The Ability of fMRI Imaging to Predict Heavy Drinking and Alcohol Problems 5 Years Later

Abstract

BACKGROUND

Low levels of alcohol responses (low LRs) are genetically influenced phenotypes that are identified before alcohol dependence and predict future heavy drinking and alcohol problems. A recent paper described 13 LR-related blood oxygen level (BOLD) response contrast patterns observed during an emotional face recognition task that might reflect cognitive processes contributing to LR and that might themselves predict adverse alcohol outcomes (Paulus et al., 2012). This paper evaluates the predictive implications of those fMRI patterns.

METHODS

Of 120 subjects from Paulus et al. (2012), 114 (57 low and high LRs; ~50% females) were interviewed 5 years later at age 25. Correlations between baseline fMRI patterns and alcohol-related outcomes were evaluated, and regressions analyses were used to determine if BOLD response contrasts incremented over LR in predicting outcomes.

RESULTS

Baseline fMRI patterns in five of 13 baseline regions of interest correlated with adverse outcomes. Such patterns in insular regions, particularly the left anterior insula, and the right frontal gyrus added to LR in predicting alcohol problems. The relationships remained robust when exact binomial procedures were used, but, reflecting the small sample size, it was not possible to adequately consider Bonferroni corrections.

CONCLUSIONS

The data suggest that fMRI BOLD response contrasts predicted heavier drinking and alcohol problems five years later, even after considering baseline low LRs. Future work will focus on whether fMRI results can predict outcomes in larger samples and among young non-drinkers, as well as how the imaging results increase understanding of the processes through which LR operates.

Introduction

The low Level of Response (low LR) to alcohol is one of several early life genetically-influenced phenotypes that predict future heavy drinking and alcohol problems (King et al., 2013; Newlin & Renton, 2010; Quinn & Fromme, 2011; Schuckit, 2014). Whether measured as lower responses to oral alcohol challenges or as self-reports of more drinks required for effects (Schuckit et al., 2010), individuals with lower LRs require higher blood alcohol concentrations (BACs) for intoxication and/or to demonstrate physiological responses to alcohol. The latter include less alcohol-induced decreases in alpha activity in background cortical electroencephalograms, less prolonged P300 waves in evoked potentials, and less change in alcohol-sensitive hormones (Ehlers et al., 1999; Schuckit & Gold, 1988; Schuckit et al., 1988a, 1988b). Low LR has a heritability of .40–.60 (Heath et al., 1999; Luczak et al., 2002; Schuckit et al., 2001; Viken et al., 2003), and predicts higher future alcohol quantities and problems, even after controlling for earlier drinking (Chung & Martin 2009; Quinn & Fromme, 2011; Schuckit et al., 2007, 2008; Volavka et al., 1996).

However, until recently little was known about the neural processing underlying the low LR. Thus, our group has searched for brain features potentially related to LR using functional magnetic resonance imaging (fMRI). We began with a study of drinking but non-alcohol dependent high school students, with results indicating that the lower the LRs the greater the prefrontal activation during a no-beverage high demand visual working memory (VWM) task above and beyond any effect of VWM performance across levels of LR (Tapert et al., 2004). Those results were corroborated in a second study with college drinkers, and expanded by including an alcohol administration session where LR-group differences diminished after a moderate alcohol dose (Paulus et al., 2006).

The third investigation, which is the basis for the current predictive analyses, generated data for a new group of drinking but not alcohol-dependent college students who participated in three cognitive tasks (Paulus et al., 2012; Schuckit et al., 2012; Trim et al., 2010). The earlier papers for this project focused on cognitive tests related to executive functioning (i.e., VWM and a stop signal task), and were published sequentially as the original data developed by using smaller subsets of the subsequent final sample (Schuckit et al., 2012; Trim et al., 2010). The manuscript for which data regarding fMRI results related to an emotional face recognition paradigm utilized the largest number of subjects and was the final paper in this series. Thus, the follow-up data in the current study focus on results from subjects who participated in the emotional face recognition paradigm as reported by Paulus et al. (2012).

Based on prior fMRI results which were obtained with smaller samples using two executive functioning tasks, we hypothesized that in the context of the emotional face recognition task, low LR subjects would demonstrate higher BOLD response contrast with placebo and lower values after alcohol, while the high LR individuals will show the opposite pattern. This may reflect a general pattern in low LR subjects of how they process information that also extends to emotional-based tasks, even after controlling for alcohol effects on cerebral blood flow (CBF) (Rickenbacher et al., 2011). This pattern was predicted in brain regions related to recognition and processing of emotional faces, such as emotion processing (e.g., insula and amygdala) and in top-down emotional processing control (e.g., medial prefrontal cortex and the anterior and perigenual cingulate). In that study LR group or LR group by placebo/alcohol condition differences were found in both right and left frontal, insula, and anterior cingulate regions.

The previous findings may indicate that individuals with low LRs need greater cognitive effort to recognize relatively subtle differences across stimuli when they are tested with placebo (Paulus et al., 2012). The neurocognitive processes related to the low LR to alcohol might more closely reflect underlying mechanisms that contribute to the low alcohol sensitivity than the potentially less precise measures of the behavioral response to alcohol during a standard alcohol challenge. In addition, the finding of the need for higher levels of neural processing resources during effortful tasks with placebo may suggest that it is possible to identify individuals with low LRs before their first drink. These issues could be addressed by following participants after imaging sessions to determine if the baseline LR and fMRI patterns during placebo and alcohol sessions predicted heavier drinking and alcohol problems.

The current paper reports the results of a five-year follow-up of 95% of the 120 individuals who completed the baseline fMRI protocol, 116 of whom were included in the analyses by Paulus et al (2012). The current paper evaluates three hypotheses: 1) the continuous measure of LR will correlate significantly with adverse alcohol outcomes in these subjects (an evaluation necessary to adequately evaluate Hypothesis 3); 2) fMRI patterns that were related to LR will predict future adverse alcohol outcomes; and 3) baseline fMRI patterns reflecting higher BOLD-response contrasts with placebo in subjects with low LR, especially in the insula, amygdala, and prefrontal cortex, will add to the prediction of adverse alcohol outcomes even when the low LR to alcohol is also considered.

Methods and Materials

Participants

Data collection for the original subjects (Paulus et al., 2012) used three steps, following approval from the University of California, San Diego (UCSD) Human Research Protections Program. First, randomly selected 18-to-25-year-old European American (EA) and white Hispanic students received emailed questionnaires. The sample did not include Asian individuals ~40% of whom experience physiological alcohol responses that might make alcohol challenges in MRI scanners unsafe (Luczak et al., 2002), and too few minorities of other ethnic/racial groups were available to create optimally matched low and high LR pairs within the projected 120 subjects. Questionnaire items were extracted from the Semi-Structured Assessment for the Genetics of Alcoholism (SSAGA) interview (Bucholz et al., 1994; Hesselbrock et al., 1999) covering demography, substance use, and major psychiatric disorders using the Diagnostic and Statistical Manual of the American Psychiatric Association, Fourth Edition (DSM-IV) (American Psychiatric Association, 2000). The initial evaluation of subjects’ LR status was based on a median split on the Self-Report of the Effects of alcohol (SRE) questionnaire (Schuckit et al., 2007, 2008). From the ~70% who responded, participants were considered for inclusion if they: 1) ever consumed a full standard drink (ethanol: 10–12 grams); 2) denied histories of brain trauma, epilepsy, or alcohol or drug dependence; 3) denied current psychiatric disorders; 4) were not pregnant; 5) were right handed; and 6) had no MRI contraindications (e.g., irremovable metal).

In step two, subjects were contacted and the study was explained, with ~90% of those eligible agreeing to participate. To corroborate their preliminary LR status, the subjects received laboratory-based alcohol challenges (males: 0.75 ml/kg ethanol; females: 0.70) with both subjects and researchers blind to the contents. LR was measured using the Subjective High Assessment Scale (SHAS) consisting of 13 items graded on a 39-point scale for both positive (e.g., feeling high or intoxicated) and negative (e.g., feeling confused or nauseous) alcohol effects (Eng et al., 2005).

In step three, two fMRI sessions scheduled within consecutive weeks on non-consecutive days were given using, in random order, alcohol (same doses as in prior sessions) or placebo. Twenty minutes after drinking, subjects entered a 3T GE CXK4 Magnet scanner (General Electric Medical Systems, Milwaukee, WI) and participated in the task near the usual time of the peak breath alcohol levels (BrAC), ~60 minutes after drinking.

Baseline Task

An emotional face-processing task (Hariri et al, 2005) was administered, in which subjects were presented, three times each in random order, a 5-second exposure to three faces (a target and two probes), and told to match emotional expressions (happy, angry, or fearful) of targets and probes, with interspersed 12-second fixation periods between each face or control task. There were 18 trials (3 blocks of 6 trials) over 512 seconds.

The imaging session consisted of a 10-second 3-plane scout scan and a sagittally acquired spoiled gradient recalled sequence (FOV = 25 cm, matrix = 192 × 256, 172 sagittally acquired slices 1-mm thick, TR = 8 ms, TE = 3 ms, flip angle = 12°). Moreover, 256 axially acquired T2*-weighted echo-planar images (EPIs) were gathered (FOV = 23 cm, matrix = 64 X 64, 30 slices 2.6-mm thick, gap = 1.4 mm, TR = 2000 ms, TE = 32 ms, flip angle = 90°) with an 8-channel brain array coil.

Follow-up Evaluations for the 120 Original Subjects

At baseline, subjects gave permission for follow up, and supplied birthdates, addresses, and how to reach additional individuals likely to know their future whereabouts. With UCSD Human Research Protections Program approval, ~5 years later individuals were re-contacted for 1-hour phone interviews regarding interval events, for which they were paid $125. A close relative or friend received $100 for a separate 1-hour phone interview about the subject.

Of the 120 participants, 114 (95%) individuals (57 matched low and high LR pairs) were evaluated. Similar to baseline, interval questions covered demography, alcohol quantities, and alcohol-related problems that included the 11 DSM-IV abuse and dependence items plus blackouts, morning drinking, efforts to stop drinking, alcohol interfering with relationships, self-help group attendance, and receiving professional help for alcohol-related difficulties.

Baseline fMRI Data Analyses

Image analysis pathway: Analysis of Functional NeuroImages (AFNI) software package was used to evaluate basic structural and functional images (Cox, 1996), while a multivariate regressor approach evaluated changes in echo planar images (EPI) intensity in relation to task characteristics (Haxby et al., 2000). Image translation and rotation relative to all other images were minimized using a 3D-coregistration algorithm for each subject (Eddy et al, 1996), including six motion regressors. EPI scan slices were temporally aligned following registration to ensure that different relationships with regressors did not reflect acquisition of different slices at different times during the repetition interval.

Within Subject Multiple Regressor Analyses: The orthogonal regressors of interest were the four images (happy, angry, fearful, and circle/oval controls) which were convolved with a gamma variate function modeling a prototypical hemodynamic response (6–8 second delay) and to account for temporal dynamics of the hemodynamic response (typically 12–16 seconds), with each convolved time series normalized as a regressor of interest (Friston et al., 1995). Along with baseline, linear, and motion regressors (roll, pitch, and yaw), the AFNI program 3dDeconvolve determined the height of each regressor of interest for each subject time series. The main dependent measure was the voxel-wise normalized percent signal change created by dividing the regressor coefficient by the zero-order regressor (i.e., baseline). A Gaussian filter (FWHM 4mm) was used to account for variations in anatomy across individuals. Spatially smoothed percent signal change data were transformed into anatomical magnetic resonance image Talairach coordinates.

Group Analysis Performed at Baseline (see Paulus et al., 2012): Anatomically-constrained functional regions of interest (ROIs) were based on a priori hypotheses regarding regions of interest in the insular cortex, medial prefrontal cortex (mPFC), and amygdala based on the Talairach Atlas (Lancaster et al., 2000). A voxel-based mixed-model analysis with fixed effects of emotion type (happy, angry, fearful), group (low and high LR), education, and response latency, with the random effects as subjects was carried out. Thirteen functionally refined clusters within these anatomical regions were defined (the volume and coordinates of the centroids in Talairach space are given in Table 2). The volume minimum for these functional ROIs were determined through threshold adjustment based on Monte Carlo simulations (conducted via AFNIs in Alpha Sim, a threshold p of .05, two-tailed) guarded against identifying false positive areas of activation. Percent signal change for the contrasts of interest were extracted from each of the 13 previously defined ROIs (Paulus et al., 2012; Table 2) and used in the statistical analysis. All analyses covaried for both cerebral blood flow (CBF) (overall blood flow differences in ml/100g/min for placebo vs. alcohol sessions) and for usual drinks per occasion.

Table 2. Emotional Face Recognition Task.

Correlations^∧ among Baseline BOLD Response Contrasts for ROIs, and Drinking Outcomes (Maximum Quantities and 1+ Alcohol Problem) 5 Years Later for 114 Subjects

Brain Region Bold Response Contrast Paulus et al, 2012	Maximum Quantity	Any of 17 Alcohol Problems
Happy Face
[1] R inferior frontal gyrus [1792,54/32/0]
Placebo	−.04	.00
Alcohol	−.07	.12
Difference√	.02	−.08
[2] R middle frontal gyrus [640,34/48/−8]
Placebo	−.08	−.18x (p=.062)
Alcohol	−.07	−.06
Difference	−.00	−.07
[3] L anterior insula [576,−46/12/0]
Placebo	−.04	−.09
Alcohol	.03	.17x (p=.072)
Difference	−.05	−.18x (p=.059)
[4] Anterior cingulate [960,2/8/−8]
Placebo	−.02	−.06
Alcohol	−.03	.06
Difference	.01	−.08
[5] R Posterior insula [512,50/−24/20]
Placebo	−.08	−.07
Alcohol	.01	.07
Difference	−.07	−.11
[6] L inferior frontal gyrus [576,−42/28/−4]
Placebo	.08	.03
Alcohol	.02	.08
Difference	.08	−.03
[7] L middle frontal gyrus [960,−34/48/−12]
Placebo	.03	.04
Alcohol	−.06	.02
Difference	.07	.02
[8] R middle insula [576,42/−4/8]
Placebo	−.09	−.17x (p=.069)
Alcohol	.07	.10
Difference√	−.11	−.19a
[9] L middle insula [576,−42/0/4]
Placebo	−.09	−.08
Alcohol	−.06	.19a
Difference	−.02	−.18x (p=.056)
Angry Face
[10] L middle frontal gyrus [1280,−50/20/32]
Placebo	.12	.09
Alcohol	.01	.00
Difference	.09	.07
[11] R middle frontal gyrus [704,46/12/40]
Placebo	.23a	.14
Alcohol	.07	.14
Difference	.16x (p=.082)	.04
Fearful Face
[12] Anterior cingulate [576,−2/40/−8]
Placebo	.06	.09
Alcohol	.07	.13
Difference	−.02	−.02
[13] L middle frontal gyrus [512,−34/52/−8]
Placebo	.01	.13
Alcohol	.02	.01
Difference	−.01	.10
LR relation to outcomes	−.20a	−.24b

^∧Pearson Product Moment Correlations

^xp ≤ .10;

^ap ≤ .05;

^bp ≤ .01;

^cp ≤ .001

ROI: Regions of Interest related to low vs. high LR in prior paper (Paulus et al., 2012)

ROI details format as per Paulus et al., 2012 [in parentheses are volumes in mm³ ; x/y/z (centroid locations in Talairach coordinates)]

LR: Level of Response

L = Left, R = Right

Outcome time frames are the same as for Table 1

Shaded rows highlight ROIs that relate at p<.10 or better to 1+ outcomes

^√The difference score was computed by subtracting the Alcohol BOLD response contrast value from the Placebo BOLD response contrast value.

Follow Up Statistical Analyses

Using the Statistical Package for the Social Sciences (SPSS), comparisons of items across low and high LR subjects and between baseline and follow up data used dependent samples t-tests for continuous variables, and the Wilcoxon signed ranks test (reported as a Z statistic) for dichotomous items. Relationships between baseline BOLD response contrasts and outcomes were evaluated through scatter plots and with Pearson’s product-moment correlations. Logistic hierarchical and two types of backward elimination regression analyses ( used to evaluate a modest-sized set of prior-selected variables allowing identification of the most robust predictors of outcomes) evaluated Hypothesis 2 regarding whether BOLD response contrast for any ROIs predicted alcohol outcomes above and beyond the variability predicted by LR. This included an augmented backward elimination (ABE) regression (Dunkler et al., 2014) that uses both significance and change-in-estimate criteria to diminish any bias from selecting only one among a set of correlated variables. The alcohol outcomes were not always normally distributed, a problem addressed by square root transformations.

Recognizing both that multiple testing was involved in these analyses and that the sample was modest sized, Type 1 errors were evaluated using the exact binomial procedure. This approach was chosen to correct for multiple testing because Bonferroni corrections are overly conservative (Bland, 2000), especially for modest sized samples.

Results

Subjects at baseline and outcome

Table 1 presents the demography and substance use characteristics for the 114 followed participants divided into high and low LR groups at baseline. At study entry, the LR groups were similar on demography and most alcohol use characteristics. Regarding baseline alcohol problems, while no person met criteria for alcohol dependence, 40% ever experienced alcohol-related blackouts, 6% needed more alcohol to experience effects (tolerance), and ~3% each ever consumed more drinks than intended, missed social occasions because of drinking, and/or had alcohol interfere with school or work. During the follow-up, both lower and higher LR subjects increased drinking frequencies, maximum quantities, and alcohol problems. Regarding problems during the follow-up, 70% had 1+ blackouts, 45% drank more than intended, ~25% each spent more time drinking than planned or drank before noon, and ~15% each reported tolerance, noted that alcohol had interfered with work/school, or that they ever skipped important activities in order to drink.

Table 1.

Baseline and Follow Up Demography and Substance Use for 114 Subjects Divided (Median Split) into Alcohol Challenge-Based Low and High Level of Response (LR) to Alcohol

	Low LR (N=57)		High LR (N=57)		t-test/χ2 Low vs. High LR Baseline	t-test/χ2 Low vs. High LR Follow Up	t-test/Z Alcohol Related Changes from Baseline to Follow Up
	Baseline	Follow Up Interval	Baseline	Follow Up Interval
Demography
Age	19.8 (1.54)	24.9 (1.78)	19.9 (1.37)	25.2 (1.66)	−0.58	−0.87	NA
Female %	50.9		54.4		0.14	NA	NA
Years Education	13.5 (1.12)	16.4 (1.13)	13.8 (1.21)	16.3 (1.17)	−1.28	0.16	NA
							Low LR	High LR
Alcohol Use
Usual Frequency	6.9 (4.33)	11.5 (6.59)	6.5 (5.11)	11.0 (7.47)	0.70	0.57	−4.81c	−4.49c
Usual Quantity	3.7 (1.61)	3.5 (1.80)	3.3 (1.86)	3.3 (1.39)	1.29	0.60	−2.72 b	−4.03c
Maximum Quantity	10.1 (3.81)	10.9 (4.69)	8.5 (4.13)	10.1 (4.18)	2.40a	0.98	−2.97c	−4.80c
% Any of 17 Alcohol Problems	49.1	86.0	38.6	77.2	1.28	1.46	−4.38c	−4.49c
Mean # 17 Problems	0.7 (0.92)	2.6 (2.24)	0.4 (0.60)	2.5 (2.77)	1.46	1.09	−7.28c	−7.50c

^ap ≤ .05,

^bp ≤ .01,

^cp ≤ .001

Continuous variables presented as means and standard deviations

Z = Wilcoxon signed ranks test

Alcohol outcomes for drinking frequencies and usual quantities are for the 6 months prior to the baseline assessment and to follow up

Maximum drinks and problems are lifetime before baseline and for the entire follow up interval at follow up

Follow-up alcohol outcomes were similar for low and high LR subjects based on the original LR median split. However, consistent with Hypothesis 1, as shown at the bottom of Table 2, when baseline LRs were computed as continuous measurers, lower alcohol responses correlated with higher later maximum alcohol quantities and alcohol problems.

Baseline BOLD response contrasts predicting outcomes

Table 2 addresses Hypothesis 2 regarding whether BOLD response contrasts in the 13 ROIs identified for this task in the Paulus et al. (2012) cross-sectional paper correlated with future alcohol quantities and/or problems. Focusing on the two outcomes that correlated significantly with LR in Table 2 (maximum drinks and alcohol problems) among the nine happy faces-related regions, four (ROIs 2, 3, 8, and 9) correlated at p≤.072 with alcohol problem outcomes, and among angry faces one ROI baseline BOLD response contrast (ROI 11) correlated with future maximum alcohol quantities at p<.05 and p=.082 for two baseline measures. Using the exact binomial, the probability that five ROIs would correlate with the two outcomes across 13 regions is p<.05. Within the five regions that correlated with an adverse outcome (ROIs 2,3,8,9 and 11) there were 30 evaluations (five ROIs with three conditions each [placebo, alcohol, and difference scores] across two outcomes), nine of which correlated with baseline fMRI results at p≤.082, creating a significant exact binomial that the outcome was beyond chance (p<.001). Looking at the outcome with the highest number of notable correlations with baseline BOLD response contrast values, alcohol problems, among 39 possible evaluations (13 ROIs with three conditions each) the chances of finding seven correlations that related to baseline BOLD response contrast at p≤.072 was p<.02.

Regression analyses

The analyses next addressed Hypothesis 3 regarding whether the variances in relevant drinking-related outcomes were accounted for by baseline brain imaging even after considering baseline LR (see Table 3). These evaluations began with seven separate hierarchical regression analyses considering the seven baseline conditions (placebo, alcohol, or their difference score) in five ROIs that related to alcohol problems (the most consistent outcome). Each of these seven regressions entered both LR the baseline BOLD response contrast value to predict follow-up alcohol problems. The top of Table 3 describes the four regions where the baseline BOLD response contrast values incremented over LR in predicting problems. The second step considered the fact that, with rare exceptions (data not shown), the seven ROI BOLD response contrasts correlated significantly with each other with r’s between .29 and .76. Therefore, the four baseline ROI values that incremented over LR and related to future alcohol problems were entered along with LR into a single simultaneous entry regression analysis, with the result that only LR entered significantly, with R² =.21, p=.008. Third, we noted the high correlations among the fMRI measures and the fact that when entered into a regression along with LR four ROIs incremented over LR in predicting alcohol outcomes, which raised the possibility that in simultaneous regressions the overlap between ROIs might disallow any one of them to enter as a significant contributor to the equation. Therefore, we carried out a backward elimination regression analysis to better determine which ROI values remained related to future alcohol problems. The results in Table 3 indicate that LR, ROI 2, and ROI 3 significantly contributed to the regression analysis, with R² =.21, p=.001. The backward elimination ABE approach yielded the same model with R² =.21, p=.001.

Table 3.

Logistic Regressions Predicting having any of the 17 Outcome Alcohol Problems

Simultaneous Entry
Variable	Odds Ratio (95% Confidence Interval)
LR	0.47 (0.29 – 0.79)b
[2] R middle frontal gyrus (placebo)	0.64 (0.34 – 1.17)
[3] L anterior insula (alcohol)	1.57 (0.62 – 4.00)
[3] L anterior insula (difference)	0.93 (0.40 – 2.15)
[9] L middle insula (alcohol)	1.11 (0.46 – 2.67)
R2 = .21p=.008
Backward Elimination
Variable	Odds Ratio (95% Confidence Interval)
LR	0.47 (0.28 – 0.77)b
[2] R middle frontal gyrus (placebo)	0.61 (0.36 – 0.99)a
[3] L anterior insula (alcohol)	1.78 (1.02 – 3.06)a
R2 = .21p=.001

Abbreviations are the same as in Tables 1 and 2

Independent variables were standardized for ease of interpreting Odds Ratios

The last step in these analyses focused on searching for commonalities among the baseline BOLD response contrasts for the ROIs that correlated with adverse alcohol outcomes. In Table 4, columns 1–3 describe the four baseline conditions and related ROIs that were entered into the regression analysis in Table 3. Columns 4 and 5 present the baseline BOLD response contrast characteristics for those brain regions listed in the original Paulus et al. (2012). As shown in columns 4 and 5, in all relevant baseline conditions low LR subjects demonstrated baseline BOLD response contrasts that were higher for placebo than for alcohol, while for high LRs none demonstrated this pattern.

Table 4.

Summary of Baseline Patterns^√ of LR Group Differences for 4 fMRI Happy Face ROIs

ROI	ROI Name	Condition	Low LR Placebo>Alcohol	High LR Placebo>Alcohol
2	R middle frontal gyrus	1. Placebo*	Yes	No
3	L anterior insula	2. Alcohol*	Yes	No
		3. Difference
9	L middle insula	4. Alcohol	Yes	No

Abbreviations are the same as in Tables 1 and 2

ROI: Regions of Interest from Tables 2 and 3

^*Contributed significantly to the Backward Elimination Regression

^√Paulus et al., 2012

Discussion

The data for the current analyses utilized 240 baseline placebo and alcohol fMRI sessions with 120 drinking but not alcohol dependent young adults, which had been collected ~5 years previously, to determine whether fMRI BOLD contrasts predicted follow-up drinking-related patterns. The baseline analyses (Paulus et al., 2012) were consistent with prior investigations (Paulus et al., 2006; Tapert et al., 2004) where individuals with low LR tended to demonstrate greater fMRI BOLD response contrasts during cognitive or emotional-based tests with placebo compared to alcohol, while high LR subjects showed the opposite pattern, with no task performance differences across LR groups. The current work examines whether the neural activation to affective tasks as a function of LR status predicted drinking outcomes, and, if so, if the fMRI results added to LR in those predictions. The answer to both these questions was affirmative, although the small sample requires that these results be considered preliminary.

The first step in the current stage of this work was to determine if a lower LR related to future heavier drinking and alcohol problems (Hypothesis 1) in a manner consistent with prior work (e.g., Chung & Martin, 2009; Quinn & Fromme, 2011; Schuckit et al., 2007, 2008; Volavka et al., 1996). The results using a baseline continuous measure of LR revealed that the lower the intensity of reaction during the alcohol challenge, the higher the follow-up maximum drinks and the greater the number of alcohol problems (see Table 2). These two outcomes have been more consistently related to earlier LR than drinking frequency or usual drinks per occasion in prior studies. The fact that a median LR split did not identify those findings may reflect the limited power of dichotomous measures.

The second hypothesis was that baseline fMRI patterns that related to a low baseline LR would predict future higher alcohol quantities and associated problems. The pattern of the current data lends support to this prediction in that baseline BOLD response contrasts in five of 13 regions related to at least one alcohol outcome five years later. The outcome most consistently related to baseline fMRI results was alcohol problems.

Those findings allowed for testing Hypothesis 3. Here, in four of the seven baseline conditions where ROIs related to future alcohol problems, baseline BOLD response contrast predicted that adverse alcohol outcome above and beyond the outcome predicted by LR alone. Backward elimination regression analyses that included LR and four ROIs (right middle frontal gyrus, left anterior insula, and left middle insula) indicated that activation to affective faces in the right middle frontal gyrus and the left anterior insula were the strongest predictors of future alcohol problems.

The involvement of the insular cortex in predicting drinking outcomes adds to the growing literature indicating that this region, an integrator of the body-brain connection, may be related to the motivation to use substances (Paulus & Stewart, 2014; Verdejo-Garcia et al., 2012), especially in the context of approach or avoidance behaviors. Potentially pertinent to LR processing, the insula influences the nucleus accumbens and striatum downstream, with possible impacts regarding reward-related behaviors and craving (Cho et al., 2012; Goldstein et al., 2009). The frontal systems are important in cognitive effort (Corbetta & Shulman, 2002) and affective modulation of stress (Miller & Cohen, 2001), two processes that may be vital to alcohol effects and recovery from their influences (Franken, 2003).

There are several potential implications of these data. First, they indicate that BOLD response contrast results related to the low LR to alcohol can predict drinking behaviors over extended periods of time. The modest proportion of the variance in predicting outcomes reported here is consistent with prior studies of heavy drinking and AUDs that underscore the importance of many genetic and environmental factors in the genesis of adverse drinking outcomes, with no one element both necessary and sufficient (Schuckit, 2014). Regarding LR, the finding that perturbations in the prefrontal and insular regions appear to be associated with both a low LR and future heavy drinking may improve our ability to more accurately identify the LR-related phenotype compared to results from alcohol challenge-related changes in subjective feelings of intoxication and related characteristics (see Paulus et al., 2012). This, in turn, may open the door to a more optimal multi-level process model that helps explain at least one aspect of the vulnerability toward problem drinking associated with LR itself.

Another potential implication of these data is highlighted in Table 4. The most consistent baseline fMRI-related pattern was that subjects with low LR were more likely than those with high LR to demonstrate higher BOLD response contrast under placebo conditions than when after drinking alcohol. This is consistent with lower alcohol responses for subjects with low LRs in prior studies evaluating alcohol responses for EEG, ERP, ACTH and prolactin (Ehlers et al., 1999; Schuckit & Gold, 1988; Schuckit et al., 1988a, 1988b). While from this perspective the fMRI results were similar to other measures of alcohol sensitivity, it is important to note that the fMRI BOLD response contrasts added significantly to LR in predicting outcomes.

The data are also consistent with our previous interpretation (Paulus et al., 2012) that the BOLD response contrasts might indicate that individuals with a low LR may need greater cognitive effort to recognize relatively subtle differences across stimuli. This increase of neural processing resources during effortful tasks when not exposed to alcohol might indicate that it is possible to identify individuals prone to low LRs to alcohol and consumption of higher alcohol quantities even before their first drink. As shown in Table 4, among the three ROIs that related to future adverse alcohol outcomes, for low LRs all three demonstrated the pattern of placebo values being greater than alcohol values while this was not true for any of the ROIs for high LRs.

It is important to consider potential future steps for these studies. First, in light of the great expense and large amounts of staff and subject time involved in fMRI measures, it will be worthwhile to continue to search for alternate, less costly approaches that might measure phenomena that overlap with the fMRI patterns. Second, additional work is needed to further elucidate the brain mechanisms that contribute to these fMRI results and how they relate to drinking practices. Third, the current data neither definitively support nor refute the possibility that the relatively high BOLD response contrasts during a cognitive task after placebo can by themselves predict future low LRs and/or future heavier drinking in subjects even before their first drink. Perhaps that might be best evaluated as an add-on to prospective investigations of fMRI in the future.

This investigation has several strengths, including a 95% five-year follow up, and careful measures of both LR and BOLD response contrasts. At the same time it is important to recognize that the reaction to alcohol is only one of several intermediate phenotypes related to the AUD risk (e.g., impulsivity), and that there are several additional characteristics related to how a person reacts to alcohol, including an exaggerated alcohol response at rapidly rising BACs (King et al, 2013; Schuckit, 2014). It will be interesting to establish whether functional brain imaging might have a role in predicting drinking outcomes related to other types of LR measures and additional risk factors for AUDs. A second important caveat is that the current sample of 114 subjects is underpowered for analyses that consider 13 ROIs generating 39 values across placebo, alcohol, and a score that reflects the difference between those two conditions. Thus, more conservative steps to control for Type 1 errors could not be applied, and we used the less conservative exact binomial approach. As a consequence, the current results may be an overly optimistic estimate of data fit, and must be considered preliminary until larger studies are carried out. Third, we only studied young EA drinkers from a single university in Southern California. More work is needed to see whether similar results are applicable to other markers of risk and to alternative populations. Fourth, the focus of the work is on alcohol, as low LRs have only been found to relate to adverse alcohol outcomes and not to the use of or problems with other drugs. Finally, it is important to remember that there may be additional brain regions of interest to LR and to the related prediction of adverse alcohol outcomes.