Negative Emotion Differentiation Promotes Cognitive Reappraisal: Evidence From Electroencephalogram Oscillations and Phase‐Amplitude Coupling

ABSTRACT

Cognitive reappraisal, an effective emotion regulation strategy, is influenced by various individual factors. Although previous studies have established a link between negative emotion differentiation (NED) and cognitive reappraisal, the underlying neural mechanisms remain largely unknown. Using electroencephalography, this study investigates the influence and neural basis of NED in cognitive reappraisal by integrating aspects of event‐related potentials, neural oscillation rhythms, and cross‐frequency coupling. The findings revealed that individuals with high NED demonstrated a significant decrease in parietal late positive potential amplitudes during cognitive reappraisal, suggesting enhanced cognitive reappraisal abilities. Moreover, high NED individuals displayed increased γ synchronization, parietal α–γ coupling, and frontal θ–γ coupling when reappraising negative emotions than those with low emotion differentiation ability. Machine learning analysis of these neural indicators highlighted the superior classification and predictive accuracy of multimodal indicators for NED as opposed to unimodal indicators. Overall, this multimodal evidence provides a comprehensive interpretation of the neurophysiological mechanisms through which NED influences cognitive reappraisal and provides preliminary empirical support for personalized cognitive reappraisal interventions to alleviate emotional problems.

This study provides multimodal evidences about how negative emotion differentiation (NED) impacts cognitive reappraisal. High NED (HNED) individuals demonstrated enhanced cognitive reappraisal abilities, characterized by a significant decrease in parietal LPP amplitudes. They displayed increased frontal θ–γ coupling during cognitive reappraisal.

1. Introduction

Cognitive reappraisal (CR), an antecedent‐focused emotion regulation strategy, has been demonstrated to promote mental health and well‐being (Gross and John 2003). This strategy modifies emotions by shifting the individual's perspective and their interpretations of emotional events (Gross and John 2003). Research has shown that CR can have enduring and potent effects on easing negative emotions (Troy et al. 2018). However, its effectiveness is influenced by various factors such as emotional intensity (Shafir et al. 2015), stress (Zhan et al. 2017), sleep deprivation (Walker and van der Helm 2009), and social support (Metts and Craske 2023). Additionally, individual differences such as age (McRae et al. 2012; Livingstone and Isaacowitz 2021), reappraisal or worry traits (Moser et al. 2014), alexithymia (Pollatos and Gramann 2012), attachment orientations (Domic‐Siede, Sánchez‐Corzo, and Guzmán‐González 2024; Mikulincer, Shaver, and Pereg 2003), and personality traits (Hughes et al. 2020) also play a significant role in the variability of how helpful CR can be (McRae and Gross 2020). This variability underscores the need for an in‐depth exploration of the individual factors that influence the effectiveness of CR. This investigation is a critical step towards the development of personalized CR interventions.

Emotion differentiation, the ability to discern subtle differences between similar emotions (Kashdan, Barrett, and McKnight 2015), may act as an individual factor that influences CR effectiveness. High‐emotion differentiators, for instance, use specific emotional words and make fine distinctions in their feelings, such as feeling “sad” or “anxious.” Conversely, low differentiators use broad terms such as “happy” or “unhappy” or being in a “good” or “bad” mood, distinguishing emotions solely based on valence and neglecting specific categories. Prior research has found that fine emotion differentiation, particularly in distinguishing negative emotions, can enhance psychological well‐being (Erbas et al. 2022; Kashdan, Barrett, and McKnight 2015) and promote emotion regulation (Brown et al. 2021; Kashdan, Barrett, and McKnight 2015).

Emotion differentiation facilitates CR, which can be understood through the Conceptual Act Theory of Emotion (Lindquist and Barrett 2008). This theory posits that emotional experiences emerge from a process termed “situated conceptualization” (Barrett 2006). In this process, individuals use their emotional concept knowledge—encompassing scenarios that trigger emotions, accompanying physiological responses, and rules for emotional expression—to categorize their emotions within specific contexts. CR can be reconceptualized as a shift from one situated conceptualization to another, a process that depends on having a rich repertoire of emotional concepts. Individuals with high emotion differentiation possess detailed or rich knowledge of emotional concepts (Lindquist and Barrett 2008), which is instrumental in shifting situated conceptualizations and promoting the reinterpretation of emotional stimuli. Additionally, individuals with high negative emotional differentiation (HNED) use emotional concepts to label and conceptualize their core affect. This knowledge provides critical information on the causes and consequences of negative emotions, including the context in which they arise and their associated cognitive and physiological responses (Barrett et al. 2001; Ruys et al. 2012). These emotional insights may enhance their capacity to effectively employ CR (Kalokerinos et al. 2019; Kashdan, Barrett, and McKnight 2015).

Empirical research also indicates that HNED individuals tend to employ CR strategies more frequently (Barrett et al. 2001), and they are able to downregulate the intensity of negative emotions more effectively (Kalokerinos et al. 2019). However, current research on the impact of negative emotional differentiation (NED) on CR has primarily focused on behavioral outcomes, while the underlying neural mechanisms—particularly how cognitive processes support emotion regulation—remain underexplored. To address these gaps, we decoded the neural mechanisms through which NED affects CR, integrating event‐related potentials (ERPs) and oscillatory dynamics from a multimodal perspective. Specifically, ERP analysis reveals the temporal processes involved in the neural effects of NED on CR, while the investigation of oscillatory dynamics explores electroencephalography (EEG) oscillatory features and the transmission of information and interactions between brain regions, which previous research has not addressed. To the best of our knowledge, this is the first study to use such a multimodal approach. This multimodal approach overcomes the limitations of using single neural measures (such as ERP or oscillations alone) and demonstrates how NED influences the coordination of different neural processes during CR. This comprehensive framework provides a neurophysiological foundation for understanding the role of NED in CR, allowing for future intervention strategies based on neural mechanisms.

ERPs research has identified late positive potential (LPP) as the ERP component most closely related to CR. Specifically, the amplitude of the parietal LPP primarily responds to arousal level rather than valence (Cuthbert et al. 2000; Hajcak and Foti 2020), making the LPP a well‐established neurophysiological marker of arousal in emotion regulation (MacNamara, Joyner, and Klawohn 2022). In relation to CR, the LPP provides insight into how regulatory strategies modulate emotional responses, and its changes serve as a common measure for assessing emotion regulation efficacy (Hajcak, MacNamara, and Olvet 2010; He et al. 2020). Previous studies have found that when CR is used to downregulate emotions, the amplitude of the parietal LPP decreases (Cao, Li, and Niznikiewicz 2020; He et al. 2020; Moser et al. 2014). Thus, the relationship between LPP and CR lies in attenuating arousal through the top‐down cognitive control processes involved in reappraisal. Moreover, certain studies found that an increase in the frontal LPP amplitude is also linked with CR (Moser et al. 2014; Yuan et al. 2015), reflecting enhanced attentional control during CR (Bernat et al. 2011; Moser et al. 2014).

In addition to LPP, stimulus‐preceding negativity (SPN) is another event‐related potential that has been investigated in emotion regulation. The SPN emerges in anticipation of emotionally salient stimuli and is linked to the cognitive and affective processes involved in preparing for upcoming events (Brunia, Boxtel, and Böcker 2012). Early SPN enhancement reflects the orienting and processing of the cue, whereas later SPN increases indicate preparation for and anticipation of the impending stimulus (van Boxtel and Böcker 2004). Previous studies have demonstrated that enhanced early and late frontal SPNs are associated with CR, suggesting increased anticipation and preparation for implementing reappraisal (Moser et al. 2009, 2014; Thiruchselvam et al. 2011; Qi et al. 2017). Therefore, the SPN serves as a valuable complement to LPP in understanding the temporal dynamics of CR.

Previous research has found that individual factors such as alexithymia (Pollatos and Gramann 2012) modulate the LPP amplitude induced by CR. Similar to alexithymia, NED refers to the ability to discern subtle differences among similar emotions and has been linked with CR in behavioral outcomes (Barrett et al. 2001; Kalokerinos et al. 2019). Moreover, individual differences in emotion differentiation modulate LPP amplitudes (540–570 ms) during affective picture processing (Lee, Lindquist, and Nam 2017). Thus, further investigations are required to ascertain how NED modulates the ERPs elicited by CR.

Research on neural oscillations found that CR was associated with increased frontal θ oscillation power (FZ, 3.5–8 Hz), with sources in the left middle/inferior frontal gyrus and anterior dorsal cingulate cortex (Ertl et al. 2013; Zouaoui et al. 2023). In addition, CR has been linked to increased parietal gamma oscillatory power (P3/P4, 30–55 Hz, Kang et al. 2012). Based on the findings of LPP research, neural oscillation rhythms during the CR process can also be modulated by individual differences in factors, such as depression (Parvaz et al. 2015) and so on. Given that NED reflects individual specificity in experiencing and representing emotions, it was found to modulate the alpha (8–12 Hz) and gamma power (30–50 Hz) neural oscillations during the processing of affective pictures. Emotion generation and emotion regulation are not separate processes but interconnected aspects of emotional functioning (Gross and Barrett 2011). Therefore, understanding how NED shapes emotion processing may provide valuable insights into its role in emotion regulation. However, while previous research (Lee, Lindquist, and Nam 2017) establishes the influence of NED on emotion processing, its application to CR‐specific contexts remains underexplored.

Neural oscillations are interconnected rather than isolated, interacting, or connected, through a process known as cross‐frequency coupling. This process involves the cross‐modulation between the oscillation rhythms of different neuronal clusters, shedding light on their mutual influence, regulation, and mechanisms of information transmission and exchange (Zhang et al. 2017). Cross‐frequency coupling encompasses three types of coupling: phase–phase coupling (PPC), phase–amplitude coupling (PAC), and amplitude–amplitude coupling (AAC; Zhang et al. 2017). δ–β coupling is interpreted as the interaction between subcortical and cortical regions of the brain, particularly between the limbic system (e.g., amygdala) and the prefrontal cortex (Schutter and Knyazev 2012). Thus, it reflects top‐down emotions (Knyazev 2011; Schutter and Knyazev 2012) and stress regulation (Poppelaars et al. 2018; Morillas‐Romero et al. 2015). However, Poppelaars et al. (2021) did not confirm that this coupling was a neural indicator of stress regulation. The inconsistency between these research findings may be because individual difference factors modulate the δ–β coupling associated with emotion regulation. Poppelaars et al.'s (2021) studies employed healthy subjects, whereas Poppelaars et al. (2018) included participants with social anxiety. From a behavioral standpoint, NED is associated with CR (Barrett et al. 2001; Kalokerinos et al. 2019), necessitating further exploration of NED's role in modulating the cross‐frequency coupling during CR.

Finally, machine learning, a method that learns from existing data to uncover underlying patterns and applies these patterns to the analysis and prediction of unknown data, has become a pivotal tool in modern research (Senders et al. 2018). Notably, features that significantly contribute to the prediction results could be identified as potential biological markers. This approach has gained significant traction in recent years, particularly in predictive research on personalized diagnosis and treatment strategies for mental disorders. For example, neurophysiological markers such as LPPs (Stange et al. 2017), γ oscillations (Fitzgerald and Watson 2018), θ oscillations (Fernández‐Palleiro et al. 2020), and delta–beta coupling (Miskovic et al. 2011) have been identified in the diagnosis and treatment of depression and anxiety. However, research on multimodal indicators related to emotional regulation interventions remains limited. Building upon the exploration of the neurophysiological mechanisms through which NED influences CR from a multimodal perspective, it is imperative to further explore whether these neural indicators (e.g., LPP, θ or γ oscillation power, and PAC) can serve as markers of the influence of NED on CR.

In summary, the present study explored the impact of NED on CR and its neural underpinnings from a multimodal perspective. Using ERP, time–frequency, and cross‐frequency coupling analyses, we respectively explored the influence of NED on the LPP, neural oscillation rhythms, and PAC during CR. Furthermore, ERP analysis offers high‐temporal precision evidence from the temporal process perspective and time–frequency analyses capture a more comprehensive characterization of the EEG oscillatory features, providing a closer interpretation of neurophysiological mechanisms. Furthermore, cross‐frequency coupling can provide evidence from the perspective of transmission and interaction of information through brain oscillations. Finally, these indicators were synthesized using machine learning analysis to model NED's influence on CR, creating multi‐dimensional neural markers.

Based on existing research, our preliminary hypothesis was that individuals with HNED would present diminished parietal LPP amplitudes when using CR to alleviate negative emotions and exhibit higher frontal LPP amplitudes and SPNs during the CR process. These expectations align with previous research indicating that individuals with higher emotion differentiation are more effective at downregulating emotional responses (Kalokerinos et al. 2019; Kashdan, Barrett, and McKnight 2015), which may be reflected in reduced parietal LPP activity. Furthermore, the expected increase in frontal LPP amplitudes and SPNs corresponds to the cognitive demands of reappraisal, consistent with findings that individuals with higher emotion differentiation exhibit greater N2 amplitudes—typically linked to cognitive control—during emotion processing (Lee, Lindquist, and Nam 2017). Given the limited research on neural oscillations and cross‐frequency coupling in CR, we used a data‐driven approach to investigate their modulation by the NED.

2. Method

2.1. Participants

The initial participants of this study consisted of 220 healthy university or graduate students, aged 18–24 years, with normal or corrected vision and no history of mental disorders or brain injuries. Based on the measurement results of emotional differentiation levels, the top and bottom 10% of the participants were selected as subjects for the HNED and low NED (LNED) groups, respectively. Four participants were excluded from subsequent analyses due to excessive EEG artifacts (with fewer than 30 valid trials), and 2 withdrew due to discomfort, leaving a final sample of 38 participants (22 females, age: 22.68 ± 1.73 years; males, age: 22.62 ± 1.93 years, 19 individuals in each of the HNED and LNED groups). After removing trials with EEG artifacts, the average number of trials per condition was 36.34 for Neutral‐View (NV) (HNED: 36.32; LNED: 36.37), 36.87 for Unpleasant‐View (UV) (HNED: 36.84; LNED: 36.89), and 37.29 for Unpleasant‐Reappraisal (NR) (HNED: 37.32; LNED: 37.26). Each condition was maintained for a sufficient number of trials. Using G*power 3.1 (Faul et al. 2007), with an alpha level of 0.05 and statistical power of 0.80, we calculated an effect size of 0.52, confirming that the sample size was sufficient to detect the effects under investigation. Before starting the experiment, the experimental procedures and protocols were reviewed and approved by the institution's academic ethics committee, and informed consent was obtained from all participants. Upon completion, they received compensation of 100 CNY (approximately 14 USD).

2.2. Experimental Design and Materials

This experiment utilized a mixed experimental design structured as 2 (group: HNED and LNED) *3 (task: NV, UV, and UR). Here, group was a between‐subjects variable, and task was a within‐subject variable.

A total of 120 images (80 negative and 40 neutral; matched for content, e.g., figure or scenery; see Supporting Information 1) were selected from the International Affective Picture System (Lang, Bradley, and Cuthbert 2008). Negative images were split into two sets, A and B. Participants were categorized into two groups: one viewed Set A, and the other reappraised Set B, and vice versa. No significant difference was found between sets A and B in terms of valence (t ₇₈ = 0.49, p = 0.63) or arousal (t ₇₈ = −0.02, p = 0.98). However, both sets significantly differed from neutral images in valence (Set A: t ₇₈ = −18.66, p < 0.001, Cohen's d = −4.29; Set B: t ₇₈ = −17.05, p < 0.001, Cohen's d = −3.92) and arousal (Set A: t ₇₈ = 17.59, p < 0.001, Cohen's d = 4.04; Set B: t ₇₈ = 17.57, p < 0.001, Cohen's d = 4.04). The NED experimental materials, also derived from the International Affective Picture System, included 10 additional images. Accompanying these images, 10 negative emotion words were chosen: fear, anxiety, anger, disgust, depression, sadness, loneliness, shame, frustration, and guilt (Erbas et al. 2013).

2.3. Experimental Procedure

2.3.1. Overview Procedure

Firstly, participants were recruited by the posting advertisements, and those recruited underwent measurements of NED levels. Based on these measurements, the top and bottom 10% of the participants were selected for subsequent experiments. In the formal experiments, the participants initially completed demographic and control variable questionnaires upon arrival at the laboratory. Subsequently, they undertook a CR task while their EEG activities were recorded.

2.3.2. NED Procedure

NED was assessed using the Photo Emotion Differentiation Task (Erbas et al. 2013) in E‐Prime 2.0 (Psychology Software Tools Inc., Sharpsburg, PA, USA). Throughout the task, participants viewed 10 emotional images, each followed by 10 emotion category words presented in a random order. When the emotional photos were presented, the participants were asked to rate the extent to which they felt or experienced each emotion category using a 7‐point Likert scale ranging from 0 (not at all) to 6 (very much). The instructions were as follows: “When viewing this emotional photo, please rate the extent to which you feel/experience [emotion category]?” After the experiment, a manipulation check was conducted. Participants were required to identify the emotion words used in the NED measurement from the three presented words, and their attention was assessed based on selection accuracy. NED ability was quantified using the intraclass correlation of emotion ratings across images, with lower coefficients indicating higher NED ability. Considering the non‐normal distribution of the intraclass correlation coefficient, it was converted to Fisher's Z scores for further analysis. The final NED score was obtained by subtracting the Fisher's Z score from 1, where higher scores reflect greater NED ability.

2.3.3. Cognitive Reappraisal Procedure

The experimental procedure for CR, based on Baur et al. (2015), is depicted in Figure 1. Participants followed the cue, either viewing or reappraising the image and then rated the emotional valence and arousal on a 9‐point scale (1—very weak or very unpleasant and 9—very strong or very pleasant). In the viewing condition, participants were instructed to simply focus their attention on observing the image, maintaining their initial reaction to what the image appeared to be, without any need to alter their emotional response. In the reappraisal condition, participants were asked to diminish their negative emotions by reinterpreting the image from a new perspective. To familiarize participants with the experimental procedures, practical trials involving viewing and reappraisal were conducted prior to the main experiment. Following the practice trials, participants were asked how they engaged in reappraisal to ensure that they correctly understood and utilized the CR strategy. The experiment comprised three trial types—NV, UV, and UR—each containing 40 trials, for a total of 120. These trials were categorized into three blocks, with breaks provided after each block. The entire experimental procedure lasted for a total of 40 min.

FIGURE 1.

Schematic representation of CR procedure.

2.4. EEG Data Recording and Analysis

2.4.1. EEG Data Recording and Preprocessing

EEG data were collected using a 64‐channel NeuroScan 4.5 system (NeuroScan Inc., El Paso, Texas, USA), with electrode placement based on the international 10–20 system. The reference electrode was positioned at the left mastoid (M1), and the ground electrode was positioned between FPZ and FZ during data acquisition. Electrooculograms were situated on the outer sides of both eyes (horizontal) and above and below the left eye (vertical). The recording protocol employed a DC to 100 Hz band‐pass filter and a 1000 Hz sampling rate, maintaining scalp impedance below 5 kΩ. During preprocessing, the data were re‐referenced to the bilateral mastoids and underwent band‐pass filtering. ERP analysis used a 0.01–24 Hz filter, while time–frequency and coupling analyses utilized a 0.01–100 Hz filter. Upon preprocessing and epoching the data, we visually inspected it to remove epochs with gross artifacts and then ran the ICA. EEG artifacts, such as ocular and muscle electrical activity, were eliminated through a combination of independent component analysis and visual inspection. After removing trials with artifacts, the remaining trials for each task were deemed adequate (NV: 36.34 ± 2.85; UV: 36.86 ± 2.89; UR: 37.29 ± 2.63). There were no significant differences in the number of trials across the three task types (F _2,72 = 2.05, p = 0.14), and no significant differences in the number of trials between the HNED and LNED groups under the task types (F _2,72 = 0.08, p = 0.99).

2.4.2. ERP Analysis

According to previous studies (Moser et al. 2014; Thiruchselvam et al. 2011), SPNs were identified and quantified at the fronto‐central electrodes (FZ and FCZ) during two‐time windows: early SPN (300–2000 ms) and late SPN (3800–4500 ms, corresponding to 700 ms immediately preceding picture onset). In the LPP analysis, the data were segmented into epochs defined from −200 ms before stimulus (baseline) to 3000 ms after stimulus. For the frontal LPP, we used the average amplitude at FZ and FCZ electrodes within an 850–1250 ms post‐image stimulus onset window based on the grand average graph and prior research (Moser et al. 2014). As for the parietal LPP, we calculated the mean amplitude from five surrounding electrodes in the PZ (CPZ, P1, PZ, P2, and POZ). Based on the two phases of cognitive reappraisal—emotional processing and emotional regulation—we selected the following time windows for CR modulation: 0.4–0.6, 0.6–0.8, 0.8–1, 1–1.2, 1.2–1.7, and 1.7–3 s. A 0.2 s segmentation was applied to maintain consistency with previous studies (Foti and Hajcak 2008; Thiruchselvam et al. 2011) and to facilitate the comparison of results.

2.4.3. Time–Frequency Analysis

Time–frequency analysis was performed using a fixed 200 ms Hanning window for the short‐time Fourier transform. This analysis computed the oscillation power at each time–frequency point (0–6000 ms, 1–80 Hz at 1 Hz intervals) for each participant. Each time–frequency point obtained was subjected to point‐by‐point baseline correction (as shown in Formula (1)). Here, P (t, f) denotes the oscillation power at each time–frequency point (t, f), and R(f) signifies the average oscillation power during the baseline period from −800 to −200 ms. The oscillation power for each trial was first calculated and then averaged to derive the mean power for each experimental task. Increases in neural oscillations relative to the baseline state are referred to as event‐related synchronization (ERS), whereas decreases are called event‐related desynchronization (ERD).

$$\mathrm{ER}\left(\mathrm{t},\mathrm{f}\right)=\frac{\mathrm{P}\left(\mathrm{t},\mathrm{f}\right)–\mathrm{R}\left(\mathrm{f}\right)}{\mathrm{R}\left(\mathrm{f}\right)}$$ (1)

Given the limited research on NED time–frequency analysis, our study employed a data‐driven approach using variance analysis and nonparametric permutation testing (Maris and Oostenveld 2007) to identify time–frequency regions of interest (ROI). The process was as follows: Initially, a 2 (group: HNED and LNED) * 3 (task: NV, UV, and UR) repeated measures analysis of variance was conducted for each time–frequency point, generating three time–frequency maps. These maps illustrated the main effects of group, task, and their interactions, with our study primarily focusing on the latter.

The selected ROIs for further analysis met these criteria: (1) They exhibited a significant interaction between group and task (p ≤ 0.01) and included at least 300 continuous time–frequency points. (2) To address the multiple‐comparison problem in point‐by‐point variance analysis, cluster‐based multiple‐comparison corrections were employed. The cluster's effect intensity was represented by the sum of F‐values (∑_F) for the time–frequency points within each cluster. For each participant, labels were randomized, and a permutation test was performed 5000 times, each generating a cluster‐level statistic ${{\sum }_F}^{\prime }\left(\mathrm{i}\right)$ with each iteration. (3) Based on the distribution of ${{\sum }_F}^{\prime }\left(\mathrm{i}\right)$ from the permutation tests, a two‐tailed normal distribution test was used to determine the p‐value of ∑_F from step (2). The final ROI for time–frequency analysis was the cluster with a p‐value ≤ 0.01 and the highest ∑_F value in the permutation test.

2.4.4. Cross‐Frequency Coupling Analysis

PAC, the modulation of the high‐frequency oscillation amplitude by the low‐frequency oscillations phase, is vital for cognitive control and working memory (Riddle, Mcferren, and Frohlich 2021). Most emotional regulation research employed PAC as the cross‐frequency coupling indicator (Popov et al. 2012; Poppelaars et al. 2018; Schutter and Knyazev 2012); thus, this study analyzed the PAC by calculating the modulation index (MI).

The computation steps were as follows: (1) preprocessed data were band‐pass filtered to produce low‐frequency (1–29 Hz) or high‐frequency (7–80 Hz) signals. (2) The Hilbert transform was applied to obtain the instantaneous phase of low‐frequency bands and the amplitude of high‐frequency bands, respectively. (3) The low‐frequency phase was divided into 18 bins of 20° each, and the mean high‐frequency amplitude was computed for each phase bin. (4) Normalization was done by dividing each phase bin's amplitude by the total amplitude across the 18 intervals, resulting in a distribution of P(j). (5) The Kullbeck–Leiber (K–L) distance was computed to evaluate the average amplitude distribution in each phase bin across the 18 phase intervals, as shown in Formula (2). Here, P represents the average amplitude distribution, Q denotes a uniform distribution, and N denotes the phase bin count. The MI is the normalized K–L distance, as depicted in Formula (3), where N signifies the number of phase bins. (6) To avoid potential extremes in the PAC resulting from amplitude or phase value fluctuations (Cohen 2014), the time segment of the high‐frequency signal was split into two, and permutation tests were performed 100 times. The distribution resulting from the permutation test was then acquired, and the PAC was converted into a Z‐score based on this distribution (i.e., PAC_Z), which served as the final PAC value.

$$D\left(P,Q\right)=\sum _{j=1}^{N}P\left(j\right).\log \left(\frac{P\left(j\right)}{Q\left(j\right)}\right)$$ (2)

$$\mathit{MI}=\frac{D\left(P,Q\right)}{\log \left(N\right)}$$ (3)

For PAC analysis, a data‐driven method similar to the time–frequency analysis was used to identify the ROIs. The selected ROIs for further analysis must meet these criteria: a significant interaction between group and task (p ≤ 0.01), and at least 10 continuous coupling regions.

2.5. Machine Learning Analysis

Given that EEG signals are characterized by high‐dimensional feature vectors and are prone to contamination from various noises and artifacts, and considering that the support vector machine (SVM) is adept at handling high‐dimensional data—even with a small dataset—and demonstrates strong classification capabilities in multimodal data (Cervantes et al. 2020; Sha'abani et al. 2020), we employed SVM for our analysis. An SVM with a radial basis function kernel was used for the classification task. The feature variables included the arousal ratings, parietal LPP indicators in the UR condition, time–frequency, and PAC indicators from the significant interaction ROIs. These features predicted the classifications for HNED and LNED. Participants were randomly assigned to the training and testing sets in a 4:1 ratio. The model was trained on the training data using the fitcsvm function in MATLAB, with the data standardized and the kernel scale set to “auto.” The hyperparameters of the SVM were optimized to enhance the model's ability to generalize new data. Cross‐validation using a fivefold partitioning strategy was used to assess the model's performance. This was achieved using a crossval function on the trained SVM model. Classifier performance was evaluated using precision, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC).

2.6. Statistical Analysis

Statistical analyses were performed using SPSS 19.0 (IBM, Somers, USA). Two‐factor mixed‐design ANOVAs were conducted on self‐reported ratings (i.e., valence and arousal), along with the frontal LPP amplitude, ROIs for time–frequency analysis, and PAC, with group (HNED and LNED) as the between‐subject variable and task (NV, UV, and UR) as the within‐subject variable. A three‐factor mixed‐design ANOVA was conducted on the parietal LPP and SPN amplitudes, with the added factor of time window as the within‐subject variable. For all variance analyses, the significance level was set at p < 0.05. Descriptive data are presented as Mean ± Standard Deviation. Greenhouse–Geisser corrections were applied if sphericity was violated. The Bonferroni correction was used for multiple comparisons.

3. Result

3.1. Manipulation Check

The valence ratings results revealed a significant main effect of task, F (2, 72) = 99.73, p < 0.001, ${\mathrm{\eta }}_p^2$ = 0.74. The UR condition (4.25 ± 0.16) had higher valence ratings than the UV condition (3.47 ± 0.09, t = 5.52, p < 0.001, Cohen's d = 1.03), while the latter had lower ratings than the NV condition (5.35 ± 0.06, t = −20.23, p < 0.001, Cohen's d = −4.43). Similarly, arousal ratings displayed a significant main effect of task, F (2, 72) = 12.93, p < 0.001, ${\mathrm{\eta }}_p^2$ = 0.26. Arousal was lower in the UR (5.47 ± 0.14) compared to the UV condition (5.75 ± 0.14, t = −3.35, p = 0.006, Cohen's d = −0.34), but higher in the UV compared to the NV condition (5.11 ± 0.12, t = 4.32, p < 0.001, Cohen's d = 0.85).

The parietal LPP amplitude further revealed a significant main effect of task, F (2, 72) = 5.40, p = 0.007, ${\mathrm{\eta }}_p^2$ = 0.13. The LPP amplitude was notably higher in the UV (2.41 ± 0.57) compared to the NV condition (0.85 ± 0.67, t = 2.99, p = 0.015, Cohen's d = 0.42). These findings confirmed the effectiveness of our experimental manipulation.

3.2. Behavioral Results

No interaction was observed between group and task for either valence ratings (F _2,72 = 0.24, p = 0.73) or arousal ratings (F _2,72 = 1.70, p = 0.20).

We then examined group differences in both regulatory (difference between the UV and UR conditions) and emotional effects (difference between the UV and NV conditions) for valence and arousal ratings. No significant group difference was found in either the regulatory effect (t ₃₆ = −0.403, p = 0.69) or the emotional effect (t ₃₆ = −0.362, p = 0.72) of valence ratings.

However, a significant group difference (t ₃₆ = −2.20, p = 0.034, Cohen's d = 0.73) in the regulatory effect for arousal was found between the HNED (0.47 ± 0.46) and LNED groups (0.10 ± 0.57, see Supplementary Figure 1). This implies that individuals with HNED were more effective in reducing their arousal. However, the emotional effects revealed no significant group differences (t ₃₆ = 1.48, p = 0.15).

3.3. ERP Results

In the parietal LPP amplitude, a significant three‐way interaction was observed among group, task, and time window, F (10, 360) = 3.58, p = 0.004, ${\mathrm{\eta }}_p^2$ = 0.09. Simple effects analysis revealed that the HNED group demonstrated a regulation effect within the T2 (0.6–0.8 s; t = −2.80, p = 0.024, Cohen's d = −0.36), T3 (0.8–1 s; t = −2.96, p = 0.016, Cohen's d = −0.44), and T4 (1–1.2 s; t = −3.73, p = 0.002, Cohen's d = −0.63) time windows. This effect was characterized by a lower LPP amplitude under UR conditions than under UV conditions (Figure 2A,C). Conversely, the LNED group did not exhibit this regulation effect (Figure 2B,D). Moreover, the HNED group displayed an emotional effect in the T1 (0.4–0.6 s; t = 2.99, p = 0.015, Cohen's d = 0.43) and T2 (0.6–0.8 s; t = 3.78, p = 0.002, Cohen's d = 0.58) time windows, characterized by a higher LPP amplitude in the UV compared to the NV condition. However, for the LNED group, this emotional effect was only evident at T2 (0.6–0.8 s; t = 3.19, p = 0.009, Cohen's d = 0.46).

FIGURE 2.

ERP results. Figures (A and B) illustrate the LPP amplitude changes across the three experimental tasks for the high and low negative emotion differentiation groups, respectively. Significant differences are observed between the Unpleasant‐Reappraisal and Unpleasant‐View conditions. Figures (C and D) display brain topographies for the high and low negative emotion differentiation groups during the time windows with significant differences between experimental tasks. Figure (E) shows a scatter plot correlating arousal ratings with average LPP amplitude under the Unpleasant‐Reappraisal condition, while Figure (F) shows the same for the Unpleasant‐View condition. Figure (G) presents the SPNs across tasks. Shaded areas highlight significant differences between the Unpleasant‐Reappraisal and Unpleasant‐View conditions. HNED, high negative emotion differentiation; LNED, low negative emotion differentiation.

Regarding the LPP regulatory or emotion effect, we further conducted the group difference analysis. Significant differences were observed between the HNED and LNED groups for regulatory effect across several time windows, including T2 (0.6–0.8 s; t ₃₆ = 2.08, p = 0.045, Cohen's d = 0.69), T3 (0.8–1.0 s; t ₃₆ = 2.21, p = 0.034, Cohen's d = 0.74), T4 (1–1.2 s; t ₃₆ = 2.82, p = 0.008, Cohen's d = 0.94), T5 (1.2–1.7 s; t ₃₆ = 2.35, p = 0.025, Cohen's d = 0.78), and T6 (1.7–3 s; t ₃₆ = 1.90, p = 0.066, Cohen's d = 0.63, marginally significant). In these windows, the HNED group exhibited a greater regulatory effect than the LNED group. However, no significant differences were noted in the emotional effects between groups (ps > 0.1).

Correlation analysis of LPP amplitude under different tasks and ratings revealed that the LPP amplitude under the condition of UR (r = 0.36, p = 0.027, Figure 2E) or UV (r = 0.36, p = 0.026, Figure 2F) was respectively significantly positively correlated with arousal ratings in these tasks. These results suggest that the parietal LPP amplitude reflects emotional arousal, providing a significant neural marker for the intensity of emotional responses. The significance of this correlation lies in its potential to bridge the gap between neural activity and subjective emotional experience, highlighting the LPP's role as a measurable indicator of emotional intensity.

A significant interaction between group and task was also found, F (2, 72) = 3.86, p = 0.026, ${\mathrm{\eta }}_p^2$ = 0.10. Simple effect analysis demonstrated that HNED individuals presented a significantly lower LPP amplitude (t = −2.88, p = 0.020, Cohen's d = −0.41) in the UR condition (0.52 ± 0.98) compared to the UV condition (2.25 ± 0.81). However, no such differences were noted in LNED individuals. We also identified significant primary effects of time window and interactions between time window and group. However, these were not the primary focus of this study and will not be discussed further. Descriptive statistics for the parietal LPP amplitude are presented in Table 1.

TABLE 1.

Descriptive statistical results of the parietal late positive potential amplitude.

Time window	HNED			LNED
	UR (M ± SD)	UV (M ± SD)	NV (M ± SD)	UR (M ± SD)	UV (M ± SD)	NV (M ± SD)
T1	5.30 (6.02)	6.06 (5.46)	3.65 (5.98)	3.91 (4.00)	3.78 (3.80)	2.09 (3.81)
T2	2.26 (5.23)	3.95 (4.35)	1.31 (5.15)	3.80 (3.30)	3.72 (2.96)	1.49 (3.55)
T3	0.66 (4.69)	2.52 (4.04)	1.17 (4.69)	3.24 (4.22)	3.14 (2.92)	1.98 (2.79)
T4	−0.98 (4.69)	1.81 (4.37)	0.14 (4.86)	2.50 (4.67)	2.31 (3.19)	0.63 (2.98)
T5	−1.71 (4.81)	0.01 (4.31)	−0.91 (5.80)	2.27 (4.90)	1.43 (3.64)	−0.33 (2.97)
T6	−2.42 (4.75)	−0.83 (4.51)	−0.21 (6.61)	1.87 (4.38)	1.01 (4.01)	−0.78 (4.26)

Note: The amplitude value of the late positive potential was averaged across five electrode sites (CPZ, P1, PZ, P2, and POZ). T1, 0.4–0.6 s. T2, 0.6–0.8 s. T3, 0.8–1 s. T4, 1–1.2 s. T5, 1.2–1.7 s. T6, 1.7 –3 s. HNED, high negative emotion differentiation. LNED, low negative emotion differentiation.

Abbreviations: NV, Neutral‐View; UR, Unpleasant‐Reappraisal; UV, Unpleasant‐View.

For SPNs, no significant three‐way interaction was observed between group, task, and time window; F (2, 72) = 0.55, p = 0.579. However, there was a significant two‐way interaction between task and time window, F (2, 72) = 3.57, p = 0.033, ${\mathrm{\eta }}_p^2$ = 0.09. Simple effects analysis revealed that for the late SPN, UR (−1.874 ± 0.72) presented larger SPN amplitudes than UV (−0.397 ± 0.67, t = −3.11, p = 0.01, Cohen's d = −0.33, Figure 2G), while no difference was found between tasks for early SPN (t = 0.53, p = 1.00). Finally, no significant effects were identified in the frontal LPP results (ps ≥ 0.099).

3.4. Time–Frequency Analysis

Exploratory data analysis identified four ROIs with significant interaction effects, primarily characterized by high‐ or low‐frequency γ waves across mid‐late time windows (see Figure 3A–D). Repeated measures variance analysis of these ROIs' average oscillations found a significant interaction between group and task. However, no significant main effects were observed for either task or group (Table 2).

FIGURE 3.

ROIs with Interaction Effect: Task × Group (Time–frequency Analysis). UR, Unpleasant‐Reappraisal; UV, Unpleasant‐View; NV, Neutral‐View. In Figures (A–D), significant interaction time–frequency points from the point‐to‐point variance analysis are illustrated. Rectangular boxes emphasize points selected for further analysis (take one of the electrode sites as an example). Figures (E–H) present bar charts of the mean power value relative to the baseline at these significant time–frequency points in the subsequent analysis. HNED, high negative emotion differentiation. LNED, low negative emotion differentiation. **p < 0.01, *p < 0.05. The error bars represent standard deviations.

TABLE 2.

Interaction: Task × Group (Time–frequency Analysis).

ROI	Time windows (ms)	Frequency (Hz)	Electrode site	Interaction effect	Main effect
ROI1	2909–3406	34–45	C3 P7 P5 P3 P1 PZ P8 PO3 POZ PO4 PO6 PO8 O1 OZ O2	F (2, 72) = 13.01, p < 0.001, ${\mathit{\eta }}_{\mathbf{p}}^{\mathbf{2}}$ = 0.27	task: F 2,72 = 1.13, p = 0.310 group: F 1,36 = 3.86, p = 0.057
ROI2	2830–3006	64–80	FC3 C1 PO5 PO7	F (2, 72) = 11.69, p < 0.001, ${\mathit{\eta }}_{\mathbf{p}}^{\mathbf{2}}$ = 0.25	task: F 2,72 = 1.45, p = 0.240 group: F 1,36 = 3.47, p = 0.071
ROI3	3333–3454	67–75	FCZ CZ CPZ CP1	F (2, 72) = 9.53, p < 0.001, ${\mathit{\eta }}_{\mathbf{p}}^{\mathbf{2}}$ = 0.21	task: F 2,72 = 2.59, p = 0.080 group: F 1,36 = 0.34, p = 0.565
ROI4	4540–4609	63–71	F1 F2 FZ	F (2, 72) = 8.38, p = 0.001, ${\mathit{\eta }}_{\mathbf{p}}^{\mathbf{2}}$ = 0.19	task: F 2,72 = 0.26, p = 0.770 group: F 1,36 = 0.95, p = 0.340

Abbreviation: ROI, regions of interest.

The simple effects analysis consistently revealed that the HNED group showed stronger gamma ERS compared to the LNED group across the UR condition. Specifically, in ROI 1, the HNED group exhibited significantly stronger gamma ERS than the LNED group (t = 2.82, p = 0.011, Cohen's d = 0.90, see Figure 3E), with similar patterns observed in ROI 2 (t = 2.13, p = 0.029, Cohen's d = 0.75, see Figure 3F), ROI 3 (t = 2.80, p = 0.006, Cohen's d = 0.98, Figure 3G), and ROI 4 (t = 2.50, p = 0.014, Cohen's d = 0.86, Figure 3H). In contrast, the UV condition showed no significant differences between groups for ROIs 1 (t = −0.54, p = 0.568; see Figure 3E) and 2 (t = −0.60, p = 0.531; see Figure 3F), while ROI 3 (t = −2.20, p = 0.042, Cohen's d = −0.71) and ROI 4 (t = −2.78, p = 0.010, Cohen's d = −0.91; see Figure 3H) demonstrated weaker ERS in the HNED group.

The grand time–frequency representation is illustrated in Supplementary Figure 2. Exploratory analysis also identified ROIs with significant task main effects. These findings, which were not the primary focus of this study, are detailed in Supplementary Table 1 and Supplementary Figure 3.

3.5. Cross‐Frequency Coupling Results

An exploratory analysis of cross‐frequency coupling revealed four ROIs with significant interactions, particularly involving α–β, α–γ, and θ–γ couplings. Specifically, α–β coupling has been linked to maintaining working memory (Daume et al. 2017), while θ–γ coupling is associated with encoding and storing working memory information (Alekseichuk et al. 2016). Additionally, α–γ coupling plays a role in gating external information processing and inhibiting irrelevant inputs (Bonnefond and Jensen 2015). Repeated‐measures variance analysis of these ROIs revealed significant interactions between group and task but no significant main effects for task or group (Table 3).

TABLE 3.

Interaction: task × group (cross‐frequency coupling).

ROI	Time window (ms)	Coupling	Electrode site	Interaction effect	Main effect
ROI 1	400–1700	α–β	P8	F (2, 72) = 11.21, p < 0.001, ${\mathbf{\eta }}_{\mathit{p}}^{\mathbf{2}}$ = 0.24	task: F 2,72 = 6.29, p = 0.003 group: F 1,36 = 6.22, p = 0.017
ROI 2	1700–3000	α‐γ	FPZ	F (2, 72) = 9.31, p < 0.001, ${\mathbf{\eta }}_{\mathit{p}}^{\mathbf{2}}$ = 0.21	task: F 2,72 = 0.21, p = 0.810 group: F 1,36 = 0.82, p = 0.372
ROI 3	3000–6000	θ–γ	P8	F (2, 72) = 7.30, p = 0.001, ${\mathit{\eta }}_{\mathbf{p}}^{\mathbf{2}}$ = 0.17	task: F 2,72 = 2.68, p = 0.075 group: F 1,36 = 0.01, p = 0.971
ROI 4	3000–6000	δ–γ	FPZ	F (2, 72) = 8.84, p < 0.001, ${\mathit{\eta }}_{\mathbf{p}}^{\mathbf{2}}$ = 0.20	task: F 2,72 = 0.80, p = 0.456 group: F 1,36 = 0.16, p = 0.695

Note: Coupling refers to the coupling between the phase of low‐frequency oscillations and the amplitude of high‐frequency oscillations. For instance, α–β coupling indicates the coupling between the phase of α and the amplitude of β. In the ROI 1 with significant interactions, the task main effect was found. Since this ROI is also an area of interest with a significant task main effect, refer to supplementary Figure 1 for details.

Abbreviation: ROI, region of interest.

In the simple effect analysis, the following results were observed:

The HNED group exhibited stronger α‐β coupling than the LNED group in the UR condition (t = 4.14, p < 0.001, Cohen's d = 1.38), with no significant difference in the UV condition (t = 1.22, p = 0.229, Figure 4A).

Similarly, the HNED group showed higher α‐γ coupling than the LNED group in the UR condition (t = 4.13, p < 0.001, Cohen's d = 1.38). Again, no significant differences were found in the UV condition (t = −0.94, p < 0.354, Figure 4B).

The HNED group displayed stronger θ‐γ amplitude coupling than the LNED group during UR (t = 2.35, p = 0.024, Cohen's d = 0.78), with no significant difference in UV (t = 0.66, p = 0.512, Figure 4C).

No significant differences were found between the groups in either of the experimental tasks.

FIGURE 4.

ROIs with Interaction Effect: Task × Group (Cross‐frequency Coupling). Figures (A–C) respectively illustrate the coupling of α phase and β amplitude (0.4–1.7 s), α phase and γ amplitude (1.7–3 s), and θ phase and γ amplitude (3–6 s). Black‐outlined regions denote areas with significant coupling differences between high and low negative emotion differentiation groups when reappraising negative images. HNED, high negative emotion differentiation. LNED, low negative emotion differentiation. UR, Unpleasant‐Reappraisal. UV, Unpleasant‐View. NV, Neutral‐View.

A significant positive correlation was found between α–β coupling at the P8 electrode under the UR condition and arousal ratings in regulation effect (r = 0.37, p = 0.024), suggesting its link with CR effects.

Although the exploratory analysis revealed ROIs with significant task main effects, this was not the central focus of the study. Please refer to Supplementary Table 2 and Supplementary Figure 4 for further details.

3.5.1. Control Analysis

As the power of the frequency bands can directly affect their modulation range, changes in the PAC may be due to changes in power (Aru et al. 2015). To exclude the possibility that the observed PAC interactions resulted from power spectrum shifts, we conducted a control analysis correlating high‐frequency band power values with PAC for each group under all conditions (Voloh et al. 2015). The results suggested no association between power shifts and PAC differences between HNED and LNED groups in the UR condition or UV condition (ps ≥ 0.09). Likewise, no link was found between power shifts and the regulatory (ps ≥ 1.00) or emotional effects (ps ≥ 0.074) identified in the task main effects. Refer to Supplementary Table 3 and Supplementary Figure 5 for detailed correlation data and power spectrum changes.

3.6. Machine Learning Results

The machine learning findings showed that multimodal indicators, which fully utilize data from various modalities, demonstrated superior classification performance compared to unimodal indicators (see Table 4). From the perspective of the AUC, the prediction result of the multimodal indicators was the highest, at 0.955 (see Figure 5).

TABLE 4.

Predictive result of machine learning.

Indicators	Accuracy	Sensitivity	Specificity	AUC
Arousal ratings	0.63	0.64	0.63	0.63
LPP	0.66	0.69	0.65	0.61
TF	0.63	0.63	0.64	0.70
PAC	0.77	0.75	0.79	0.88
Arousal ratings +LPP + TF + PAC	0.91	0.88	1	0.99

Note: LPP, indicators of late positive potentials, including parietal LPP in the Unpleasant–Reappraisal condition, were observed from T2–T6 time windows. TF, time–frequency indicators, encompass ROIs from significant interactions in the time–frequency analysis. PAC, phase‐amplitude coupling indicators, including ROIs of significant interactions in phase‐amplitude coupling analysis. Behavioral indicators refer to arousal ratings in regulatory effects that present a significant group difference.

FIGURE 5.

ROC for classification.

4. Discussion

This study utilized various EEG analysis methods to investigate the impact of individual differences in NED on cognitive reappraisal and the underlying neural mechanisms. Our findings revealed that HNED individuals exhibited a significant reduction in LPP amplitudes when reappraising negative images, implying more efficient CR. They also showed heightened γ ERS in posterior and midline electrode sites during CR, with increased frontal α–γ and parietal θ–γ coupling. All effects identified were of medium size or larger.

4.1. NED's Modulation of ERP Amplitude in Cognitive Reappraisal

As anticipated, this study found a significant regulatory effect of HNED during the early‐to‐mid time window, characterized by a significant reduction in LPP amplitude in the UR condition compared with the UV condition. However, this effect was not observed in the LNED group. These findings suggest that NED can enhance the effects of CR, corroborating the findings of a previous questionnaire‐based study (Kalokerinos et al. 2019).

Our findings also align with those of Wang et al. (2020), who found increased LPP amplitudes during positive reappraisal in individuals with high positive emotion differentiation. However, this modulation occurred later (after 3 s), whereas our study found NED's modulation of LPP during CR in the early to middle stages (before 3 s). This supports Schwarz's (2012) information theory model of emotions, which posits that positive emotions convey feelings of safety and satisfaction, whereas negative emotions signal threats and discomfort in the environment. Through precise emotional differentiation, individuals with HNED rapidly identify these threatening signals and promptly initiate emotion regulation, leading to an observed modulatory effect in an earlier time window.

For the late SPN, we found that CR elicited a greater frontal SPN than the viewing condition, which aligns with previous findings (Moser et al. 2009, 2014; Thiruchselvam et al. 2011; Qi et al. 2017). The late SPN reflects preparation for and anticipation of the upcoming stimulus (van Boxtel and Böcker 2004). This result suggests a heightened anticipation of and preparation for implementing reappraisal. However, no significant differences in the SPN were observed between the emotion differentiation groups across the tasks. One possible explanation is that individuals with high emotion differentiation excel at extracting more emotional information, which facilitates emotion regulation. This advantage may only manifest after the emotional stimulus appears and may not be evident during the cue phase. Consequently, these individuals may not exhibit greater readiness or anticipation before implementing reappraisal. Additionally, note that our sample size was relatively small, so these findings should be interpreted with caution. Previous studies have also found that SPN related to CR is not modulated by individual differences, such as trait reappraisal (Moser et al. 2014) or trait anxiety (Qi et al. 2017). This suggests that reappraisal‐related SPN may not be sensitive to individual differences, which warrants further exploration in future research.

4.2. NED's Modulation of Neural Oscillation in Cognitive Reappraisal

Although ERP analysis offers high temporal precision and accuracy, it overlooks certain aspects of the EEG signal, such as information from non‐phase‐locked oscillations (Cohen 2014). To complement this, time–frequency analyses were employed to provide a more comprehensive characterization of the EEG's oscillatory features, capturing the intricacies of frequency, power, and phase over time (Morales and Bowers 2022). Moreover, certain neural intervention techniques, such as repetitive transcranial magnetic stimulation (rTMS) and transcranial alternating current stimulation (tACS), require this frequency band information to conduct interventions (Liu et al. 2018; Rosa and Lisanby 2012). Time–frequency analysis also offers empirical evidence for the application of rTMS or tACS in CR interventions. Thus, we further conducted a time–frequency analysis.

Time–frequency analysis demonstrated that HNED individuals showed enhanced γ ERS during the reappraisal of negative images compared to LNED individuals. γ oscillations, typically linked with cognitive functions like selective attention, stimulus selection, memory formation, and feature integration (Knyazev 2007), also have a strong association with attention in emotional regulation processes (Kaiser and Lutzenberger 2003). This implies greater allocation of attentional resources and integrative functions during CR by individuals with HNED.

This can be explained by the constructionist theory of emotion (Barrett, Wilson‐Mendenhall, and Barsalou 2014), which posits that emotional experiences result from a process termed “situated conceptualization.” This process refers to individuals using emotional concept knowledge—including scenarios that trigger emotions, physiological responses associated with emotions, and rules of emotional expression—to categorize their present emotional states, thereby generating emotional experiences. Within this framework, CR can be perceived as a transformation of these situated conceptualizations (Barrett, Wilson‐Mendenhall, and Barsalou 2014). Individuals with high emotional differentiation are capable of drawing from a rich repository of emotional concept knowledge to conceptualize situations or transform situated conceptualizations. This complex process requires more attentional resources, stimulus selection, and meaning integration. Consequently, those with HNED display elevated γ oscillations during the CR process. In contrast, individuals with LNED may struggle with the situated conceptualization transformation owing to a deficit in emotional concept knowledge, leading to diminished γ oscillations.

However, γ ERS pattern changes under viewing of negative image conditions: HNED individuals exhibited lower γ ERS than LNED individuals. This finding aligns with a study by Lee, Lindquist, and Nam (2017) that reported lower gamma ERS during emotional processing in individuals with high emotional differentiation. These individuals, who possess a broad spectrum of emotional concept knowledge (Barrett, Wilson‐Mendenhall, and Barsalou 2014), can assign emotional significance to stimuli without additional cognitive effort or expenditure. In contrast, individuals with LNED, who are less equipped with emotional concept knowledge, require more cognitive resources for emotional processing, resulting in heightened γ synchronization. While under CR conditions, γ synchronization increases in HNED individuals. The elevated gamma ERS during CR is attributed to the greater cognitive demands of actively reappraising emotions compared to passive viewing. Although HNED individuals may process emotions more efficiently in passive contexts, CR necessitates additional cognitive resources to regulate and reinterpret emotional stimuli, leading to increased synchronization. This finding supports the “neural efficiency” account (Grabner, Neubauer, and Stern 2006; Neubauer and Fink 2009), which suggests that, for simple tasks (e.g., viewing negative images), high‐efficiency individuals (e.g., HNED) can complete tasks using fewer neural resources, as indicated by lower EEG activity. However, as task difficulty increases (e.g., reappraising negative images), complex tasks demand more cognitive resources, leading even high‐efficiency individuals to exhibit greater neural activity due to the need for enhanced cognitive control and resource allocation.

In contrast to previous studies reporting a significant increase in frontal θ (3–6 Hz) oscillations during CR (Ertl et al. 2013; Domic‐Siede, Sánchez‐Corzo, and Guzmán‐González 2024), this effect was not observed in this study. Both studies used a hypothesis‐driven approach, whereas we employed an exploratory data analysis method, which may account for the discrepancy in the results. Although we did not find an enhancement in θ oscillations, we observed an increase in the coupling of θ phase and γ amplitude. γ is typically linked with the allocation of attentional resources in emotional regulation processes (Kaiser and Lutzenberger 2003; Knyazev 2007), while frontal θ activity suggests enhanced cognitive control engagement (Cavanagh and Frank 2014; Domic‐Siede et al. 2021). The observed coupling of γ and θ may indicate that these two processes interact with each other—one involving the allocation of attentional resources and the other involving top‐down control mechanisms during CR (Kaiser and Lutzenberger 2003; Ertl et al. 2013; Domic‐Siede et al. 2021). The functional significance of γ–θ coupling will be discussed later.

4.3. NED's Modulation of Cross‐Frequency Coupling in Cognitive Reappraisal

Neural oscillations are interconnected via a process known as cross‐frequency coupling, which reveals the mechanisms of neutral information transmission and exchange. Cross‐frequency coupling results showed that HNED individuals had a stronger α phase and γ amplitude coupling than LNED during the reappraisal of negative images. The α–γ coupling, associated with the gating of external information processing, can reduce interference from task‐irrelevant stimuli (Bonnefond and Jensen 2015). An increase in α–γ coupling suggests an enhanced inhibitory capacity, restricting distractors from entering visual encoding and memory storage (Bonnefond and Jensen 2015; Tzvi et al. 2018). This heightened inhibitory ability likely supports HNED individuals in CR by allowing them to better suppress interference from previous interpretations of emotional images. This process aids selective attention toward reappraisal‐relevant information and reduces the influence of maladaptive thoughts, potentially enhancing CR efficiency. This finding is supported by prior research. For instance, Lee, Lindquist, and Nam (2017) reported that individuals with high emotion differentiation exhibited an enhanced N2 component, a typical inhibitory control element, during emotional stimuli processing.

The study also found that during the reappraisal of negative images, individuals with HNED demonstrated stronger parietal θ phase and γ amplitude coupling, as well as α phase and β amplitude coupling, compared to those with LNED. Parietal θ–γ coupling is associated with the encoding and storage of working memory information (Alekseichuk et al. 2016), while α–β coupling is related to the maintenance of working memory (Daume et al. 2017). Aspects such as the updating function of working memory (Zhao, Cai, and Maes 2023), working memory load (Adamczyk, Wyczesany, and van Peer 2022), and working memory training (Long et al. 2024) contribute significantly to effective CR. The enhanced θ–γ and α–β coupling in HNED individuals suggests that their stronger working memory capacity may facilitate sustained CR strategies. Working memory is essential for CR as it enables individuals to hold, manipulate, and update emotional information, supporting goal maintenance, suppression of maladaptive thoughts, and effective reappraisal (Adamczyk, Wyczesany, and van Peer 2022; Long et al. 2024). While preliminary, these findings suggest that working memory‐supported coupling mechanisms may underlie the observed CR effectiveness in HNED individuals.

The nuanced ability of high‐emotion differentiators to distinguish emotions can be attributed to two primary factors: a rich reservoir of emotional conceptual knowledge and the capacity to utilize this knowledge, such as working memory. Enhanced working memory is particularly critical for HNED individuals. It enables them to retain information about their current emotional state while simultaneously drawing on multiple emotion concepts to interpret and reframe it, a process called situated conceptualization (Barrett, Wilson‐Mendenhall, and Barsalou 2014). As previously discussed, CR represents a transformation between situated conceptualizations. In each of these processes, individuals with high emotional differentiation exhibit strong working memory capabilities, which could explain the observed strong θ–γ and α–β coupling during the CR process.

In summary, HNED individuals demonstrated effective CR, possibly because they could initially suppress irrelevant stimulus interference. Additionally, their effective CR ability may be linked to their robust working memory, which aids in the encoding, storage, and maintenance of information, allowing them to access emotional conceptual knowledge and potentially enhance their CR. However, further investigation is required to confirm these mechanisms.

It is noteworthy that the behavioral data for valence and arousal ratings did not show any interaction effect between group and task, contrary to our predictions. Despite the absence of a significant interaction, there were group differences in the arousal ratings for the regulation effect. Specifically, the HNED group exhibited a more pronounced decrease in emotional intensity under the UR condition compared with the UV condition. However, the results revealed a significant group difference in the arousal ratings of regulatory effects; no such difference was observed in the valence ratings. Similar findings were reported in a study on emotion regulation in individuals with alexithymia, where those with high levels of alexithymia were associated with effective arousal reduction but not changes in emotional valence (Pollatos and Gramann 2012). Similarly, research on positive emotion differentiation found no significant effect on changes in emotional valence (Wang et al. 2020). This may be because valence ratings are more subjective, as participants may vary in their sensitivity and criteria for evaluating emotional valence. In contrast, arousal ratings, reflecting physiological activation levels, align more closely with the neural markers observed in our LPP and α‐β oscillation coupling results. Thus, arousal, unlike valence, may serve as a more objective and stable behavioral marker of CR effectiveness. This also aligns with the dimensional perspective of emotions (Barrett and Russell 2014), which posits that individuals can experience unpleasant emotions with varying arousal levels. In this context, CR strategies may effectively reduce arousal without necessarily altering emotional valence. Thus, individuals with HNED may lower their arousal levels, maintaining the experience of an unpleasant emotion but with reduced intensity. The machine learning results further suggest that behavioral indicators alone may not reliably predict outcomes, thus emphasizing their limited accuracy. Moreover, emotion differentiation captures the specificity with which people represent or experience valenced emotions (Barrett et al. 2001; Kashdan, Barrett, and McKnight 2015). High levels of emotion differentiation are related to lower levels of valence focus (Erbas et al. 2015), leading individuals to focus more on arousal indicators during emotion regulation. ERP findings also showed that emotion differentiation modulated arousal‐related markers (e.g., the LPP amplitude). Given that changes in valence are crucial in emotion regulation, further research is required to explore whether high emotion differentiation can enhance changes in emotional valence.

4.4. Machine Learning Findings

We explored the neural mechanisms by which emotion differentiation influences CR from the perspective of temporal processes, oscillatory features, and the transmission and interaction of information through brain oscillations. To integrate these multimodal research findings, we incorporated these indicators as feature variables into a pattern recognition classifier to examine their predictive accuracy for NED groups. The results showed that multimodal indicators achieved better classification performance than unimodal indicators. From the machine learning results, the predictive accuracy of single indicators, such as behavior, LPP, time frequency, and PAC for NED grouping, was below 80%. However, when these indicators were combined, the accuracy increased significantly to over 90%, with notable improvements in both sensitivity and specificity. A multimodal approach would allow for a more comprehensive and integrated understanding of how emotion differentiation influences CR. Similar to our study, Xie et al. (2023) have also employed machine learning methods to distinguish brain activity patterns associated with different emotion regulation strategies and explored the differences in neural dynamics between high‐ and low‐depression groups during the emotion regulation process. Machine learning analysis allowed fully leveraging the data, providing a comprehensive understanding of the emotion regulation process and the impact of individual differences. Recently, machine learning has been applied to predict personalized treatment strategies for mental disorders, offering opportunities for the development of precision psychiatry (Bzdok and Meyer‐Lindenberg 2018). This approach can also be applied to intervention studies on CR, such as identifying individuals who are likely to benefit more from CR interventions.

4.5. Significance

This study provides multimodal evidence of the neural mechanisms by which NED influences CR from a comprehensive perspective. The ERP analysis helped obtain dynamic evidence of high temporal precision. Time–frequency and cross‐frequency coupling allowed interpretation regarding the neurophysiological mechanisms of neural oscillations, providing a link to multiple disciplines of neurophysiology (e.g., single‐cell recordings, intracranial EEG, and MEG).

From a clinical perspective, emotional problems such as depression and anxiety have emerged as significant and severe facets of psychological problems (Huang et al. 2019; Sheldon et al. 2021). Cognitive‐behavioral therapy (Beck 2005), psychodynamic therapy (Palmieri et al. 2022), and other interventions often employ CR strategies to address these problems (Clark 2022). Although CR is effective in regulating negative emotions (Gross and John 2003), its effects vary among individuals. We found that a higher level of negative emotion differentiation enhanced the effectiveness of CR. Therefore, strengthening the NED abilities may enhance the effectiveness of emotional interventions. For instance, emotional conceptual knowledge training—such as learning about specific categories of emotions and conducting comparative learning of emotion categories—could improve NED by enabling individuals to better distinguish and process their emotional states (Vedernikova, Kuppens, and Erbas 2021), facilitating effective CR application. Other techniques, such as mindfulness training (van der Gucht et al. 2019), also show promise in increasing NED, thus potentially enhancing CR efficacy in individuals with emotional disorders.

Furthermore, the time–frequency and cross‐frequency coupling results can inform personalized CR intervention for individuals with LNED. These findings have substantial implications for transitioning theoretical research into clinical practice. For example, time–frequency results suggest that using tACS to induce γ‐band neural rhythm oscillations in the posterior brain may enhance CR effects. Cross‐frequency coupling results indicate that tACS can be used to induce compound rhythms (like γ or α nested within θ cycle peaks or troughs) to further improve CR in emotionally disordered individuals. However, the use of these neuro‐intervention techniques to enhance the CR effects in individuals with low emotional differentiation requires further validation in subsequent research.

4.6. Limitation and Future Directions

First, when exploring the neural mechanisms through which emotion differentiation influences CR, NED was measured as an individual difference construct. This approach focuses exclusively on “trait” differentiation, neglecting “state” differentiation. However, recent studies (Erbas et al. 2022; Mikkelsen, O'Toole, and Mehlsen 2020) have suggested that emotion differentiation fluctuates with time and context and is influenced by factors such as stress (Erbas et al. 2018) and interventions (van der Gucht et al. 2019). Future research should manipulate state emotion differentiation to study its impact on CR, potentially by asking participants to deliberately differentiate emotions (Cameron, Payne, and Doris 2013). Second, although we provided evidence from temporal processes, oscillatory features, and PAC when exploring the neural mechanisms through which emotion differentiation influences CR, we could not establish causality. Based on the time–frequency study results, future research could explore whether using tACS to induce gamma‐band neural rhythmic oscillations in the posterior brain might enhance CR effectiveness in individuals with LNED. Similarly, based on the cross‐frequency coupling results, it would be worthwhile to test whether using tACS to induce composite rhythms (such as gamma or alpha sub‐cycles nested within a single theta cycle peak or trough) could improve CR effectiveness in individuals with LNED. Finally, the cognitive reappraisal instructions in this study asked participants to reinterpret emotional stimuli from a new perspective, which led most participants to adopt a situation‐focused reappraisal strategy. Consequently, this study only examined the impact of emotion differentiation on situation‐focused reappraisal, leaving the self‐focused strategy unexplored. Future research should investigate how emotion differentiation affects both reappraisal strategies and explore differences in their underlying neural mechanisms.

5. Conclusion

In this study, we used event‐related potentials, time–frequency analysis, and cross‐frequency coupling analysis of neural oscillations to investigate the influence and neural mechanisms of NED on CR. Our key findings are as follows: (1) Individuals with HNED individuals can effectively diminish the amplitude of the LPP during negative reappraisal, suggesting more effective CR. (2) During negative reappraisal, HNED individuals exhibit stronger α‐β coupling, which may indicate enhanced working memory capabilities that contribute to more effective CR. (3) Multimodal indicators related to CR had better classification prediction effects for NED than unimodal indicators. These findings suggest that NED may help optimize CR processes, offering valuable insights into the neurophysiological mechanisms by which NED influences CR. Understanding NED's impact on CR could inform broader theories in emotion regulation and cognitive neuroscience, highlighting potential pathways for enhancing emotion regulation interventions through targeted training in NED.

Supporting information

Data S1 Supporting Information.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.