Connectome-wide brain signature during fast-food advertisement exposure predicts BMI at 2 years

Abstract

Food advertisements target adolescents, contributing to weight gain and obesity. However, whether brain connectivity during those food advertisements can predict weight gain is unknown. Here, 121 adolescents [14.1 ± 1.0 years; 50.4% female; body mass index (BMI): 23.4 ± 4.8; 71.9% White] completed both a baseline fMRI paradigm viewing advertisements (unhealthy fast food, healthier fast food, and nonfood) and an anthropometric assessment 2 years later. We used connectome-based predictive modeling to derive brain networks that were associated with BMI both at baseline and the 2-year follow-up. During exposure to unhealthy fast-food commercials, we identified a brain network comprising high-degree nodes in the hippocampus, parahippocampal gyrus, and fusiform gyrus rich with connections to prefrontal and occipital nodes that predicted lower BMI at the 2-year follow-up (r = 0.17; P = .031). A similar network was derived from baseline BMI (n = 168; r = 0.34; P < .001). Functional connectivity networks during exposure to the healthier fast food (P = .152) and nonfood commercials (P = .117) were not significant predictors of 2-year BMI. Key brain regions in our derived networks have been previously shown to encode aspects of memory formation, visual processing, and self-control. As such, the integration of these regions may reflect a mechanism of adolescents’ ability to exert self-control toward obesogenic food stimuli.

Introduction

Adolescents are one of the primary targets of the food marketing industry (Truman and Elliott 2019). The advertisement of unhealthy foods to adolescents has been associated with higher liking of the advertised products, increased fast-food intake, poor diet quality, and greater rates of obesity (Folkvord et al. 2016, Harris et al. 2022, Bagnato et al. 2023). Fast-food companies are among the top advertisers to adolescents with the majority of advertisements in this age group promoting unhealthy fast-food products (Kelly et al. 2019, Pauzé and Potvin Kent 2021). US adolescents are exposed to an average of 9.4 food-related TV ads per day, summing to over 3500 food-related ad views per year (Frazier and Harris 2018). Moreover, adolescents are constantly exposed to food marketing campaigns through social media with 6.2 million adolescents following 27 of the most highly advertised fast-food, snack, and drink brands in 2019 on various social media platforms (Rummo et al. 2020).

Food commercials have the power to influence adolescents’ eating behavior via a series of psychological and physiological responses that render food advertisements highly reinforcing, leading to increased food intake (Folkvord et al. 2016), potentially by influencing reward-related networks in the developing brain (Casey and Jones 2010). Two meta-analyses have concluded that acute exposure to unhealthy food advertising significantly increases food intake in children and adolescents (Boyland et al. 2016, Russell et al. 2019). In addition to behavioral observational studies, functional magnetic resonance imaging (fMRI) studies have examined how brain responses can predict behaviors, offering insights into the neural basis of eating behaviors (Giuliani et al. 2018). Exposure to food vs. nonfood commercials is associated with greater brain responses in regions including the cerebellar culmen, middle occipital gyrus, and superior parietal lobule, which are commonly activated in response to both natural (food and sex) and drug cues highlighting their role in reward processing, emotional responses, and habit formation (Noori et al. 2016, Yeung 2021). Additionally, exposure to dynamic food commercials, compared to static food ads, elicits greater blood-oxygen-level-dependent (BOLD) response in regions that encode aspects of reward and gustatory processes (ventral tegmental area, substantia nigra, amygdala, and insula), suggesting that a more naturalistic presentation of food cues, as the ones seen in TV ads, might have a greater impact in engaging dopaminergic reward pathways (Yeum et al. 2023). These brain regions are thought to underlie hedonically motivated food behaviors such as food cravings, food-approach behaviors, and appetitive motivation in both animal and human research (Nummenmaa et al. 2012, Siep et al. 2012, Dietrich et al. 2016, Douglass et al. 2017). In support, a cross-sectional analysis of the data in the current sample demonstrated that greater BOLD activation in reward-related regions (nucleus accumbens and caudate) in response to both unhealthy and healthier fast-food vs. nonfood commercials is associated with greater food intake (Gearhardt et al. 2020). Furthermore, less activation in the bilateral fusiform gyrus to both food and nonfood brands (vs. control images) is associated with greater energy intake from branded meals (Masterson et al. 2019). These results suggest that brain responses to branding in general, and not just food brands, could be associated with the energy intake-promoting effects of marketing.

Despite these advances in knowledge, prospective investigations of the impact of food and nonfood commercials on future weight change among adolescents are lacking. Adolescence is a critical developmental period marked by extensive changes in the structure and function of the adolescent brain that refine higher-order cognitive functions including decision-making (Blakemore and Mills 2014, Larsen and Luna 2018). Thus, understanding how exposure to environmental food cues interacts with functional brain connectivity during this sensitive period can provide insight into the brain pathways through which food ad exposure leads to increased food intake and weight gain in adolescents. Advances in fMRI data-driven machine learning-based connectivity analyses have yielded tools, e.g., connectome-based predictive modeling (CPM), that allow us to identify brain circuits predictive of certain behaviors and traits across individuals to develop models of brain–behavior relationships “i.e., neural signatures” (Finn et al. 2015, Rosenberg et al. 2016, Shen et al. 2017). The ability to identify brain functional connectivity patterns that make some adolescents more susceptible to food advertising content than others could ultimately aid in the development of policies that tightly regulate advertisment exposure among the younger population.

Here, we utilized CPM to examine whether functional connectivity networks during passive viewing of food and nonfood advertisements could be derived from baseline body mass index (BMI) and predict future BMI in adolescents. Based on previous work studying BOLD responses to food commercials (Gearhardt et al. 2014, 2020, Masterson et al. 2019, Arrona-Cardoza et al. 2022, Yeum et al. 2023), we hypothesized that increased functional connectivity between regions in dopaminergic networks (dorsal striatum, ventral striatum, and thalamus) and salience networks (insula and amygdala) in response to fast-food commercials compared to nonfood commercials would predict greater BMI in adolescents at a 2-year follow-up.

Materials and methods

Sample

The current sample was drawn from a larger study detailed here (Gearhardt et al. 2020, Hoover 2022). In brief, participants (n = 193) were recruited from southeast Michigan between 2015 and 2017 and asked to complete an fMRI task measuring BOLD responses to fast-food (unhealthy and healthier) and nonfood commercials at baseline, and complete height and weight measurements at baseline and the 2-year follow-up. Participants were English-speaking adolescents between 13 and 16 years of age. Adolescents were excluded if they were using psychotropic medications or illicit drugs, had a lifetime psychiatric disorder, a BMI percentile of <5%, or fMRI contraindicators (e.g., presence of metal implants). Of the 193 adolescents in the full sample at baseline, 186 completed the fMRI scan, of which 9 showed excessive movement (i.e., head motion of >3 mm or degrees in any direction within two or more runs), 2 had scan data with acquisition errors, and 2 did not have the correct onset timing files for the fMRI task, resulting in a baseline sample of 173 (n = 124 with the 2-year follow-up). Anthropometric measurements for five participants from baseline and three from the 2-year follow-up were flagged as statistical outliers and removed, resulting in a final analytic sample of 168 adolescents at baseline and 121 at the 2-year follow-up.

Overview of study procedures

During the baseline assessment, parents provided written informed consent. Adolescents had their height and weight measured and completed the fMRI commercials neuroimaging paradigm. Adolescents completed a 2 -year follow-up visit, where they repeated the height and weight measurements. Study procedures were approved by the University of Michigan Institutional Review Board.

Measures

Anthropometrics

Participants’ height and weight were collected in light clothing without jackets, socks, or shoes. Height was measured using an O’Leary Acrylic Stadiometer (in centimeters, to the nearest tenth); weight was measured using a Detecto Portable Scale (in kilograms, to the nearest tenth). Height and weight measurements were repeated twice and the averages were used for subsequent analyses. BMI values (kg/m²) were calculated at baseline and the 2-year follow-up. Participants reported their pubertal development at baseline and follow-up using a series of sex-specific line drawings at various stages of pubertal development (Bonat et al. 2002).

Scan procedures

For the fMRI scan, participants were asked to refrain from eating or drinking (except water) following their last meal before the scan (last meal consumed 1–7 h prior to scan). Participants rated their hunger on a scale from 0 (not hungry at all) to 100 (extremely hungry) before the scan. As described in Gearhardt et al. (2020), if a rating of ≥70 was indicated, participants were offered a small snack (e.g., crackers, fruits) to normalize their hunger to a neutral state. From the individuals included in the baseline analysis (n = 168), a total of 7.7% received a snack (n = 13). From the individuals included in the 2-year follow-up analysis (n = 121), a total of 5.8% received a snack (n = 7). Hunger ratings were repeated after the snack and hunger was decreased by 53.6% in those who consumed the snack.

fMRI data acquisition

Data were acquired with a GE Discovery MR750 3T scanner. An 8-channel head coil was used to acquire data from the entire brain. Functional data were acquired using a spiral sequence with the following parameters: repetition time (TR) = 2000 ms, echo time (TE) = 30 ms, flip angle = 90°, field of view (FOV) = 22 × 22 cm², acquisition matrix 64 × 64, 3-mm slice thickness with no gap, 43 axial slices, and voxel size = 3.44 × 3.44 × 3.0 mm³. In total, 156 volumes were collected during each of four functional runs. Anatomical scans were acquired using a high-resolution T1-weighted spoiled-gradient-recalled acquisition (TR = 12.3 ms; TE = 5.2 ms, inversion time (TI) = 500 ms, flip angle = 15°, FOV = 22 × 22 cm², slice thickness = 1.0 mm, voxel size = 1 × 1 × 1 mm³).

fMRI commercial paradigm

The fMRI commercial paradigm was developed to include commercials commonly viewed by young adolescents based on Nielsen Gross Ratings Point national data (Harris 2013). Details on the commercial selection as well as a full list of the commercials used can be found in Gearhardt et al. (2020). Briefly, the stimuli in the fMRI task included 20 fast-food commercials promoting unhealthy food from Wendy’s and McDonald’s (e.g., Quarter Pounder with Cheese), 20 fast-food commercials for healthier food options from McDonald’s and Wendy’s (e.g., Southwest Salad), and 20 nonfood phone (control) commercials from AT&T and Verizon (e.g., iPhone). Foods depicted in the commercials were evaluated as unhealthy or healthier using the Nutrient Profile Index (NPI), a measure of overall nutrition quality used in the UK (Ofcom 2006, Rayner et al. 2009). NPI scores range from 1 to 100 (higher score indicates healthier items), with foods in the unhealthy fast-food commercials averaging 44.05 ± 4.21 and foods in the healthier fast-food commercials averaging 70.20 ± 3.86. Each commercial lasted ∼15 s and was shown only once. Between each commercial was a jittered fixation cross (4–8 s) (Fig. 1). The paradigm consisted of four 7-min runs with 15 commercials per run. The order of commercials was randomized in each of the runs and the order of the four runs was randomized over the participants. Participants were instructed to watch the commercials. After the scan, self-reported liking ratings of the commercials in the paradigm were collected. Participants were asked “How much do you LIKE the following products?” for each of the advertised items (i.e., unhealthy fast-food, healthier fast-food, and phones). Responses were provided on a 5-point scale ranging from Dislike Extremely (1 point) to Like Extremely (5 points). In addition, both at baseline and the 2-year follow-up, participants answered questions about the frequency of visiting Wendy’s and McDonald’s in a typical week to assess their overall fast-food exposure, with 63.7% and 78.5% of adolescents reporting visiting those restaurants at least once per week at baseline and the 2 year follow-up, respectively.

fMRI preprocessing

Following previous work (Burger and Stice 2013), all images were manually realigned to the anterior commissure - posterior commissure line and skull stripped using the Brain Extraction Tool in the FMRIB Software Library FSL; v5.0.9; FMRIB Analysis Group, Oxford, UK). Neuroimaging data were preprocessed and analyzed primarily using Statistical Parametric Mapping (SPM12, Functional Imaging Laboratory, University College of London) in MATLAB (R2015a; MathWorks, Inc., Natick, MA, USA). Anatomic images were segmented and normalized to Montreal Neurological Institute (MNI) space with the use of the DARTEL toolbox. Anatomical data were coregistered to the mean functional image and segmented into six tissue types using unified segmentation approach (Ashburner and Friston 2005). Functional images were realigned to the mean functional image, coregistered with the anatomic images, normalized to MNI space with the use of DARTEL, and smoothed with an 8-mm full-width at half-maximum isotropic Gaussian kernel. DARTEL was used to create a group anatomical template, transformations from which were applied to warp functional data to the ICBM-152 template supplied with SPM12 (Ashburner 2007). Artifact Detection Tools (http://www.nitrc.org/projects/artifact_detect/) software package was used for automatic detection of spikes and motion in the functional data. Motion parameters (three translations and three rotations) were included as regressors in the design matrix at individual-level analysis. Additionally, functional image volumes where the z-normalized global brain activation exceeded 3 SDs from the mean or showed >1.5 mm of composite movement were flagged as outliers and deweighted during individual-level model estimation.

Statistical analyses

Behavioral analysis

Raw BMI scores were used to assess BMI changes, as they are considered better than BMI-for-age percentiles or BMI z-scores for modeling change over time (Adise et al. 2024). Paired sample t-tests were used to compare BMI across the two timepoints, and repeated measures ANOVA was used to explore biological sex by timepoint differences and differences in the liking ratings by commercial type.

Functional connectivity

Following data preprocessing, we used a whole-brain functional atlas defined on a separate group of healthy adult subjects (Shen et al. 2013) to parcellate the brains of each subject into 268 regions of interest. We chose the Shen 268-node parcellation because it covers the whole brain, including subcortical areas and cerebellum, and has been used in previous CPM studies (Rosenberg et al. 2016, Yip et al. 2019, Farruggia et al. 2020, Finn and Bandettini 2021). Here, the mean timecourses per run per participant (i.e., average BOLD signal of all voxels within the node during each task run) for each of the 268 nodes were extracted using Nipype 1.8.6 implemented in Python 3.11.2. To minimize the impact of task-induced co-activation in the functional connectivity analyses (Cole et al. 2019) and capture mostly “background” fluctuations in brain activity induced by the different commercial types, after modeling the task effects within a general linear model design, we extracted the timecourses on the residuals of this regression (Al-Aidroos et al. 2012). Timecourses were concatenated across all available runs per participant, standardized within subject, and detrended to remove linear trends from the signal. Functional connectivity matrices for each subject were created by correlating each node’s timecourse with every other node’s timecourse to construct 238 × 238 matrices, one per subject (all 30 nodes that were part of the cerebellar canonical network, as defined by Noble et al. (2017) and Greene et al. (2018), were dropped to focus the analyses on networks more commonly associated with BMI; note that 28 nodes in the cerebellum that were part of the other canonical networks were kept in). Correlation coefficients in the 238 × 238 connectivity matrices were transformed to z-scores using Fisher’s transformation.

Connectome-based predictive modeling

CPM is a data-driven approach that uses whole-brain functional connectivity to develop models of brain–behavior relationships using machine learning (Finn et al. 2015, Shen et al. 2017). CPM analysis was conducted using previously validated custom scripts in MATLAB (Shen et al. 2017, Boyle and Weng 2023). A detailed description of the CPM protocol can be found in Shen et al. (2017). An overview of the method is as follows:

• Without any prior scaling or standardization, data, comprising each subject’s functional connectivity matrix and the target behavioral variable, were divided into training and test sets for 10-fold cross-validation, where the total number of subjects was binned into 10 equal size groups, and for each iteration subjects from nine bins (90% of the participants) were used for training and subjects from the one left-out independent bin (10% of participants) were used for testing. We ran two separate models each with a different target variable: one with baseline BMI and the other with 2-year BMI (controlling for baseline BMI). For each model, we then repeated the 10-fold cross-validation process 100 times to create a distribution of the test statistic and ensure its stability across different train-test folds.

• For each model, in the training set, each edge in the connectivity matrix was correlated with the target variable. Target variables (baseline BMI, 2-year BMI) were normally distributed, and thus we used partial Pearson’s correlation to control for age, biological sex, baseline pubertal stage for both models, and baseline BMI for the 2-year BMI model only. In post hoc analyses, additional covariates (pre-scan snack intake, pre-scan hunger ratings, self-reported liking ratings for the commercials, change in pubertal stage from baseline to the 2-year follow-up, and fast-food exposure) were included in the models but results remained similar in statistical significance across all models and types of commercials, so they were omitted from the results presented here.

• A feature selection threshold was applied to select the most relevant edges for use in the predictive model based on the correlation coefficients calculated in (ii). Here, we retained edges with |r| > 0.25 (corresponding to a two-tailed P-value of approximately .006) for the correlation between connectivity and the target variable. Edges were separated into a positive network (edges where connectivity is positively correlated with the target variable) and a negative network (edges where connectivity is negatively correlated with the target variable).

• For each subject in the training set, we calculated the summed connectivity strength across all retained edges in the positive and negative networks separately.

• Still using the training set only, we fitted two linear models: one using the positive network and the other using the negative network, by regressing network strength on the target variable. The model parameters were extracted and saved.

• For each subject in the test set, we calculated positive and negative network strength using the same method described in (iii). These values were used as input to each of the linear models estimated in (v) to create the predicted scores for the target variable for each subject.

Evaluating model performance and statistical significance

To assess prediction accuracy of each linear model (positive and negative), we correlated the predicted (model generated) and observed target variable scores across subjects using Pearson’s correlation. The coefficient of determination (r²) was also calculated using the sum-of-squares approach to assess the proportion of variance in the observed values explained by the predicted values. Given that correlation is a relative measure of accuracy rather than an absolute one, the mean absolute error (MAE) between the predicted and observed target variable scores was also calculated and reported below.

In cross-validation, regression folds are not independent of one another, and as such, we used nonparametric permutation testing to assess the statistical significance of prediction accuracy. We generated a null distribution for the correlations between the predicted and observed target variable scores by randomly shuffling the target variable with respect to connectivity matrices and repeating the entire CPM pipeline 10 000 times. The nonparametric P-value for each network strength was calculated as the proportion of permuted correlation coefficients (i.e., from the null distribution) that are greater than or equal to the true prediction correlation coefficient (i.e., mean across the 100 true models) for that network strength model.

Data visualization

Only edges selected in at least 90% of all folds across all 100 iterations were extracted and graphed for the significant networks. Figures were constructed using BioImage Suite (Joshi et al. 2011). The 238 nodes included in the analyses were assigned to nine canonical networks as defined in Greene et al. (2018) emphasizing different properties of the human brain and consist of medial-frontal (MF), fronto-parietal (FP), default mode network (DMN), motor/somatosensory (Mot), primary visual (VI), secondary visual (VII), visual association (VAs), salience (SAL), and subcortical (SC) networks.

Results

Participant characteristics at baseline (n = 168) and the 2-year follow-up (n = 121) are summarized in Table 1. The subsample that did not complete the 2-year follow-up (n = 47) was significantly older at baseline compared to the participants that completed the 2-year follow-up (t = 2.70, P = .008), but showed no other differences.

BMI was significantly higher at the 2-year follow-up compared to baseline [t(121) = 6.53, P < .001]. Differences among self-reported liking of the three types of advertised items were observed [F(2, 334) = 33.74, P < .001]; nonfood (phone) items were rated higher compared to both unhealthy and healthier fast-food items (P < .001 for both). Unhealthy fast-food items had higher liking ratings than healthier food items (P = .001; Table 1).

Pubertal stage at baseline was significantly higher in female vs. male participants at baseline (χ² = 12.37, P = .006), but not at the 2-year follow-up (χ² = 0.32, P = .853). Pubertal stage was also significantly different by race at baseline (Black adolescents reported higher pubertal stage; χ² = 56.56, P < .001), but not at the 2-year follow-up (χ² = 13.18, P = .356).

Brain networks in response to unhealthy fast-food commercials

Baseline

We first examined whether functional connectivity in response to each type of commercial is significantly associated with BMI at baseline, controlling for age, sex, and baseline pubertal stage. We identified a functional connectivity network in response to passive viewing of “unhealthy fast-food” commercials where stronger connectivity among the nodes of the network was associated with lower BMI at baseline (r = 0.341, nonparametric permutation P < .001, r² = 0.116, MAE = 3.722; Fig. 2). This “negative” network contained 450 edges (1.6% of all possible connections among the nodes in the atlas used). Highest-degree nodes (i.e. nodes with the greatest number of nodal connections contributing to the network) were found in the fusiform gyrus, cerebellum, occipital cortex, hippocampus, and anterior prefrontal cortex (Table 2). These nodes are summarized based on their overlap with the nine canonical networks and can be seen in Fig. 2b and c. We observed connections within the SC canonical network, as well as connections between the SC and VAs, SAL, and DMN canonical networks.

Two-year follow-up

We identified a functional connectivity network in response to passive viewing of “unhealthy fast-food” commercials where stronger connectivity among the nodes predicted lower BMI at 2 years (r = 0.174, nonparametric permutation P = .031, r² = 0.031, MAE = 4.164; Fig. 3), controlling for age, sex, baseline pubertal stage, and baseline BMI. This network contained 132 edges (0.5% of all possible connections). Highest-degree nodes were found in the parahippocampal gyrus, hippocampus, posterior cingulate cortex, superior temporal gyrus, and insula (Table 3). These nodes are summarized based on their overlap with the nine canonical networks and can be seen in Fig. 3b and c. Here, nodes within the SC canonical network were interconnected with nodes within the MF, FP, DMN, and Mot networks.

Brain networks in response to healthier fast-food commercials

Baseline

A functional connectivity network in response to passive viewing of “healthier fast-food” commercials was identified, where stronger connectivity among the nodes of the network was associated with lower BMI at baseline (r = 0.366, nonparametric permutation P < .001, r² = 0.134, MAE = 3.672; Fig. 4), controlling for age, sex, and baseline pubertal stage. This negative network contained 566 edges (2.1% of all possible connections). Highest-degree nodes for the network were found in the fusiform gyrus, cerebellum, occipital cortex, and hippocampus (Table 4). These nodes are summarized based on their overlap with the nine canonical networks and can be seen in Fig. 4b and c. Here, we observed connections within the FP and within the SC canonical networks. There were also connections between the FP, VAs, SAL, and SC networks.

Two-year follow-up

Functional connectivity networks in response to passive viewing of the “healthier fast-food” commercials did not significantly predict BMI at the 2-year follow-up (r = 0.103, nonparametric permutation P = .152, r² = 0.011, MAE = 4.194; Fig. 3e), controlling for age, sex, baseline pubertal stage, and baseline BMI.

Brain networks in response to nonfood commercials

Baseline

There was also a significant functional connectivity network in response to passive viewing of “nonfood” commercials where stronger connectivity among the nodes of the network was associated with lower BMI at baseline (r = 0.331, nonparametric permutation P < .001, r² = 0.110, MAE = 3.732; Fig. 5), controlling for age, sex, and baseline pubertal stage. This negative network contained 666 edges (2.4% of all possible connections). Highest-degree nodes were found in the fusiform gyrus, cerebellum, and occipital cortex (Table 5). These nodes are summarized based on their overlap with the nine canonical networks and can be seen in Fig. 5b and c. We found connections within the SC and within the Mot networks, as well as connections between the SC, VAs, and FP canonical networks.

Two-year follow-up

Functional connectivity networks in response to passive viewing of the “nonfood” commercials did not significantly predict BMI at the 2-year follow-up (r = 0.115, nonparametric permutation P = .117, r² = 0.015, MAE = 4.215; Fig. 3f), controlling for age, sex, baseline pubertal stage, and baseline BMI.

Discussion

During adolescence, individuals develop food choice autonomy, forming eating habits and behaviors that continue into adulthood (Videon and Manning 2003, Bassett et al. 2008). Adolescents are exposed to thousands of food commercials increasing their risk of developing unhealthy eating habits that lead to overeating and weight gain (Scully et al. 2012, Zimmerman and Shimoga 2014). In recent meta-analyses, food (vs. nonfood) commercials elicit increased brain activation in regions shown to encode self-control behaviors, modulation of food cravings, reward processing, and visual processing (Yeung 2021, Arrona-Cardoza et al. 2022). However, there is little evidence to understand whether brain network activation patterns to food commercials are associated with future weight change. Here, we observed that a brain network rich in connections in regions previously implicated in stimuli processing and self-control predicted lower current and future BMI. Interestingly, these networks were specific to when the participants were viewing the unhealthy fast-food commercials.

In this investigation, we employed CPM that uses cross-validation to protect against overfitting and ensure that models generalize to novel individuals in the same dataset, increasing the likelihood of replicating the results in novel datasets (Shen et al. 2017). We observed that stronger baseline functional connectivity within subcortical network nodes as well as between subcortical, salience, visual association, and frontoparietal canonical networks was associated with lower BMI at baseline. The most highly interconnected nodes associated with current BMI across all three types of commercials included regions in the fusiform gyrus, hippocampus, and cerebellum (regions assigned in the visual association, subcortical, and frontoparietal canonical networks, respectively), with connections to other visual association, prefrontal, temporal, and subcortical brain regions. We also observed that stronger functional connectivity among regions that encode aspects of memory formation, visual processing, and self-control predicted lower BMI at the 2-year follow-up. The most highly interconnected nodes when predicting lower future BMI included regions in the parahippocampal gyrus and hippocampus, with connections to prefrontal and temporoparietal brain regions. This may suggest that increased interplay among these regions may have a protective effect against the detrimental influences of fast-food marketing on eating behaviors.

The fusiform gyrus was involved in the baseline CPM networks with more connections from this region to inferior frontal gyrus, prefrontal cortex, and caudate being associated with lower baseline BMI. The fusiform gyrus is involved in semantic memory (Mion et al. 2010) but also appears to play a role in the visual processing of food cues in both adults and adolescents (Huerta et al. 2014, van Meer et al. 2015). Children have shown increased bilateral fusiform gyri activation in response to both food and nonfood brand images, while less activation in this area to both food and nonfood brands was associated with greater energy intake of the branded foods (Masterson et al. 2019). The fusiform area, part of the ventral visual pathway, contains subregions with a neural response that is highly selective to images of food (Khosla et al. 2022, Jain et al. 2023). Earlier results (Bruce et al. 2013, Maynard et al. 2017) suggest that the fusiform gyrus might play a more general role in the visual processing of all salient stimuli, both food and nonfood. In agreement with the literature, we corroborate that this is a region implicated in the processing of multiple types of visual cues and highlight that increased functional connectivity of the fusiform gyrus to the frontoparietal and subcortical networks is associated with lower current BMI, a relationship possibly mediated by suppressing the salience of rewarding stimuli commonly shown in fast-food commercials.

The parahippocampal gyrus was a key node in the significant 2-year negative CPM networks with more connections from this region to the prefrontal cortex, orbitofrontal cortex, inferior frontal gyrus, and middle temporal gyrus being linked to lower future BMI. This region plays an important role in memory and learning as well as in the hedonic processing of feeding and encoding of salient stimuli, allowing for food-related recollection (Berthoud 2002, Bragulat et al. 2010). Greater activation in the parahippocampal gyrus correlates with greater food cravings in young women (Chen et al. 2017). Individuals with obesity show increased parahippocampal activation when exposed to naturalistic food odors (Bragulat et al. 2010) and differences in parahippocampal resting-state functional connectivity with increased connections to reward processing regions (Zhang et al. 2020). Furthermore, individuals with obesity show increased activation in the parahippocampal gyrus and decreased activation in the dorsolateral prefrontal cortex during exposure to food images compared to healthy controls (Brooks et al. 2013). In contrast, our results show that stronger connectivity between the parahippocampal gyrus and the prefrontal cortex during viewing of unhealthy fast-food commercials is associated with lower future BMI, suggesting that it is not only the activation within each region that plays a role in the neural processing of food cues but also the connections between different parts of the brain that influence food cue encoding. In this case, strong connections between the parahippocampal gyrus, implicated in cognitive evaluation of salient stimuli, and the prefrontal cortex, implicated in cognitive control, might make individuals to assign a lower motivational value to unhealthy foods shown in commercials and potentially lead to inhibiting consumption of those foods.

A greater number of connections from the hippocampus to prefrontal cortex and inferior frontal gyrus were also linked to both lower current and lower future BMI. The hippocampus, through its involvement in the formation and retrieval of declarative memories (Bird and Burgess 2008), plays a critical role in food intake regulation (Stevenson and Francis 2017) by allowing us to consciously evaluate when, what, and how much to eat. Research suggests that recollection of past eating events acts as a form of self-control to inhibit further food intake (Higgs et al. 2008, Collins and Stafford 2015). Lesions in this area lead to impaired body weight regulation in animal models, suggesting that the hippocampus mediates inhibition of food intake (Davidson et al. 2009, Henderson et al. 2013). During satiation, the hippocampus appears to encode an inhibitory form of learning that allows us to suppress consummatory behaviors despite the presence of pleasant food cues in the environment (Davidson et al. 2009, 2005, 2014). Similarly, human neuroimaging studies have shown that in response to the consumption of a satiating liquid meal, individuals with current and past obesity have decreased regional cerebral blood flow to the hippocampus (DelParigi et al. 2004). Our results are in line with these previous findings and suggest that increased involvement of the hippocampus in frontoparietal, subcortical, motor, and default mode networks is associated with better future weight outcomes. People with obesity have shown reduced BOLD signal intensity in the hippocampus and decreased functional connectivity in the frontal gyrus of the default mode network compared to a control group (Chao et al. 2018), both findings that align with the results presented here. In contrast, BOLD signal intensity in a sample of young adults with obesity was higher in the hippocampus and lower in the prefrontal cortex compared to adults with normal weight, with resting-state activity in the hippocampus mediating task-induced hippocampal activation in response to high-calorie food cues in participants with obesity (Li et al. 2021). Although these results might appear contrasting, they could also suggest the heterogeneity of hippocampal function in the regulation of food intake and the importance of studying whole-brain functional connectivity in parsing out the complex relationships between all the brain networks involved in eating regulation.

It is important to consider the limitations of this study. First, we lacked power to test possible racial/ethnic differences. There is evidence to suggest that Black adolescents are exposed to a higher number of TV ads for snacks and sugary drinks compared to White youth (Miller et al. 2021), placing this group at a greater risk of consuming unhealthier foods. Imaging data for this study were collected in an MRI scanner using an 8-channel head coil, which, compared to the more widely used 32-channel head coil, has shown increased noise amplification and reduced signal-to-noise ratio during data acquisition (Panman et al. 2019). The machine learning approach implemented here provides a more conservative estimate of the strength of the brain–behavior relationship compared to traditional correlation approaches. However, the ultimate test of the generalizability of our results would be whether the predictive networks identified here would relate to body weight measures in an entirely separate sample. Given that we have not performed this external validation, it limits the generalizability and replicability of our results. Additionally, it is important to note that we lacked information on socioeconomic status in this sample. There is evidence to suggest that lower socioeconomic status is associated with altered neural responses to food cues in reward processing and executive control regions (Zhang et al. 2023), with socioeconomic status influencing the relationship between brain food cue reactivity and weight. As such, future studies should include a comprehensive assessment of socioeconomic status. Another variable that would have added further nuance to the results would be recall of the commercials after the fMRI paradigm to assess participants’ attention to the task. However, this was not directly tested here. Lastly, over the past several years, there has been a considerable shift from more traditional to digital marketing with the food industry advertising foods and beverages on social media apps, streaming platforms, and video games, with equally alarming effects on children’s and adolescents’ food choices as traditional TV commercials (Mc Carthy et al. 2022), highlighting the need for further studies.

In sum, this study found that stronger connectivity between regions implicated in memory formation, learning, visual processing, and self-control in response to unhealthy fast-food commercials predicts lower BMI at a 2-year follow-up. These results suggest that some adolescents might be able to exert more self-control over the influence external food stimuli have on their food choices, making them less vulnerable to overeating and weight gain later in life. Further examining the biological and environmental factors that lead to stronger connectivity in this protective network in some individuals could be crucial for understanding the neuropsychological underpinnings of eating behaviors, while longer follow-up studies across developmental transitions would be an important future direction. Although further research on individual-level differences in eating behavior and food choices is warranted, it is imperative to consider the need for population-level changes in policy that strictly regulate fast-food advertisements targeting adolescents.

Acknowledgements

The authors would like to thank the study participants and the research team at the University of Michigan for completion of data collection and initial analyses.

Author contributions

Afroditi Papantoni (Conceptualization; Formal Analysis; Methodology; Visualization; Writing – original draft; Writing – review & editing), Ashley N. Gearhardt (Conceptualization; Funding acquisition; Investigation; Project administration; Resources; Writing – review & editing), Sonja Yokum (Conceptualization; Investigation; Project administration; Writing – review & editing), Lindzey V. Hoover (Data curation; Writing – review & editing), Emily S. Finn (Conceptualization; Formal Analysis; Methodology; Writing – review & editing), Grace E. Shearrer (Supervision; Writing – review & editing), Lindsey Smith Taillie (Supervision; Writing – review & editing), Saame Raza Shaikh (Supervision; Writing – review & editing), Katie A. Meyer (Supervision; Writing – review & editing), Kyle S. Burger (Conceptualization; Formal Analysis; Methodology; Supervision; Writing – review & editing).

Conflict of interest:

None declared.

Funding

This work was supported by the National Institute of Diabetes and Digestive and Kidney Diseases [R01DK102532 (Principal Investigator: A.N.G.)].

Data availability

The imaging data used in the manuscript are publicly available at: https://doi.org/10.7302/5vve-yz30 (Hoover 2022). The code book and analytic code will be made available upon request from the corresponding author.

References