The impact of elevated body mass on brain responses during appetitive prediction error in postpartum women

Abstract

Repeated exposure to highly palatable foods and elevated weight promote: 1) insensitivity to punishment in striatal regions and, 2) increased willingness to work for food. We hypothesized that BMI would be positively associated with negative prediction error BOLD response in the occipital cortex. Additionally, we postulated that food reinforcement value would be negatively associated with negative prediction error BOLD response in the orbital frontal cortex and amygdala. Postpartum women (n=47; BMI= 25.5±5.1) were ‘trained’ to associate specific cues paired to either a highly palatable milkshake or a sub-palatable milkshake. We then violated these cue-taste pairings in 40% of the trials by showing a palatable cue followed by the sub-palatable taste (negative prediction error). Contrary to our hypotheses, during negative prediction error (mismatched cue-taste) versus matched palatable cue-taste, women showed increased BOLD response in the central operculum (pFWE=0.002; k=1680; MNI: −57, −7,14) and postcentral gyrus (pFWE=0.006, k=1219; MNI: 62, −8,18). When comparing the matched sub-palatable cue-taste to the negative prediction error trials, BOLD response increased in the postcentral gyrus (r = −0.60, pFWE=0.008), putamen (r = −0.55, pFWE=0.02), and insula (r = −0.50, pFWE=0.01). Similarly, viewing the palatable cue vs sub-palatable cue was related to BOLD response in the putamen (pFWE=0.025, k=53; MNI: −20, 6, −8) and the insula (pFWE=0.04, k=19, MNI:38, −12, −6). Neither BMI at 6-month postpartum nor food reinforcement value was related to BOLD response. The insula and putamen appear to encode for visual food cue processing, and the gustatory and somatosensory cortices appear to encode negative prediction errors. Differential response in the somatosensory cortex to the matched cue-taste pairs to negative prediction error may indicate that a palatable cue may dull aversive qualities in the stimulus.

1. INTRODUCTION

Increased energy needs and weight gain characterize pregnancy [1]. However, in the last two decades pregnancy weight retention has become a risk factor for future development of obesity in both the mother and child [2–7]. Despite substantial evidence of the negative consequences of postpartum weight retention, interventions have been generally unsuccessful [8,9]. Desire to consume palatable foods is a noted barrier to weight loss in women [10] and overconsumption of palatable foods has been linked overweight and obese status [11]. Understanding the desire and motivation to over consume palatable foods may inform the development of more effective intervention strategies.

Intake of palatable food, a hedonic reward, is a multidimensional behavior comprised of learning, incentive motivation, and pleasure [12]. While the expected value of a stimulus [13], salience of both a cue and a stimulus [13] are important components of learning; Mirenowicz and Schultz’ prediction error theory posits that the brain learns about rewarding stimuli through coding unexpected outcomes as pleasurable or unpleasant if they are better or worse than the expected respectively [14,15]. As a simple example, consider a tray of cookies that appear to be chocolate chip. However, upon biting into a cookie one realizes that it is in fact an oatmeal raisin cookie, a flavor the individual does not enjoy. As the person dislikes oatmeal raisin cookies, but was expecting a highly pleasant chocolate chip cookie, she/he experienced negative prediction error; characterized as an anticipated pleasant outcome that, when experience, is worse than expected. As a function of this reinforcing learning experience, the person now has learned that this tray of cookies contains different cookies and cookie selection requires closer attention. As such, prediction errors act to update the brain with new information based on reinforcement experienced and disregard expected responses (those that are not better or worse than expected) [15]. Within the brain, this can be quantitatively measured via dopamine signaling from the ventral tegmental area [16,17]. When a stimulus is better than expected (positive prediction error) it results in higher dopamine response [18], and when a stimulus is worse than expected (negative prediction error) it attenuates the dopamine response [18]. Therefore, the effect of a stimulus when it is different than expected is a useful measure of outcome-based learning [19].

Preclinical rodent models indicate that dopamine underpins cue-reward learning and modulates habit formation [20]. Ad libitum access to a cafeteria diet, mimicking an obesogenic environment, was found to decrease dopamine signaling in the striatum with concurrent insensitivity to environmental cues predicting adversity [21]. These rodent models suggest that an obesogenic environment may induce habit formation via dopaminergic signaling; behaviorally, this may present as maintaining a response despite negative outcomes. Together, these findings show that an obesogenic environment is enough to induce habit formation through altered dopaminergic functioning. Functional magnetic resonance imaging (fMRI) offers a non-invasive method of assessing the relationship between prediction errors and neural correlates of dopaminergic function in humans [22]. For example, blood oxygen level dependent (BOLD) response changes to prediction errors in the striatum [23,24]. Previous work has demonstrated individuals with obesity compared to healthy weight peers showed lower incorporation of negative prediction errors to improve reward learning on a non-food related task [23]. This is consistent with a large body of research showing that individuals with obesity show dysfunctional striatal response when presented with a sweet reward [25,26]. Alterations in dopaminergic signaling is not limited to the striatum. The occipital cortex, predominately considered a ‘feature detection’ region, has been shown to encode mismatches between predictions and inputs [27,28]. In particular, the occipital cortex has been shown to be sensitive to aversive prediction errors, with increased BOLD response in this region to an aversive shock when a monetary reward was expected [29] and this sensitivity is due to midbrain dopaminergic signaling [30]. Further, participants with obesity showed increased BOLD activity in the occipital cortex to a probabilistic prediction error paradigm [31]. Insensitivity to negative reward learning in the striatum appears to be related to habit formation and weight gain, and may be reflected in increased BOLD response in the occipital cortex with increasing BMI [20,21,23,31].

While much work has been done associating reward learning to dopaminergic functioning, less is known about how insensitivity to negative reward learning is associated with food-related behaviors. This is despite the fact both humans and animals will work harder to obtain rewards [32]. Preclinical rodent models show that increased motivation to work for food is associated with decreased dopamine signaling via dopamine 2 receptors (D2) [33]. Rodent models have also shown the importance of the amygdala in maintaining responses during a progressive ratio task [34]. The amygdala has been shown to be important for associative learning and is modulated by prediction errors through dopaminergic axons [35,36]. Connections between the orbital frontal cortex (OFC) and the amygdala further promote reward-seeking behaviors [37]. However, individuals with substance dependency were less sensitive to loss in the modified Iowa Gambling Task, and fMRI showed the BOLD response in the OFC did not track prediction errors compared to healthy individuals [38]. Therefore, this suggests that decreased BOLD response in the OFC may indicate an insensitivity to punishment, and a higher motivation to work for a food reward and the amygdala may modulate this relationship. In human studies, motivation for food rewards can be assessed via a progressive ratio task, in which one must work (via pressing a button) more each session to earn the same amount of food [39]. Greater motivation to work for food has been previously associated with higher BMI [40,41]. The relationship of insensitivity to negative reward learning with motivation for food reward and BMI is unknown. Thus, the present study sought to evaluate how both BMI and food reinforcement value at 6 months postpartum are related to neural response to negative prediction error. Postpartum women are uniquely susceptible to weight retention and obesity, and thus, represent an important population for examining the relationship of prediction error with BMI. We hypothesized that BMI would be positively associated with negative prediction error BOLD response in the occipital cortex. Additionally, we postulated that food reinforcement value would be negatively associated with negative prediction error BOLD response in the orbital frontal cortex and amygdala.

2. METHODS

2.1. Recruitment

A convenience sample of women from the Pregnancy Eating Attributes Study (PEAS) cohort were recruited to participate within two-weeks of the their six-month postpartum visit [42]. All participants that met inclusion criteria for fMRI procedures (e.g., no fMRI contra-indicators such as metal implants) were given the option to participate in the fMRI sub-study (Figure 1). Women were excluded if they reported use of medications that impact body weight (e.g., insulin, glucocorticoids) or anti-depressants. One-hundred and sixty-one women were emailed invitations to join the study. A total of 75 women expressed interest in participating. Twelve women missed the study participation window, one was disqualified due to medication, and two chose not to participate. Of the 60 women assessed, 50 had complete scanning data for this analysis. All participants provided written informed consent. The University of North Carolina at Chapel Hill’s Institutional Review Board approved all methods. All study visits took place at the University of North Carolina at Chapel Hill’s (UNC) Biomedical Research Imaging Center (BRIC).

Figure 1. Diagrams of the prediction error paradigm.

A. Diagram representing the possible cue/tastant combinations.

B. Diagram outlining a block of the prediction error paradigm. A cue of either a sub-palatable or palatable milkshake, was followed by either the expected (matched) tastant or in 40% of trials the palatable cue was matched with the sub-palatable milkshake resulting in a prediction error. All taste events were followed by a fixation cross, a neutral rinse, and a variable length jitter before the next block.

2.2. Procedures

Participants were asked to refrain from eating or drinking for at least 4h before their imaging session. The study visits began between 5am and 3pm. Hunger, fullness and degree of thirst were assessed before and after the scan via visual analogue scale (VAS). Body mass index (BMI kg/m²) was calculated with height and weight that were measured using a digital scale and wall-mounted stadiometer, to the nearest 0.1 kg and 0.1 cm.

Scanning was performed on a Siemens Prisma 3T scanner at the BRIC to collect functional and anatomical imaging data. Visual stimuli were presented with a digital projector/reverse screen display system. Tastants were delivered using programmable syringe pumps (Braintree Scientific BS-8000, Brain-Tree, MA) operated through a in house-developed program written in python (available here: https://github.com/niblunc/taste_task) to ensure consistent volume, rate, and timing of taste delivery. A manifold attached to the head coil fit into the participants’ mouths and delivered the tastes to a consistent tongue segment.

A susceptibility weighted single shot echo planar sequence imaged the regional distribution of the BOLD signal with TR = 2000 ms, TE = 20ms, flip angle = 80°, with an inplane resolution of 3.0 × 3.0 mm² (64 × 64 matrix; 192 × 192 mm² field of view) was acquired during functional runs. To cover the whole brain, 32 4mm slices (interleaved acquisition) were acquired along the AC-PC transverse, oblique plane as determined by the midsagittal section. Slices were acquired in an interleaved mode to reduce cross talk of the slice selection pulse. High resolution, T1 weighted 3D volume was acquired with a TR/TE of 2100ms/2.4ms, flip angle of 15°, TI of 1100ms, matrix size of 256×256, FOV of 22cm, slice thickness of 1mm was acquired for the anatomical scan.

2.3. Training paradigm

We used an adapted block version of the taste receipt paradigm, which assesses activation in response to receipt of a palatable chocolate milkshake (3.6g fat, 2.8g sugar/fl oz; made with half and half, chocolate Nesquik© powder, and simple syrup), a sub-palatable (unsweetened; unappealing) chocolate milkshake (0.4g fat, 1.5g sugar/fl oz; made with skim milk, unsweetened cocoa powder, and soy emulsifier) and a tasteless, odorless solution containing the main ionic components of saliva (25 mM KCl, 2.5 mM NaHCO₃ in distilled water) as a control contrast. Subjects randomly received the 3 fluids through individual beverage tubes, which were anchored to the head coil. During tastant delivery block, a taste-paired cue (2 sec) was presented followed by a blank screen; the paired tastant was then delivered (3 sec; 0.5mL). This was followed by a waiting period (2 sec), a rinse with the neutral solution (3 sec; 0.5mL), and variable jitter (4–9 sec). Each run contained 45 blocks, 15 blocks of each taste randomly presented. Two runs (11 min, 40 secs per run) were performed for a total time just over 23 minutes.

2.4. Prediction error paradigm

We used a probabilistic reward paradigm to assess activation in response to negative prediction error (Figure 1). The paradigm was similar to the training paradigm except in 40% of the blocks, the cue associated with the palatable milkshake was paired with the sub-palatable milkshake, creating a negative prediction error event. Each run contained 24 blocks. In 12 blocks the cue accurately predicted the taste (as trained in the previous training paradigm). These are considered matched palatable and matched sub-palatable depending on the taste given. In 10 blocks the cue for the palatable milkshake was presented with the taste of the sub-palatable milkshake; this is considered a prediction error event. All blocks were randomly presented. Two runs (6 min and 30 sec per run) were performed for a total time of 16 minutes.

2.5. Perceptual hedonic and internal state measures

Explicit perceived ‘pleasantness’, ‘sweetness’, ‘familiarity’, and ‘intensity’ of the milkshakes were measured using adapted labeled hedonic scales after the scan. Scales were tailored to the beverages and were anchored by (−100) ‘least imaginable’ to (100) ‘most imaginable’, and ‘neutral’ (0) in the middle. Hunger, fullness, and thirst was assessed in a similar fashion before and after the scan assessment.

2.6. Food reinforcement task

After the scan participants preformed a progressive food reinforcement value task [43]. In this task, participants worked to earn points toward a snack food reward of their choice (e.g., Doritos® nacho cheese tortilla chips, Reese’s® mini peanut butter cups, Skittles® fruit candies, or Smartfood® white cheddar popcorn). Participants first performed a taste test of 1g of each food and rated each food on a cross-modal visual analogue scale. Participants then selected the snack food they wanted to earn. In the second phase, three boxes varying in shape and color were displayed on a computer screen. Points were earned each time the shapes match in color and shape after the participant made a button press. The task started at a variable ratio 1/4 (VR4) schedule meaning that, on average, one point is awarded for four button presses. The progressive ratio schedule for the food item doubled (VR8, VR16, VR32) each time they earned ten points. A total of ten points was rewarded with a 1/2 standard portion of the food (per the nutritional information for the snack). Participants played for as long as they liked. The break point at which the participant stops button pressing for food was used as the behavioral measure of anticipatory food reward (i.e., how many button presses are made in total before the subject stops). This food reinforcement value task shows 2–7 day test-retest reliability (r = .80) [44]. Validity of the task is supported by findings that food reinforcement value as measured by the task is associated with higher hedonic ratings of snack foods, greater ad libitum food intake, and obesity; additionally subjects work harder for food when food deprived [43–45].

2.7. fMRI Preprocessing and Modeling

Functional imaging data was processed and analyzed with FSL (FMRIB’s Software Library, www.fmrib.ox.ac.uk/fsl). BET was used to extract the brain from the skull, and MCFLIRT was used for motion correction. A participant’s run was excluded for movement greater than 3.0 mm in any direction. Spatial normalization was performed using nonlinear registration with FSL’s FNIRT tool. Individual’s data was then smoothed with a 6-mm full-width at half-maximum isotropic Gaussian filter. Confound motion parameter regressors were created for TRs with a framewise displacement (FD) greater than 0.9. These FD-based motion regressors were modeled along with the 6 motion parameters (from MCFLIRT), and the derivatives of each. A functional run for a subject was excluded if more than 20% of all the volumes were tagged as high motion. Three subjects were removed from analysis due to high motion. Within-run parameters were prewhitened and estimated at the first level using FSL’s FILM. Contrasts of interest were: 1) anticipatory palatable milkshake cue vs. sub-palatable milkshake cue; 2) matched palatable cue and milkshake taste vs. matched sub-palatable cue and milkshake taste; 3) negative prediction error (response to the sub-palatable taste that was preceded by the palatable milkshake cue) vs. matched palatable cue and taste, which tests the effect of sub-palatable taste independent of cue; and 4) negative prediction error vs. matched sub-palatable cue and taste, which tests for negative prediction accounting for the overall effect of sub-palatable taste. Contrast 1 was modeled during the cue presentation, whereas contrasts 2 through 4 were modeled during the taste administration. A fixed-effects model estimated within-subject effects at the second level and group effects were estimated using a mixed-effects model (FSL’s FLAME1) at the third level, which appropriately addresses the potential presence of differences in variance across groups. At the third level, the following covariates were assessed: BMI, average response to the reward reinforcement task, and the multiplicative interaction of BMI and average response to the reward reinforcement task. Voxel wise nonparametric permutation testing was preformed using FSL’s Randomise, with 5000 permutations. All statistical maps additionally are based on the threshold-free cluster enhancement (TCFE) generated from Randomise [46].

Non-imaging data was analyzed using R (version 3.3.1, R Foundation for Statistical Computing Platform, Vienna, Austria). Paired student’s T-tests were used to compare pre and post scan fullness, hunger, and thirst; and to compare familiarity, pleasantness, desire to consume, and pleasantness of the palatable and sub-palatable milkshakes. One-way analysis of variance (ANOVA) tests were used to assess the relationship between BMI and pre and post scan fullness, hunger, and thirst, as well as the relationship between BMI and pleasantness ratings of both the palatable and sub-palatable milkshakes. All analytic code is available on github (https://github.com/grace-shearrer/PEAs).

2.8. Post Hoc Analysis

After the initial analyses, we performed post hoc analyses to assess the effect of the perceived pleasantness of both the palatable and sub-palatable on BOLD response in the significant contrasts. Pleasantness ratings of both the milkshakes were modeled in the third level model (described in detail above) with the original BMI, average response to reward reinforcement, and the interaction between BMI and average response to reward reinforcement covariates. Pleasantness ratings were added to the model if the taste was given in the contrast. Therefore, the sub-palatable and palatable pleasantness ratings were separately modeled in the prediction error > matched palatable taste and cue contrast, as this contrast contains both the sub-palatable and palatable milkshakes. However, in the prediction error > matched sub-palatable taste and cue, only the pleasantness ratings of the sub-palatable milkshake were modeled as only the sub-palatable taste is included in this contrast.

3. RESULTS

Demographics of the participants are summarized in Table 1. Generally, the sample was predominately white with an above average household income for 2017 [47]. Participants were, on average, 25.2± 2.4 weeks postpartum and 70% reported still breastfeeding. They reported hunger of 36.3± 37.6 (scale −100 to 100) at the time of the scan. There was no difference in hunger, fullness, or thirst before or after the scan (p’s 0.3 – 0.8). Participants rated the palatable milkshake significantly more pleasant than the sub-palatable milkshake (mean and standard deviation of palatable 26.6± 34.4 and sub-palatable −47.0± 29.7; p<0.001). Hunger before the scan and pleasantness of the palatable milkshake were significantly related to BMI (β = 0.37, p=0.01 and β = .40, p=0.006). The average food reinforcement value across the sample was 54.9± 24.1 (range 32.0–143.5). BMI was not significantly related to prescan fullness or thirst, nor post scan hunger, fullness, thirst, hedonic ratings or food reinforcement value (p values = 0.09–0.7).

Table 1.

Demographics

Demographic n= 47	Mean± SD	Range
Age (years)	31.5± 4.5	(19, 42)
BMI	25.6± 5.1	(17.4, 37.2)
Average Reinforcers (points)	9.7± 0.6	(8, 10)
Average Response (clicks)	54.9± 24.1	(32.0, 143.5)
Session Count (block of 10 points)	1.7± 0.9	(1, 4)
Hunger
Before Scan	36.6± 37.6	(−100, 100)
After Scan Fullness	35.2± 34.2	(−100, 99)
Before Scan	−55± 44.6	(−100, 16)
After Scan Thirst	−57.6± 37.9	(−100, 32)
Before Scan	20.2± 33.7	(−79, 100)
After Scan Socio-economic status	26.6± 31.4	(−88, 83)
Average household income	$74777.78± $31729.60	($15000, $105000)
SNAP1 Race1	6%
White	77%
Non-Hispanic Black	13%
Native American	4%
African American/ White	4%
Asian/ Pacific Islander Ethnicity1	2%
Hispanic	15%
Non-Hispanic	85%
Sleep quality (previous evening) 1
Very bad	4%
Fairly bad	17%
Fairly good	59%
Very good Breastfeeding status1	19%
Currently breastfeeding	70%
Not breastfeeding	30%
Menses returned?1
Yes	57%
No	0%
I don’t know	38%

¹Percent frequency

3.1. BOLD response to palatable and sub-palatable taste and cues

When examining the contrast of anticipatory palatable milkshake cue > sub-palatable milkshake cue we observed significant BOLD response in putamen (r= 0.50, pFWE= 0.025), insula (r= 0.50, pFWE= 0.039), and the supramarginal gyrus (r= 0.51, pFWE= 0.039; Table 2). When examining the contrast of the matched palatable cue and taste > matched sub-palatable cue and taste, we observed significant decreases in BOLD response in regions associated with somatosensation and gustatory processing: postcentral gyrus (r= −0.70, pFWE= 0.001), insula (r= −0.50, pFWE= 0.02), and operculum (r= −0.51, pFWE= 0.001; Table 2). BMI and food reinforcement value were not significantly related to BOLD response in the above contrasts.

Table 2.

Significant BOLD response to Cues, Taste, and Prediction Error¹

Contrast	K2	Z-score	r3	pFWE2	X4	Y	Z
Response to Cue and Taste
Palatable milkshake cue > Sub-palatable milkshake cue
Putamen	53	3.4	0.50	0.025	−20	6	−8
Supramarginal gyrus	35	3.5	0.51	0.039	−52	−22	29
Thalamus	28	3.4	0.50	0.04	−18	−24	0
Middle frontal gyrus	8	3.3	0.48	0.042	−32	0	50
Insula	19	3.4	0.50	0.044	38	−12	−6
Superior parietal lobule	5	2.9	0.43	0.048	−3	−50	50
Postcentral gyrus	18	3.1	0.46	0.049	52	−18	28
Postcentral gyrus	6	3.3	0.49	0.048	42	−28	42
Palatable taste > Sub-palatable taste
Postcentral gyrus	823	−4.8	−0.70	0.001	63	−8	25
Lateral occipital	25	−3.2	−0.46	0.038	36	−58	42
Postcentral gyrus	748	−3.9	−0.58	0.01	−60	−7	16
Central operculum	35	−3.6	−0.52	0.01	38	−8	15
Insula	44	−3.5	−0.51	0.02	−35	−8	9
Lingual gyrus	11	−3.5	−0.51	0.006	16	−56	−2
Fusiform gyrus	15	−3.3	−0.48	0.006	25	−59	−9
Response to Prediction Error
Prediction error > Palatable taste
Central operculum	1680	5.0	0.73	0.002	−57	−7	14
Post central	1219	4.9	0.72	0.006	62	−8	18
Prediction error > Sub-palatable taste
Postcentral gyrus	198	−4.1	0.60	0.008	−58	−16	28
Superior parietal lobule	187	−3.5	0.51	0.006	−34	−46	46
Thalamus	58	−3.7	0.54	0.038	−18	−23	1.9
Insula	43	−3.4	0.50	0.01	39	−10	−5
Postcentral gyrus	22	−3.5	0.52	0.013	39	−28	39
Putamen	47	−3.8	0.55	0.02	−21	8	−7
Post Hoc Response to Prediction Error controlling for pleasantness5
Prediction error > Palatable taste controlling for sub-palatable pleasantness
Post central	461	4.4	0.65	-	60	−10	30
Central operculum	429	4	0.59	-	−56	−6	16
Prediction error > Sub-palatable taste controlling for palatable pleasantness
Insula	26	−3.5	−0.51	-	38	−10	−4
Prediction error > Sub-palatable taste controlling for sub-palatable pleasantness
Insula	6	−3.3	−0.49	-	44	−8	−4

¹Prediction error was defined as response to the sub-palatable taste that was preceded by the palatable milkshake cue (negative prediction error).

²P values and cluster size (k) were calculated with FSL Randomise threshold free cluster enhancement and cluster respectively.

³Effect size (r) was calculated $\frac{Zvalue}{\sqrt{n}}$

⁴Coordinates are in MNI space

⁵n=46, one person did not fill out the pleasantness survey. Due to the post hoc nature of the analysis there are no p values.

3.2. BOLD response to negative prediction error

Prediction error taste was analytically defined as response to the sub-palatable taste that was preceded by the palatable milkshake cue. The negative prediction error > matched sub-palatable cue and taste contrast was used to assess the effect of the cue in prediction error as it accounts for the overall effect of sub-palatable taste. This contrast elicited decreased BOLD response in the postcentral gyrus (r= −0.60, pFWE= 0.008), putamen (r = −0.55, pFWE= 0.02), insula (r = −0.50, pFWE= 0.01), and superior parietal lobule (r = −0.52, pFWE= 0.006; Figure 2A). To assess the effect of the violated taste expectation, we examined BOLD response to negative prediction error > matched palatable cue and taste. This contrast resulted in significant BOLD response in regions related to somatosensation postcentral gyrus (r= 0.72, pFWE= 0.006) and the central opercular cortex (r= 0.73, pFWE= 0.002; Figure 2B). Neither BMI nor food reinforcement value were significantly related to BOLD response to the above contrast.

Figure 2. BOLD response to prediction error.

A. Z statistical maps representing greater BOLD response in the insula (pFWE=0.01, MNI: 39, −10, −5) and putamen (pFWE=0.002, MNI: −21, 8, −7) (y=62, z=−8) to matched sub-palatable cue and taste compared to negative prediction error in 47 postpartum women.

B. Z statistical maps representing greater BOLD response in the postcentral gyrus (pFWE=0.006, MNI: 62, −8,18) and central operculum (pFWE=0.002, MNI: − 57, −7, 14) (y=62, z=−8) to negative prediction error compared to matched palatable cue and taste in 47 postpartum women.

3.3. Post hoc analysis of BOLD response to negative prediction error controlling for milkshake pleasantness

After separately controlling for the pleasantness of the sub-palatable milkshake, we found that the BOLD response associated with the prediction error > matched sub-palatable taste and cue contrast was attenuated (Table 2). After controlling for the pleasantness of the sub-palatable milkshake the BOLD response to the prediction error > matched palatable cue and taste was attenuated in the post central gyrus and the central operculum (Table 2). Controlling for the pleasantness of the palatable milkshake had no effect on the BOLD response in the prediction error > matched palatable taste and cue in the post central gyrus (k=1036; z-stat = 4.8; x= −56, y= −10, z= 24) nor in the central operculum (k= 789; z-stat= 4.7; x= 58, y= −6, z= 26).

4. DISCUSSION

Here we sought to evaluate characteristics of whole brain BOLD response to negative prediction error using primary reinforcers of palatable and sub-palatable tastes in women at 6 months postpartum. Specifically, our primary hypotheses were that BMI would be positively associated with negative prediction error BOLD response in the occipital cortex, and food reinforcement value would be negatively associated with negative prediction error BOLD response in the orbital frontal cortex and amygdala. While the current data did not support either of these hypotheses, the present study provides valuable information regarding taste processing of palatable and sub-palatable foods and the neural encoding of negative prediction error. Here, negative prediction error was defined as response to the sub-palatable taste that was preceded by the palatable milkshake cue. We contrasted the negative prediction error event (sub-palatable taste with the palatable cue) versus matched sub-palatable cue and taste. Using this contrast, we observed decreased BOLD response in the postcentral gyrus, putamen, insula, and superior parietal lobule. Whereas, when contrasting negative prediction error relative to a matched palatable taste and cue regions thought to encode gustatory and somatosensory processing (central opercular cortex and postcentral gyrus) increased BOLD response.

Generally, somatosensory processing encodes representation of taste [48,49]. More specifically for the present study, BOLD response in the somatosensory region has been shown to be sensitive to prediction errors [50,51], is affected by stimulus probability [52], and is associated with salience prediction errors [53]. In the present study we found the somatosensory cortex was sensitive to negative prediction error when compared with either a matched palatable taste and cue, or with a matched sub-palatable taste and cue. The seemingly similar results may be due to modeling the prediction error contrasts during the taste administration. However, between the two contrasts the valence of activity differed. When the prediction error trials were compared to the matched palatable taste and cue, BOLD activity in the postcentral gyrus increased. Conversely, when comparing the prediction error response and the matched sub-palatable taste and cue, BOLD response in the postcentral gyrus decreased. The sub-palatable taste compared to the palatable taste (as was tested in the prediction error compared to matched palatable taste and cue) may have been more salient due to its mild aversive taste and therefore increased response in the postcentral gyrus. This is supported in the post hoc analyses. When controlling for the pleasantness of the sub-palatable flavor, BOLD response was attenuated in the post central gyrus and the central operculum in the prediction error versus matched palatable taste and cue contrast. The decrease in response when controlling for the of the sub-palatable milkshake, but not when controlling for the pleasantness of the palatable milkshake, suggests the perceived pleasantness of the sub-palatable milkshake drives the BOLD response in the post central gyrus and the central operculum. Furthermore, when the taste was held constant (as in the prediction error compared to the matched sub-palatable cue and taste) the palatable cue may have lessened the aversive qualities of the sub-palatable taste compared to when the participants knew a sub-palatable taste was imminent. When controlling for the pleasantness of the sub-palatable milkshake this effect was attenuated. Further highlighting the importance of the sub-palatable taste even when paired with a palatable cue. A pleasant cue, such as food advertisements or logos, may have the potential to influence intake of food items that are seemingly less desirable.

Further, we observed a robust effect of negative prediction error in the contrast between prediction error relative to the matched sub-palatable cue and taste in areas including the putamen, insula, and superior parietal lobule. This contrast is of particular interest as the taste that was analyzed here was the same and thus controlled for; the only difference being the preceding cue. This points to a large, differential impact of expectation on the subsequent taste. The insula has previously been shown to be important for ‘bottom-up’ processing (from sensory processes up to cognitive processes) of reward cues and events [54], and for stimuli salience [55]. The decrease in insula BOLD activity when comparing prediction error to the matched sub-palatable cue and taste may reflect reduced ‘bottom-up’ processing and a decrease in the stimuli salience. This aligns with the above hypothesis that stimuli salience is decreased after a palatable cue in the somatosensory cortex. Therefore, interference of ‘top-down’ processes (from cognitive processes to sensory processes) such as prediction and expectation after a palatable cue may be altering the salience of the sub-palatable taste. Unlike the insula and the somatosensory cortex, the putamen is known to deactivate with lack of reward [56]. Similar decreased response in the putamen has been shown when a cue predicting a juice taste was followed by no juice compared to a lack of cue followed by no juice [57]. This is a similar contrast to the one used presently, with similar results in the putamen. Furthermore, decreased BOLD response in the putamen to a negative prediction error has been associated with the dopamine 2 receptor (D2R) Taq1A polymorphism [58]. Persons with the A2/A2 variant of the Taq1A allele have lower D2R concentrations in striatal regions, and this polymorphism has been related to BMI [25] and weight gain [59]. Although no relationship between BMI and BOLD response was observed presently, the interaction of the Taq1A allele with the putamen suggests the importance of dopamine on prediction error response. Specifically, dopaminergic brain regions receive cortical inputs from the somatosensory cortex [60] and dopamine is thought to be dependent for reward based decision making in the sensory cortex [61] and in the insula [62].

Possibly, for postpartum women, repeated exposure to visual cues predicting a palatable stimulus may blunt the effect of an aversive stimulus. In the context of an obesegenic environment: this may be related to a predisposition to like foods associated with rewarding cues, such as food advertisements. Previous research has shown exposure to food cue was associated with a more pleasant perception of a food taste [63]. This is in line with a body of research which has found both children and adults consume more food following food advertisements [64–67], although this literature is mixed in regards to young/healthy weight women [68]. While further study is needed to elucidate the relationship between food cues and palatable and sub-palatable tastes in the general population, the present study highlights the importance of food related cues on neural taste perception in a specific population of postpartum women.

We observed increased BOLD response to an anticipatory palatable cue relative to a sub-palatable cue in regions thought to encode motivated behavior (putamen), gustatory and somatosensory processing (insula and postcentral gyrus), and the integration of sensory signals (thalamus). This response pattern dovetails with previous reports that use a contrast of an anticipatory palatable milkshake cue relative to a cue for a tasteless solution [49,69], and is also consistent with data from studies that examine pictures of appetizing foods, relative to lower level contrasts such as pictures of glasses of water or blurred images [70–72]. Conversely, when examining the contrast of matched palatable cue and taste relative to the matched sub-palatable cue and taste, we observed robust decreases in response in gustatory and somatosensory regions. Intuitively one would anticipate that this contrast would result in increased BOLD response in these regions as well as regions more directly related to dopaminergic functioning, i.e., the striatum. A possible explanation of this unanticipated finding is that the sub-palatable taste was a more novel taste. The sub-palatable milkshake is not something that is regularly available in the food environment, whereas the palatable milkshake was designed to mimic a standard milkshake. Moreover, if the sub-palatable milkshake was available, it is unlikely that it would be selected for consumption. Further, the milkshakes were designed to be similar in chocolate flavor and viscosity, yet there may be a detectable (unmeasured) difference in ‘natural’ fat in the milkshake relative to the emulsifier used.

The present study has a number of limitations to address. First and foremost, this study only included six-months postpartum women, limiting its generalizability to other populations. It should also be noted that the present sample is predominately white and the average household income was above the national average in 2017 [47], this may further limit the generalizability of the findings to other racial or socio-economic groups. Additionally, the weight gained in pregnancy may operate differently compared to weight gain due to overeating. In pregnancy, maternal weight gain is physiologically necessary for fetal development and subsequent breastfeeding [73]. Further, other factors may play a role in weight gain and eating behavior in the postpartum period, such as lack of sleep, increased stress. The present study utilized training runs for participants to learn the association between the symbols and tastes. However, the prediction error paradigm did not utilize an active paradigm where the stimuli have a dynamic ratio of paired and mismatched cues that is based on participant response. In order for ‘true’ prediction error, the paradigm would need to asses a behavioral response of expectation in real time. One would expect this methodical approach would elicit similar, but more robust results and likely more engagement of regions associated with motivated behavior as well as motor function. The absence of a dynamic paradigm also precludes the analysis of how well the women “learned” to associate the cue with the subsequent taste. Therefore, the lack of relationship between BMI and the task presented could be due to a lack association of the cue to the taste. Although the women were instructed to pay attention to the cues and the tastes and a similar paradigm has been used previously [57]. Moreover, the energy content of the two stimuli may have impacted the findings such that variability in energy density impacted BOLD response that was attributed to palatability. Though, in real world settings there is a strong relation between energy density and palatability meaning the stimuli varies as they would in the ‘real world’ thus increasing the generalizability of the results. The lack of relationship between BOLD response with BMI and food reinforcement value observed in this study could be a function of a number of participant characteristics and/or unmeasured confounds. A growing body of literature indicates many of the relationships between brain response to food stimuli typically seen when comparing individuals with versus without obesity may be a function of habitual eating behavior [69,74–76]. That is, a higher weight due to pregnancy might not reflect habitual patterns of eating well beyond energy needs, and thus the potential association of BMI with brain response to food may be decreased, and in this case not evident after correction for multiple comparisons. Further research examining the association of prediction error with postpartum weight change will be informative.

In sum, we did not find support for our primary hypotheses of a relationship between BOLD response to negative prediction error with maternal BMI at 6-months postpartum and food reinforcement value. Consistent with previous literature, the insula and putamen are responsive to food stimuli, bolstering the theory that these regions are associated with the underpinnings of reward-based eating. Expanding this notion, we suggest that during the evaluation of food cues a combination of reward and gustatory regions is responsive indicating initial engagement of decision-making processes that drive (food-related) hedonically motivated behavior. This approach provides novel information about reinforcement learning, in which gustatory and somatosensory cortices appear to encode negative prediction errors. This indicates that an unexpected deviation in the palatability of a cue taste pair is processed in both gustatory and salience regions of the brain, highlighting the importance of expectations in cue-taste pairing in driving eating behavior.