Generalizable prediction of stimulus-independent, task-unrelated thought from functional brain networks

Neural substrates of “mind wandering” have been widely reported, yet experiments have varied in their contexts and their definitions of this psychological phenomenon, limiting generalizability. We aimed to develop and test the generalizability, specificity, and clinical relevance of a functional brain network-based marker for a well-defined feature of mind wandering—stimulus-independent, task-unrelated thought (SITUT). Combining functional MRI (fMRI) with online experience sampling in healthy adults, we defined a connectome-wide model of inter-regional coupling—dominated by default-frontoparietal control subnetwork interactions—that predicted trial-by-trial SITUT fluctuations within novel individuals. Model predictions generalized in an independent sample of attention-deficit/hyperactivity disorder (ADHD) adults. In three additional resting-state fMRI studies (total n=1,115), including healthy and ADHD populations, we demonstrated further prediction of SITUT (at modest effect sizes) defined using multiple trait-level and in-scanner measures. Our findings suggest that SITUT is represented within a common pattern of brain network interactions across time scales, populations, and contexts.


Introduction
Humans spend a large proportion of their waking lives engaged in mind wandering, 1,2 often defined as self-generated experiences that are decoupled from immediate environmental inputs and current tasks at hand. [3][4][5] The tendency to mind wander varies substantially within and between individuals, and this variability can have severe, broad mental health consequences in conditions such as attention-deficit/hyperactivity disorder (ADHD) and Alzheimer's disease. [6][7][8] Thus, understanding the nature of mind wandering and its neural basis has emerged as a central goal within cognitive and clinical neuroscience. 4,6,9,10 Within the last decade, neuroimaging combined with online experience samplingwherein people are intermittently prompted to report their current thoughts-has become a well-established, powerful method to link brain dynamics with mind wandering episodes. [11][12][13][14][15][16] Initial functional MRI (fMRI) studies using this and related techniques confirmed an association between mind wandering and default mode network (DMN) activation, 11,17,18 a finding now also supported by fMRI 19,20 and intracranial electrophysiology 21 evidence of increased DMN activity preceding behavioral lapses (despite a nuanced role in task engagement 13,14,20 ). Subsequent research has additionally revealed the critical role of distributed, dynamic, network-level interactions within and beyond the DMN, 6,9,12,13,[22][23][24][25] including functional coupling patterns of the frontoparietal control network (FPCN) 25,26 and primary sensory/motor regions. 13,24,[27][28][29][30] Further lesion and neurostimulation evidence suggest potential causal roles of core DMN [31][32][33][34][35] and FPCN. 36,37 regions. Despite this multimodal evidence for key roles of distinct regions and their interplay, mind wandering has been defined in distinct manners and in different contexts across experiments, 3 limiting generalizability of findings. It remains unknown whether a common functional network pattern could provide a specific, generalizable marker of any specific aspect of mind wandering within and between individuals, across distinct populations, and across multiple contexts.
Predictive modeling of neuroimaging data 38,39 has emerged as a promising tool for developing generalizable functional network markers of various cognitive and affective functions. [40][41][42][43] Data-driven, multivariate approaches within a predictive modeling framework could be fruitful in the study of mind wandering, given the hypothesized role of interactions within and between multiple distributed brain networks. 6,9,10 Indeed, prior research suggests that multivariate features, based on within-and between-DMN functional connectivity, carry predictive information about task-unrelated thought in healthy adults (at the within-dataset, single-task context level). 12,44 Establishing broader generalizability of a network-based model would have significant implications for understanding the brain bases for conditions such as ADHD, where excessive mind wandering is strongly associated with clinical outcome. 8 Additionally, a predictive model may have broad value for the interpretation of functional neuroimaging data because mind wandering may be an important, but typically unexplained, source of brain variability within any given experimental or clinical session.
Here we aimed to identify and test the generalizability, specificity, and clinical relevance of a brain network-based marker for a well-defined feature that is considered central to mind wandering-stimulus-independent, task-unrelated thought (SITUT). [3][4][5] First, using experience sampling in healthy adults, we defined a functional network-based model that predicted trial-bytrial fluctuations in SITUT within novel individuals. We tested the model's validity and its specificity to the construct of SITUT relative to related, but not equivalent constructs (e.g. sustained attention) and characterized the model's features in relation to functional neuroanatomy previously linked to mind wandering. Next, we tested whether the same model predicted trial-wise SITUT within ADHD adults who reported an increased frequency of SITUT during experience sampling. Finally, in three independent resting state fMRI (rs-fMRI) datasets from distinct populations, we tested whether the model predicted individual trait and state mind wandering, as defined with multiple distinct measures of daily life and in-scanner SITUT.

1.0: A functional connectivity pattern predicts SITUT within healthy adults
To identify a functional network-based marker of SITUT, we analyzed data from healthy adults who participated in an fMRI study with online experience sampling at Massachusetts General Hospital (MGH). Participants completed the Gradual-Onset Continuous Performance Task 20 (gradCPT) while receiving intermittent thought probes every 44-60 seconds. The paradigm uniquely allowed us to capture unique neural predictors of SITUT while accounting for behavioral performance (reaction time) fluctuations. 14 Upon each thought probe (36 total trials), participants provided subjective ratings of task-related focus on a graded scale, with 0 indicating fully "on-task" and 100 indicating fully "off-task." We retained for initial analysis only those participants who reported that their "off-task" reports were due to task-unrelated thoughts that were also stimulus-independent (17 participants and a total of 612 trials; see Methods).
We applied connectome-based predictive modeling (CPM) 45 to model and predict trialwise, intra-individual fluctuations in SITUT from functional connectivity within a leave-oneparticipant-out (leave 36 trials out) cross-validation framework. For each thought probe (trial), we computed the functional connectivity matrix within the 30-second window (28 fMRI frames) preceding probe onset based on a 268-node whole-brain functional atlas (Shen268) (Fig. 1a).
Within each cross-validation fold (comprising 16 participants), we identified all node pairs (edges) exhibiting a suprathreshold-level positive or negative correlation with within-participant normalized SITUT ratings. Based on the positive minus negative edge sum scores for each trial, we constructed a linear model to predict SITUT based on all 576 trials (i.e., 16 participants x 36 trials) within a given cross-validation fold. For the held-out one participant (36 trials), we applied this linear model to compute predicted trial-wise SITUT, which we correlated with observed SITUT rating and then compared the correlation value with a null distribution (Fig. 1b).
Predicted versus observed SITUT correlations were significantly greater than mean permutation test-based null values in held-out participants (M±SD within participant r = 0.11±0. 16, P = 0.019, Wilcoxon signed rank test; MSE = 1.19±0.15) (Fig. 1c). This provided evidence that our functional network model was predictive of SITUT within novel individuals at the within-dataset level (i.e., establishing internal validation) and with an overall effect size that was on par with that typically found for functional connectivity-based prediction of self-report outcomes. 46 The edges contributing to the model (hereafter referred to as "SITUT-CPM" masks) included 258 and 139 edges, respectively, positively and negatively associated with SITUT.
These edges were distributed widely throughout the brain, with high-degree nodes (i.e., nodes involved in multiple contributing edges) situated in prefrontal, parietal and temporal cortices and cerebellum (Fig. 1d). Edges that were correlated with SITUT rating were identified at a threshold of P<0.01 (uncorrected), and a summary score was obtained for each trial based on a subtraction of positive and negative edge sums.

1.1: Validity and specificity of the SITUT-CPM.
We next performed confirmatory analyses to assess the validity and specificity of the SITUT-CPM. Despite performance of fMRI preprocessing steps that aim to limit the impact of nuisance factors on functional connectivity (as done here; see Methods), frame-wise head motion can still influence observed relationships between functional connectivity and behavior. 47,48 However, we found that SITUT-CPM compared to null model prediction remained significant when using partial correlations within held-out participants in each cross-validation fold, controlling for mean frame-wise head motion (r partial = 0.091±0.15, P = 0.039). Additionally, though our main analysis was based on data preprocessed with an aCompCor 49 noise-correction pipeline, prediction remained significant when using an alternative ICA-AROMA 50 approach to control for head motion (r = 0.10±0.14, P = 0.0086). When using an alternative 300-node functional atlas (Schaefer300), trial-wise SITUT-CPM network strength was similar to that initially estimated with the Shen268 atlas (withinparticipant trial-wise correlation between Shen268-versus Schaefer300-derived network strengths: r = 0.87±0.054), and prediction of SITUT remained significant (r = 0.12±0.23, P = 0.044).
Given that SITUT and sustained attention are inter-related, but not equivalent constructs, 51 we tested for specificity of SITUT-CPM predictions using a simultaneous online behavioral measure, reaction time (RT) variability, indicative of sustained attention 14,20,51 and previously shown to be positively correlated with SITUT. 5,14,[52][53][54] In our participants, self-reported SITUT was positively associated with pre-thought probe (30-sec mean) RT variability across trials (F 1,612 = 8.70, P = 0.0033; F test on linear mixed effects model), suggesting behavioral validation of subjective reports but also implying overlapping variance between the two measures. However, SITUT-CPM compared to null model prediction remained significant when controlling for pre-thought probe RT variability in held-out participants (partial r partial = 0.10±0.16, Moreover, we compared our newly defined SITUT-CPM to the previously defined Sustained Attention CPM (SA-CPM) shown to be predictive of gradCPT behavioral performance within and between individuals. 41,55 The SITUT-CPM and SA-CPM masks did not show significant edge overlap (Fig. S1a). Trial-wise SITUT-CPM versus SA-CPM network strengths were significantly negatively correlated at the group-level (r = -0.13±0.21, P = 0.031) (Fig. S1b), suggesting an expected relationship of higher expression of a brain marker for SITUT associated with lower expression of a marker for sustained attention. However, when controlling for SA-CPM network strength, SITUT-CPM compared to null model prediction of observed mind wandering remained significant (r partial = 0.096±0.16, P = 0.028), suggesting independence between the behavioral relevance of the SITUT-CPM versus SA-CPM. Moreover, SA-CPM predictions were not significantly associated with observed SITUT (r = -0.073±0.16, P = 0.093), despite an expected negative trend.
Additionally, although SITUT and creative thinking may draw from similar cognitive processes, 56 the SITUT-CPM showed independence from a previously published CPM 40 based on individual differences in creative ability (Cr-CPM). Though there was some significant edge overlap between the SITUT-CPM and Cr-CPM (Fig. S1c), trial-wise network strengths from the two CPMs were independent from one another (r=-0.029±0.18, P=0.94) (Fig. S1d), and SITUT-CPM prediction remained significant when controlling for Cr-CPM strength (r partial = 0.11±0.17, P=0.031). Collectively, these findings provide evidence for validity and specificity of the SITUT-CPM.

1.2: Functional neuroanatomical basis of the SITUT network. For improved
interpretation of the functional neuroanatomical basis of patterns contributing to the SITUT-CPM, we examined relationships with well-described functional networks previously linked to mind wandering. Based on the Schaefer300 atlas, with each node assigned to 1 of 7 standard Yeo-Krienen networks, 57 we quantified the number of SITUT-CPM mask edges belonging to each intra-or inter-network pair. Among edges positively correlated with SITUT, DMN withinand between-network connections contributed most strongly (Fig. 2a). The top 5 network pairs contributing to positive edges were DMN-FPCN, DMN-DMN, sensorimotor-visual, DMNsensorimotor, and sensorimotor-dorsal attention. The DMN-FPCN contribution, which was greatest (61 total pairs), was explained by anticorrelated activity during task-focused trials and lack of anticorrelation during SITUT trials (Fig. 2b). This finding aligns with prior theoretical and empirical research supporting a role of DMN-FPCN interactions in various forms of mind wandering. 6,11,58 Among edges negatively correlated with mind wandering, DMN-sensorimotor network (SMN) pairs (30 total) contributed most strongly (see Fig. 2c for other contributions). The DMN-SMN contribution was explained by positive correlation during task-focused trials and anticorrelation during SITUT trials (Fig. 2d). This finding aligns with the notion that during mind wandering, the DMN engages its capacity to draw from information unrelated to immediate sensory input. 24,59 Given that within-network functional heterogeneity is a key consideration within current theoretical models of mind wandering, 6,10 we examined the specific contributions of subnetworks based on the Yeo-Krienen 17-network atlas. This revealed that positively contributing DMN-FCPN edges were largely explained by specific interactions between DMN A (as labeled within the Schaefer atlas), a subnetwork including the medial prefrontal and posteromedial cortices (consistent with the DMN "core" 60 ), and FPCN A , 61 a subnetwork anchored in rostrolateral prefrontal cortex (Fig 2e). Negatively contributing DMN-SMN edges were largely explained by interactions between the DMN A and SMN A , a subnetwork comprised of primary somatosensory and motor cortices (Fig. 2f). In summary, although SITUT-CPM predictions were based on a complex, distributed pattern of interacting networks, key components associated with increased SITUT were a) decreased DMN A -FPCN A anticorrelation; and b) decreased DMN A -SMN A correlation (Fig. 2g).  "SITUT" and " Task

2.0: External validation #1: Intra-individual prediction of mind wandering in ADHD
Having established the neurophysiological plausibility of the SITUT-CPM, we turned next to establishing external validity within independent datasets. First, we examined generalizability in ADHD adults (n = 20) who participated in the same gradCPT experience sampling study as the healthy adult group (i.e., with the same MRI scanner at MGH). We hypothesized that a) the SITUT-CPM-generated within healthy adults-would predict intra-individual SITUT ratings in an independent sample of ADHD participants (i.e., demonstrating external validation); and b) given the association between increased mind wandering frequency and ADHD symptoms, 8,62 the SITUT-CPM network strength would be 'over-expressed', on average, within ADHD compared to healthy participants throughout task performance.
Correlations between SITUT-CPM prediction (based on the healthy participant model) and observed SITUT in ADHD participants were significantly greater than mean null values (r = 0.045±0.086, P = 0.028) (Fig. 3a), suggesting generalizability from the healthy to ADHD population. To further confirm this generalizability, we also independently generated a CPM to predict intra-individual SITUT within the ADHD, rather than healthy, sample (using leave-one-

3.0: External validations #2-4: Inter-individual prediction of trait SITUT from rs-fMRI
We have so far established a functional network model that carries predictive information about intra-individual SITUT, measured online during task performance, within healthy and ADHD individuals. In the next analyses (external validations 2-4), we focused on whether the SITUT-CPM is sensitive to trait-level measures of SITUT based on rs-fMRI, which measures spontaneous brain activity in an awake, "task-free" condition. 63 We tested whether the SITUT-CPM was predictive of individual differences in SITUT tendencies in three rs-fMRI datasets

3.1: External validation #2: Superstruct dataset (Healthy Adults). First, we examined
predictions of trait mind wandering within 911 healthy adults from the Brain Genomics Superstruct project. These participants completed one 6-minute run of rs-fMRI and, separately, the Daydreaming Frequency Scale (DDFS) to characterize individual mind wandering tendency.
We found a significant positive correlation between DDFS score and SITUT-CPM prediction from resting state functional connectivity (Spearman's ρ = 0.074, P = 0.025) (Fig. 4a), which remained significant when controlling for frame-wise head motion (partial Spearman's ρ = 0.080, P = 0.016) and age (partial ρ = 0.072, P = 0.029). Though the effect size was modest, we examined the specificity of DDFS prediction relative to a comprehensive battery of 67 distinct behavioral and self-report individual outcomes acquired among participants in the dataset.
Among this entire set of individual outcomes, SITUT-CPM prediction showed the strongest correlation with DDFS score, including in comparison to correlation with performance on a Flanker task purported to more generally capture attentional and inhibitory control abilites 67 (Fig.   4b). Further demonstrating specificity, SA-CPM-based predicted sustained attention was not significantly correlated with DDFS (Spearman's ρ = 0.022, P = 0.50).

3.3: External validation #4: MIT dataset (ADHD Adults).
As a further test of generalizability to rs-fMRI data in a clinical context, we investigated sensitivity to trait SITUT in an independent ADHD sample. Previous findings indicate that ADHD individuals with high Mind Wandering Questionnaire (MWQ) scores (> 23) exhibit increased severity of ADHD symptoms. 8 Thus, we classified 49 ADHD adults into high (n = 31) versus low (n = 18) SITUT subgroups, based on the MWQ, within a cohort of patients that underwent rs-fMRI (7 mins each) at Massachusetts Institute of Technology. These two groups did not show significant differences in head motion (P = 0.43, Wilcoxon rank sum test, M±SD of 0.092±0.027 mm for high SITUT participants, 0.088±0.027 mm for low SITUT participants), age (P = 0.097; M±SD of 32.0±7.5 for high SITUT participants, 28.2±7.0 for low SITUT participants) sex (high SITUT: 18 females, 13 males; low SITUT: 7 females, 11 males), or ADHD medication status (high SITUT: 27.6% using medication; low SITUT: 25% using medication). However, across the entire group, MWQ score was positively associated with inattention and hyperactivity symptom severity within the past 6 months (inattention: ρ = 0.34, P = 0.017; hyperactivity: ρ = 0.34, P = 0.017) and showed a weaker, but also positive association with diagnostic symptom severity (inattention: ρ = 0.23, P = 0.11; hyperactivity: ρ = 0.23, P = 0.12), in line with prior findings from a larger cohort of ADHD patients. 8 Leipzig Dataset (n=144) Healthy Adults
We further examined whether SITUT-CPM and SA-CPM network strengths were related to ADHD symptom severity (rather than MWQ scores). As shown in Supplementary Table 1, SITUT-CPM network strength was not significantly predictive of symptoms (though trends were positive), while SA-CPM showed its strongest positive relationship with diagnostic inattention severity. These results suggest that SITUT-CPM network strength was more strongly associated with a SITUT-based outcome (MWQ score) than with ADHD symptom severity, whereas SA-CPM network strength was more strongly associated with symptom severity (inattention) than with MWQ score. Taken together, in addition to SITUT-CPM sensitivity to state dynamics within ADHD adults (Fig. 3), these findings provide evidence for sensitivity and specificity to clinically-relevant individual differences within an independent ADHD cohort.

4.0: State-dependent dynamics of the SITUT-CPM during rs-fMRI
Our findings so far suggest that the SITUT-CPM is sensitive to both state SITUT within individuals and trait SITUT between individuals. One possible explanation for this  (Fig. 7, black). Post-hoc comparisons revealed a significant decrease in "surroundings" rating following run 2 compared to 1 (P FDR = 0.016; Wilcoxon signed rank test), and the overall trend suggested that the rating remained relatively stable following this post-run 1 decrease. This 'tuning out' over time was expected, given that SITUT intensity generally increases gradually following a task's onset. 68 Importantly, thoughts about surroundings were dissociable from self-reported wakefulness, which did not decrease over time (Fig. S3).
At the neural level, SITUT-CPM network strength also showed a significant interaction with rs-fMRI run number (F 1,419.73 = 14.56, P = 0.00016; F test on linear mixed effects model).
The general trend indicated an increase in network strength following the first run (Fig. 7, blue), with post-hoc tests revealing significant increases for run 3 compared to 1, 3 compared to 2, and 4 compared to 1 (P FDR = 0.0027, 0.0039, and 0.035, respectively; Wilcoxon sign rank tests).
Thus, across the time scale of ~1 hour, thought content gradually diverged from immediate surroundings while SITUT-CPM network strength increased within individuals. However, network strength was not related to run-to-run changes in the "surroundings" rating (

Discussion
Here we developed a neuroimaging-based model (SITUT-CPM) of functional brain networks that reliably predicted trial-by-trial fluctuations in stimulus-independent, task-unrelated thought-a key feature of mind wandering-within novel individuals. The SITUT-CPM was based on a distributed functional connectivity pattern that varied across 30-second windows and included major contributions of within-and between-network interactions strongly linked to mind wandering (operationalized in various ways) in prior theoretical and empirical research. 6,9,12,24,58 The SITUT-CPM predictions were generalizable across healthy and ADHD adults and were sensitive to the increased frequency of SITUT in ADHD. Moreover, in three independent resting state fMRI samples, including both healthy and ADHD populations, we found consistent evidence for further generalizability (with modest effect sizes) to predictions of trait and state mind wandering, as defined with multiple distinct measures of daily life and in-scanner SITUT.
Though prediction effect sizes were modest, they were in line with typical functional networkbased predictions of self-report outcomes in large samples 46 and were striking given the various differences among datasets (e.g. online task/experience sampling versus rs-fMRI of varying duration as well as different MRI scanners/protocols, populations, and outcome measures). Our findings suggest that SITUT is represented within a common pattern of brain network interactions across multiple time scales, populations, and contexts. We discuss how our findings fit within and advance the neuroscience of mind wandering, clinical implications of these findings, and broad implications for rs-fMRI biomarker development.

Neural basis of mind wandering
Our findings build on prior work demonstrating neural substrates of mind wandering, typically operationalized based on task-unrelated thought, 11,12,69,70 but sometimes including measurements of stimulus-independence, 22,71 multi-dimensional features of thought contents 13,29 or other single features. 15 Focusing on a SITUT-based neural model, we were able to demonstrate generalizability and robustness to multiple distinct definitions of mind wandering that are typically used in the field. We found that the distributed nature of interactions within and between distinct brain networks was critical to generalizability of SITUT-CPM predictions. Many of the intra-and inter-network interactions that strongly contributed to SITUT-CPM predictions are consistent with findings from prior neuroimaging, electrophysiological, neurostimulation, and lesion studies of mind wandering. Some less dominant, but contributing, interactions were not necessarily anticipated (Fig. 2), and an understanding of their potential significance requires further study.
Increased coupling (or reduced antagonism) between DMN and FPCN, a dominant SITUT-CPM feature, has previously been associated with goal-directed cognition with an internal focus 72 and was hypothesized to underlie the top-down control needed to deliberately maintain a stream of thought. 6 Another key SITUT-CPM component was increased anticorrelation between the DMN and primary sensory and motor regions in SMN A . Compared to other cortical association networks, the DMN is uniquely situated at the furthest geodesic distance and longest connectivity path from primary sensory regions (including SMN), potentially suggesting a capacity to draw from information untethered to immediate sensory input. 75 Thus, DMN-SMN A antagonism may reflect the perceptual decoupling (i.e., dampened processing of sensory input) and decreased motor preparedness that is characteristic of SITUT. 24,59 The coupling patterns of primary sensory/motor regions with association networks, including DMN and FPCN, have previously been linked to individual differences in specific contents of experience during mind wandering. 27,28 Within-DMN connectivity contributed both positively and negatively to the SITUT-CPM, consistent with evidence for within-DMN functional heterogeneity. 60,76 Prior research suggests that individual differences in trait mind wandering are both positively and negatively associated with within-DMN functional connectivity. 18,23,26,33 The SITUT-CPM takes this functional heterogeneity into account, which likely contributed to the success of predictions across contexts and populations. Importantly, our model was based on functional connectivity-based features only. This methodological choice was based on theoretical (i.e., hypothesized role of intra-and internetwork dynamic communication 6,9 ) and practical considerations. Consistent with predictive modeling studies in other domains, 40,41,55 our network-based approach facilitated model testing across a broad set of external datasets (e.g,. rs-fMRI scans in which baseline regional activation cannot be readily inferred). In prior research, functional connectivity, regional activity, and pupil diameter carried complementary predictive information about task-unrelated thought. 12 It is likely that inclusion of such measures and further multimodal recordings could improve model performance 44,77 ; however, large datasets that would be needed to test external validity of such richer models are currently unavailable.

Clinical implications
Our findings from two separate ADHD cohorts highlight SITUT-CPM generalizability to a clinical context and offer novel neurophysiological evidence for previously identified, clinically relevant ADHD subgroups. 8 We hypothesize that ADHD patients with high, compared to low, SITUT would more directly benefit from a therapy that targets brain network interactions contributing to the SITUT-CPM. Given that high mind wandering in ADHD is associated with more severe clinical features (e.g. inattention, hyperactivity, executive function, emotional dysregulation, quality of life 8 ), a treatment targeting SITUT could hold promise toward improving multiple, rather than single, symptoms/outcomes.
Long-term or brief behavioral interventions, such as mindfulness 78  Importantly, although we demonstrated that DMN-FPCN interaction was, overall, a key feature contributing to SITUT-CPM predictions in healthy participants, predictions in ADHD were largely based on other network pairs. A possible reason for this group difference is that ADHD, Beyond ADHD, the SITUT-CPM should be tested in other psychiatric and neurological conditions that exhibit alterations in mind wandering (see 6 ), such as rumination in depression and anxiety. Growing evidence suggests that changes in SITUT characterize a wide range of diseases and could be an early marker of Alzheimer's disease. 7 One study suggests that the relationship of mind wandering tendency with DMN-FPCN and within-DMN functional connectivity is altered in patients with neurodegeneration. 33 Thus, further study of SITUT-CPM generalizability, and potential alternative models in disease, could have far reaching implications within various clinical contexts.

Implications for resting state fMRI biomarkers
Over the past two decades, rs-fMRI has become a standard method to investigate the self-organization of intrinsic activity into large-scale functional networks. 63,86 Even though the unconstrained rs-fMRI paradigm inherently involves SITUT, 9,18 individual-specific topographic patterns of correlated brain activity are highly reliable across repeat measurements. 87 Here we have begun to uncover how SITUT may impact rs-fMRI network estimates, as we found that the same functional network pattern that was predictive of state SITUT within individuals was predictive of trait SITUT/mind wandering at rest. The latter finding may be driven by a) a trait-related, enduring change in the brain's intrinsic network architecture; b) increased SITUT at the time of rs-fMRI scanning within high trait SITUT individuals; or c) a combination of trait and state SITUT. We suggest that increased SITUT at the time of scanning is at least a likely contributor, given that within individuals, we found increased SITUT-CPM network strength across rs-fMRI runs that coincided with self-reported thought contents drifting away from immediate surroundings. In future studies, our approach could be extended to isolate an rs-fMRI "signal" component that represents relatively stable, intrinsic functions from a more statedependent "noise" component that represents SITUT. 9 Taken together, our findings call for the need to further investigate how mind wandering affects rs-fMRI intrinsic network biomarkers.

Limitations
A key limitation concerns the specific task context (gradCPT) in which we defined the SITUT-CPM. During the gradCPT, visual detection of scenes (mountains and cities) engages brain regions (e.g. medial temporal and retrosplenial 20,93 ) that could overlap, or dynamically interact, with key networks involved in SITUT (e.g. DMN). Thus, during time windows prior to thought probes, functional connectivity patterns are not only a product of an individual's level of SITUT but also various other processes such as scene detection. As such, inter-regional interactions involved in SITUT could have been missed due to masking with specific functions associated with ongoing gradCPT performance. Though the SITUT-CPM was sensitive to predictions generated from resting state data, effects were modest, and it remains possible that a different task context for model generation would result in improved out-of-sample prediction performance.
A related concern surrounds our use of 30-second pre-thought probe windows to compute functional connectivity for training and testing of the CPM. This analysis choice was based on our specific task design as well as prior evidence demonstrating that functional connectivity computed from similar durations can reliably distinguish distinct cognitive tasks 94,95 and mind wandering levels. 12 Though brain activity within the seconds that immediately precede thought probes are most relevant to self-reported experiences, 14,96 our selection of a ~30second window was based on the trade-off between the need to capture activity time-locked to experiences and the need to estimate functional connectivity within a window long enough to provide confidence in the sampling of correlations. 97 However, there is no a priori reason to suspect that 30 seconds corresponds to the time course of cognitive processes that unfold during any individual episode of SITUT. 9 In future work, applications of methods sensitive to instantaneous edge co-fluctuations, 98 as well as unsupervised methods sensitive to frame-wise network 'states', 99 could be valuable toward better resolving the time course of network dynamics that underlie SITUT.
In our experience sampling study, online thought probes assessed task focus and awareness of task focus, while retrospective ratings assessed the overall degree to which reports of off-task focus were due to SITUT, external distractions, or task-related interferences.
Though online and retrospective ratings typically show good correspondence, 100 retrospective ratings could have been associated with decreased relative accuracy in our study, as participants may not have precisely remembered the proportion of time spent focused off-task for each of the three respective categories. Prior studies, using thought probes to assess these three categories online, suggest that participants may fluctuate from trial to trial in their reasons for reporting off-task focus and that the distinct categories have different neural correlates. 71 Moreover, multidimensional experience sampling has revealed that distinct thought categories recur over time and are linked to different functional connectivity patterns across individuals. 13,27,29 Future work adopting such a multidimensional approach to online thought probes could better clarify cross-trial heterogeneity within individuals and potentially lead to development of more specific neural models of SITUT components that are either individualspecific or generalize across individuals and populations.

Conclusion
We have demonstrated that SITUT is represented within a common pattern of brain network interactions across multiple time scales, populations, and contexts. Our model-based approach can continue to advance the basic and clinical understanding of mind wandering through further testing across novel conditions, comparisons with alternative models, and continuous refinement based on newly emerging evidence.

Overview
We conducted our analyses with multiple datasets that were acquired as part of four distinct research projects. The Massachusetts General Hospital (MGH) sample included an fMRI experience sampling study of healthy and ADHD adults. The Superstruct and Leipzig samples included rs-fMRI data from healthy adults who completed questionnaires assessing trait SITUT. The Massachusetts Institute of Technology (MIT) sample included rs-fMRI data from ADHD adults who completed a questionnaire that assess trait SITUT.
We first performed analyses within the MGH healthy adult sample to define a 'normative' functional brain network model of SITUT at the within dataset/population level (internal validation) that could be then used to test in other datasets/populations (external validation). We  projection through a mirror from within the scanner bore. Scenes from a set of 10 city and 10 mountain photographs, respectively, gradually transitioned from one to another for the duration of a given run. Scenes were presented randomly, with cities and mountains, respectively, presented pseudorandomly at a rate of ~10%. Using an MRI-compatible button box, participants were instructed press a button with their right index finger each time they saw a city scene but to withhold a press when they saw a mountain scene.
Thought probes appeared after block durations of 44, 52 or 60 seconds (selected randomly). The number of mountain (target) stimuli presented prior to thought probes was matched across blocks via a biased presentation rate within the 32-second window prior to thought probe onset. In the -12 seconds before thought probe onset, no targets were presented.
In the -12 to -32 second window, the target rate was 8% (outside of the pre-thought probe window, the target rate was 13%). During the thought probe, a question was displayed: "To what degree was your focus just on the task or on something else?" On a self-paced, continuous scale (right anchor: "Only task"; left anchor: "Only else"), participants provided a rating using the middle and ring fingers to move the scale left and right, respectively. They used their thumb to submit their response. Responses were recorded on a graded scale of integers ranging from 0 to 100. Similar to previous work, 11 a second question was then presented, which assessed meta-awareness of attentional focus using the question "To what degree were you aware of where your focus was?" Participants provided a self-paced response for this metaawareness item on a continuous scale recorded on a 0 to 100 graded scale of integers (right anchor: "Aware"; left anchor: "Unaware"). After the participant responded to the second question, the next gradCPT block resumed immediately.
Following completion of the scan session, participants completed an oral interview in which they were asked to what degree their reports of attention off-task during the session was due to a) external distractions (EDs; e.g. sounds of the scanner), b) task-related interferences (TRIs), and c) stimulus-independent thought (SITUT). 5 For each item, participants provided a rating on a 7-point Likert scale (1=never; 7=always). 22 Subsequently, to confirm the numerical ratings, we asked participants to verbally provide specific examples of the stimulus-independent thoughts and external distractions that had occurred during task performance. Across participants, SITUT ratings were negatively correlated with both ED (Spearman's ρ = -0.58, P = 0.0012) and TRI (ρ = -0.47, P = 0.011) ratings, suggesting that it was rare for a participant who reported high SITUT to also report high EDs or TRIs.

2.2: Superstruct study procedures.
We downloaded openly available imaging and phenotypic data from the Brain Genomics Superstruct Project (see 67

2.3: Leipzig study procedures.
We downloaded openly available imaging and phenotypic data from the functional connectome phenotyping dataset, a component of the MPI-Leipzig Mind-Brain-Body project (see 103 (64), resolution (2.3 mm isotropic). In the first and third runs, the phase encoding direction was anterior to posterior, whereas in the second and fourth runs, the phase encoding direction was posterior to anterior.

2.4: MIT study procedures.
We analyzed neuroimaging data from a cohort of patients with ADHD who completed

fMRI preprocessing
We preprocessed each fMRI run individually using identical procedures across all datasets, based on procedures implemented in the CONN toolbox (version 19c) 105 and SPM12 in Matlab R2019a (Mathworks Inc., Natick, MA). The pipeline included deletion of the first 4 volumes, realignment and unwarping, 106 and identification of outlier frames (frame-wise displacement > 0.9 mm or global BOLD signal change > 5 SD). 107 Functional and anatomical data were normalized into standard MNI space and segmented into grey matter, white matter (WM), and cerebrospinal fluid (CSF) in a unified step. 108 Smoothing of fMRI data was performed using spatial convolution with a Gaussian kernel of 6mm full width half maximum (FWHM).
We next performed fMRI denoising based on linear regression of the following parameters from each voxel: a) 5 noise components each from minimally-eroded WM and CSF (one-voxel binary erosion of voxels with values above 50% in posterior probability maps), respectively, based on aCompCor procedures 49,109 ; b) 12 motion parameters (3 translation, 3 rotation, and associated first-order derivatives); c) all outlier frames identified within participant; and d) linear BOLD signal trend within session. In a separate step after nuisance regression, 110 data were then temporally filtered with a band-pass of 0.01-0.1 Hz.
Though these specific denoising procedures have been shown to reduce the impact of head motion on functional connectivity, 111 excessive head motion can continue to confound estimates. 47,48 We therefore set an absolute threshold such that participants with mean overall frame-wise displacement (FD) of >0.15 mm (based on the Jenkinson method 112 ) were excluded from analyses in all datasets. In the Leipzig dataset in which four rs-fMRI runs were obtained within participants, this threshold was based on the mean across runs, and we additionally excluded individual runs showing an FD value exceeding the 75 th percentile plus 1.5 times interquartile range (based on procedures implemented in fsl_motion_outliers). 113 Within retained participants, we tested for correlations between FD and behavioral outcomes of interest. Also, given that FD can influence observed relationships between functional connectivity and behavior, 47 we controlled for FD in analyses focused on relationships between functional connectivity and self-report ratings (see Methods Section 5).
Notably, ICA-based denoising represents an alternative to aCompCor that can be effective in reducing the impact of head motion on functional connectivity. 114 To further confirm the validity of our predictive model (Methods Section 5), we thus applied an alternative fMRI preprocessing pipeline based on procedures implemented in FSL, 113 in line with our previous analysis of the MGH healthy adult sample. 14 Briefly, for each run, this included brain extraction, realignment to middle volume, spatial smoothing with a 5mm FWHM kernel, application of ICA-AROMA, 115 regression of WM and CSF signals, bandpass temporal filtering (0.01-0.1 Hz), and linear registration to T1-weighted and MNI152 space.

Functional connectivity feature extraction
Within each dataset, we extracted the preprocessed BOLD time series from the mean across all voxels within each node defined based on previously described intrinsic functional network atlases in MNI space (Shen 116 and Schaefer 117 atlases of 268 whole-brain and 300 cortical regions, respectively). For the MGH experience sampling dataset only, these time series were extracted for each trial based on the ~30-second window (28 TRs) prior to thought probe onset. This window length was chosen based on the task design, which included a matched number of target/non-target stimuli across pre-thought probe windows (see Methods Section 2.1). Moreover, prior evidence suggests that ~22-40 second functional connectivity estimates can reliably distinguish distinct cognitive tasks 94,95 and intensities of self-reported task-unrelated thoughts. 12 For both the gradCPT experience sampling and rs-fMRI datasets, we also extracted time series across the whole duration of each run. We computed a matrix of functional connectivity values between all region pairs based on the Fisher z-transformed Pearson correlation coefficient of time series.

Predictive modeling analysis of SITUT in healthy adults
We adapted the connectome-based predictive modeling (CPM) method 45 to identify functional connectivity patterns predictive of within-participant SITUT. Though CPM has typically been applied to predict individual differences in behavior based on data from whole fMRI runs, 40,41,87 our approach is inspired by recent work suggesting that the method can also be sensitive to intra-individual behavioral fluctuations based on shorter time windows within fMRI runs. 55 Our CPM analysis was performed based on the experience sampling and fMRI data from healthy adults (MGH dataset). Though 29 healthy adults participated in the experience sampling study, we retained for this analysis only those who reported a) trial-by-trial variation in "off-task" ratings (1 participant excluded due to no variation), and b) a retrospective, post-scan questionnaire rating indicating that "off-task" ratings during thought probes were due to taskunrelated thought that was also stimulus-independent (SITUT; minimum Likert scale rating of 4, as in prior work, 23 resulting in exclusion of 11 participants). To further confirm that the retained cohort's "off-task" reports were largely due to SITUT, as opposed to EDs or TRIs, we performed Wilcoxon rank sum tests (two-sided) comparing SITUT ratings to those of the other two categories.
The 17 included participants (which we term "normative sample") consisted of 9 females and 8 males (M ± SD age: 25.6 ± 3.6). This sample size is very similar to those used in various previous studies for the purpose of building fMRI-based predictive models that are testable in larger samples. 41,43,55,118 Moreover, given that trial-wise data included 612 samples across and were treated as individual observations in our modeling, the sample size was in line with putative guidelines for prediction analyses. 119 For each 1 of the 17 participants in the normative sample, we generated model-based predictions of trial-wise SITUT ratings based on independent data from all other included participants (i.e., leave-one-participant-out, leave 36 trials out cross-validation). For each crossvalidation fold, we computed the correlation between each unique edge (node pair) in the functional connectivity matrix (derived from the Shen or Schaefer atlas) and within-participant zscored SITUT ratings (i.e., across 576 trials in 16 participants). We then 'masked' the brainbehavior correlations such that only edges correlated with SITUT ratings at the suprathreshold level of P < 0.01 (two-tailed) were retained, resulting in positive and negative edge masks (Fig.   1a). For each trial, we computed the dot product between the functional connectivity matrix and each mask and then summed all positive and negative edge values separately. We then calculated a single S (network strength) value, the subtraction of negative edge from positive edge sums. We then fit a linear model, based on all trials within the fold, of the form MW = β*S + c, where SITUT is SITUT rating. In the held out-participant, we performed a Pearson correlation between model-predicted SITUT ratings and observed, within-participant normalized SITUT ratings. Importantly, as normalization was performed at the within-participant level, the training and testing data were always kept fully independent. 120 To determine whether predicted versus observed correlations in held-out participants were statistically significant at the group level, we generated a distribution of null values within each fold. To do so, we repeated all of the same CPM procedures, as described, except the trial assignments of SITUT ratings were randomly permuted (1,000 iterations) to obtain null correlation values. At the group level, we performed a Wilcoxon signed rank test to compare the within-participant prediction (predicted versus observed SITUT) versus the mean of the withinparticipant null correlation values (significance set at P<0.05, two-tailed).
We repeated CPM procedures in a series of control analyses, where we performed partial correlations between predicted and observed SITUT within held-out participants for each cross-validation fold. Using these partial correlations, we controlled for the following variables:  Based on the Schaefer atlas, the SITUT-CPM included 385 and 155 edges, respectively, in the positive and negative masks. We assigned each of these edges to one of 28 within-or between-network Yeo-Krienan pairs. For positive and negative masks, we identified the top five network pairs showing the largest number of edges contributing to the masks. For illustration and the purpose of interpretation (Fig. 2b, 2d)

Within-participant SITUT prediction in ADHD
We tested generalizability of the SITUT-CPM within a population distinct from that used to derive the SITUT-CPM, namely 20 adults diagnosed with ADHD (M±SD age: 26.2±8.5; 10 females, 10 males) who underwent the same task performance and experience sampling procedures with the same MRI scanner (MGH sample). Post-scan questionnaire ratings indicated that "off-task" ratings in the ADHD group were largely due to stimulus-independent thought (M±SD rating: 5.

Comparison of SITUT-CPM network strength in healthy versus ADHD adults during task performance
In addition to testing whether the SITUT-CPM within-participant predictions were generalizable across healthy and ADHD populations, we tested whether these two groups reported different levels of SITUT during experience sampling and showed differences in SITUT-CPM network strength during task performance. We included 28 healthy adults (after excluding one participant excluded due to no variation in "off-task" ratings; 13 males, 15 females; M ± SD age: 26.2 ± 3.8) and all 20 ADHD participants in these analyses. At the behavioral level, we compared groups in terms of mean ratings across trials (computed within each participants for a) off-task ratings, and b) meta-awareness ratings . At the neural level, we compared groups in terms of mean SITUT-CPM network strength (averaged across runs). We performed statistical comparisons between groups using Wilcoxon rank sum tests (significance set at P < 0.05, two-tailed). The healthy and ADHD groups showed no significant difference with one another in mean frame-wise displacement across runs (P = 0.31, Wilcoxon rank sum test).

Prediction of trait SITUT from rs-fMRI in healthy adults
In the Superstruct and Leipzig datasets, we tested whether SITUT-CPM predictions generalize to predictions of trait SITUT from rs-fMRI data. In the Superstruct dataset, we To assess specificity of model prediction in the Superstruct dataset, we computed the Spearman's ρ value for 66 total behavioral and self-report measures that were obtained in addition to daydreaming frequency, such as performance on Flanker and Mental Rotation tasks, IQ, and self-report measures of personality, depression and anxiety (see 67 for full description; depending on data availability, these correlations were performed with sample sizes ranging from 892 to 911 participants). Finally, as a further test of specificity, we repeated analyses (Superstruct and Leipzig) using the SA-CPM to compare predicted sustained attention ability versus observed trait SITUT.

Prediction of trait SITUT from rs-fMRI in ADHD
In the ADHD rs-fMRI dataset (collected at MIT), we tested whether SITUT-CPM network strength was different between subgroups of patients with high versus low trait SITUT, based on previously established clinical relevance of this subgrouping. 8 We excluded 10 out of 59 total participants due to excessive head motion, resulting in 49 participants included. Using previously defined criteria, 8 we then classified these participants into those with high and low trait SITUT (MWQ score >23 and <24, respectively). We compared the two groups in terms of head motion and age (Wilcoxon rank sum test, two-sided) as well as sex (F/M ratio) and current ADHD medication status (proportion of patients using medications). Data were missing on ADHD medication status for 1 participant in the high SITUT and 2 participants in the low SITUT subgroups.
We used the SITUT-CPM parameters to compute network strength within each participant, and then we compared network strengths between subgroups using a Wilcoxon rank sum test (significance set at P < 0.05, two-tailed). Additionally, we performed a complementary analysis where MWQ score was treated as a continuous variable, and we computed SITUT-CPM-predicted versus observed MWQ scores (Spearman correlation, including partial correlations controlling for head motion and age). To gain insight into the clinical relevance of MWQ scores within our ADHD cohort, we performed correlations (Spearman) between MWQ and K-SADS symptom severity scores for inattention and hyperactivity (including the diagnostic scores and the current scores based on the past 6 months).

Intra-individual analysis of SITUT-CPM network dynamics during rs-fMRI
In the Leipzig dataset, we performed analyses of the temporal dynamics of subjective experience and SITUT-CPM network strength across four consecutive rs-fMRI runs. On the SNYCQ, which participants completed after each run, we focused on the item "my thoughts involved my surroundings" as an indicator of fluctuations in stimulus-independence of experience across runs. We additionally analyzed the item "I was fully awake" as an indicator of subjective wakefulness. In these analyses, we excluded 3 participants who reported 0 on at least one run for the wakefulness item. After further excluding runs due to excessive head motion, the following numbers of participants were included: run 1 (n = 140) run 2 (n = 165), run 3 (n = 140), and run 4 (n = 146).
We computed SITUT-CPM network strength based on the dot product between the within-run functional connectivity matrix and the SITUT-CPM positive and negative masks. For SNYCQ item ratings and network strength, we performed linear mixed effects model analyses with participant entered as random effect, rating (or network strength) as a the dependent variable, and run number as fixed effect. We performed F tests on the model coefficients, with significance set at P < 0.05 (Satterthwaite's approximation, two-tailed). Using post-hoc Wilcoxon rank sum tests, we performed statistical comparisons between each pair of runs, with significance set at P < 0.05 (false discovery-rate corrected, two-sided). Additionally, to test for a relationship between SITUT-CPM network strength and subjective experience across runs, we performed a linear mixed effects model analysis with participant as random effect, "my thoughts involved my surroundings" rating as dependent variable, and network strength as fixed effect.

Data Availability Statement
Data from the Brain Genomics Superstruct (https://www.neuroinfo.org/gsp) and Leipzig