Authors: Masaya Misaki, Aki Tsuchiyagaito, Salvador M. Guinjoan, Martin P. Paulus
Categories: Article, Cognitive ageing, Emotion
Source: Communications Biology
Authors: Masaya Misaki, Aki Tsuchiyagaito, Salvador M. Guinjoan, Martin P. Paulus
Emotional thought patterns such as rumination, worry, and positive reflection are rooted in autobiographical memory and influence well-being. However, their neural representations and how they vary remain poorly understood. This study examined brain activation patterns linked to rumination, worry, and positive thinking in adults aged 18–64. Participants recalled autobiographical events during fMRI, and machine learning was used to decode whole-brain activation patterns. Results showed that neural differentiation of rumination and worry increased with age. Greater age was associated with increased activation in cognitive control regions during rumination, and decreased activation in the cingulate cortex and temporoparietal junction during worry. Positive thinking showed no significant age-related effects. Greater neural discrimination between positive thinking and rumination was associated with higher well-being. These findings suggest that individual differences in the neural representation of emotional thought may reflect age-related improvements in cognitive control and reduced reactivity to distressing thoughts across mid-adulthood, highlighting a possible neural basis for enhanced emotional well-being in midlife.
Emotional processing varies widely across individuals, particularly in the context of personally meaningful experiences. Negative thought patterns such as rumination and worry, which are closely linked to mood and anxiety disorders, are often rooted in autobiographical memory and have well-documented effects on emotional well-being^1–3^. Positive thinking may likewise arise from self-relevant experiences^4^, supporting psychological resilience and adaptive functioning. Unlike emotional responses to standardized affective stimuli, which are typically impersonal, these thought patterns reflect subjective, internally generated mental states. Despite their clinical and experiential relevance, the neural representations of such emotional processes, and their relationships with individual traits, remain poorly understood. Investigating their neural discriminability may provide new insights into the neurobiological foundations of individual differences in emotional health.
Given that these clinically relevant emotional processes are grounded in personal experience, age may shape the neural representation of such emotional states. Emotional well-being tends to improve with age^5–11^, a trend attributed to divergent age trajectories for positive and negative affect. Specifically, older adults show reduced rumination about past negative events^12^, which contributes to improved well-being^13^, while positive thoughts remain relatively stable across adulthood^14^. These findings suggest that age-related changes in emotional processing may be reflected in the brain in ways that are specific to the individual and the type of emotional state. A closer look at the different types of emotional thought, their neural distinctiveness, and how they vary across individuals and age groups warrants further investigation.
Negative emotional thinking can be categorized by its temporal rumination, which involves dwelling on past negative experiences, and worry, which centers on future uncertainties^1,3,15^. Together, these processes form the core of repetitive negative thinking (RNT) and share overlapping cognitive features^16–20^. While both are associated with heightened risk for anxiety and depression, rumination tends to predict depressive recurrence and comorbidity, whereas worry is more closely linked to anxiety disorders^21,22^. Importantly, these emotional tendencies may change differently across the adult lifespan, shaped by evolving social environments, cognitive goals, and emotion regulation strategies^23–25^.
Repetitive thought (RT), more broadly, also includes positive forms, such as constructive reflection and reminiscing, which have been linked to improved psychological well-being and eudaimonic function^26–28^. However, while typically adaptive, positive RT can also be dysregulated in certain clinical contexts. For example, individuals at risk for mania may engage in overly persistent or goal-directed positive ideation, reflecting maladaptive cognitive patterns^29^.
This study aimed to investigate how emotional thought patterns, specifically rumination, worry, and positive thinking, are represented in the brain, and whether the neural distinctiveness of these states within individuals is associated with age and affective traits. We focused on individual differences in the discriminability of emotional brain states, as measured through machine learning classifiers, and examined how these differences relate to self-reported measures of rumination, worry, and positive affect. By emphasizing within-subject neural distinctiveness, this study aimed to advance our understanding of how emotional thought patterns are neurally organized in relation to individual variation in emotional well-being.
To examine neural distinctiveness, we employed a whole-brain decoding approach using individualized machine learning models. Most prior research on emotional processing has relied on self-report measures or group-level neural contrasts, which may obscure subtle differences shaped by personal experience. In contrast, emotional states such as rumination, worry, and positive reflection, particularly when based on autobiographical events, are highly self-referential and may engage distinct neural patterns across individuals.
Recent findings suggest that emotional states are better characterized by distributed whole-brain activity patterns rather than localized regional activations^30–33^. Personalized models offer a principled way to identify these individualized neural signatures while minimizing biases introduced by group-based modeling, which may reflect stereotypical activation profiles^2,34,35^.
Given the limited existing research using individualized decoding approaches for autobiographical emotional states, the present study adopted an exploratory framework. We examined whether the neural distinctiveness of rumination, worry, and positive thinking, measured as the ability of individualized machine learning models to discriminate between these states, was associated with age, sex, and affective traits, such as trait rumination, worry, and positive affect. We also explored how decoded emotional brain states unfolded over time and identified brain activation patterns contributing to these distinctions. This approach allowed us to assess whether individual variability in emotional thought processing is reflected in neural signatures, and how these signatures relate to well-being and emotion-related traits. A more nuanced understanding of how emotional thought patterns are neurally represented across adulthood may help inform personalized approaches to promoting emotional well-being and mitigating age-related risk for affective disorders.
Thirty-seven healthy individuals (24 females, aged 18–64) participated in the study. Due to excessive head motion leading to the exclusion of more than two of six runs, two participants were removed from the analysis (see Methods for the exclusion criterion). The final dataset included 35 participants. Table 1 presents the demographic and assessment statistics of the included participants. Across participants, three runs were excluded in total, and the mean censoring ratio (i.e., the proportion of TRs excluded due to motion) for the remaining data was 4.2%. There was no significant relationship between censoring ratio and age (t = –1.337, Cohen’s d = −0.093, P = 0.190), nor between censoring ratio and overall brain state discriminability (area under the receiver operating characteristic curve (AUC); t = –0.626, Cohen’s d = -0.044, P = 0.531).Table 1Participant demographics and assessment scoresCronbach’s alphaAge associationSex (n = female/male)24/11t = 1.932, P = 0.068Age (mean [SD])40.7 [14.0]RRS31.6 [7.9]0.90r = −0.278, P = 0.105PSWQ37.9 [14.6]0.95r = -0.085, P = 0.626STAI-T31.3 [11.0]0.95r = -0.461, P = 0.005SCI-QOL Positive57.1 7.9^†^r = 0.103, P = 0.557RRS Ruminative Response Style, PSWQ Penn State Worry Questionnaire, STAI-T State-Trait Anxiety Inventory Trait score, SCI-QOL Positive, Spinal Cord Injury—Quality of life index positive affect and well-being score, r Spearman correlation.^†^Cronbach’s alpha for the SCI-QOL Positive scale is taken from Bertisch et al.^41^, as raw item-level responses were not available.P values for the age association are uncorrected.
Participants completed a thought induction task involving rumination, worry, and positive thinking blocks, interspersed with attention (flanker) and rest periods (Fig. 1). To examine how these thought states were represented in the brain, we applied individualized machine learning models to decode neural patterns associated with each condition. The model was trained to classify four conditions, rumination, worry, positive thinking, and rest, based on whole-brain activity patterns. Detailed procedures for model training and evaluation are provided in “Methods”.Fig. 1Experimental design and scan session structure.The scan session began with a T1-weighted anatomical scan, followed by a resting-state run (6 min 50 s) and six runs of the Think Affective Events Task (TAET). Each TAET run began with an 18 s Rest block, followed by blocks of 22 s Think, 18 s Flanker task, and 18 s Rest. These blocks were repeated eight times per run, with each run lasting 8 min 2 s.
The resting state was clearly distinguishable from the three thinking states (rumination, worry, and positive thinking) in the whole-brain machine learning analysis. All participants demonstrated significant AUC values (P < 0.05 with Bonferroni correction) for distinguishing rest from rumination (mean ± SD AUC = 0.879 ± 0.041), worry (0.882 ± 0.048), and positive thinking (0.893 ± 0.059). However, discrimination between the three thinking states varied across individuals. Figure 2 displays the distribution of AUC values for the distinctions between rumination, worry, and positive thinking. The data used to generate Fig. 2 are provided in Supplementary Data 1. The lowest discrimination score was observed between rumination and worry (AUC = 0.648 ± 0.115), with nine participants showing non-significant discrimination. The distinction between rumination and positive thinking (AUC = 0.739 ± 0.116) and between worry and positive thinking (AUC = 0.736 ± 0.117) also showed variability, with four and three participants, respectively, showing non-significant discrimination.Fig. 2Distribution of AUC values across participants.AUCs from the whole-brain machine learning classification results are shown for each contrast among the three thinking rumination, worry, and positive events. The symbols represent the significance of AUC for individual participants. n.s. non-significant. In each box plot, the center line represents the median of the distribution. The box limits define nested quantiles, and the outermost box aligns with the upper and lower quartiles (75th and 25th percentiles).
These AUC values were not correlated with self-reported task evaluation metrics, including perceived task difficulty, successful thinking, sleepiness, emotional exhaustion, or tiredness (see Supplementary Table S1). Furthermore, AUC values were not associated with self-reported distress levels related to rumination or worry events (Supplementary Table S1).
We also examined whether repeating the same thought events across runs affected neural discriminability, as repeated recall of the same topics (each event was presented four times) could potentially reduce brain responses due to repetition-related effects^36,37^. However, a linear mixed-effects model with a random intercept for participant revealed no significant main effect of run (F(5,765) = 0.001, P = 1.000), no run-by-state interaction (F(15,765) = 0.023, P = 1.000), and no run-by-age interaction (F(5,765) = 0.001, P = 1.000).
To investigate factors contributing to individual differences in thought state discrimination, we analyzed correlations between AUC values and variables such as age, sex, and assessment scores (Table 2 and Fig. 3). The data used to generate Fig. 3 are provided in Supplementary Data 1. Age showed a significant correlation with the discrimination of all three states, rumination, worry, and positive thinking, with older participants demonstrating greater discrimination across these emotional states (Fig. 3). STAI-T^38^, a measure of general anxiety or broader negative affectivity ^39^, was also significantly correlated with all three discriminability measures, such that lower scores were associated with higher discrimination. In addition, discrimination between negative (rumination and worry) and positive states was significantly correlated with trait rumination (RRS^40^) and positive affect (SCI-QOL Positive^41^). The general worrying tendency (PSWQ^42^) was specifically correlated with rumination-positive discrimination.Fig. 3Associations between thought state discrimination (AUC) and participants’ demographic variables and assessment scores.The line represents the fitted linear association, with the shaded area indicating the 95% confidence interval. A yellow background highlights combinations with significant associations (FDR < 0.05, see Table 2). RRS Ruminative Response Style, PSWQ Penn State Worry Questionnaire, STAI-T State-Trait Anxiety Inventory Trait score, SCI-QOL Positive Spinal Cord Injury—Quality of life index positive affect and well-being score.Table 2Correlations between thought state discrimination (AUC) and participants’ demographic variables and assessment scoresRumination-worryRumination-positiveWorry-positivestatPFDRstatPFDRstatPFDRSex (chi-square)0.5270.6050.6161.4610.1600.1801.6690.1130.136Age (r)0.4940.0030.0110.4100.0140.0310.4850.0030.011RRS (r)−0.3030.0760.106−0.5890.0000.004−0.4660.0050.014* Age-corrected (t)−1.7310.0930.178−2.2930.0290.086−1.5030.1430.238PSWQ (r)−0.0880.6160.616−0.4230.0110.029*−0.2940.0870.112 Age-corrected (t)−1.1480.2590.354−2.3580.0250.086−1.3320.1920.288STAI-T (r)−0.3800.0250.040*−0.5170.0010.009*−0.4060.0160.031* Age-corrected (t)−1.8330.0760.178−2.3440.0250.086−0.7720.4460.514SCI-QOL Positive (r)0.3520.0380.0570.5140.0020.0090.3830.0230.040 Age-corrected (t)2.7110.0110.0862.3710.0240.0861.7210.0950.178Statistically significant correlations are indicated in bold (^*^FDR < 0.05). FDR correction was applied across all variables and AUCs.RRS Ruminative Response Style, PSWQ Penn State Worry Questionnaire, STAI-T State-Trait Anxiety Inventory Trait score, SCI-QOL Positive Spinal Cord Injury—Quality of life index positive affect and well-being score.Statistical (stat) values are represented by t values for Sex and Spearman correlations for all other variables.
Given that age was the most significant factor overall, we assessed whether age confounded the associations between neural discriminability and affective traits. We conducted a follow-up linear model analysis including age as a covariate for the affective trait’s associations with state discriminability (Table 2). After accounting for age and applying FDR correction, none of the associations between affective traits and state discriminability remained statistically significant.
We also examined the associations between age and self-report measures related to task evaluation and distress associated with negative events. Several significant age correlations were observed in the task evaluation measures, including lower ratings of task difficulty, sleepiness, and tiredness, as well as higher ratings of successful thinking with increasing age. However, none of these measures were significantly associated with state discriminability (see Supplementary Table S1).
We further considered the possibility that the content of recalled episodic events may have systematically differed with age, potentially contributing to differences in brain activation discriminability. Although we did not collect detailed information about the content of the recalled events due to privacy considerations, participants were asked to select one of three prompts for each negative event, which may indirectly reflect the nature of the memories. For instance, older adults might be more likely to reflect on broader life implications (e.g., “Why did that happen to you?” or “How would this affect your life?”), while younger adults may focus more on their own actions or emotional responses (e.g., “Why did you do that?” or “What would you do if that happened?”). To evaluate this possibility, we examined whether instruction choices were systematically associated with age, sex, or affective trait scores (RRS, STAI-T, PSWQ, and SCI-QOL Positive Affect). No significant associations were found (all uncorrected P > 0.05), suggesting that the selection of thinking prompts was not systematically related to these individual difference variables (see Supplementary section “Frequency of Negative Thinking Instructions” and Table S3).
We examined differences in the time course of decoded states between younger and older participants. Figure 4 presents the mean time course of classifier outputs in young and old participants (divided by median age = 39). The data used to generate Fig. 4 are provided in Supplementary Data 1. Linear mixed-effects (LME) model analyses (see Supplementary Table S3 for its ANOVA table) revealed a significant main effect of age for all three states, with older participants exhibiting higher decoding probabilities for the corresponding emotional state. A significant age × time interaction was also observed for the worry condition. This interaction reflected an age-related enhancement in decoder output over worry-related state probabilities increased more prominently between 6 and 14 s in older participants, with the effect gradually diminishing thereafter. This pattern suggests that older individuals sustained distinctive worry-related brain states for a longer duration, whereas younger individuals showed earlier plateaus or less pronounced neural engagement.Fig. 4Temporal dynamics of decoded thought state probabilities across age groups.Mean time courses of classifier output probabilities are shown for rumination, worry, positive thinking, and rest during the Think and Flanker task periods, separately for younger and older participant groups. Individual participant data points at each time point are overlaid as semi-transparent dots to illustrate the data distribution. Solid lines represent the group mean, and shaded bands indicate the 95 percent confidence interval. Vertical dashed black lines mark the transition from the Think block to the Flanker task. A horizontal blue bar along the time axis indicates the time points included in decoder training for the leave-one-run-out cross-validation procedure.
We investigated brain activation patterns during each thought state by mapping beta coefficients for the decoded time series of mental state in a general linear model (GLM) analysis (Fig. 5). Peak coordinates of significant clusters are detailed in Supplementary Data 3–8. All three thought states were associated with activity in the lateral frontal regions, medial premotor and supplementary motor areas (SMA), superior temporal regions, visual cortex, hippocampus, putamen, and cerebellum. Reduced activity was observed in the anterior portion of the posterior cingulate region and the supramarginal gyrus. Both worry and positive thinking states showed positive associations with activation in the precuneus and the posterior part of the posterior cingulate, regions typically linked to self-related thoughts. In rumination, there was more activation in the inferior parietal lobule and less suppression in the supramarginal gyrus, regions also associated with self-referential thoughts, alongside lower activation in the ventromedial prefrontal cortex and visual cortex compared to other states. Worry states showed greater engagement of dorsal frontal regions compared to both rumination and positive thinking. Positive thinking involved higher activation in the retrosplenial cortex, hippocampus, and amygdala.Fig. 5Brain activation patterns associated with the thought states and their contrasts.The maps are displayed on the inflated cortical surface, with subcortical and cerebellar regions are shown on axial slice maps. Clusters with voxel-wise P < 0.001, corrected for cluster extent at P < 0.05, are highlighted with opaque colors. Unthresholded maps are also shown with lower opacity in the plots. The images are displayed in neurological orientation, with the left side corresponding to the left side of the brain.
Figure 6 shows age effects in each thought state and their contrasts. The peak coordinates of significant clusters are provided in Supplementary Data 9–13. For the rumination state, age was positively associated with increased activation in broad brain regions, with significant clusters in the right middle frontal gyrus, right superior frontal gyrus, fusiform gyrus, and cerebellum. In contrast, in the worry state, most brain regions showed a negative association with age, with significant clusters observed in the right temporoparietal junction (TPJ) region and anterior and middle cingulate cortex. The positive thinking state did not show a significant association with age.Fig. 6Age-related modulation of brain activation during emotional thought states.The maps are displayed on the inflated cortical surface, with subcortical and cerebellar regions shown on axial slice maps. Clusters with voxel-wise P < 0.001, corrected for cluster extent at P < 0.05, are highlighted with opaque colors. Unthresholded maps are presented with lower opacity in the plots. The images are displayed in neurological orientation, with the left side corresponding to the left side of the brain.
A contrast of age effects between the rumination and worry states revealed that several key regions associated with self-referential thinking and cognitive control showed greater age-related increases in rumination relative to worry. These include the right middle and inferior frontal gyrus, right supramarginal gyrus, middle temporal and occipital gyri, and the precuneus. The cerebellum and cingulate cortex also exhibited heightened engagement in rumination with advancing age, compared to worry.
To confirm that the decoder-derived time series reflected meaningful emotional brain states, we compared brain activation patterns derived from two (1) a conventional block-based GLM using “Think” periods modeled as boxcar regressors (Supplementary Fig. S1), and (2) a GLM using the decoder output time series as regressors (Fig. 5). The spatial activation patterns observed in these two analyses were broadly consistent across conditions and contrasts, supporting the validity of the decoder-derived time series in capturing relevant emotion-related brain activity during the task. However, age-related effects were only observed in the decoder-based GLM analysis, not in the block-based model, suggesting that individualized decoding may offer greater sensitivity to interindividual differences in the neural dynamics of emotional thought.
The brain states of rumination and worry, two core components of RNT, were distinguishable from each other based on whole-brain activation patterns, although the discriminability varied depending on the participant’s age. Older participants exhibited more distinct activation patterns, characterized by increased activation in cognitive control regions during rumination and decreased activation in the TPJ and anterior and middle cingulate cortex during worry. No significant age-related changes were observed in brain activation for the positive thinking state. These results highlight the importance of considering age when investigating the neural underpinnings of emotional states and the differentiation between rumination and worry. While the commonality of these RNT components has been demonstrated in psychological assessments and factor analyses^16,17,19,20,43^, incorporating a lifespan perspective on these emotional states may yield new insights into the factors driving their differentiation.
The absence of significant age effects on positive thinking-related brain activity observed in this study aligns with previous findings on lifelong changes in emotional states. Older adults often report higher levels of well-being compared to younger adults, which is linked to a reduction in negative thinking as they age^11,13^. However, this reduction does not extend to positive repetitive thoughts, which remain stable across age^14^. In the current study, while no significant correlation was found between age and positive affect or well-being scores (SCI-QOL Positive), the separability of brain activity between positive thinking and rumination was significantly associated with well-being scores (Table 2). These findings suggest that age-related changes in brain activation patterns during negative thinking may play a key role in enhancing well-being in older adults, supporting the view that modifications in negative emotional processing are crucial for improved emotional well-being with age.
Moreover, the current results showed that rumination and worry were associated with distinct alterations in brain activity with age among negative thought patterns. A significant increase in brain activity was observed during rumination, particularly in regions related to cognitive control, including the lateral frontal and superior parietal areas (Fig. 6). In addition, a greater distinction between rumination and positive thinking was linked to reduced levels of rumination and anxiety (Table 2). These findings suggest that, with age, individuals may respond more adaptively to negative memories by enhancing cognitive control, thereby reducing the rumination and anxiety associated with these memories. This aligns with the suggestion that older adults achieve well-being by optimizing the emotion regulation process^44^ and the view that aging is a process of adaptive change, rather than a decline in cognitive function^45^.
In contrast to the findings on rumination, there was a decline in brain activity associated with worry as individuals aged, particularly in the right TPJ and anterior and middle cingulate regions. The TPJ is a crucial area involved in social cognition and self-referential processing, with the right TPJ playing a central role in understanding others’ mental states and evaluating social information^46–48^. Its engagement in worry and anxiety may reflect the self-focused and socially evaluative nature of future-oriented concerns. Similarly, activity in the anterior and middle cingulate cortices has been previously associated with anxiety^49,50^ and social distress^51^. Thus, the observed age-related decline in activation across these regions may reflect a reduction in social-evaluative concern and emotional reactivity to anticipated events, hallmarks of worry, across adulthood.
These spatial activation patterns were complemented by the results of the time-resolved decoding analysis, which revealed a significant age × time interaction in the worry condition. Specifically, older adults showed a more sustained increase in worry-related decoder output over time during the Think blocks. This temporal signature may reflect a shift toward more deliberate or stable suppression of TPJ and middle cingulate activities in older individuals.
Together, these findings point to a nuanced age-related trajectory for worry-related thought. This pattern may reflect an adaptive shift in emotional processing with age, away from reactive, socially evaluative distress and toward more stable internal regulation of emotional thought. This interpretation aligns with prior work suggesting that affective reactivity decreases^11^ and emotional empathy increases with age^52,53^, supporting the notion that the reduction in social-anxiety-related neural activity may be part of a broader emotional adaptation across the lifespan. Importantly, while initial correlations suggested that discrimination between emotional states was associated with affective trait scores, these associations did not remain significant after accounting for age, indicating that age likely contributes substantially to both neural and behavioral variance in these traits.
Changes in emotional processing with age may not reflect cognitive decline but rather adaptive shifts in the functional significance of emotions^25^ and age-related goals influenced by changes in the social environment. Socioemotional Selectivity Theory (SST) suggests that as we age, we prioritize emotionally meaningful goals, savoring positive emotions while avoiding negative ones^24^. SST posits that as individuals perceive the end of life approaching, emotionally meaningful goals take precedence over exploration. The present findings complement this theory by demonstrating that the age-related decline in negative thinking is driven by distinct neural changes in rumination and worry. These results highlight the importance of examining emotional aging not only through the general valence of emotions but also by considering how aging affects individual emotions as discrete processes^25^. This perspective offers a more nuanced understanding of emotional aging, suggesting that the decline in negative thinking is tied to specific adaptive processes for each emotional state, rather than a simple shift toward positive or negative emotions.
Although the present results were based on healthy subjects, age influences the neural substrates relevant to depression, with previous research indicating distinct variations across age groups in major depressive disorder (MDD). Specifically, alterations in the global network organization of structural and functional brain connectomes in MDD vary according to the age of onset^54^. Late-onset depression, in particular, is more strongly associated with anatomical brain alterations compared to early-onset depression^55–57^. The present findings demonstrate that the RNT process also exhibits age-related variations, with distinct effects on rumination and worry. Research suggests that these two processes differentially impact mood rumination is associated with depressive comorbidity and recurrence, while worry is more closely linked to anxiety disorders^21,22^. Therefore, age-related differences may also influence the relationship between RNT and depression. Future studies should explore how age interacts with these processes to better understand their role in mood disorders.
It is important to highlight the methodological strengths of this study. We analyzed brain activity using decoding probability outputs at each time point as regressors. The observed average activity patterns (Fig. 5) closely matched those from traditional block-model analyses (Supplementary Fig. S1), validating our approach. Moreover, utilizing decoding probability allowed us to capture dynamic changes in brain activity, offering more precise and individualized modeling of emotional state transitions compared to static, predefined state models. The effectiveness of machine learning in detecting and tracking changes in mental states over time has been demonstrated in other studies^58,59^. In the present study, this approach successfully highlighted individual differences in the time course of decoded states and revealed variability within individuals across sessions and trials. By accounting for these variations, our regression analysis provided a more nuanced detection of brain activity, likely contributing to the identification of age-related effects in this study.
This study has several limitations. First, the small sample size limits the generalizability of the findings, particularly for between-subject associations such as those involving age and affective traits. While the within-subject decoding approach enhances sensitivity to individual neural patterns, larger and more diverse samples are needed to provide greater statistical power and to robustly evaluate how emotional brain states vary across adulthood. These findings should therefore be interpreted as exploratory and in need of replication.
Second, the participant pool was female-biased, which may influence the results, as sex differences have been observed in emotional processing^60^. The sample also ranged in age from 18 to 64 and therefore did not include individuals typically classified as older adults. As such, findings related to age should not be interpreted as representative of older adulthood. Future studies should aim to include a more balanced and demographically diverse sample to better understand the role of gender and age across the adult lifespan. In addition, this study was cross-sectional, making it difficult to draw conclusions about the temporal dynamics of age-related changes in emotional processing. Longitudinal studies are needed to track how rumination, worry, and their neural correlates evolve over time and to determine whether the observed changes are causally linked to aging.
Another potential consideration is the use of the flanker task interleaved between the emotional thinking blocks. While the flanker task served as a neutral attentional control and helped structure the session, we did not include an explicit measure of participants’ ability to disengage from their emotional thoughts during this task. However, the task was designed to be non-frustrating and did not include performance feedback, minimizing the likelihood of it eliciting negative affect. Supporting this, the classifier showed an increase in Rest classification probability during the flanker task (Fig. 4), suggesting that participants’ brain states shifted away from emotional episodic thinking during these periods. This pattern indicates that the flanker task likely served its intended purpose of facilitating a return to a neutral cognitive state before the next emotional induction block.
While our use of autobiographical prompts enhanced ecological validity by capturing personally meaningful emotional states, it may also limit generalizability to studies employing standardized or externally cued emotional stimuli. The operational definitions of rumination, worry, and positive thinking were tailored to elicit self-relevant, episodic thought patterns, which likely shaped the brain activation patterns used for model training. As such, the decoding performance and neural representations observed in this study should be interpreted within the context of these individualized, self-referential experiences. While this design choice may diverge from conventional definitions of emotional states, it represents a key strength of the study by allowing us to investigate the neural signatures of internally generated, naturalistic emotional thought.
In addition, as detailed in the Supplementary Materials (“Model Profiles in Auto-sklearn Classifier Training”), the classifiers were primarily linear models selected through an automated process, and model choice was not systematically related to age or classification performance. These factors should be considered when interpreting the scope and transferability of our findings.
In conclusion, the present results suggest that emotional brain states evolve across the lifespan, enhancing emotional well-being in older adults, particularly through changes in negative thinking processes. Notably, these changes are not uniform across individual negative emotional states such as rumination and worry. Aging appears to enhance cognitive control during rumination and reduce neural responses to anxiety-provoking events during worry. The individuality of emotional brain states and their age-related evolution should be a key consideration in promoting lifelong mental health and developing personalized interventions for mental disorders. Understanding these dynamics could pave the way for more targeted and effective mental health strategies across different life stages.
Thirty-seven healthy individuals (24 females, aged 18–64) participated in the study. Exclusion criteria included a positive screen for recent use of alcohol or recreational drugs, a lifetime diagnosis of psychiatric disorders, and significant issues that could affect participation, such as uncorrectable vision or hearing problems and MRI contraindications. Individuals with moderate to severe traumatic brain injury, active suicidal ideation, recent changes in medication affecting brain function, or prescriptions outside of accepted ranges were also excluded. All participants provided informed consent in accordance with the principles outlined in the Declaration of Helsinki. The study procedures were approved by the WCG IRB (https://www.wcgclinical.com; tracking 20224917). All ethical regulations relevant to human research participants were followed.
The study sessions consisted of three one preparation session and two MRI scanning sessions. During the first visit, participants received study information, provided informed consent, and completed a battery of assessments measuring tendencies toward negative and positive thinking. Although a comprehensive set of affective trait measures was collected for use in other study components, the present analysis focused specifically on traits related to rumination, worry, and positive thinking.
Rumination was assessed using the Ruminative Responses Scale (RRS)^40^, which measures the tendency to engage in repetitive thoughts about the causes, meanings, and consequences of one’s distress. Worry was assessed using the Penn State Worry Questionnaire (PSWQ)^42^, which evaluates the general tendency to engage in excessive and uncontrollable worry. To capture broader anxiety-related traits, we also administered the State-Trait Anxiety Inventory—Trait measure (STAI-T)^38^, which assesses general distress or negative affect^39^, often comorbid with worry and rumination. Positive affect was assessed using the SCI-QOL Positive Affect and Well-Being scale^41^. Internal consistency values (Cronbach’s alpha) for each of these scales are reported in Table 1. We also collected several additional questionnaires related to depression, state rumination, state anxiety, and emotion regulation for exploratory purposes. These included the Quick Inventory of Depressive Symptomatology (QIDS)^61^, the Brief State Rumination Inventory (BSRI)^62^, the State-Trait Anxiety Inventory—State measure (STAI-S)^38^, the Metacognitive Questionnaire-30 (MCQ-30)^63^, the Thought Control Questionnaire (TCQ)^64^, the Emotion Regulation Questionnaire (ERQ), the Difficulties in Emotion Regulation Scale (DERS)^65^, and the Cognitive Emotion Regulation Questionnaire (CERQ)^66^. Although these measures were primarily collected for a separate component of the study involving an emotion regulation task, their correlations with AUC-based neural discriminability are reported in Supplementary Table S2 for completeness.
During the second visit, participants were asked to recall twelve autobiographical events, each clearly associated with one of three different thinking rumination (negative thoughts about the past), worry (anxious thoughts about the future), and positive thinking (positive thoughts about the past). Participants generated four events for each category and wrote down short descriptions (keywords) to help them recall these events during the MRI session. The Supplementary Materials (“Instructions for collecting personal events related to rumination, worry, and positive thinking”) provide the exact instructions given to participants for generating these events. To maintain confidentiality, no details of the personal events were requested, allowing participants to honestly recall their negative personal experiences. Participants evaluated the distress levels associated with each of the rumination and worry event using a seven-point Likert scale (1: Not distressing at all, to Extremely distressing).
During the MRI session, participants performed the Think Affective Events Task (TAET) with functional MRI (fMRI), as detailed below. In the third visit, participants completed another emotion regulation task, which is not included in the present paper and will be reported separately.
The MRI scan was performed using a 3 T Discovery MR750 scanner (GE Healthcare, Milwaukee, WI, USA) with a 32-channel head coil. Brain functional images of blood oxygenation level-dependent (BOLD) signals were collected using gradient echo-planar imaging of T2*-weighted signals with the following TR = 2.0 s, TE = 25 ms, FA = 90°, SENSE acceleration factor (R) = 2, number of axial slices = 40, slice thickness = 2.9 mm, FOV = 240 × 240 mm, and matrix = 96 × 96, resampled to 128 × 128.
The anatomical reference image of the brain was obtained using T1-weighted imaging with the MPRAGE sequence with TR = 6 ms, TE = 2.92 ms, SENSE acceleration factor (R) = 2, flip angle = 8°, inversion time = 1060 ms, sampling bandwidth = 31.25 kHz, FOV = 256 × 256 mm, 208 sagittal slices, slice thickness = 1.0 mm, and a scan time of 6 min 11 s.
After preparation runs for locating the imaging position, the scan session began with the T1-weighted anatomical scan, followed by the resting-state run (6 min 50 s), and then six runs of the Think Affective Events Task (TAET), each lasting 8 min 2 s (Fig. 1). Each run of the TAET began with an 18-s Rest block, followed by a 22-s Think block and an 18-s Flanker task block. These blocks were repeated eight times per run and concluded with an 18-s Rest block.
During the Rest block, a white cross was displayed on the screen, and participants were instructed not to think about anything in particular. During the Think block, keywords representing one of the personal events collected prior to the scan were shown along with an instruction sentence, such as “Why did that happen to you?”, “Why did you do that at the moment?”, or “Why did you react that way?” for rumination, and “How would this affect your life?”, “How would you feel if that happened?”, or “What would you do if that happened?” for worry. The instruction sentence appropriate for each event was selected by the participants during the thought collection process. Participants were instructed to engage with the thought related to the event’s keywords according to the instruction sentence.
In the Flanker task block, five arrowheads were presented, and participants were instructed to respond to the direction of the center arrow by pressing the left or right button as quickly as possible. The four flanker arrows pointed in the same direction, while the center arrow could point in the same (congruent) or opposite (incongruent) direction. Once the button was pressed, the arrows disappeared without any feedback. Trials were presented every 3 s, regardless of response time, and six trials were performed in each block. This task was designed to impose an attentional load and was placed after the Think block to help participants clear their thoughts before the next Rest block.
Six TAET runs were performed during the session. Each of the 12 thought events was presented four times during the session, and the order of events was pseudorandomized with the following events from the same category (rumination, worry, positive) did not repeat consecutively, each run included approximately the same number of events (two to three) from each category, and the same event did not repeat within the same category.
At the end of each task run, participants answered five questions about their mental states during and after the task using a seven-point Likert “How difficult was it to engage with your thoughts during the Think blocks?” (difficulty), “How successful were you in engaging with your thoughts during the Think blocks?” (success), ‘How sleepy were you during the scan?’ (sleepiness), “How emotionally drained do you feel right now?” (emotional exhaustion), and “How tired are you right now?” (tiredness).
We used AFNI (https://afni.nimh.nih.gov/) to process the MRI image data. The initial three fMRI volumes were discarded to allow the signal to reach a steady state. The following preprocessing steps were despiking, RETROICOR^67^ and RVT^68^ physiological noise correction, slice-timing correction, motion correction, spatial normalization to the MNI template brain with resampling to a 3 × 3 × 3 mm voxel size using Advanced Normalization Tools (ANTs)^69^, spatial smoothing with a 6 mm full-width at half maximum (FWHM) kernel within the brain, and scaling to percent signal change in each voxel.
Noise components were regressed out using a general linear model (GLM) analysis. The noise components included low-frequency drift modeled by up to third-order Legendre polynomials, six motion parameters (three rotations and three translations) and their temporal derivatives, three principal components of the ventricle signals, and local white matter average signals (ANATICOR)^70^. Response models for each of the Flanker task trials were also included, with congruent and incongruent trials modeled separately using a canonical hemodynamic response function (HRF) convolved with a delta function at the onset of each event.
Volumes with frame-wise displacement greater than 0.3 mm, along with their preceding volumes, were censored in the GLM analysis. Task runs in which more than one-third of time points were censored were excluded from analysis. The residuals from the GLM analysis were used as input for the decoding analysis of thought states.
The thought state at each time point during the Think and Rest blocks was classified using a machine learning model trained to distinguish among four rumination, worry, positive thinking, and rest. This four-class design allowed the model to learn not only the distinctions between the three target emotional states, but also to differentiate them from a neutral baseline (rest). Including rest as a fourth category was critical for preventing the decoder from learning only relative contrasts between emotional states (e.g., labeling the least negative state as “positive” by default), which could result in biased or miscalibrated outputs when applied across the full time course. This approach improved the decoder’s ability to capture meaningful distinctions in internal thought states and enabled its application to the continuous time series of brain activity throughout the entire scan.
The input signals were extracted from the gray matter voxels of the preprocessed images across the whole brain. The data for each thought state were sampled during the Think block, excluding the initial 3 time points (6 s) to account for the hemodynamic response delay, providing 8 time points per block and 128 samples per thought category. Data for the rest state were sampled during the Rest block, excluding the initial 6 time points (12 s) to avoid any residual effects from the Flanker task and to approximately equalize the number of samples with the other thought states. This provided three time points per block and 144 samples in total. Censored time points were excluded from the input to the decoding analysis.
The classification analysis was performed using the Auto-sklearn package (https://automl.github.io/auto-sklearn/)^71^. This package runs an AutoML analysis, which performs Combined Algorithm Selection and Hyperparameter optimization (CASH) to find the optimal model. Auto-sklearn creates ensembles of data preprocessors, feature preprocessors, and machine learning models from the scikit-learn (https://scikit-learn.org) machine learning package in Python. The model training was performed using default parameters, except for the time limit for model training. In the default configuration, the Auto-sklearn search was time-limited to 6 min per model training call and 1 h for the entire task, given the extensive parameter space in the CASH process. We extended these limits to 3 h per model training call and 30 h for the total task to increase the likelihood of identifying better-performing models. In addition, we tested limits of 6 h per model call and 60 h in total, but observed no significant performance improvement, suggesting that model performance plateaued with the 3-h and 30-h configuration. The training script is provided in Supplementary Data 2.
Training and evaluation were performed using a leave-one-run-out cross-validation scheme. In this setup, a model was trained on data from five TAET runs and tested on the remaining run. This process was repeated such that each of the six runs served once as the test set. To optimize model parameters, we applied a nested leave-one-run-out cross-validation procedure within the training set. The held-out test run was never used during training or optimization and was reserved exclusively for model evaluation.
Model performance was quantified using the area under the receiver operating characteristic curve (AUC). We did not use classification accuracy as a performance metric because it can be biased in the presence of class imbalance, which may arise from motion-related censoring of samples across categories. Such imbalances can make accuracy scores less comparable across participants or conditions. The AUC was calculated using data points from the same time window used for extracting training data, specifically, from the Rest and Think blocks, after removing the initial TRs to account for the hemodynamic response delay and any residual activation from the preceding Flanker task.
To identify brain regions associated with decoded thought states, we used the classifier-derived time series as parametric regressors in a GLM. For each test run, the trained classifier was applied to every time point (TR), generating continuous output probabilities for each thought category (e.g., rumination, worry, positive thinking). These time-resolved model outputs captured the BOLD response to the moment-to-moment evolution of internal mental states throughout the run. Rather than averaging brain activity within fixed block durations, this approach leveraged the full prediction time course to characterize internal dynamics with greater temporal precision, capturing intra-block fluctuations and individualized timing that may not align precisely with task blocks^59^.
The GLM included the decoded time series for all three thought categories, along with the same nuisance regressors described above. It was applied to the original preprocessed fMRI data without prior noise regression. This analysis was distinct from the GLM used in the decoding pipeline, which was conducted solely to remove noise components prior to classification.
As a complementary exploratory analysis, we also ran a traditional GLM using block-based regressors, boxcar functions corresponding to each Think block convolved with a canonical hemodynamic response function (HRF). This allowed us to compare activation maps derived from the decoded time series with those based on the task timing alone, thereby evaluating the added value of data-driven, individualized decoding over conventional block modeling.
The significance of AUC for thought state discrimination was assessed for each participant using a normal approximation of Mann–Whitney U statistics^72^. The associations between thought state discriminability (AUC) and each of the demographic variables and assessment scores were tested using the Spearman correlation, except for sex, where a t test was used. The P values across these variables were corrected by false discovery rate (FDR)^73^.
The effect of age on the temporal pattern of decoded state probability was tested using a linear mixed-effects (LME) model analysis. The dependent variable was the decoder output probability corresponding to the instructed thought condition on each trial (i.e., the probability assigned to rumination, worry, or positive thinking, depending on the target state). Fixed effects included Time (TR), Age, and their interaction, with a random intercept for each participant. The Time variable was entered as a factor variable since the probability time series was not linearly evolving. This analysis focused on time points used during decoder training (from 6 s after block onset to the end of the Think block), allowing consistent comparison across trials while accounting for the BOLD response delay.
Group analysis of the BOLD signal change associated with thought states was performed using LME model analysis. Beta maps for each thought condition and run, obtained from the first-level GLM analysis with decoded time series, were used as the dependent variable. Fixed effects included thought condition, participants’ age, and their interaction, with participant included as a random effect on the intercept. Voxel-wise LME analyses were conducted using the lme4 package^74^ in the R language for statistical computing (version 4.3.2), and post-hoc contrasts were estimated using the emmeans package^75^. This analysis was implemented via a custom script that automated voxel-wise data extraction and model fitting. The statistical map from the LME analysis was thresholded at P < 0.001 voxel-wise, and a cluster extent threshold of P < 0.05 was then applied. The cluster extent threshold was evaluated using AFNI 3dClustSim with an improved spatial autocorrelation function^76^.
This study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki. The study protocol was reviewed and approved by the WCG IRB (https://www.wcgirb.com). All participants provided written informed consent prior to their inclusion in the study.
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3-13 Reporting Summary