Abstract

Objectives:

The interplay between emotion and memory is a central topic in cognitive neuroscience, with open questions about the underlying neuronal mechanisms. This article aims to study the effects of order and intensity of emotional information on associative memory encoding. To this aim, we employ dynamic causal modeling to model the dynamic network composed of the hippocampus, amygdala, and orbitofrontal cortex during an fMRI associative memory encoding task and apply graph and control theory tools to obtain novel insights.

Methods:

Participants were clustered into three condition groups, neutral–neutral, neutral–emotional, or emotional–emotional, and viewed image pairs associated with their assigned condition. Using the dynamic causal modeling framework, we explore several dynamic models and show that a stochastic bilinear state-space model best describes the neuronal dynamics in all conditions. Furthermore, we use graph and control theory techniques to both validate and analyze the model. Particularly, we analyze the network dynamics of each condition using tools from graph theory and stability theory and discuss the differences in the strength and direction of connectivity as well as stability of each of these networks.

Results:

We confirm the prior finding that memory is enhanced in the neutral-emotional condition. In our work, this enhanced memory is associated with the increased hippocampus–amygdala coupling strength in this condition. Moreover, we show that in the emotional–emotional condition, coupling of hippocampus and amygdala, as well as the whole network connectivity increases. We further predict that the hippocampus–amygdala connectivity in this condition increases, when the first image's valence is substantially less negative rated than the second image, but decreases otherwise. This pattern mirrors the neutral–emotional condition, where the first image is emotionally neutral compared with the second. Moreover, our model-based analyses suggest that the amygdala predominantly influences the other two regions in the neutral–emotional condition.

Conclusion:

Combined data-driven DCM modeling, stability analyses, and graph-theory tools led to new insights and enhanced the mechanistic understanding of dynamics of emotional associative memory. We discuss these insights, utilize these analytical tools to generalize our findings to some unmeasured conditions, and highlight the potential of these techniques to inform the development of future technological or pharmacological approaches targeting regulatory mechanisms.

1 Introduction

The intricate relationship between emotions and memory has long fascinated researchers due to its profound implications for understanding human cognition and behavior, central to cognitive neuroscience and psychiatric research (Ochsner and Schacter, 2000). Emotional experiences are often remembered more vividly than neutral ones, shaping how individuals perceive and interact with the world (Cahill and McGaugh, 1996; LaBar and Cabeza, 2006). Yet, the neural mechanisms that govern emotional associative memory remain partially understood.

The brain regions involved in memory-emotion interactions include the amygdala, the hippocampus, and specific prefrontal areas. The amygdala is mainly involved in assigning emotional value to sensory stimuli and in modulating memory consolidation and retrieval (MorÉn, 2001) while the hippocampus supports the formation and retrieval of declarative, associative memory (Eichenbaum, 2004). The orbito-frontal cortex (OFC) within the PFC exerts top-down control by integrating emotion regulation, attention, working memory, and reward processing to support flexible, goal-directed behavior (Miller and Cohen, 2001). To reveal the mechanisms of brain networks in cognition, data collected by functional magnetic resonance imaging (fMRI) has often been used. The technique offers non-invasive high-resolution imaging capable of capturing all brain regions, their structures, connectivity, and activity. Although some interactive mechanisms have been delineated, a comprehensive understanding has not yet been achieved.

To identify and predict underlying neural dynamics, several approaches have been proposed in the literature, including generative models, which are typically formulated as systems of differential equations or density dynamics and are often represented within a state-space framework (Ramezanian-Panahi et al., 2022). These models enable the study of dynamic interactions, support mechanistic interpretations of observed data, and provide a foundation for brain simulation and stimulation. Generative models vary in their level of abstraction, the extent of biological detail incorporated, the amount of prior knowledge required, and the size of the datasets they can accommodate (Bakels et al., 2025; Jafarian et al., 2023).

Detailed biophysical models require substantial prior knowledge and involve the estimation of a large number of parameters, specially for large scale neural dynamics, often resulting in considerable complexity in capturing underlying biological processes (Izhikevich and Edelman, 2008). This complexity can hinder parameter estimation, analysis, and inference. While purely data-driven methods minimize reliance on prior biological knowledge (Chen et al., 2018; Deco et al., 2017), hypothesis-driven models lie between detailed biophysical and purely data-driven approaches. These methods, such as Dynamic Causal Modeling (DCM), aim to estimate both the parameters of partially known biophysical models as well as the remaining model structure.

Dynamic Causal Modeling, in particular, is a widely used framework for modeling effective connectivity, that is, the causal influence that neuronal systems exert on each other (Friston et al., 2003). It describes the temporal evolution of latent neuronal states and associated hemodynamic responses under prior assumptions, enabling the estimation of endogenous, modulatory, and driving influences within a specified network. DCM has been extensively applied to investigate the effective connectivity in brain networks (Daunizeau et al., 2011; Lohmann et al., 2012; Friston et al., 2013; Lohmann et al., 2013), with extensions incorporating stochastic dynamics (Li et al., 2011), excitatory and inhibitory populations (Marreiros et al., 2008) and nonlinear formulations (Stephan et al., 2008).

Prior work has utilized DCM to study emotional memory mainly focusing on the amygdala and hippocampus interactions. Using deterministic two-state bilinear DCM, where each brain region is modeled with two interacting neural states (typically excitatory and inhibitory), and inter-regional couplings are modulated by external inputs, Fastenrath et al. (2014) reported that the strength of the connection from the amygdala to the hippocampus was rapidly and robustly increased during the encoding of emotional pictures compared to the neutral ones. Another fMRI study employed classical DCM, i.e., inter-regional couplings are modulated by external inputs, to investigate the effective connectivity among the amygdala, hippocampus, and dorsolateral prefrontal cortex during a memory–emotion task (Gagnepain et al., 2017). Their analysis showed that the suppression of distressing memories requires PFC regions to inhibit both amygdala and hippocampus activities. Using classical DCM, a further study on emotional associative memory retrieval demonstrated that the OFC modulates amygdala-hippocampus interactions during the recall of emotional context (Smith et al., 2006).

These studies demonstrate the utility of DCM in revealing effective connectivity in brain regions involved in emotional memory. To date, to the best of our knowledge, the obtained models have been mainly used for inferring effective connectivity and not for qualitative analysis of the network behavior. In fact, differential equations representing the emotional memory dynamics are capable of providing insights on the whole network performance and can be used for prediction and regulation. In this work, we aim at addressing this gap by modeling and analyzing the emotional memory encoding network in a task where the order and valence of negative emotions matter.

The recent study in Zhu et al. (2023) has performed fMRI data analysis from an experiment involving the memorization of neutral and emotionally charged image pairs and explored how emotional stimuli influence memory integration. Participants were clustered to three condition groups and viewed image pairs that were either neutral–neutral, or neutral–emotional, or emotional–emotional. The study has used task-dependent functional connectivity that allows measuring correlated activities among brain regions. Their findings suggested that emotional information facilitated memory integration with related neutral information but disrupted the integration with other emotional information.

In this work, we develop a dynamic model of the emotional associative memory encoding task reported in Zhu et al. (2023). Our aims are to: (1) use Dynamic Causal Modeling (DCM) to derive a set of differential equations, i.e., dynamic model, that reproduce the data and captures the underling neuronal dynamics of emotional associative memory; (2) validate the dynamical properties of the model through stability and controllability analyses; (3) infer effective connectivity among the amygdala, hippocampus, and orbitofrontal cortex (OFC); (4) apply graph-theoretic measures to assess network connectivity; and (5) evaluate how emotional valence, as the model input, affects the model's dynamic properties, particularly stability. We expect the model to generalize to some unmeasured conditions. We discuss the results, limitations, and directions for future research.

2 Materials and methods

This section details the experimental paradigm, data acquisition and processing procedures, modeling framework and the model space employed to obtain the associative memory encoding dynamics.

2.1 Experimental paradigm

The data used in this research have been derived from the associative encoding phase of a functional magnetic resonance imaging (fMRI) experiment designed to investigate how emotional information influences associative memory (Zhu et al., 2023). During this phase, participants were shown 48 different “ABC” image triplets, where a single spatial location cue (A) was paired first with one image (B), and then the same location was paired with a second image (C), forming AB and AC pairs. The images B and C carried emotional valence; either neutral or negative. Participants were instructed to vividly imagine the relationship between the location and the associated image to facilitate memory formation.

A total of 70 healthy young adults completed the experiment. They were randomly assigned to one of three condition groups, based on the emotional valence of the image stimuli they were exposed to:

Neutral–Neutral (NN) group with 25 participants: both associated images (B and C) in a triplet were neutral,

Neutral–Emotional (NE) group with 21 participants: the first image (B) was neutral and the second image (C) was emotional (negative),

Emotional–Emotional (EE) group with 24 participants: both images (B and C) were emotional (negative).

The 48 ABC triplets were split into four sets of 12 triplets. During the study, each participant completed four runs of the experiment, with a different set of 12 used per run. In each run, the AB and AC pairings were displayed in a blocked and repeated manner: 12 AB pairs were shown in consecutive encoding trials (explained below), then 12 AC pairs, followed by a repetition of the same AB and AC pairs, concluding a run. Throughout the experiment, participants were scanned using fMRI, and their blood-oxygen-level-dependent (BOLD) responses were recorded. As shown in Figure 1, each encoding trial (AB or AC) consisted of:

A brief display of a cartoon map (0.5 s),

A highlighted location on that map (1.0 s), and

A simultaneous display of the location cue (A) and the associated item (B or C) (2.5 s).

Illustration of the associative encoding trial paradigm used in the experiment. Adapted from Zhu et al. (2023).

Trials were separated by a jittered inter-trial interval ranging from 0.5 to 1.5 seconds in which a fixation cross was displayed. As a result, each trial lasted approximately 5 seconds and each run around 4 minutes and 30 seconds. The experiment procedure also included recall and behavioral analysis phases. However, this study focuses exclusively on the encoding phase, as it provides the most direct insight into the neural mechanisms of emotional associative memory formation.

2.1.1 Data acquisition

MRI data were acquired using a 3.0 T Siemens Skyra (Siemens Medical, Erlangen, Germany) with a 32-channel head coil system at the Donders Institute, Centre for Cognitive Neuroimaging in Nijmegen, the Netherlands. Functional images were collected using a multi-band echo-planar imaging (mb-EPI) sequence (slices, 66; multi-slice mode, interleaved; slice thickness, 2 mm; TR, 1,000 ms; TE, 35.2 ms; flip angle, 60°; multiband accelerate factor, 6; voxel size, 2 × 2 × 2 mm; FOV, 213 × 213 mm). To correct for spatial distortions, fieldmap images were acquired (slices, 66; multi-slice mode, interleaved; slice thickness, 2 mm; TR, 500 ms; TE1, 2.80 ms; TE2, 5.26 ms; flip angle, 60°; voxel size, 2 × 2 × 2 mm; FOV, 213 × 213 mm). Structural images were acquired using a three-dimensional sagittal T1-weighted magnetization-prepared rapid gradient echo (MPRAGE) sequence (slices, 192; slice thickness, 1 mm; TR, 2300 ms; TE, 3.03 ms; flip angle, 8°; voxel size, 1 × 1 × 1 mm; FOV, 256 × 256 mm).

2.2 ROIs and DCM data processing

This section describes the data processing pipeline for DCM. For preliminary processing of fMRI images, we refer to (Zhu et al., 2023). The present study focuses on region-of-interest (ROI) including the amygdala (Amy), hippocampus (Hip) and orbitofrontal cortex (OFC), restricted to the left hemisphere. The selection of ROIs was guided by prior work (Zhu et al., 2023) using the same dataset, which focused on the hippocampus and the left amygdala. In line with our hypothesis-driven modeling approach, we therefore restricted the analysis to left-lateralized ROIs. We further reflect on this choice in Section 4. These anatomical regions are identified using the Automated Anatomical Labeling (AAL) atlas via the WFU_PickAtlas toolbox in MATLAB, which generates binary brain masks.

The fMRI dataset consists of voxels, units on a 3D grid, in the brain regions and a corresponding time series for each voxel. The goal of DCM data processing is to identify the task-relevant voxels within specific brain regions (Volumes of Interest, VOIs), and then average their time series to obtain a single representative signal per VOI. Task-relevant voxels are identified by fitting a Generalized Linear Model (GLM) to the BOLD signal at each voxel. A voxel is considered significant if its activity correlates with the experimental design, here, showing significant responses to both inputs. The GLM is defined as Ashburner et al. (2014):

where yvoxel(t) is the voxel's observed BOLD signal over time, X is the design matrix, i.e., input signal convolved with a Hemodynamic Response Function(HRF), β are the parameter estimates, ϵ is the residual noise. The procedure for reducing voxel-level data to a single time series per VOI is provided in the Supplementary material. For DCM data processing, we have used the Statistical Parametric Mapping (SPM12) software package running in MATLAB (Ashburner et al., 2014). At the end of pre-processing, participants with no significant voxels in at least one VOI were excluded leaving a final sample of 65 participants. To obtain DCM models, we used the data corresponding to first of four experimental runs per participants in Zhu et al. (2023), allowing us to focus on neural dynamics during the early stages of learning, before participants develop strategies across repeated exposures. We also note that the number of participants in each condition is sufficient for reliable group-level modeling. Moreover, the first 10 initial values from each run has been discarded due to the scanner setup.

The timing parameters for fitting data to DCM models were chosen to correspond to the experimental conditions. The echo time was to 40ms matching the property of the fMRI machine used (Zhu et al., 2023). In addition, slice-timing correction was applied to compensate for the delays within the Temporal Resolution (TR) of 1 sec. The time series were realigned to a reference time at the TR midpoint, following SPM12 recommendations.

2.3 Modeling framework

As motivated in the introduction, we chose a DCM framework for developing a state-space model that captures the dynamics of a three-region network comprising the Amygdala (Amy), Hippocampus (Hip), and Orbitofrontal Cortex (OFC). To this aim, we need to choose a model structure, e.g. bilinear, nonlinear, etc, external input signals, and the assumed connection across the network nodes, i.e., brain regions, as well as the manner that the exogenous inputs affect the connections or region dynamics. DCM estimates effective connectivity among brain regions using variational Bayesian inference under free-energy principle (Friston et al., 2003), briefly, a unified theory of how the brain combines prior knowledge with stimuli from the environment to learn and adapt.

The models' outputs are the averaged time series per VOI obtained after the processing steps explained in Section 2.2. The inputs to our models are the deterministic signals defined as a step function takes the value 1 during external stimulation and 0 otherwise. As described in Section 2.1, each encoding trial has three stages: map, location cue, and simultaneous display of the item pair AB or AC. Because associative encoding is expected to occur only during co-presentation of AB or AC, the inputs are set to 1 only for this 2.5s and to 0 during the map, cue, and the jittered inter-trial interval. To distinguish first (AB) and second (AC) associations within each triplet, two separate input signals are defined: u1 for AB and u2 for AC. These functions are illustrated in Figure 2.

Example of input signals u1 and u2 used for fitting DCM in a dummy case with only two triplets in condition NE.

Models are identified using SPM12 on MATLAB. The algorithm expects the user to define which model connections are assumed. These connections are usually defined by using biological assumptions. If a connection is assumed, its value will be updated during model fitting, otherwise it will remain 0. In what follows, we provide a review on model structures, classic and stochastic DCM. An overview on nonlinear and two-state DCM, used in model comparison, is provided in Supplementary material.

2.3.1 Classical (Bilinear) DCM

DCM modeling is composed of two dynamics: neuronal and hemodynamic states. For neuronal dynamics both model structure, inputs and parameters need to be identified, whereas the identification of hemodynamic only requires parameter identification. The neuronal dynamics are expressed by a nonlinear function that can be approximated by the Taylor Series expansion. Only the first-order derivatives are included in the classical DCM. The neuronal dynamics representing the Classical (Bilinear) DCM are:

with z as states, u as inputs and α as the parameters of the model. Matrix A represents anatomical connections between the brain regions, Bi the change in coupling caused by the j-th input and B the direct influence of inputs on neuronal dynamics. Figure 3 shows the structure of the model.

Model space for classical DCM. Blue arrows show couplings between nodes (A); dots show input-dependent modulation of couplings (Bi); yellow arrow with question mark indicates direct driving effects of the external input on node states (B).

In addition to these dynamics, the DCM framework uses the Balloon model (Buxton et al., 1998; Mandeville et al., 1999; Buxton and Frank, 1997) for the hemodynamic which includes vasodilator signals, inflow, blood volume and normalized deoxyhemoglobin content. The dynamics explains how the activity of neural regions influence hemodynamic responses. When combined with the neuronal dynamics, the full model of the system is obtained. The final model then contains x as the states of both models, u the inputs and θ all of the parameters to be estimated for both models:

where y is the output, h(u, θ) is the estimated BOLD response, X captures the confounding effects, usually defined as a low-order discrete cosine that models low-frequency response drifts, with unknown coefficient β and ε is the error (Friston et al., 2003).

When using this formulation, noise and parameter priors are assumed Gaussian. The assumption is adopted to ease computation rather than to reflect biology (Lohmann et al., 2012). With this model definition in mind, we choose the model structure of classical DCM as:

where * denotes the parameter which needs to be identified. Here, all nodes are assumed to be connected with A matrix having all ones. This assumption is supported by previous studies on the connectivity of these brain regions (Fastenrath et al., 2014; Ćurcić Blake et al., 2012; Gagnepain et al., 2017; Nawa and Ando, 2019; Smith et al., 2006). Most of these studies found a bidirectional connection between the nodes which matches our assumptions for this matrix.

For the Bi matrices, we allow both inputs (u1, u2) to modulate all couplings, so differences between emotional and neutral inputs are expressed in the degree to which they modulate each connection. A similar assumption for Bi matrices have been imposed in another study on emotional associative memory (Ćurcić Blake et al., 2012).

In the case of matrix B, it is unclear which brain regions receive a direct input. However, related studies (Fastenrath et al., 2014; Ćurcić Blake et al., 2012; Gagnepain et al., 2017; Nawa and Ando, 2019; Smith et al., 2006) found that the input acts directly on only one node which then influences the other nodes. Therefore, three different matrices are proposed, with inputs given solely to either Amy, Hip, or OFC.

2.3.2 Stochastic DCM

Deterministic variations of DCM omit the random firing of neurons and variations in transmissions between neurons due to the stochasticity at the cellular and molecular level (Destexhe and Rudolph-Lilith, 2012; Guo et al., 2018; Stein et al., 2005; Faisal et al., 2008). To account for these, stochastic DCM includes an additive term to model neuronal noise (Li et al., 2011; Friston et al., 2011; Daunizeau et al., 2012). The resulting neuronal model, with ϖ as the noise, becomes:

The final model is then obtained by including the same hemodynamic model as the other DCM formulations. In the end, the stochastic DCM includes two noise terms: the measurement noise ε which is common for all types of DCM utilized in this paper, and the neuronal noise ϖ introduced in the stochastic DCM. The node connections of Section 2.3.1 are kept for the stochastic DCM as well, yielding three models.

3 Results

This section first presents the result of model comparisons among four model structures. We show that the stochastic DCM performed best. Thereafter, we discuss the choices of external inputs, i.e., which of the three brain regions receives the external input per condition, and provide a comparison between the model output and the measured data. Next, we explore the system properties, stability and controllability, of the obtained models in order to validate them from a dynamical systems' perspective.

3.1 Model selection

This research aims to identify a single best-fitting DCM for each condition (NN, NE, EE) that explains the data across participants. To this end, we first fitted every model of the model space to each participant's data for each of the 65 participants. By using four different model structures of DCM, we have examined 18 models to be evaluated: 3 classical DCM, 9 nonlinear DCM, 3 two-state DCM, and 3 stochastic DCM.

DCM models were estimated using variational Bayesian inference, which depends highly on prior specification. For all models, we adopted the default SPM12 priors without modification. After obtaining the subject-level models, we performed Bayesian Model Selection (BMS) separately for each condition (NN, NE, and EE). Model comparison used log model evidence. We adopted fixed-effects BMS, which assumes a single best model for all participants corresponding to one condition given the identical task and conditions. Under this assumption, the winning model is common across participants, only the parameter estimates differ. Finally, we performed Bayesian Parameter Averaging (BPA) to obtain group-level parameter estimates for the winning model structure.

To quantify goodness of the fitting, we compute the coefficient of determination (R2) between the DCM-predicted time series and group-average measured data. The measure is used to compute fraction of variance in the measured time series that can be explained by the model, i.e.,

where yi is the measured signal at time i, fi is the model prediction and ȳ is the mean of the measured signal. Positive (R2) values closer to 1 indicate better fit, whereas negative values arise when the model performs worse than a mean-only (flat) fit. The results of this metric are displayed on Table 1. For classical, nonlinear and two-state models the (R2) values are negative or close to zero. In contrast, the stochastic DCM yields mostly positive (R2) with larger magnitudes, except for the OFC in the EE condition, and slightly in the NE condition. These results, together with the qualitative observations from the figures, indicate that the stochastic DCM provides the best overall account of the data.

Classical DCMCondition NN–6.7550–7.8580–3.1585Condition NE–2.9134–2.0622–1.5884Condition EE–3.3053–3.2494–0.0100Nonlinear DCMCondition NN–0.0411–0.8667–0.2948Condition NE–0.0422–0.0120–0.0272Condition EE0.00750.01540.0005Two-State DCMCondition NN–0.0411–0.8667-0.2948Condition NE0.02410.00650.0088Condition EE–1.9679–2.1009–0.0840Stochastic DCMCondition NN0.64930.76260.3101Condition NE0.38890.2011–0.0344Condition EE0.48020.5599–0.6269

The R-squared between the response of the optimal average models and the average measured data, calculated per condition, per brain region, and per DCM variation.

3.2 Stochastic DCM model: structure and accuracy

The stochastic DCM provided the best overall fit to the data, so we focus on this variant and examine its model structure, estimated parameters, state evolution properties and network connectivity.

3.2.1 Model structure: Bayesian model selection

As outlined in Section 2.3.1, we evaluate three stochastic models that share the same structure for A and Bi matrices of neuronal state equations but differ in their B matrices to identify which of the three nodes, Amy, Hip, and OFC, the external input stimulates. Bayesian Model Selection (BMS) compares candidate models using the log model evidence. This value combines how well a model fits the data with how simple it is, by penalizing deviations from priors. The log evidence values for the stochastic DCM models are reported in Table 2. It is easier to evaluate these values on a scale relative to the lowest value. The relative log evidence values are obtained by MErelative, model i = MEabsolute, model i−min. A relative log evidence of 3 or more is often considered as a strong indication for that model to be the optimal one. For each of the conditions, one of the models presents a relative log evidence of at least 3 compared to the other two models for that condition. These models are:

Condition NN: Model 1 - input to Amy

Condition NE: Model 3 - input to OFC

Condition EE: Model 2 - input to Hip

ModelNN/NE/EENNNEEELog model evidenceModel 1: Input to Amy–0.0002–32.0000–11.4785Model 2: Input to Hip–32.0002–32.0000–0.0242Model 3: Input to OFC–8.51190.0000–3.7351Posterior probabilityModel 1: Input to Amy0.99980.00000.0000Model 2: Input to Hip0.00000.00000.9760Model 3: Input to OFC0.00021.00000.0240

Absolute log model evidence values and posterior model probabilities obtained during Bayesian Model Selection (BMS).

Looking across the three conditions, the log model evidence is highest for the NE condition. For the EE condition, model comparison is less decisive, as the difference in log evidence between Models 2 and 3 is only slightly above 3, indicating moderate evidence in favor of the winning model. Posterior probabilities, computed separately for each condition, nevertheless indicate a clear preference for a single model within each condition.

3.2.2 Model accuracy

Figure 4 shows the output of the group DCM model vs. the data. In DCM for fMRI, neuronal states typically start at their prior mean, commonly zero, and hemodynamic states at the steady states. Initial states are not usually estimated as free parameters. For the stochastic DCM, although the response also starts at zero, it quickly shifts to the level set by the data and then tracks it without a sustained offset.

Stochastic DCM results for Amy, Hip, and OFC: blue line is the response of the average optimal model obtained for each condition; red line is the average signal for each condition.

To assess node-wise fit, we examined coefficients of determination (R2) for all candidate models rather than only the BMS winners. Table 3 reports these values separately for each node under each input mapping for each condition, NN, NE, and EE. Higher R2 values indicate better fit to the data. Comparing the models selected by BMS, R2 is highest in the NN condition. This can be explained by emotional images introducing more complex, less predictable neural dynamics that are harder to capture.

Condition NNModel 1: Input to Amy (optimal)0.64930.76260.3101Model 2: Input to Hip0.46310.69960.0277Model 3: Input to OFC0.46320.69960.0278Condition NEModel 1: Input to Amy0.27190.0304–0.5625Model 2: Input to Hip0.27170.0305–0.562516-7.4,-13.5242ptModel 3: Input to OFC (optimal)0.38900.2011–0.0344Condition EEModel 1: Input to Amy0.47300.5534–0.5530Model 2: Input to Hip (optimal)0.48020.5599–0.6269Model 3: Input to OFC0.47300.5535–0.5518

R-squared values in each node (