Body mass index–dependent shifts along large-scale gradients in human cortical organization explain dietary regulatory success

Significance

Over 18% of the global adult population is projected to be obese by 2025, showcasing widespread difficulties in adopting healthy diets. Why? Using a gradient approach, we examined how brain states change when making natural and regulated dietary choices in an established food task. Individuals with lower weight status could successfully modify their eating behaviors while maintaining similar modes of brain activity. Individuals with higher weight status could not rely on this mechanism, requiring more reconfigurations of food-evoked activation patterns to respond to the challenge of dietary control. Requiring more or fewer reconfigurations of large-scale brain patterns to align behaviors across contexts might explain why some people struggle with dietary control in their daily life, and others do not.

Abstract

Making healthy dietary choices is essential for keeping weight within a normal range. Yet many people struggle with dietary self-control despite good intentions. What distinguishes neural processing in those who succeed or fail to implement healthy eating goals? Does this vary by weight status? To examine these questions, we utilized an analytical framework of gradients that characterize systematic spatial patterns of large-scale neural activity, which have the advantage of considering the entire suite of processes subserving self-control and potential regulatory tactics at the whole-brain level. Using an established laboratory food task capturing brain responses in natural and regulatory conditions (N = 123), we demonstrate that regulatory changes of dietary brain states in the gradient space predict individual differences in dietary success. Better regulators required smaller shifts in brain states to achieve larger goal-consistent changes in dietary behaviors, pointing toward efficient network organization. This pattern was most pronounced in individuals with lower weight status (low-BMI, body mass index) but absent in high-BMI individuals. Consistent with prior work, regulatory goals increased activity in frontoparietal brain circuits. However, this shift in brain states alone did not predict variance in dietary success. Instead, regulatory success emerged from combined changes along multiple gradients, showcasing the interplay of different large-scale brain networks subserving dietary control and possible regulatory strategies. Our results provide insights into how the brain might solve the problem of dietary control: Dietary success may be easier for people who adopt modes of large-scale brain activation that do not require significant reconfigurations across contexts and goals.

Choosing when, what, and how much to eat requires integrating numerous considerations whose relative importance can depend on individual preferences, dietary goals, and context. Making appropriate dietary choices can be challenging. Over one billion people are obese globally (1), with severe long-term consequences for personal and public health (2, 3). Nearly half (42%) of all adults in the United States report actively trying to lose weight at any given time (4). Yet individuals differ greatly in their ability to implement health goals (5–9). What distinguishes people who successfully implement their dietary goals from those who fail?

Neuroimaging studies have started to address this question, identifying a cacophony of brain areas reliably activated during dietary regulation, including the dorsolateral prefrontal cortex, supplemental motor cortex, anterior insula, inferior frontal gyrus, parietal areas, and their modulation of the brain’s value system, to name a few (10). However, variation in the activity in these brain regions has been only inconsistently linked to individual differences in regulatory success (RS). Why? Given the complexity of food choice—which draws on visual, olfactory, interoceptive, and tactile cues along with semantic and episodic memory and higher-level goal states—this might not be surprising. What determines RS (or failure) might depend on the particular strategy or sets of strategies each individual adopts. Region-based analyses at the group level—which examine whether variance in brain responses in a given area predicts people’s dietary success—might struggle to account for the heterogeneity of regulatory tactics. Thus, it should come as little surprise that the most consistent regions to correlate with individual differences in dietary success (i.e., ventromedial and dorsolateral prefrontal cortex) likely represent integrative endpoints in the valuation process (5–7, 11). These regions are densely interconnected with sensory, association cortices, and motor areas, further underscoring the notion that modulation of the specific inputs these regions integrate might contribute to the successful alteration of decision-making. However, this leaves open important questions about whether and how shifts in the larger landscape of neural processing might support successful self-regulation. Although, in theory, those changes should reflect variance in dietary success, previous work has failed to identify such mechanisms.

Here, we provide evidence that understanding how individuals cope with the challenge of dietary control might require drawing on analytical approaches capable of identifying large-scale dynamic changes across the entire suite of complex processes linked to dietary behaviors and potential regulatory strategies. To this end, we propose using an increasingly popular framework that examines spatial transitions (or gradients) of macroscale brain organization. Gradients (or manifolds) allow for lower dimensional representations of functional or structural properties of the brain. The principal gradients provide an organizing spatial framework for multiple large-scale networks. They characterize an intrinsic coordinate system of topographical brain organization that may play a role in constraining functional dynamics (12). Gradients have been identified across different brain modalities (13–20), with a high degree of convergence of described gradients (21, 22) and species (17, 23). To characterize the intrinsic gradient space, our study relied on an established decomposition of the resting-state functional magnetic resonance imaging (fMRI) data from the Human Connectome Project (15, 17, 24). Thus, the principal axes of topographical organization of brain function (17) were completely independent of dietary brain states or behavior related to regulatory dietary success. We focused on the first three dimensions of brain variation that differentiate unimodal and heteromodal (association) cortices (Gradient 1), visual and sensorimotor cortices (Gradient 2), and the default mode network (DMN) and frontoparietal systems (FP) (Gradient 3). Describing dietary brain states in this three-dimensional gradient space can help understand how cognitive phenomena such as dietary control can arise from the macroscale topographical organization of the cortex.

Adopting the analytical gradient framework, we examined how regulatory goals alter whole-brain patterns of cortical activity (brain states) during dietary choices. Thus, we tested whether individual differences in dietary success might be better predicted not by activity in one or a small set of brain regions, but rather by the dynamic reconfiguration of large-scale neural networks along principal gradients of cortical organization. This raises an important question: Do successful regulators require more or less network reorganization? Greater differences between “normal” and regulated eating behaviors would point to distinct, reconfigured brain modes for dietary control. In contrast, successful dieters might possess more efficient patterns of whole-brain function (brain states), that require little to no changes to modify eating behaviors to align them with current dietary goals. Furthermore, we explored the moderating role of people’s weight status in this relationship, captured in their body mass index (BMI). Do low BMI individuals need more or fewer changes in large-scale brain configurations to regulate their diets than their high BMI counterparts? We examined these questions using data from three fMRI studies on dietary choice that captured dietary brain states in unconstrained contexts and when participants actively tried to make healthier choices. We tested whether regulatory changes in brain states—indicating large-scale reconfiguration of brain function along established gradients—predict variance in dietary success across people and contexts, depending on their weight status.

Methods

Participants.

This study combined data from 137 volunteers with a BMI below 35 at the time of data collection from three previously published fMRI studies on dietary choice (7, 25, 26). We excluded data from four participants because of missing BMI data and ten participants during outlier removal (defined using the MATLAB R2021b is outlier function as data exceeding three median absolute deviations from the group median [RS (N = 1) and Age (N = 9)]). The final sample included data from 123 individuals (84/39 females/males, age: 26.94 ± 6.2, all BMI < 35, 22.8 ± 3.1, mean ± SD) (Table 1). All participants had normal or corrected-to-normal vision, no history of substance abuse or neurological or psychiatric disorder, and no medication or metallic devices.

Table 1.

Demographic sample characteristics and study information

	Study 1	Study 2	Study 3	Combined dataset
N (final N)	36 (36)	59 (56)	41 (31)	123
Sex at birth (Female/Male)	16/20	37/19	31/0	84/39
BMI (Mean ± SD)	23.3 ± 2.7	22.6 ± 3.5	22.3 ± 1.8	22.8 ± 2.9
Age (Mean ± SD)	29.1 ± 5.5	23.1 ± 4.1	31.8 ± 6.9	27.1 ± 6.5
Published work	(7)	(26)	(25)	–
Ethics committee	Caltech Internal Review Board	University of Toronto Research Ethics Board	Institut National de la Santé et de la Recherche Médicale (INSERM)	–
Scanner	Siemens 3T	Siemens Prisma 3T	Siemens Verio 3T	–
MPRAGE sequence	TR = 1.5 s, TE = 2.91 ms, matrix size 256*256, 176 slices, voxel size 1 × 1 × 1 mm	TR = 2 s, TE = 2.4 ms, matrix size 256*256, 176 slices, voxel size 1 × 1 × 1 mm	TR = 2.3 s, TE = 2.98 ms, matrix size 240*256, 176 slices, voxel size 1 × 1 × 1 mm	–
EPI sequence	TR = 2.5 s, TE = 30 ms, matrix size 64 × 64, 47 slices, voxel size 3 × 3 × 3 mm, flip angle = 85°	TR = 2 s, TE = 30 ms, matrix size 96 × 96, 69 slices, voxel size 2 × 2 × 2 mm, flip angle = 30°	TR = 1.92 s, TE = 28 ms, matrix size 68 × 68, 36 slices, voxel size 3 × 3 × 3 mm, flip angle = 80°	–
Task condition	Natural (NC), Health-focus (HC), Taste-focus (TC)	Natural (NC), Health-focus (HC), Distance (DC)	Natural (NC), Health-focus (HC), Taste-focus (TC)	Natural (NC), Health-focus (HC)
Decision value (DV)	DV (4-point, “strong no” to “strong yes”)	DV (4-point, “strong no” to “strong yes”)	DV (4-point, “strong no” to “strong yes”)	DV (4-point, “strong no” to “strong yes”)
Food ratings	Health, Taste	Health, Taste	Health, Taste	Health, Taste

Experimental Design and Statistical Analysis.

All participants performed a version of a well-established laboratory food choice task while their brain responses were collected using fMRI (5, 6). In each trial, participants viewed a food picture for up to 4 s and indicated whether they wanted to eat the displayed food by pressing one of four buttons of a button box placed in their hand (“strong no,” “weak no,” “weak yes,” and “strong yes”). At the end of the study, one trial was randomly selected and implemented (incentivized choice). Thus, responses had real consequences, ensuring that participants’ choices reflected their actual preferences.

Participants made food choices under three different conditions implemented in separate task blocks. In studies 1 and 3, participants made choices while being asked to focus on the foods’ tastiness (taste-focus condition, TC), healthiness (health-focus condition, HC), or as they naturally would (natural condition, NC). NC served as a baseline representing individuals’ natural dietary decision processes. Participants in study 2 also completed the HC and natural conditions (NC) but were instructed to distance themselves from their food cravings in a third condition (distance, DC). Our study focuses on dietary choices in the two conditions shared across all studies: natural choice settings (NC) and in regulatory dietary settings (HC) (Fig. 1).

Fig. 1.

fMRI food task. Each task block started with a condition-specific instruction screen asking to make food choices as they naturally would (natural condition, NC, Upper row) or consider foods’ healthiness (health-focus condition, HC, Lower row). In each trial, participants saw a food image (4 s) and indicated their choice (illustrated on a brief feedback screen), followed by a fixation period before the subsequent trial (inter-trial interval, ITI, mean ± SD = 6.61 ± 0.31 s). One choice was randomly selected and implemented after the task to incentivize food choices.

A screen at the beginning of each block instructed participants about the upcoming task condition (Fig. 1). In studies 1 and 2, participants performed nine blocks of 10 trials for each of the three task conditions, totaling 270 trials (90 trials per condition). The overall task consisted of nine runs. In study 3, participants performed six blocks of 10 trials per condition, yielding a total of 180 trials (60 per condition). The overall task was divided into three runs. Each run contained two blocks of each condition. In all studies, the order of blocks was counterbalanced across runs and participants, except for the first block of the first run, which was preassigned to the natural choice setting (NC).

Displayed foods varied in their tastiness and healthiness. All participants rated the perceived healthiness and tastiness before the fMRI choice task outside of the scanner. This enabled quantifying the effects of a food’s healthiness and tastiness on participants’ dietary choices in each task condition (see Eq. 1 below).

All participants fasted for 3 to 4 h before the experiment to ensure the motivational saliency of the foods in the laboratory food task. Food pictures were presented on a computer screen as high-resolution pictures (72 dpi). Matlab and Psychophysics Toolbox extensions were used for stimulus presentation and response recording. Participants saw the stimuli via goggles or a head coil-based mirror and indicated their responses using a response box system. Further details on each study can be found in the previous publications (7, 25, 26).

MRI Acquisitions and Data Analysis.

Acquisition.

Table 1 summarizes the acquisition parameters for functional and structural brain data. Further details are provided in prior publications (7, 25, 26).

Preprocessing.

Functional and structural neuroimaging data of each study (Table 1) were preprocessed separately using the FMRIPREP software version 20.2.1 (27). The detailed description of the standardized analysis steps is provided in SI Appendix, Supplementary Note N1.

General linear model (GLM) of condition-wise brain states during dietary choice.

For each participant, we estimated the following neural GLM to identify brain states associated with dietary choices in different choice settings. The model included two regressors of interest for each functional run: R1) choice periods in the natural condition (NC) of the laboratory food task and R2) choice periods in the HC (Fig. 1). A third regressor-of-no-interest varied across studies: For studies 1 and 3, R3 captured choice periods in the taste-focus condition (TC). For study 2, R3 modeled food choices in the distance condition (DC). R1 to R3 were modeled as a boxcar function time-locked to the onset of the food picture and duration corresponding to the participant’s time to make a choice (button press) on that trial. The model contained eight additional regressors-of-no-interest: R4: missed trials (if applicable), R5-R11: motion parameters estimated during the preprocessing of functional brain data (three for translations and three for rotations; one for the framewise displacement, FD), and run-wise session constants. (For number of runs per study, see Methods above.) Note that across all participants, missed responses represent less than 1% of the data (mean ± SD = 2.3 ± 3.9 trials). For each participant, we then created two contrast images based on all the corresponding run-wise regressors of interest: dietary choices in natural settings [R1 > implicit baseline] and dietary choices in a regulatory health-focus condition [R2 > implicit baseline]. The resulting contrast images represent individuals’ brain states in different dietary contexts: natural and regulatory food choice (NC, HC) (Fig. 2A). These contrast images were then masked with the independent gradient mask (see below), coregistered, and resliced to the gradient maps to allow computing the similarity between task-related brain state maps and gradient maps (see below).

Fig. 2.

(A) For each participant (n = 123), we estimated two activation maps of dietary choices in natural (NC) and health-focus (HC) conditions of the fMRI food task (representing whole-brain dietary brain states in unconstrained and regulatory choice settings). (B) Illustration of the three principal axes of hierarchical brain organization based on independent resting-state fMRI data [Human Connectome Project (17, 24)]. (C) To project individuals’ neural brain state maps (panel A) into a three-dimensional space of previously identified principal cortical gradients (panel B), we correlated brain state maps and gradient maps using Pearson correlation coefficients (Fisher z-transformed). Each point represents a participant-specific neural activation map (n = 123) in the three-dimensional gradient space. Purple represents dietary brain states in HC; gray represents brain states in NC. (D) For each participant, we calculated the Euclidean Distance (ED) between brain states in the natural vs. regulatory health-focus condition. For illustrative purposes, we display two participants with a small (green) or a large (red) shift between these two condition-specific dietary brain states. Participant-specific shifts of brain states (from NC to HC) in the gradient space were then issued as a predictor of individuals’ dietary regulatory success (RS) observed in the food task.

Principal dimensions of brain variation (gradient maps).

Gradients quantify core topographic principles of macroscale organization of the brain (12). Brain areas that are more similar regarding the feature of interest occupy similar positions along a principal axis of variance (gradient). Here, we use a set of established whole-brain gradient maps derived from a connectivity matrix for resting-state fMRI data (1 h) acquired through the Human Connectome Project (17, 24). Thus, gradient maps are based on independent fMRI brain data obtained during task-free rest periods (17), unrelated to the current participant sample or food choice task (available at https://neurovault.org/collections/1598/). These principal gradients provide a framework for the spatial ordering of large-scale neural networks. Here, we focus on the first three gradients that were previously found to explain most of the variance (≈50%) and generalize across species (17). Gradient 1 represents a principal axis of variation that distinguishes unimodal and heteromodal (association cortex) brain systems. Gradient 2 characterizes an axis that differentiates visual and sensorimotor networks. Gradient 3 represents a principal component of brain variation that dissociates the DMN and FP network. Fig. 2B visualizes these previously identified gradient maps. Multiple reasons motivated our decision to focus on the first three gradients. First, the first three gradients explain around 50% of the variance in the brain data (17). The variance explained by gradients four and five was considerably smaller (~5%). Second, the three first gradients have been linked to important features of cognition (16, 28–30). Third, the lower-dimensional space (3D vs. 5D) facilitates the interpretation of the effect and visualization. Fourth, and most importantly, when we repeated our core analysis and calculated regulatory shifts in a 4- or 5-dimensional gradient space (based on other gradients identified and shared by ref. 17), the results remained qualitatively unchanged, pointing toward the robustness of our findings (SI Appendix, Tables S6 and S7). We focus our analysis on voxels shared across these three established gradients (92,130 voxels) using SPM12’s conjunction function (http://www.fil.ion.ucl.ac.uk/spm) (Cortical Gradient Map 1 ∩ Cortical Gradient Map 2 ∩ Cortical Gradient Map 3). This was motivated by the concern that a subset of voxels (included in one pre-established gradient map, but not the other two) might bias the results of our core analyses (see below). Thus, we opted to focus on voxels shared across the three-dimensional gradient space. This approach led to the exclusion of 245 voxels out of 92,130 voxels. A supplemental analysis confirmed that including those voxels did not qualitatively alter the main results.

Projecting task-based brain states during natural (NC) and health-focused (HC) dietary choices into the gradient space.

Gradients have previously been proposed to describe an intrinsic coordinate system of neural organization (12). We used the following approach to project our task-based neural activation maps—observed in the laboratory food task—into the gradient space. Separately for each participant and task condition, we computed Pearson correlation coefficients between the individual’s condition-specific brain map (e.g., NC, whole-brain contrast image estimated in the neural GLM using all voxel-wise values) and an independent gradient map (17) (Fig. 2 A and B). Pearson correlation coefficients were then Fisher-Z-transformed using MATLAB R2021b (MathWorks). The three similarity measures (one per gradient map) allowed the positioning of the individual’s condition-specific brain state in a three-dimensional gradient space. We repeated this procedure separately for both task conditions (NC, HC), yielding two data points per participant in this low-dimensional space that represents principal dimensions of variance in brain function (Fig. 2C). This enabled examining shifts of dietary brain states in the gradient space as a function of individuals’ task goals (i.e., position changes in the three-dimensional space when actively paying attention to foods’ healthiness in HC compared to natural choice settings in NC, Fig. 2D). To quantify regulation-related shifts, for each participant, we computed the Euclidean distance (ED) between an individual’s brain state in the natural choice (NC) and regulatory condition (HC) using MATLAB’s pdist function (Fig. 2D).

Behavioral Data Analysis.

Task-based measures of dietary choice and RS in the laboratory.

Do goal-related shifts of dietary brain states in the gradient space predict variance in people’s dietary behaviors in the laboratory food task? Answering this question requires behaviorally characterizing an individual’s dietary RS in the food choice task. To this end, we computed an established behavioral model to capture idiosyncratic differences in dietary behavior in natural and health-focus settings (11). The behavioral model is based on two fundamental notions: First, dietary choices result from considering and integrating different food attributes (i.e., a food’s taste and health). People may consider additional attributes (e.g., price); however, prior evidence suggests that models based on food’s health and taste attributes capture choices and reaction time in the food task well (7, 11, 31). Second, the weight of health and taste on food choices may differ across people and settings. For instance, people may put more weight on health considerations while dieting. In contrast, people may prioritize taste over health during a celebratory dinner. The laboratory food task simulates different choice settings (natural [NC] vs. health-focus contexts [HC]). The behavioral model describes how people can achieve goal-consistent RS in HC (i.e., healthier food choices): by caring more about a food’s health in HC, less about food’s taste, or a combination of both.

Matching previous implementations of the model (11), for each participant, we fitted a behavioral GLM to the decision values (DV) of all trials (i.e., participants’ responses of how much they want to eat the food shown on that trial scaled as 1 = “strong no,” 2 = “weak no,” 3 = “weak yes,” and 4 = “strong yes”). The behavioral GLM is described by Eq. 1, as follows:

$$\begin{array}{c}\hfill \mathrm{DV} = {\mathrm{\beta }}_0 + {\mathrm{\beta }}_{\mathrm{HC}} \mathrm{HC} + {\mathrm{\beta }}_{\mathrm{HR}} \mathrm{HR} + {\mathrm{\beta }}_{\mathrm{TR}} \mathrm{TR} \hfill \\ \hfill + {\mathrm{\beta }}_{{\mathrm{HC}}^{*}\mathrm{HR}} {\mathrm{HC}}^{*}\mathrm{HR} + {\mathrm{\beta }}_{{\mathrm{HC}}^{*}\mathrm{TR}} {\mathrm{HC}}^{*}\mathrm{TR} + \mathrm{\epsilon }.\hfill \end{array}$$ [1]

HC indicates trials in the health-focus condition (dummy coded). HR and TR correspond to the health and taste ratings for the food shown in the trial (participant-wise ratings obtained before the scanning session). Thus, regression coefficients β_HR and β_TR indicate the weight that an individual puts on a food’s perceived health (HR) and taste (TR) to guide dietary choices in natural settings. The behavioral GLM also includes two interaction terms: an individual regression coefficient for β_HC*HR, reflecting how a food’s healthiness (HR) guides choices in the regulatory health-focus condition (HC), and β_HC*TR, reflecting how a food’s tastiness (TR) guides dietary behavior in HC.

Finally, for each participant, we estimated their dietary RS by combining the goal-consistent changes in the weight on food’s healthiness and tastiness in HC (11) as follows:

$$\mathrm{RS} = {\mathrm{\beta }}_{{\mathrm{HC}}^{*}\mathrm{HR}} \mathrm{–} {\mathrm{\beta }}_{{\mathrm{HC}}^{*}\mathrm{TR}}.$$ [2]

RS integrates estimates of how much more a participant considered a food’s healthiness (β_HC*HR) and how much less they weighted its tastiness (β_HC*TR) in the regulatory condition while actively considering foods’ healthiness. Note that both interaction terms are negatively correlated at the group level [matching previous findings (11)], indicating that many individuals use a combination of both processes to regulate food choices in health-focus settings (Pearson’s r = −0.75, P < 0.001, 95% CI [−0.82, −0.66]). Higher positive RS values correspond to participants’ better ability to regulate themselves in health-focus trials (i.e., alter their decision process in a goal-consistent manner and make healthier food choices).

Goal-related shifts of brain states in the gradient space predict variance in dietary RS.

Having quantified the dietary success (RS scores, see above) of each participant, we proceeded to the core question of this study that we tested at the group level: Do goal-related changes of brain states in the gradient space (in HC relative to NC) reflect changes in the observed dietary behaviors (in HC relative to NC)? To address this question, we fitted a mixed generalized linear model using Matlab fitglme function (Eq. 3).

$$\text{RS} = {\mathrm{\beta }}_0 + {\mathrm{\beta }}_1 {\text{ED}}_{\text{HC}-\text{NC}} + {\mathrm{\beta }}_2 {\text{ED}}_{\text{HC}-\text{NC}} * \text{BMI} + {\mathrm{\beta }}_3 \text{BMI}$$ [3]

Participant-specific behavioral RS scores served as the dependent variable (Eq. 2). As predictors, the model included peoples’ shifts (Euclidean distance, ED_HC-NC) between the two brain states observed in both dietary settings (HC and NC) (Fig. 2D) and the interaction between the ED and BMI terms. The model also controlled for the overall effects of BMI, age, sex at birth (males = 1/females = 0), and study id (random effects).

To assess the predictive value of these variables, we conducted an out-of-sample (leave-one-participant-out) analysis for all participants using Eq. 3. To this end, we used 122 participants in the training set and one left-out participant in the test set. Estimates derived from fitting Eq. 3 to data in the training set were used to predict the dietary success of the left-out participant. We repeated this procedure 123 times, always leaving out a different participant for testing (yielding one predicted RS score per individual). We quantified the association between predicted and observed RS scores using Pearson’s correlation and a permutation test. The permutation test involved estimating the null distribution of correlation coefficients achieved by chance (by randomly resampling with replacement 1,000 observations for observed and predicted RS) and comparing our “real” correlation with this empirical null distribution (P < 0.05).

We also performed two sanity checks: First, we examined the overall position of the projected dietary brain states in NC and HC on each gradient (Fig. 3). Here, we tested whether the brain state maps were significantly different from zero on each gradient, using one-sample t tests [false discovery rate corrected, FDR (32), using the fdr-bh function in MATLAB R2021b]. Second, we examined whether the distribution of participants’ brain states in HC differed from that observed in NC, using a paired sample t test. We hypothesized that regulatory, health-focused settings in HC would elicit greater engagement of the FP network (33–35). Thus, while our main analysis used the ED between brain states in a three-dimensional space, this control analysis here focused on gradient 3, which represents the principal axis of variance that dissociates the DMN and FP network.

Fig. 3.

Projected dietary brain states in NC and HC along the three principal axes of brain variation (gradients). Histogram of projected participant-specific brain states during natural (NC, gray) and regulated health-focus (HC, purple) dietary choices (means are presented by dotted gray and purple lines). On average, dietary brain states in both food choice settings can be characterized as more unimodal (than heteromodal), visual (than motor), and FP (than DMN). Comparisons of condition-specific brain states revealed that the regulatory health-focus (HC relative to NC) was related to increased activity in the FP network relative to the DMN (gradient 3; P = 0.027, 95% CI [−0.02, −0.00], t₍₁₂₂₎ = −2.24).

Traditional univariate analysis approaches of variance in dietary RS.

For comparison, we also explored neural correlates of individuals’ RS using classic univariate analysis approaches. For each participant, we estimated one contrast image representing goal-dependent changes in brain activation (i.e., responses during choice periods in HC minus NC, see Methods: neural GLM). At the group level, we then tested whether these neural responses reflected variance in individuals’ RS (simple t test as estimated in SPM12, modulated by participant-specific RS scores). Regulatory changes in brain activation did not systematically covary with people’s dietary success scores (P < 0.05, FWE corrected, as implemented in SPM12). In other words, localized regulation-related brain responses at the level of individual voxels/brain regions did not reveal individual differences in dietary control, highlighting the benefit of our analytical gradient framework.

Results

Behavior in the Food Task.

We used an established behavioral model to capture variance in dietary decision-making and RS across people and context (Eq. 1). As expected, in natural choice settings (NC), participants relied more on food’s taste (β TR = 0.58 ± 0.29, mean ± SD, P < 0.001, 95% CI [0.53, 0.64], t₍₁₂₂₎ = 22.8, one-sample t test against zero) than health (β HR = 0.05 ± 0.18, P = 0.006, 95% CI [0.01, 0.08], t₍₁₂₂₎ = 2.78, one-sample t test) to guide their dietary decisions (paired t test, P < 0.001, 95% CI [0.47, 0.60], t₍₁₂₂₎ = 15.60). When participants’ attention was directed to health goals in HC (relative to NC), participants increased weights on health (β HC*HR = 0.41 ± 0.32, P < 0.001, 95% CI [0.35, 0.47], t₍₁₂₂₎ = 14.39, one-sample t test) and decrease the weights of taste on food choices (β HC*TR = −0.35 ± 0.34, P < 0.001, 95% CI [−0.41, −0.29], t₍₁₂₂₎ = −11.52, one-sample t test), mirroring previous results (11). The RS score integrates these changes in individuals’ ability to up-regulate health concerns and down-regulate taste considerations (Eq. 2). Average RS scores were positive (RS = 0.76 ± 0.61, P < 0.001, 95% CI [0.65, 0.87], t₍₁₂₂₎ = 13.80, one-sample t test), suggesting that participants successfully implemented goal-consistent changes in their health- and taste considerations, yielding healthier dietary behaviors in the food task’s HC condition. These model-based RS scores were also highly correlated with the observed increased number of healthy food choices (HC vs. NC), see SI Appendix, Fig. S1 for an illustration (r = 0.826, P < 0.001, 95% CI [0.76, 0.88]).

Importantly, we observed a considerable variance in these estimates (SI Appendix, Fig. S2). Thus, all model-based estimates of goal-consistent dietary changes–including individuals’ RS scores–varied sufficiently, allowing subsequent analyses of differences in dietary choices across people and contexts.

Projecting Goal-Dependent Dietary Brain States into the Gradient Space.

We first projected participants’ dietary brain states (whole-brain activation patterns observed when participants made food choices) into the three-dimensional gradient space, separately for both choice contexts (NC and HC). Fig. 3 visualizes the distribution of participants’ projected brain states for natural and regulatory food choices. Dietary brain states in both conditions were characterized as more unimodal (than heteromodal, gradient 1; NC: mean ± SD = −0.11 ± 0.06, P < 0.001, 95% CI [−0.12, −0.09], t₍₁₂₂₎ = −18.49; HC: −0.09 ± 0.07, P < 0.001, 95% CI [−0.11, −0.09], t₍₁₂₂₎ = −16.34, one-sample t tests against zero); more visual (than motor, gradient 2; NC: 0.14 ± 0.06, P < 0.001, 95% CI [0.12, 0.15], t₍₁₂₂₎ = 24.53; HC: 0.14 ± 0.06, P < 0.001, 95% CI [0.13, 0.16], t₍₁₂₂₎ = 25.76), and more FP (than DMN, gradient 3; NC: 0.08 ± 0.05, P < 0.001, 95% CI [0.07, 0.09], t₍₁₂₂₎ = 16.52; HC: 0.09 ± 0.05, P < 0.001, 95% CI [0.08, 0.10], t₍₁₂₂₎ = 19.97).

As expected, regulatory dietary brain states in HC (relative to NC) became more similar to the FP end of gradient 3 (P = 0.027, 95% CI [−0.02, −0.00], t₍₁₂₂₎ = −2.24). This result confirmed the engagement of the FP network when individuals make goal-directed, health-focused dietary decisions, supporting the notion that our analytical framework can replicate well-established findings in the field (36, 37). We did not find evidence that regulatory task goals (HC vs. NC) significantly affected the position along gradient 1 (P = 0.082, 95% CI [−0.02, 0.00], t₍₁₂₂₎ = −1.75) or gradient 2 (P = 0.309, 95% CI [−0.01, 0.00], t₍₁₂₂₎ = −1.02).

Goal-Related Shifts of Brain State Similarity with Gradients Predict Regulatory Dietary Success.

Next, we proceeded to our key question: Do goal-dependent changes in individuals’ brain states (in HC relative to NC) reflect variance in regulatory dietary success (in HC relative to NC)? To address this question, we fitted a model to predict individuals observed RS based on the ED between brain states in natural and regulated choice settings (ED _HC-NC) while controlling for age, sex at birth, BMI, and study id (Eq. 3). The overall model significantly predicted variance in individuals’ RS scores (R² _adjusted = 0.261, P = 0.015, F_5,117 = 2.96) (Table 2). We found a significant main effect of the ED between dietary brain states (β ED _HC-NC = −24.83, P = 0.009, 95% CI [−43.45, −6.22], t₍₁₁₇₎ = −2.64) (Fig. 4A). Better RS was predicted by a smaller change in task-evoked brain states. In other words, larger goal-dependent changes in observed behavior were linked with more similar brain states in terms of their position in the three-dimensional space of topographical axes of brain organization. Importantly, we also identified a significant interaction effect between the Euclidean distance and BMI (β ED _HC-NC * BMI = 0.99, P = 0.017, 95% CI [0.18, 1.81], t₍₁₁₇₎ = 2.41). To illustrate this interaction effect, we performed a supplemental slope analysis using the jtools (version 2.0.0) package implemented in R Stat (38). We first created three groups according to the BMI based on the terciles: a low BMI group (N = 41, mean BMI ± SD = 19.7 ± 1.3), a medium BMI group (N = 41, 22.4 ± 0.7), and a high BMI group (N = 41, 26.0 ± 1.9). We then fitted the same model using the categorial BMI variable (instead of the continuous BMI) and performed a simple slope analysis to compare the BMI groups. We found that the linear fit was significantly different across the three BMI groups (P = 0.035, 95% CI [0.81. 2.13], t₍₁₁₇₎ = 2.13) and that the effect was driven by the low BMI group (P = 0.010, 95% CI [−9.60, −1.76], t₍₁₁₇₎ = −2.80). No significant effect was observed for the medium BMI group (P = 0.060, 95% CI [−9.18, 0.41], t₍₁₁₇₎ = −1.88), or for the high BMI group (P = 0.790, 95% CI [−2.92, 5.61], t₍₁₁₇₎ = 0.26) (Fig. 4B). Comparison of the Pearson correlation between the groups using Fisher’s r-to-z transformation revealed significant differences in the fitting between the medium BMI and high BMI groups (P = 0.010), as well as between the low BMI and high BMI groups (P < 0.001). No differences were found between the low and medium BMI groups (P = 0.750). Finally, age was also a significant predictor of observed dietary RS, with older individuals having higher RS estimates than youth (β Age = 0.02, P = 0.047, 95% CI [0.00, 0.04], t₍₁₁₇₎ = 2.01).

Table 2.

Goal-dependent shifts in brain states between HC and NC predict individuals’ dietary RS

Predictors	Estimates	95% CI	P
Intercept	1.72	[0.23, 3.21]	0.024
ED HC-NC	−24.83	[−43.45, −6.22]	0.009
ED HC-NC * BMI	0.99	[0.18, 1.81]	0.017
BMI	−0.06	[−0.11, 0.00]	0.065
Age	0.02	[0.00, 0.04]	0.047
Sex at birth [female]	−0.05	[−0.27, 0.18]	0.679
Study [1]	0.39	[0.02, 0.77]	0.041
Study [2]	−0.10	[−0.48, 0.27]	0.583
Study [3]	−0.29	[−0.67, 0.09]	0.134
Observations	123
R2 adjusted	0.261
P (model)	0.015

Note: ED _HC-NC = Euclidean distance between task-evoked brain states in natural (NC) and health-focused (HC) dietary choices; BMI = body mass index, 95% CI = confidence interval at 95%.

Fig. 4.

Goal-dependent shifts in brain states predict variance in dietary RS. (A) Smaller goal-dependent changes of dietary brain states predict better RS (i.e., larger positive RS scores) (β = −24.83, P = 0.009, 95% CI [−43.45, −6.22], t₍₁₁₇₎ = −2.64). Values on the x-axis represent the ED between dietary brain states observed in NC and HC; smaller values represent more similar whole brain activation patterns across regulated and natural dietary settings. (B) Illustration of the interaction effect between changes of brain states and BMI using slope analysis (38). The relationship between brain state shifts and RS is plotted for individuals with a high (dotted light blue line), medium (dotted dark blue line), and low weight status (solid blue line). The negative link between RS and regulatory brain state shifts was strongest in individuals in the low BMI group (P = 0.010, 95% CI [−9.60, −1.76]), negative in the medium BMI group (marginally significant, P = 0.060, 95% CI [−9.18, 0.41]), and reversed toward a slightly positive trend in the high BMI group (P = 0.790, 95% CI [−2.92, 5.61]). (C) Correlation between the observed and predicted RS scores in the out-of-sample prediction (Pearson’s r = 0.462, P < 0.001, permutation test, 95% CI [0.31, 0.59], t₍₁₂₁₎ = 5.73). BMI = body mass index.

Robustness Checks.

We ran several supplemental analyses and sanity checks. First, the results of a formal model comparison confirmed our primary model (Eq. 3) outperformed reference models (SI Appendix, Fig. S3, Table S1, and Note N2). We found that the best model includes the regulatory brain state shift (represented by the ED) and its interaction with individuals’ BMI scores (P = 0.017, compared to the null model) as described in Eq. 3 (based on models weighted Akaike Information Criteria, wAIC) (39).

Second, we tested the model’s generalizability across people and time points. Following recommendations on how to establish robust brain-behavior relationships without thousands of individuals (40, 41), we tested the predictive power in out-of-sample data using a leave-one-participant-out cross-validation procedure. We found a significant association between the predicted and observed RS scores in left-out-participants, demonstrating that our model reliably predicts the dietary RS of “new” individuals not used for model training (Fig. 4C, Pearson’s r = 0.462, P < 0.001, 95% CI [0.31, 0.59], t₍₁₂₁₎ = 5.73). We also implemented a permutation test to confirm the out-of-sample prediction's statistical robustness. Specifically, we ran 1,000 permutations of our model (breaking up the matching between the observed RS scores and predictor variables in the training set) to create an empirical null distribution of predictions achieved by chance. Comparisons against this null distribution confirmed that the accuracy of our out-of-sample prediction is unlikely due to chance (P < 0.001, 95% CI [0.20, 0.21], t₍₉₉₉₎ = −125.25, permutation test) (SI Appendix, Fig. S4). Given differences across participants regarding the food task, scanning protocols, and test settings (data were collected in France, Canada, and the US), this out-of-sample generalization speaks to our findings’ robustness.

Third, we tested whether the prediction of RS scores was driven by the ability to suppress taste considerations or enhance health considerations while pursuing health goals (the two processes used to characterize people’s RS; Eq. 2). To this end, we ran two modified versions of our model (Eq. 3): We replaced RS scores as the dependent variable with model-based estimates of altered taste- or health-inputs on choices in HC (β HC*TR or β HC*HR). We found that shifts in the neural state space predict both the ability to suppress taste and increase health considerations to pursue dietary goals (SI Appendix, Fig. S5 and Tables S2 and S3). Thus, while the relevance of both subprocesses differed across participants, shifts in brain states reflected the altered inputs of both underlying processes that yielded healthier dietary behaviors in HC.

Fourth, we examined whether the predictive information in brain state shifts (captured in the ED between goal-specific brain states, ED _HC-NC) was due to changes along one (or more) individual gradients. To this end, we implemented an additional linear-mixed model (LMM) that decomposed the ED variable (SI Appendix, Supplementary Note N4 and Eq. S8). Specifically, we substituted participant-specific ED _HC-NC scores with three separate predictor variables that represented that individual’s goal-related shifts (HC—NC) along gradients 1 (Δ G1 _HC-NC), 2 (Δ G2 _HC-NC) and 3 (Δ G3 _HC-NC). We found that changes along the three individual gradients did not significantly predict variance in RS across people (P’s > 0.10 for all three main effects; see SI Appendix, Table S4). Instead, individuals’ dietary success was predicted by a triple interaction of gradient-specific brain state shifts (SI Appendix, Fig. S6 and Note N4). This finding indicates that RS is not enabled by changes in whole brain activation patterns along one particular topological axis of brain organization (e.g., the FP circuit captured in gradient 3). Instead, it relies on different conditional rules and interactions of multiple gradients that facilitate dietary control–multiple configurations can produce the same functional output. For example, achieving effective RS can involve reconfigurations of large-scale brain states to become more similar to the FP end of gradient 3, less visual/more motor (gradient 2), and more unimodal. Alternatively, effective RS might be achieved by brain states becoming more FP (gradient 3), more visual (gradient 2), and less unimodal/more heteromodal (gradient 1).

Fifth, we examined whether our findings were robust when controlling for dietary behaviors in natural, unconstrained choice settings. Put differently, people who naturally tend to make more healthy choices may not be able to increase their behaviors as strongly when asked to focus on foods’ health in HC. To examine the impact of natural eating behaviors, we added another predictor variable to Eq. 3. Mirroring our approach to quantify RS (Eq. 2), for each participant, we calculated a natural dietary success (NDS) score based on the degree to which the individual used foods’ taste and health to guide eating behaviors in natural choice settings (NDS = β_TR HR – β_HR TR, estimated in Eq. 1). Higher positive NDS scores represent healthier dietary decision-making in natural choice settings (NC). As expected, we found that NDS was a negative predictor of individuals’ RS scores (SI Appendix, Table S5). Thus, people that already make healthier dietary choices in NC tended to show less goal-consistent changes in HC (β _NDS = −0.88, P < 0.001, 95% CI [−1.11, −0.66], t₍₁₁₆₎ = −7.75). However, controlling for this baseline dietary control did not affect our key results: Smaller changes in brain states in the gradient space (captured in the ED) still predicted variance in people’s RS (β _ED = −18.12, P = 0.036, 95% CI [−34.98, −1.25], t₍₁₁₆₎ = 2.13), with a positive interaction effect between ED and BMI (β _ED*BMI = 0.83, P = 0.033, 95% CI [0.06, 1.54], t₍₁₁₆₎ = 2.16). In other words, the observed key effects of our study are not merely a reflection of people’s natural (better or worse) eating habits.

Finally, we repeated the analysis and calculated regulatory shifts in a four- and five-dimensional gradient space (based on other gradients identified and shared by ref. 17). We replicated our key results irrespective of the dimensionality of the gradient space, pointing toward the robustness of our findings (SI Appendix, Tables S6 and S7).

Note that this work was not preregistered, but we strictly follow previously established procedures and best practices. First, at the behavioral level, we used a pre-established, peer-reviewed analytical pipeline to compute participants’ RS score (11) to reduce the analytical degrees of freedom of the authors and the risk of the “garden of forking paths.” Second, for maximum transparency, we report the full set of predictive models of RS explored in this paper (SI Appendix, Supplementary Notes N2–N5) and utilize formal model comparison to select the winning model (42). Third, we build on prior research investigating brain reconfiguration across conditions along principal gradient axes to analyze the brain data (16, 28–30). Fourth, we use the ED to measure the metric of the brain state shift as it is widely used to unravel the connectivity patterns underlying brain organization (43–45). Finally, data and analysis scripts are openly available on GitHub to facilitate open science efforts (https://github.com/Remi-Janet/Brain-state-shift).

Discussion

How does the brain solve the problem of dietary self-control? Given the ongoing obesity health crisis, addressing this question is of paramount importance. Projections estimate that over 18% of the global adult population will be obese by 2025 (46–49), a designation that comes with significant health risks and economic costs (50, 51). Using an analytical framework that examined food-evoked brain states under different dietary goals in a low-dimensional gradient space provided several insights. First, successful regulators (those making healthier choices when focused on healthy eating) needed less reconfiguration of large-scale brain states to modify their behaviors. Second, weight status (as captured by BMI) moderated this relationship: Low BMI individuals maintained more similar modes of brain activity while achieving larger goal-consistent changes in dietary behaviors; individuals with a higher BMI did not rely on this mechanism. Third, when we examined correlates of RS along each principal axis of functional brain organization (gradients), we found a complex three-way interaction predicting dietary RS. Thus, dietary success (and failure) were characterized by significant heterogeneity in neural reorganization, potentially reflecting the heterogeneity in regulatory strategies across individuals. Together, our results highlight the complexity and importance of goal-dependent reconfigurations of macroscale networks for adaptive behaviors and provide insights into why dietary self-control may be more difficult for some people than others.

Our work capitalizes on an increasingly popular framework that focuses on a series of primary dimensions of variation in large-scale hierarchical brain organization (14, 15, 17, 52). Notably, these principal gradients have been identified using several measures, including the brain’s cortical morphology, microstructure, and functional and structural connectivity (12–19, 52–55). Atypical macroscale organization of brain function along these axes captured at rest may also be suitable for identifying neurobiomarkers and overcoming the limitations of traditional edge-based connectivity measures (56). However, to date, few studies have used this framework to characterize task-based activation patterns by exploring the role of study- and condition-specific gradients (57, 58). Here, we used established functional gradients to characterize a low-dimensional, intrinsic coordinate system of large-scale functional brain organization that may play a role in constraining functional dynamics related to dietary control. Our work demonstrated that projecting task-evoked brain states into an established gradient space may represent a powerful approach to examining cognitive phenomena like self-control and adaptive behaviors.

Supporting the validity of our analytical framework, we replicated established findings, showing enhanced activation of the FP network (FPN) in regulatory, health-focused dietary settings (indicated by regulatory shifts of dietary brain states toward the FPN end of gradient 3). This finding is consistent with a large body of work implicating this brain circuit in goal-related behaviors and cognitive control (59–61). In dietary choice contexts, the FPN is reliably activated when people deliberately try to reduce food cravings (10, 62, 63). The FPN plays a central role in executive control (64), which might be impaired in individuals with high weight status and obesity (65). Aberrant functioning in the FPN is linked to dysregulated eating associated with obesity risk in youth (66). Moreover, food-evoked activation and connectivity in core areas of this brain network, like the dorsolateral prefrontal cortex (dlPFC), correlated with weight loss from dieting and bariatric surgery (67–70) (for reviews on the relationship between dlPFC and dietary behavior, see refs. 36 and 37). Yet other studies using traditional univariate approaches often fail to link variance in localized brain activation with variance in people’s dietary RS in the FPN or elsewhere (including relatively well-powered task-based fMRI studies like the current one). Our results can shed light on this inconsistency: We found that regulatory changes in FPN activation alone did not reflect dietary success. Instead, RS was predicted by complex reconfigurations of brain patterns along all three gradients, as captured in a triple interaction. This finding suggests that the benefit of recruiting executive control regions of the FPN for dietary RS might depend on activation patterns in other brain networks, underscoring the potential heterogeneity of regulatory strategies and “avenues” to dietary success. Future work will be needed to unpack this intriguing result further.

This brings us to our main finding. Our analysis revealed that better regulators—who made healthier food choices in regulatory settings—expressed smaller brain state shifts between natural and regulatory contexts. Put differently, their patterns of food-evoked brain activity remain relatively similar across different contexts, as captured by their similarity with large-scale topographical axes of brain organization. This relationship also held when we controlled for variance in people’s baseline dietary control in unconstrained settings. Importantly, this does not imply that successful regulators do not shift brain activation patterns at all. Rather, they appear to require relatively less reconfiguration of their brain modes than their less successful counterparts, implying greater efficiency of the changes that do occur. From a biological perspective, less rearrangement of brain activity across dietary contexts and goal states might suggest that regulation in these individuals requires less energy expenditure (71, 72). Future studies will be needed to investigate the energetic costs of brain state transition in relation to dietary RS (71). If true, the ability of successful regulators to adopt brain states enabling context-dependent behavioral changes without significant neural reorganization or increased energy demands might contribute to their sustained ability to make appropriate dietary choices over time.

Notably, however, this mechanism was disrupted in high-weight individuals, for whom we observed a positive (albeit not significant) trend toward larger shifts in brain states to accomplish dietary goals. Put differently, their natural dietary brain patterns did not allow flexible dietary behaviors and required more reconfiguration to pursue health goals compared to lower BMI individuals. Interestingly, a prior study examining macroscale functional organization along hierarchical axes (gradients) at rest revealed a BMI-related increase in differentiation between heteromodal and unimodal brain networks [akin to our gradient 1 (54)]. This finding suggested disrupted segregation of low-level and higher-order brain systems, which could impact the ability to adopt modes of brain activation that enable adaptive regulation across dietary contexts. Our findings also bear on current questions in the decision sciences about the neural and computational basis of successful self-regulation (73). Based on the prominent activation of lateral prefrontal areas during self-regulation, some theories have argued that activation of executive control might be necessary to either inhibit reward and valuation regions (e.g., refs. 74 and 75), or to modulate their sensitivity to rewarding properties of foods (5, 76). Other theories have proposed that multiple regions, including lateral prefrontal areas, might play distinct roles in valuation at different times, with differing implications for choices (7, 8, 26, 77). Our results are perhaps more consistent with the latter sets of theories: Not only was greater RS on average associated with less reconfiguration of neural gradients, but regulation was also predicted by complex interactions among these brain states, suggesting that multiple combinations of configurations could yield larger behavioral change. Such a finding is less consistent with the idea of a single route for self-regulation operating through executive inhibition of reward-based regions. We note, however, that we apply this conclusion narrowly within the context of the deliberate adoption of self-regulatory goals (e.g., by deliberating about health aspects of dietary choice). Somewhat more consistent links have been drawn between spontaneous neural responses within reward- and regulation-related regions during unregulated food viewing (78–80), suggesting that the mechanisms producing healthy behavior across these contexts might differ.

Our findings were derived from a relatively large sample of task-based fMRI data [the median sample size in openneuro.org studies as of September 2021 was n = 23 (41)] and confirmed using internal validation approaches (k-fold out-of-sample predictions), following recommendations for establishing generalizable brain-behavior relationships (40). However, future work needs to replicate our results in samples with thousands of individuals (41). Including more carefully designed, task-based fMRI paradigms—that closely map evoked neural activation with specific behaviors of interest captured in performance-based variables—in large human neuroimaging datasets will be crucial to further verify robust associations between brain function and cognitive phenomena like dietary self-control.

Some limitations and associated future directions need highlighting. First, while BMI is the cornerstone of the classification system for obesity, it has limitations as a surrogate measure of body composition (81, 82). Future research will need to validate our results using standardized assessments of body fat. Second, our results cannot speak to the direction of causality: Is variance in context-dependent reconfigurations of dietary brain states a consequence of weight status, or does it represent a vulnerability that might precede and predict weight gain over time? Data from longitudinal or intervention studies (e.g., before and after bariatric surgery) might illuminate this issue. Third, more work is required to understand the precise mechanisms underlying the regulatory changes in dietary brain states. Future research will require integrating our results (and analytical framework) with other lines of work on obesity and dietary success, spanning different neurobiological properties (83, 84). For example, prior work has shown that variation in BMI relates not only to differences in brain organization in terms of whole-brain structure, function, and connectivity (83), but also to low-grade inflammation and cognitive impairment (85–87) (for an immunologic model of self-regulatory failure, see ref. 88). These neurobiological changes in individuals with a higher weight status could contribute to the BMI-dependent ability to adopt brain states enabling dietary control and context-dependent adaptive behavioral changes, and potentially also their ability to benefit from a specific intervention, regardless of whether it utilizes pharmaceutical, educational, or social means. Fourth, while the established food choice task steers individuals toward health-conscious decision-making by emphasizing the health attributes of foods and encouraging participants to bring their behaviors into line with this focus, there were no explicit instructions for participants to make healthy choices during health blocks. We thus cannot say for sure whether, for any individual participant, a lack of change in behavior results from a true regulatory failure (i.e., an unfulfilled goal to eat healthily) or from a lack of motivation to translate thoughts into action (i.e., a lack of goal). Indeed, this may, in part, be why we observe a complex interaction among gradients predicting behavioral change: For some individuals, regulatory change may occur through explicit goals to change behavior, whereas for others, it may occur through simple attentional shifts in the absence of explicit goals. Different mechanisms might ultimately be associated with different combinations of changes in gradients. Future research will be needed to identify a means for dissociating goal-related and non-goal-related behavior change to assess whether these are associated with different neural pathways (77, 89).

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Previously published data were used for this work (7, 25, 26). All other data are included in the manuscript and/or SI Appendix.