneity of variance between each pairwise comparison of the 3 groups, and recalculated the ANOVA without this outlier. The equality of variance test (testing if the variance ratio was significantly different from 1) found no significant differences among the groups (CTL/AUD-L F(15,8)=0.64, p=.48; CTL/AUD-H F(15,5) = 0.39, p= 0.22; AUD-L/AUD-H F(8,5) = 0.60, p = 0.55). Without the outlier, the ANOVA again found a significant main effect of groups (F(2,18) = 8.39, p = 0.0014), with the AUD-H mean₂ time = 23.74s +/−5.74. To examine the possibility that the two subjects who received nicotine patches could have influenced the results, we also recalculated the ANOVA after excluding these subjects, and again found significant group differences (F(2,27) = 5.98, p = 0.007). We then performed, across all subjects, a regression analysis of mean₂ time as a function of the number of lifetime drinks, and found a significant positive correlation (r = 0.63, p < 0.0001, Figure 3). A similar regression of mean₂ as a function of months abstinence in the AUD subjects trended toward a reduction of mean₂ with longer periods of abstinence, but did not reach significance (r = 0.41, p = 0.114). The mean₂/abstinence correlation was not significant when the CTL and AUD subjects were combined (r = 0.182, p = 0.32), or when the CTL subjects were examined alone (r = 0.086, p = 0.75).

Figure 3.

Regression results showing mean time of the second gamma function (i.e., mean₂), corresponding to the HRF undershoot, as a function of number of lifetime drinks consumed (r = 0.63, p < 0.0001).

To explore possible vascular and/or structural factors underlying group differences in mean₂ values, one way ANOVAs were computed for CBV, CBF, breath hold, and gray matter values extracted from a sphere of 10 mm radius centered on the X,Y,Z finger tap coordinate for each subject. These analyses revealed no significant differences for CBV (F(2,28) = 0.87, NS), CBF (F(2,26) = 0.22, NS), or bold breath hold (F(2,28) = 0.06, NS), but a significant difference was found for gray matter (CTL = 0.31 +/− .072; AUD-L = 0.278 +/− .078; AUD-H = 0.214 +/− .013; F(2,29) = 5.0, p = 0.014). Analysis of simple main effects indicated that CTL and AUD-H subjects were significantly different (p = 0.004), and the difference between AUD-L and AUD-H approached significance (p = 0.073). Regression analyses confirmed the latter ANOVA results, showing that the number of lifetime drinks was negatively correlated with gray matter (r = 0.40, p = 0.025, Figure 4A), and that lower mean₂ values were associated with greater amounts of gray matter (r = 0.35, p = 0.048, Figure 4B).

Figure 4. A.

Regression results depicting gray matter (within the finger tap ROI) as a function of number of lifetime drinks (r = 0.40, p = 0.025). B. Regression results showing mean time of the second gamma function (i.e., mean₂), as a function of gray matter (r = 0.35, p = 0.048).

Because group differences were observed in gray matter, we performed an analysis of covariance (ANCOVA) to determine if group differences in the HRF mean₂ value would be eliminated if gray matter was included as a covariate. This analysis revealed that the group main effect approached significance (F(2,26) = 3.21, p = 0.057) as did the main effect of gray matter (F(1,26) = 3.68, p = 0.066). When the one mean₂ outlier subject described above was removed, both of these main effects were significant (group: F(2,25) = 3.94, p = 0.033; gray matter: F(1,25) = 4.30, p = 0.049).

Comparison of custom-fitted HRF versus canonical HRF for fMRI analyses

Spatial extent comparison

The results of the conjunction analysis illustrating differences in regions found to be significant using the canonical vs custom fitted HRFs are depicted in Figure 5A for the healthy subjects, and in Figure 5B for the AUD subjects. The red voxels indicate regions in which both the canonical HRF and the custom-fitted HRF analyses reached p < 0.0001 significance. Yellow voxels represent regions in which the custom fitted HRF reached p < 0.0001 significance but the canonical HRF did not. Blue voxels depict regions in which the canonical HRF reached p < 0.0001 significance but the custom fitted HRF did not.

In Figure 5A, of the 60,957 voxels reaching p < 0.0001 significance for the canonical and/or the custom fitted HRF for the healthy subjects analysis, 62.4% of the voxels were common to both HRFs (red), 36.6% were significant for the custom fitted but not canonical HRF (yellow), and 1% were significant for the canonical but not custom fitted HRF (blue). For the AUD subjects depicted in Figure 5B, of the 59,997 significant voxels, 70.5% were common to both HRFs (red), 22.7% were significant for the custom fitted but not canonical HRF (yellow), and 6.8% were significant for the canonical but not custom fitted HRF (blue). Interestingly, the main regions in which the latter voxels were found were in the left and right basal ganglia.

Activation magnitude comparison

The comparison of mean t values within each spherical ROI for the canonical versus custom fitted HRF is summarized in Figure 6. A repeated measures ANOVA with between subjects factor group (CTL, AUD-L, AUD-H) and within-subjects factor HRF type (canonical, custom fitted) was calculated for each ROI. If no significant effect was found for group or the group x HRF type interaction, groups were pooled and a paired t-test was calculated to assess the difference between HRF types. All ROIs except the left and right basal ganglia (LBG, RBG) showed only a significant effect for HRF type. After pooling all subjects, the paired t-test results for these ROIs were as follows: LSMC: t(31)0=04.67, p < .0001; RSMC: t(31) = 2.80, p = 0.009; LIP: t(31) = 2.89, p = 0.007; RIP: t(31) = 2.86, p = 0.008; LCBL: t(31) = 1.92, p = 0.064; RCBL: t(31) = 1.97, p = 0.058; LVPMC: t(31) = 2.21, p = 0.035; and SMA: t(31) = 1.70, p = 0.099. For LBG, the test for the group x HRF type interaction yielded F(2,29) = 4.78, p = 0.02. Inspection of the interaction indicated that CTL subjects had similar activations for both HRF types, but both AUD-L and AUD-H showed decreased activations for the custom fitted HRF relative to the canonical HRF. The AUD-L and AUD-H groups were therefore pooled, and paired t-tests for HRF type were conducted separately for CTL and AUD subjects. These tests revealed a significant difference for the AUD subjects (t(15) = 3.80, p = 0.002) but not the CTL subjects (t(15) = 0.88, NS), as shown in Figure 6. A similar pattern was found for RBG. The test for the group x HRF type interaction yielded F(2,29) = 7.07, p = 0.012. Again, CTL subjects were at comparable activation levels for both HRF types, but both AUD-L and AUD-H subjects showed greater activations for the canonical HRF than for the custom fitted HRF. After pooling the AUD subjects, paired t-tests revealed a significant HRF difference for the AUD subjects (t(15) = 2.93, p = 0.01), but not for the CTL subjects (t(15) = 0.59, NS).

Discussion

The results of the present investigation indicate that lifetime alcohol consumption influences the topography of the HRF. Specifically, while timing of the initial peak appeared to be unchanged, the timing of the HRF undershoot, as measured by the mean of the second gamma function in our curve fitting procedure, shifted to later values with greater amounts of lifetime drinking. Our group comparisons, which were equated on age, education, sex distribution, smoking behavior, and finger tap reaction times, showed the greatest change in undershoot timing in subjects with more than 50,000 lifetime drinks, but the regression analysis of mean₂ time as a function of number of lifetime drinks (Figure 3) and the average fitted HRFs for the CTL, AUD-L, and AUD-H groups (Figure 2B) suggest a gradual linear shift in the undershoot timing as a function of lifetime drinking.

To investigate if group differences in HRF mean₂ time could be related to baseline vascular or structural group differences, we tested for differences in CBV, CBF, breath hold response, and gray matter values within a 10 mm radius sphere centered on the coordinate from which the finger tap HRF was extracted. Group differences were found only for gray matter, and gray matter values were shown to decrease as a function of lifetime drinking (Figure 4A). Gray matter was also shown to be negatively correlated with mean₂ time, raising the possibility that group differences in mean₂ could be a consequence of decreased gray matter. However, an analysis of covariance indicated that group differences in mean₂ time continued to remain significant when gray matter was entered as a covariate. This analysis suggests that, while gray matter is reduced by lifetime number of drinks, this structural change is not sufficient to account for the group differences in mean₂.

To provide a possible explanation for the alcohol-related change in the HRF, we consider the factors contributing to the BOLD signal, which consist of CBF, CBV, metabolic rate of oxygen (CMRO₂) and the hematocrit fraction. As reviewed by van Zijl et al. (van Zijl et al., 2012), the origin of the HRF undershoot in the BOLD signal has been a source of considerable controversy, with the two most dominant explanations being a delayed vascular compliance based on delayed CBV recovery, or a sustained increase in CMRO₂ after stimulus or response cessation. In a previous study, we investigated these contributions using multimodal monitoring of BOLD, CBF, total CBV, and arterial CBV in human visual cortex after brief breath hold and after visual stimulation (Hua et al., 2011). We found that after visual stimulation, CBF, total CBV, and arterial CBV all recovered to baseline prior to BOLD undershoot recovery. During the BOLD undershoot, postarterial CBV was slightly elevated and CMRO₂ was above baseline by approximately 11%. In contrast, during breath hold, CBF, total and arterial CBV, and BOLD signals all returned to baseline in approximately 20 s, and there was neither a BOLD undershoot nor any elevation of either CBV or CMRO₂. These results suggested that both a delayed postarterial CBV recovery and a sustained increase in oxidative metabolism affect the HRF undershoot. Using a biophysical model, we estimated the contributions of CBV and CMRO₂ to the undershoot to be 19.7% and 78.7%, respectively (Hua et al., 2011). In our data in the current study, no significant differences were found in all vascular measures including CVR, CBV and CBF between groups. Thus, a possible explanation of the effects of higher alcohol consumption on the HRF undershoot would be a delayed or sustained timecourse of oxidative metabolism relative to that found in individuals with lower alcohol consumption.

To investigate the impact of using a custom fitted HRF versus a canonical HRF on fMRI analyses, we created two finger tap activation maps for each subject, one using the canonical HRF and the other using the custom fitted HRF. These activation maps were used to conduct two comparisons. For the spatial extent comparison (Figure 5), random effects analyses were used to create average activation maps for the canonical HRF and for the custom fitted HRF. These averages were computed separately for the CTL and AUD subjects. For the CTL subjects, the spatial extent of significant activations increased in all regions of the brain for the custom fitted HRF, and there was no apparent advantage of the canonical HRF in any region. For the AUD subjects, the same was true except for the left and right basal ganglia, in which a greater spatial extent of activation was observed for the canonical HRF.

A consistent pattern of results was obtained in our activation magnitude comparison (Figure 6), which examined if subjects’ mean t statistical values obtained using the custom fitted HRF would be larger than those obtained using the canonical HRF. We compared activations in 10 regions identified in a meta analysis of finger tapping functional activations, and found that activations for both CTL and AUD subjects were enhanced in nearly every region of the brain, with most of these increased significantly, and others approaching significance. The one exception was the basal ganglia for AUD subjects, in which the canonical HRF resulted in significantly greater activations. In contrast, for CTL subjects, the custom fitted HRFs resulted in slightly (but not significantly) greater activations than the canonical HRF in both the left and right basal ganglia.

The widespread advantage across the brain of the custom fitted HRF over the canonical HRF is relevant for fMRI investigations, as it indicates that the increased activations obtained from a subject-specific HRF measurement will not be confined to the region in which the measurement was made. Visual inspection of the HRFs in AUD-H subjects (Supplementary Figure 1) suggests that long mean₂ times observed in the precentral gyrus were also present in regions distant from the motor cortex, but the basal ganglia undershoot did not seem to follow this pattern. Both the initial rise time and peak, along with the undershoot pattern, of the basal ganglia HRF in AUD subjects appeared to be reflected well by the canonical HRF. We speculate that factors contributing to the undershoot may be different in the basal ganglia relative to other brain regions, or that the effects of alcohol on CMRO₂ may be different in the basal ganglia relative to other brain regions, possibly related to dopaminergic dysfunction observed in chronic detoxified alcoholics (Heinz et al., 2004; Heinz et al., 2005). Regardless of the underlying cause, our data suggest that a customized HRF computed from a motor cortex location in response to a finger tap event is not optimal if the basal ganglia is the primary region of interest for the study. It should be noted that our results may be specific to a motor evoked HRF, and future research will be needed to determine if BOLD hemodynamic responses derived from an auditory, visual, or somatosensory stimulus would result in the same alcohol dependent shifts in mean₂ observed in the present study, or if such HRFs would exhibit significant differences between canonical and custom fitted HRFs in activation spatial extent and magnitude that we observed for our motor-evoked HRF.

The present study is the first to report systematic changes in the shape of the hemodynamic response function that may be linked to total alcohol consumption, and there are several limitations that must be considered in the interpretation of this result. The first is that our breakdown of subjects into three groups, CTL, AUD-L, and AUD-H, resulted in samples of relatively small size and unequal N. Our AUD patients were carefully screened, through both structured interview and drug testing, to exclude comorbid factors and other possible non-alcohol related explanations, and subjects meeting the strict eligibility criteria were challenging to find. Our rationale for dividing the AUD group based on alcohol consumption was to visualize the progressive effect of alcohol consumption on the HRF waveform, which is illustrated in Figure 2. Larger group sizes would increase stability and power of the results, but we note that the ANOVA results, showing group differences in mean₂ time, are consistent with the results of the regression analysis illustrated in Figure 3 that is based on a larger overall sample of 32 subjects. Another limitation is that factors other than alcohol consumption that were not measured in the present investigation could have contributed to group differences in HRF. For example, neither our history and physical exam, nor the blood tests that we performed on our subjects, included quantitative measures of vascular health, and this factor could have differentially affected the BOLD response in CTL and AUD groups. Sleep quality, including apnea events, which has been shown to alter regional CBF (Baril et al., 2018; Elvsåshagen et al., 2019), was also not measured in our participants. Although male/female sex ratio was balanced across our groups, sex hormone levels, which have been shown to affect BOLD responses (Gleason et al., 2006; Goldstein et al., 2010), could have also influenced HRF topography. Finally, we note that one interpretation of our data is that increasing alcohol consumption causes progressive abnormalities in the HRF that manifests in progressive shifts in the time of undershoot, but as a limitation of a cross sectional study, we cannot rule out the possibility that an unknown factor caused both the delay in the undershoot and increased levels of drinking.

With these limitations in mind, our results can be interpreted in the context of the major factors influencing the BOLD response and the possible changes that could occur from alcohol. The BOLD effect is a consequence of local changes in the magnetic field arising from changes in the concentration of deoxygenated hemoglobin during brain activation. Local increases in CBF that exceed the increase in CMRO₂ cause a reduction of deoxyhemoglobin leading to a lengthening of the transverse relaxation time T2* and increased MRI signal, and it can be shown that the BOLD signal is related to these parameters via the expression CBF/(CBV * CMRO₂) (Donahue et al., 2009). As described above, the HRF undershoot has been linked to CMRO₂ in healthy individuals, but the multimodal imaging experiments employed by Hua et al. (2011) to examine the timecourse of CMRO₂ have never been performed in AUD subjects. Our baseline measures of CBF and CBV within the ROI from which HRF changes were measured did not show group differences. However other studies have shown CBF changes in AUD subjects (Erbas et al., 1992; Rogers et al., 1983; Melgaard et al., 1990; Lotfi and Meyer, 1989; Suzuki et al., 2002). Interestingly, studies using acetazolamide administration, which increases CO₂ in arterioles and causes dilation thereby increasing rates of blood flow, revealed lower rCBF in alcoholics before but not after acetazolamide administration, suggesting that decreased rCBF in AUD subjects was due to reduced brain metabolism (Suzuki et al., 2002; Oishi et al., 1997). Metabolic changes associated with chronic alcohol consumption have been documented in PET studies (Adams et al., 1993; Dao-Castellana et al., 1998) and in proton magnetic resonance spectroscopy studies (Ende et al., 2005; Parks et al., 2002). Future research will be needed to determine if brain metabolic changes from alcohol, including the timecourse of CMRO₂, underlie the changes in HRF topography that we observed.

In summary, our results indicate that the shape of the hemodynamic response to a brief finger tap event changes as a function of the total alcohol consumed over the lifetime. This may provide a biomarker for measuring changes in brain function resulting from chronic alcohol consumption. With the exception of basal ganglia regions in AUD subjects noted above, using the custom fitted HRF showed advantages in fMRI analyses for controls as well as for AUD subjects, resulting in significantly greater measures of activation in many regions of the brain.