Abstract

Introduction:

Ruminative brooding is a transdiagnostic symptom defined as repetitive dwelling on thoughts and emotions, and is linked to emotion dysregulation, maladaptive metacognitive beliefs, and abnormal interoception. The relative contributions of these factors and their neural mechanisms remain unclear. In this exploratory study, we mapped these processes onto directed cross-frequency coupling (CFC) networks using resting-state electroencephalography.

Methods:

We first identified symptoms of interest for CFC analyses by employing regularized symptom networks, revealing two clusters relevant to brooding: one dominated by interoceptive/mindfulness dimensions and another by metacognitive/emotional dysregulation, with brooding belonging to the latter. We then examined links between representative symptoms from each cluster and resting-state cross-frequency phase–amplitude coupling (PAC) using partial least squares correlation (PLS-C).

Results:

Emotional dysregulation and brooding dimensions co-varied with delta-beta PAC (representing a “brooding/dysregulation” neural signature), whereas mindfulness symptoms co-varied with beta-gamma and theta-gamma PAC (representing a “mindfulness/interoception” neural signature). More specifically, for the brooding/dysregulation signature, prefrontal and cingulate phase activity modulated amplitudes in regions implicated in emotion regulation and interoception. In contrast, the mindfulness/interoception signature reflected coupling within circuits supporting emotion regulation/interoception.

Discussion:

Overall, our results indicated that brooding was more tightly linked to maladaptive metacognitive beliefs and emotional dysregulation than to mindfulness/interoception, consistent with resistance toward one’s thoughts and emotions. Neurally, as reflected through multivariate PLS-C covariance patterns, this may be linked to compensatory top-down control from prefrontal and cingulate areas over interoceptive, affective, and self-referential systems, pointing to the potential value of therapies that cultivate self-acceptance and modify maladaptive metacognitive beliefs for reducing rumination.

1 Introduction

Rumination is a transdiagnostic symptom common to mood, anxiety, trauma-related and sleep disorders, to name a few (Ehring and Watkins, 2008; Watkins, 2009). This symptom involves repetitive dwelling on thoughts, moods or past events (Nolen-Hoeksema, 2000). In depression, ruminative thought content is often negative and self-deprecating, separating into two distinct subtypes: ruminative brooding and reflective pondering (Treynor et al., 2003). These two subtypes differ in content and functional significance. Reflective pondering is analytical in nature, involves problem-solving to understand causes of negative mood, and is thought to be adaptive as it negatively predicts future depressive episodes (Treynor et al., 2003). Ruminative brooding, on the other hand, involves the passive dwelling on abstract and critical thoughts, and is considered the maladaptive form, as it predicts future depressive episodes, depression severity and recurrence (Joormann et al., 2006; Treynor et al., 2003). The cognitive and underlying neural mechanisms of ruminative brooding remain unclear; understanding these mechanisms could help inform treatment selection and design.

Mechanistic accounts of rumination, and other forms of repetitive negative thinking such as worry in anxiety, implicate both “bottom-up” emotion-driven processes and/or “top-down” cognitive control mechanisms (Ikani, 2021). “Bottom-up” processes are typically automatic in nature, requiring little cognitive effort as they are driven by cognitive biases and emotions. In contrast, “top-down” processes require effort and rely on cognitive systems such as those involved with attention, working memory, and cognitive control. While previous work has mainly focused on rumination from a “top-down” perspective, revealing deficits in inhibitory control (Singh et al., 2025; Whitmer and Banich, 2007), poor working memory updating (Joormann and Gotlib, 2008; Zetsche et al., 2012), and attentional disengagement (Koster et al., 2011; Whitmer and Gotlib, 2013), the role of “bottom-up” factors remains unclear.

Integrative constructs that combine both top-down and bottom-up processes, such as emotional regulation, mindfulness, interoceptive awareness, and metacognitive beliefs, have been independently linked to rumination. Each of these constructs reflects how an individual relates and responds to their inner experiences, and have significant top-down (e.g., appraisal, cognitive control) and bottom-up (e.g., affect- and sensory-driven) components. These factors play a significant role in rumination, especially given that initial conceptualizations of rumination in both anxiety and mood disorders highlight this process as a stress-reactive emotional regulation process, in which individuals repetitively evaluate the causes and meanings of their moods and thoughts (Davey, 2006; Mennin et al., 2005; Nolen-Hoeksema, 1991, 2000). It is therefore crucial to consider mechanisms that integrate both top-down and bottom-up processes when studying rumination.

Mindfulness involves deliberately focusing one’s attention on current internal and external experiences, observing thoughts and emotions without judgment (Chems-Maarif et al., 2025; Kabat-Zinn, 1994). Relatedly, interoceptive awareness reflects an individual’s awareness of their internal bodily sensations (e.g., heartbeat, organ function, respiration, satiety) along with autonomic nervous system activity related to emotions (Barrett et al., 2004; Cameron, 2001; Craig, 2002; Vaitl, 1996). Low mindfulness and poor interoceptive awareness have been linked to heightened ruminative brooding (Alleva et al., 2014; Lackner and Fresco, 2016), reflecting difficulties with emotion regulation and, in turn, heightened depression and anxiety symptoms (Lackner and Fresco, 2016). Notably, some individuals have reported heightened sensitivity to interoceptive cues (e.g., awareness of their heartbeat) during episodes of rumination (Schlinkert et al., 2020). These mixed results may be explained by stress response system (SRS) dysregulation theories, such as the Allostatic Load Model, in which initial stress exposure leads to hypervigilance, increasing awareness of interoceptive cues (Del Giudice et al., 2011; Ellis et al., 2011; Juster and Misiak, 2023). Chronic stress exposure, however, may excessively burden the SRS and lead to subsequent down-regulation (Del Giudice et al., 2011; Juster and Misiak, 2023), leading to insensitivity to internal states and their causes (Schultchen et al., 2019). Rumination after acute vs. chronic stress exposure may therefore be differentially related to interoceptive cue sensitivity. Regardless of an individual’s trait interoceptive sensitivity, mindfulness- and interoceptive awareness-based treatments aim to balance one’s awareness of their internal states while encouraging reappraisal and goal-directed behaviour change, ultimately reducing distress (Price and Hooven, 2018). Mindfulness-based therapies have been shown to reduce rumination by improving attentional control and promoting adaptive awareness of bodily sensations (Hammerdahl et al., 2025; Heeren and Philippot, 2011; Perestelo-Perez et al., 2017; van der Velden et al., 2023).

The role of metacognitive beliefs in rumination both complements and diverges from mindfulness and interoceptive awareness. Like mindfulness and interoception, metacognitive processes involve heightened awareness of internal experiences, yet they diverge in being explicitly evaluative: whereas mindfulness emphasizes nonjudgmental observation and interoceptive awareness emphasizes sensitivity to bodily states, metacognition entails the appraisal, monitoring, and control of one’s thoughts (Flavell, 1979; Moritz and Lysaker, 2018). Metacognitive beliefs represent relatively stable assumptions about the meaning and function of one’s cognitive processes. These beliefs can either be adaptive (e.g., “catastrophizing is unhelpful”), or maladaptive (e.g., “worrying now will help me later”), with the latter being consistently linked to psychopathology (Capobianco and Nordahl, 2023). According to one account, repetitive negative thinking, including depressive rumination, is reinforced by maladaptive metacognitive beliefs concerning the function and consequences of such thinking (Wells and Matthews, 1996) (e.g., “If I didn’t ruminate about my feelings of depression, they would take over me/never end” (Papageorgiou and Wells, 2001)). Subsequent work supported this account, showing that individuals with recurrent depression often endorsed rumination as a coping strategy to manage overwhelming emotions, while also perceiving it as uncontrollable and thus contributing to feelings of personal failure (Papageorgiou and Wells, 2001). Importantly, interventions designed to modify maladaptive metacognitive beliefs have been shown to reduce rumination across multiple psychopathologies (Sharma et al., 2022), highlighting the importance of maladaptive metacognitive beliefs in rumination.

Although these constructs have been studied in isolation in brooding, their relationship to brooding after considering construct interdependence has not yet been investigated. Mindfulness and metacognitive beliefs may be functionally related, with evidence demonstrating that high mindfulness is associated with less dysfunctional metacognition, likely due to reducing the negative appraisal of internal thoughts (Solem et al., 2017). In addition, metacognition may facilitate mindfulness by enabling the monitoring of internal experiences (Jankowski and Holas, 2014; Norman, 2017). However, it remains unclear which of these factors plays the most salient role in rumination. A promising approach for studying the interdependence between these factors, along with identifying central features for ruminative brooding, is by estimating regularized partial correlation networks (also known as “symptom” or “psychological” networks) (Epskamp et al., 2018; Epskamp and Fried, 2018). These networks are composed of nodes (the “symptoms”) and edges (strength of mutual associations between symptoms); the resulting structure reveals relationships and patterns between symptoms while controlling for all others. Although these networks are primarily used to map relationships between symptoms of a disease or condition (i.e., at the population level), this approach also offers a valuable exploratory approach for identifying candidate features of interest for subsequent analyses. Specifically, here we are interested in linking these candidate features of rumination to their neural substrates; the potential to ground these psychological processes in brain networks and circuits not only provides us with information on the neural underpinnings but may also reveal possible targets for neurostimulation treatments (e.g., repetitive transcranial magnetic stimulation, or rTMS, and deep-brain stimulation).

Electroencephalography (EEG) offers a non-invasive method for identifying neural substrates of cognitive phenomena, while also enabling analyses that require high temporal resolution, such as those involving neural oscillations. Neural oscillations span a wide range of frequency bands, with each thought to play a distinct functional role (Fries, 2015). Furthermore, the functional coupling between oscillations of different frequency bands, known as cross-frequency coupling (CFC), is thought to be a marker of large-scale coordination between and within brain networks (Canolty and Knight, 2010). CFC occurs when the phase or amplitude of one oscillation is functionally coupled to the phase or amplitude of another. For example, phase-amplitude CFC occurs when the amplitude of a high-frequency oscillation is synchronized and/or modulated by the phase (e.g., the peak or trough) of a low-frequency oscillation. CFC may facilitate cross-brain region coordination by enabling two oscillations, localized to different brain areas, to interact. When computed across pairs of brain areas, CFC can be used to construct functional connectivity networks.

CFC features have been linked to emotional regulation, mindfulness/interoceptive awareness, and cognitive functions. Interestingly, distinct CFC signatures are thought to reflect bottom-up and top-down processes (Fries, 2015). Delta (2–4 Hz)-beta (13–30 Hz) coupling has been presented as a marker of emotional (dys)regulation (Myruski et al., 2022; Poppelaars et al., 2021) and a neural predictor of cortisol response under stress (Wang X. et al., 2024). Theta (4–8 Hz) oscillations have been implicated across many functions, including mindfulness (Duda et al., 2024; Lomas et al., 2015), with theta-gamma (30–80 Hz) CFC involved in functions relevant to ruminative brooding, including memory, attention, and emotion (Ursino and Pirazzini, 2024). Both trait mindfulness/interoceptive awareness and rumination have been associated with beta oscillations in brain regions involved with self-referential processing and attentional control (Benschop et al., 2021; Ferdek et al., 2016; Ng et al., 2021). The role of beta activity in supporting these seemingly opposing processes (i.e., mindfulness and rumination) remains unclear. One possibility is that the brain accommodates both by reconfiguring functional networks according to current demands (Reinhart and Woodman, 2014), for example, by switching the directionality of oscillatory coupling. Since gamma oscillations are thought to reflect local circuit computations (Fernandez-Ruiz et al., 2023; Fries, 2009), studying beta-gamma CFC will enable the investigation of information flow within rumination- and mindfulness-related circuits. Given this putative function of gamma oscillations, the beta phase may reflect modulatory influences, whereas the gamma amplitude may capture localized computations, enabling identification of regions that act as “modulators” vs. “processors” within these networks. Using CFC to compute functional connectivity networks therefore enables the study of how slower rhythms in one region organize fast local computations in another, capturing neurocomputationally relevant directed information flow that single-band connectivity measures typically obscure. Examining functional networks derived from CFC may therefore offer a novel method for linking symptom network organization with ruminative neural dynamics.

There are a number of regions distributed amongst the default mode, salience, central executive, and sensorimotor networks that may be relevant for brooding, emotion regulation, interoception, metacognition and mindfulness (Table 1). Notably, both the posterior parietal cortex and posterior cingulate cortex have been heavily implicated in ruminative brooding (Andersen et al., 2009; Benschop et al., 2021; Bocharov et al., 2021a; Fink et al., 1996; Forner-Phillips et al., 2020a; Kircher et al., 2002), with these regions forming a circuit with medial temporal lobe structures, such as the parahippocampal gyrus, supporting autobiographical memory recall and self-referential processes (Aminoff et al., 2013; Kobayashi and Amaral, 2003; Leech and Sharp, 2014; Suzuki and Amaral, 1994). Mindfulness and interoceptive awareness may require brain regions integrating interoceptive and somatosensory cues, such as the insula (Berntson and Khalsa, 2021; Cauda et al., 2011; Khalsa et al., 2009) and somatosensory cortex (Cameron and Minoshima, 2002; Critchley et al., 2004; Tamè et al., 2016). Both metacognition and mindfulness, along with the processing of bottom-up emotional cues, may require the top-down appraisal of thoughts facilitated by the dorsolateral prefrontal cortex, which is well known to be involved in executive functions (Friedman and Robbins, 2022). Finally, many of these regions, such as prefrontal and limbic areas, work together to facilitate emotional regulation and processing (Banks et al., 2007; Delgado et al., 2008; Johnstone et al., 2007; Siegle et al., 2007; Urry et al., 2006). For example, the subcallosal cingulate plays an important role in negative emotional processing (Hamani et al., 2011; Vogt, 2014; Vogt and Vogt, 2009), and has been associated with dorsolateral prefrontal cortex activity in depression (Benschop et al., 2022). How these sub-circuits interact to facilitate both brooding and self-regulatory processes remains unclear. We propose that investigating this question through CFC networks may be a useful first step.

RegionAbbreviationDestrieux label(s)Functional relevanceAnterior cingulate cortexACCG_and_S_cingul-AntPart of the limbic system and the salience network. Known to play a role in both cognitive control and emotional stability (Bush et al., 2000).Posterior cingulate cortexPCCG_cingul-Post-dorsal
G_cingul-Post-ventralInvolved in self-referential processing and rumination (Benschop et al., 2021; Fink et al., 1996; Kircher et al., 2002). Strongly connected to the parahippocampal gyrus and entorhinal cortex, therefore related to the hippocampal memory system (Kobayashi and Amaral, 2003; Leech and Sharp, 2014; Suzuki and Amaral, 1994).Subcallosal cingulate cortex (also known as subgenual cingulate)SCCG_subcallosalInvolved in emotion regulation and is heavily implicated in depression (Hamani et al., 2011; Vogt, 2014) and previously associated with rumination in remitted major depression (Benschop et al., 2021).Parahippocampal gyriPHGG_oc-temp_med-ParahipEpisodic memory (Aminoff et al., 2013) and depressive symptoms (Zamoscik et al., 2014)InsulaInsulaG_Ins_lg_and_S_cent_ins
G_insular_shortEmotional processing and interoceptive awareness (Berntson and Khalsa, 2021; Cauda et al., 2011; Khalsa et al., 2009)Ventromedial prefrontal cortexvmPFCS_orbital_med-olfact
S_orbital-H_Shaped
G_orbitalEmotion regulation of negative emotions (Delgado et al., 2008; Johnstone et al., 2007; Urry et al., 2006), self-referential processing (D’Argembeau, 2013) and autobiographical memory recall (McCormick et al., 2020)Dorsolateral prefrontal cortexdlPFCG_front_middle
S_front_middle
S_front_supExecutive functions (Friedman and Robbins, 2022), and top-down regulation of limbic activity (Banks et al., 2007; Siegle et al., 2007). Compensatory connectivity at rest demonstrated in depression (Iseger et al., 2017; Wang Y. et al., 2024).Somatosensory cortexSSCG_postcentralInteroceptive awareness, sensory perception and integration, and bodily awareness (Cameron and Minoshima, 2002; Critchley et al., 2004; Tamè et al., 2016).Posterior Parietal cortexPPCG_pariet_inf-Supramar
G_parietal_supWorking memory, attention, spatial cognition (Marek and Dosenbach, 2018), and autobiographical memory (Brown et al., 2018). Has been implicated in rumination (Andersen et al., 2009; Bocharov et al., 2021b; Forner-Phillips et al., 2020b).

ROI Destrieux labels and justification for inclusion.

The present study aims to explore the neural correlates of brooding and concurrent self-regulatory processes to move towards an integrative neural theory. We address this aim through the following two approaches: (1) by presenting a symptom network analysis of mindfulness, interoceptive awareness, metacognitive beliefs and ruminative brooding to identify potential patterns of interdependence among symptoms (e.g., clustering signatures) used primarily to guide subsequent EEG analyses; and (2) by linking those symptom network-derived features with functional connectivity markers from CFC-derived neural networks. We aim to test the hypotheses that (1) ruminative brooding is associated with diminished mindfulness and heightened maladaptive metacognitive beliefs, and that (2) similar circuits will underlie mindfulness, metacognition, and ruminative brooding, differing in the frequency of oscillations participating in CFC and directionality of the frequency pairing. By bridging ruminative brooding-related symptoms with functional neural networks, the present study aims to explore CFC patterns that reflect the interaction between top-down and bottom-up processes. Characterizing these network dynamics and symptom interactions may help to guide future research by identifying targets for neuromodulatory or cognitive interventions that enhance regulatory control and reduce ruminative brooding.

2 Methods2.1 Experiment2.1.1 Study design and participants

The data used for the analyses presented in this paper are from a larger study on the neural dynamics of depressive rumination. This larger study included questionnaire measures of rumination, mood, emotional regulation, mindfulness, interoceptive awareness, dissociation, metacognitive beliefs, sleep (described in detail below), along with handedness; a cognitive control assessment via the Stroop task; and finally 5 min of resting-state EEG followed by a task-switching EEG protocol adapted from a previous study from our lab (Shaw et al., 2021). Participants completed rumination and mindfulness measures in person, prior to the EEG session, and completed the remaining self-report measures using a take-home survey. The present study includes results from the resting state EEG data only, as well as a selection of questionnaire measures from the full study (i.e., all questionnaires except for the handedness assessment). We focussed on the resting state EEG data only for this study, as the task-based EEG data was aimed at testing different hypotheses regarding dynamical systems in rumination, which are fundamentally unrelated to the aims of the present study.

We aimed to recruit 40 undergraduate students through McMaster University’s SONA research participant recruitment system. We decided on this sample size based upon previous studies of rumination including EEG connectivity, source-localization and spectral analyses, which have reported samples ranging between N = 26–45 (Bocharov et al., 2021a; Ferdek et al., 2016; Forner-Phillips et al., 2020a; Reiser et al., 2012). Our exclusion criteria included a self-reported history of traumatic brain injury, any current or previous mental health diagnoses (e.g., major depressive disorder, bipolar disorder, post-traumatic stress disorder, etc.), and current engagement in more than 5 min of weekly mindfulness practice. We excluded participants with >5 min of weekly mindfulness practice as we aimed to assess mindfulness as an inherent, trait-like feature, rather than one that has been actively learned or manipulated by the participants over time. All participants provided informed consent, and protocols were approved by the McMaster Research Ethics Board (MREB number 5987). We collected data from 48 participants, of whom all provided complete symptom data, and 31 provided complete and usable EEG data (i.e., correctly saved, loadable EEG files, with data demonstrating repairable artifact contamination and/or noise). The sample comprised individuals aged 18–25, of whom 60% identified as female, 11% as male, and 29% did not report their sex and/or gender.

2.1.2 Questionnaire measures

We assessed trait-level rumination, depression, anxiety and stress levels using self-report measures as follows. Our primary measure of interest, ruminative brooding, was assessed using the brooding subscale of the Ruminative Response Scale (RRS) (Nolen-Hoeksema, 2000; Nolen-Hoeksema and Morrow, 1991). Although we aimed to recruit participants without mental health diagnoses, it is possible that some exhibited subclinical symptoms of depression or anxiety (Beiter et al., 2015), or had not yet received a formal diagnosis. In addition to the RRS, we therefore used the Depression Anxiety Stress Scale (DASS-21) to assess levels of depression and anxiety symptoms along with perceived chronic stress (Lovibond and Lovibond, 2011). The DASS-21 is grounded in a dimensional model of depression, anxiety, and stress, which views differences between clinical and nonclinical populations as variations in symptom severity. The DASS-21 cannot, therefore, provide a clinical diagnosis of depression, anxiety or a stress-related disorder, but can be helpful to probe the severity of psychopathological symptoms. We therefore used the recommended cut-off scores to assess symptom severity (i.e., normal, mild, moderate, severe and extremely severe) (Lovibond and Lovibond, 2011).

To measure integrative (i.e., top-down and bottom-up) mechanisms of rumination, including metacognitive beliefs, emotional regulation, interoceptive awareness and mindfulness, we used the following self-report measures. To assess participant metacognitive beliefs, we included the Metacognitions Questionnaire (MCQ) to assess the following beliefs: positive beliefs about worry, negative beliefs about worry, cognitive confidence, need to control thoughts, and cognitive self-consciousness (Cartwright-Hatton and Wells, 1997). Furthermore, we assessed emotional regulation using the Difficulties in Emotional Regulation Scale (DERS), which consists of 6 subscales: nonacceptance of emotional responses, difficulties engaging in goal-directed behaviour, impulse control difficulties, lack of emotional awareness, limited access to emotional regulation strategies, lack of emotional clarity (Gratz and Roemer, 2004).

To measure participant levels of trait interoceptive awareness and mindfulness, we included the following measures: the second version of the Multidimensional Assessment of Interoceptive Awareness (MAIA-2) (Mehling et al., 2018) and the Five Facet Mindfulness Questionnaire (FFMQ-39) (Baer et al., 2006; Shallcross et al., 2020). The MAIA-2 consists of 8 subscales: noticing (awareness of uncomfortable, comfortable and neutral body sensations); non-distracting (tendency not to ignore or distract oneself from sensations of pain or discomfort); not-worrying (tendency not to worry or experience emotional distress with sensations of pain or discomfort); attention regulation (ability to sustain and control attention to body sensations); emotional awareness (awareness of the connection between body sensations and emotional states); self-regulation (ability to regulate distress by attention to body sensations); body listening (active listening to the body for insight); and trusting (experience of one’s body as safe and trustworthy). The FFMQ-39 consists of 5 subscales: observing, describing, acting with awareness, non-judging of inner experience, and nonreactivity to inner experience. Given the similarities between rumination and dissociation in terms of a lack of connection to the present environment, we also assessed trait dissociation using the total score from the second version of the Dissociative Experiences Scale (DES-II) (Carlson and Putnam, 1993).

Along with mindfulness/interoceptive awareness, emotional regulation, depression and anxiety, rumination is also known to interfere with sleep (Morin et al., 2003); as such, sleep quality may be a marker of physiological distress. We assessed participant sleep quality using the Sleep Quality Scale (SQS), comprising 6 subscales: daytime dysfunction, restoration after sleep, difficulty falling asleep, difficulty getting up, satisfaction with sleep, and difficulty maintaining sleep (Yi et al., 2006).

2.1.3 Resting state EEG acquisition

Resting-state EEG data were collected for 5 min using a BIOSEMI ActiveTWO system with 128 wet, gel-based electrodes in a sound-attenuated, dimly lit room. Since active electrodes provide impedance transformation on the electrode, generating an output impedance of <1 Ω, the level of DC offset is typically used to evaluate quality of electrode contact rather than impedance values. We ensured that electrode offsets were kept within ± 20 μV. Data were collected at a sampling rate of 2048 Hz. Participants were instructed to close their eyes and minimize movement for the duration of the 5 min.

2.2 Subscale analyses2.2.1 Feature selection for symptom networks

Features derived from symptom network analyses are sensitive to sample size, with large samples (i.e., N > 100) typically required to support adequate stability analyses (Epskamp et al., 2018). Due to our relatively small sample size for this analysis (N = 48), we decided a priori to select the top 15 symptoms that were the most predictive of participant brooding scores using a regularized regression (“Elastic Net”) (Zou and Hastie, 2005). This initial regression analysis was conducted in Python using the scikit-learn package (v. 1.3.2), and was not aimed at analyzing predictors of ruminative brooding, but rather, at identifying symptoms with the highest-magnitude non-zero coefficients for subsequent analyses. Elastic Net regression combines both lasso and ridge regularization to shrink very small (i.e., irrelevant) coefficients by placing a penalty term in front of coefficients during model training as a part of the loss function. Lasso shrinks very small coefficients to zero, which is excellent for feature selection, but may erroneously shrink some non-zero coefficients if there is a high degree of multicollinearity between measures (e.g., between our measures of depression, stress, and rumination, or mindfulness and interoceptive awareness). Ridge regularization will shrink coefficients evenly without eliminating them, which is a good strategy for handling multicollinearity, but is not ideal for feature selection purposes. Elastic net combines both of these regularization techniques, enabling feature selection while handling multicollinearity (Zou and Hastie, 2005). The Elastic Net regression model was trained using 5-fold cross-validation on the standardized symptoms to predict ruminative brooding. The mixing parameter was set to 0.5 to balance the L1 (lasso) and L2 (ridge) penalties equally.

Our features of interest included the subscale and total scores of the RRS, DASS-21, DERS, MAIA-2, MCQ, FFMQ-39, DES-II, and SQS. Both total and subscale scores were entered as predictors, as Elastic Net is ideally suited to adjudicate between correlated variables to identify whether overall severity or specific symptom dimensions contributed more strongly to the prediction of brooding, as indexed by their retention and relative coefficient magnitude in the Elastic Net model. We excluded the DASS-21 and RRS depression subscales as the RRS brooding subscale may already be confounded with depression symptomatology (Treynor et al., 2003). In total, we assessed 39 scale-based scores as predictors of brooding severity. We selected symptoms with the top 15 largest magnitude regression coefficients for our subsequent exploratory network analyses (Figure 1).

Overview of study methodology. Participants first completed questionnaire measures of rumination, metacognitive beliefs, interoceptive awareness, and mindfulness, followed by 5 min of resting-state EEG using a 128-lead system. Relationships among symptoms were analyzed by constructing regularized partial correlation networks (“symptom networks”) and assessing network structure and clustering behaviour of symptoms. EEG data were analyzed by first performing source localization to map signals onto regions of interest (ROIs), followed by assessing phase–amplitude cross-frequency coupling between theta-gamma (θ/γ), delta-beta (δ/β), and beta-gamma (β/γ) oscillatory bands. During the second stage of analysis, a set of symptoms from the symptom network was selected and mapped onto cross-frequency coupling scores using partial least squares correlation.

2.2.2 Estimating symptom networks

We estimated sparse Gaussian graphical models using the extended Bayesian information criterion applied to the graphical lasso (EBICglasso) algorithm (Epskamp and Fried, 2018). We implemented these models using the bootnet and psych packages in R (Epskamp and Fried, 2018). This algorithm involves first computing a correlation matrix across the 15 previously selected symptoms, followed by applying a Lasso penalty on the inverse covariance matrix to encourage sparsity. 100 network configurations were assessed, with the EBIC applied to choose the best-fitting sparse model (prioritizing goodness of fit and model simplicity while accounting for the small sample size). We used a moderate EBIC model selection penalty (γ = 0.1). The network with the lowest EBIC score was then selected. The resulting network was characterized by a set of nodes (symptoms) and edges (non-zero partial correlations, or the most meaningful conditional dependencies).

For interpretability purposes, we computed two separate networks: one including the total MCQ score and one excluding it. The total MCQ is thought to be a general measure of dysfunctional metacognitive beliefs, with a higher score indicating greater dysfunction. The network, including the total MCQ score, demonstrated close relationships between the total MCQ score and other MCQ symptoms due to inherent multicollinearity, which may obscure other symptom relationships (Supplementary Figure 3). We therefore re-estimated the network after excluding the total MCQ score to aid with interpretability (Figure 2).

Interoceptive/emotional awareness symptoms and negative metacognitive symptoms form separate clusters in a network of brooding-associated symptoms. (A) Network using an unregularized correlation matrix. (B) Regularized sparse network of partial correlations. (C) Cluster assignment from the community detection algorithm.

2.2.3 Assessing edge weight stability and symptom clustering

We expected our estimated network to be unstable due to our limited sample size. Since the goal of this analysis was to generate hypotheses and identify candidate features for subsequent EEG analyses, rather than to derive precise population-level network parameters, we do not view edge-level instability as a severe limitation. Nonetheless, as it is best practice when reporting symptom networks (Epskamp et al., 2018; Epskamp and Fried, 2018), we assessed the stability of edge weights by constructing confidence intervals around the sample edge weights using nonparametric bootstrapping. Here, we resampled the data 5,000 times with replacement. To aid in interpreting patterns within high-level network structure, a community detection algorithm was used to identify symptom clusters using the igraph package in R (version 2.1.4) (Csárdi et al., 2025). The community detection algorithm performed short random walks from each node and computed how often two nodes co-occured in the same walk. If two nodes co-occured frequently, they were assigned to the same cluster.

2.3 EEG analyses

EEG preprocessing and source-localization were conducted using the MNE package in Python (v. 1.6.0) (Gramfort, 2013; Larson et al., 2025).

2.3.1 Preprocessing

To remove high-frequency noise and low-frequency drift effects, we applied high (80 Hz) and low (1 Hz) pass windowed-sinc finite impulse response (FIR) filters. Next, we used a notch filter (defined in MNE as a symmetric FIR band-stop filter) to remove the electrical line noise artifact at 60 Hz. An independent components analysis (i.e, FastICA algorithm) was then used to identify eye blink and muscle movement related artifacts, with components removed based on visual inspection. On average, 4.7 out of 20 components were removed per participant, with compo