Prediction of stimulus-independent and task-unrelated thought from functional brain networks

Mind wandering, often defined as self-generated experiences decoupled from the environment and current tasks, varies markedly within and between individuals and relates to mental health, including ADHD and Alzheimer’s disease. Neuroimaging with experience sampling has linked mind wandering to default mode network (DMN) activity and to distributed network interactions, including the frontoparietal control network (FPCN) and sensory/motor systems. However, heterogeneous definitions and contexts limit generalizability. The study asks whether a common, specific, and generalizable functional network pattern can serve as a marker for a central aspect of mind wandering—stimulus-independent, task-unrelated thought (SITUT)—that predicts fluctuations within individuals, generalizes across groups (including ADHD), and extends to resting-state and trait measures. This has implications for understanding brain variability in cognitive and clinical settings.

Prior work using fMRI and experience sampling shows increased DMN activation during mind wandering and before behavioral lapses, with nuanced roles in task engagement. Studies emphasize dynamic interactions among large-scale networks—DMN, FPCN, dorsal attention, salience, and primary sensory/motor networks—in mind wandering. Lesion and neurostimulation studies implicate causal roles for DMN and FPCN regions. Multivariate predictive approaches have begun to relate within- and between-DMN connectivity to task-unrelated thought but largely within single datasets or task contexts. Mind wandering is related yet distinct from sustained attention and creative thinking, suggesting a need for specific neuromarkers. Given clinical links (e.g., excessive mind wandering in ADHD), a generalizable network-based model of SITUT could clarify brain-behavior variability and inform biomarkers.

Participants: Healthy adults with online experience sampling during fMRI at MGH performed the Gradual-Onset Continuous Performance Task (gradCPT) with thought probes every 44–60 s. Of 29 healthy participants, 17 were retained for model derivation based on retrospective confirmation that off-task reports predominantly reflected SITUT (612 trials; 36 trials per participant). An independent ADHD group (n=20) completed the same protocol. Three independent resting-state fMRI (rs-fMRI) datasets were used for external validation: Superstruct (n=911 healthy adults; 6-min rs-fMRI; Daydreaming Frequency Scale, DDFS), Leipzig (n=144 healthy adults; four ~15-min rs-fMRI runs; Mind Wandering Deliberate, MW-D, and Spontaneous, MW-S scales), and MIT (n=49 adults with ADHD; 7-min rs-fMRI; Mind Wandering Questionnaire, MWQ). Imaging: 3T systems across sites; preprocessing in CONN/SPM with aCompCor denoising (and an ICA-AROMA pipeline for robustness checks), motion outlier censoring, bandpass 0.01–0.1 Hz, and exclusion of participants with mean frame-wise displacement >0.15 mm. Functional connectivity: Time series were extracted using Shen268 and Schaefer300 atlases; for experience sampling, connectivity was computed from the 30-s window (28 TRs) preceding each probe. Connectome-based predictive modeling (CPM): In leave-one-participant-out cross-validation within the healthy MGH cohort, edges whose connectivity correlated with within-participant z-scored SITUT ratings (P<0.01, two-tailed) formed positive and negative masks. A summary network strength (positive minus negative edge sums) was used in a linear model to predict trial-wise SITUT in held-out participants. Statistical significance was assessed versus permutation-derived nulls. Control analyses included partial correlations controlling for head motion, pre-probe reaction time variability (sustained attention), and strengths from published CPMs (sustained attention, SA-CPM; creativity, Cr-CPM). Neuroanatomical characterization assigned CPM edges to Yeo-Krienen 7- and 17-network systems to identify dominant intra/inter-network contributors and tested sufficiency/necessity of DMN-FPCN edges via CPM restricted to or lesioned of these edges. External validation: (1) Test SITUT-CPM within-individual predictions in ADHD (MGH); (2) Correlate SITUT-CPM predictions from rs-fMRI with DDFS (Superstruct) and MW-D/MW-S (Leipzig), including partial correlations controlling for motion and age; (3) Compare SITUT-CPM network strength between high vs low MWQ ADHD subgroups (MIT) and relate strength continuously to MWQ. State dynamics: In Leipzig, assess within-participant changes across four rs-fMRI runs in SITUT-CPM strength and retrospective ratings of stimulus-involvement ("my thoughts involved my surroundings").

- Internal validation (healthy adults, MGH): Trial-wise predictions of SITUT in held-out participants were significantly above permutation null (mean within-participant r = 0.11 ± 0.16; P = 0.019; MSE 1.19 ± 0.15). SITUT-CPM contained 258 positive and 139 negative edges (Shen268), distributed across cortex and cerebellum. Robustness: Predictions remained significant controlling for head motion (partial r = 0.091 ± 0.15; P = 0.039), with ICA-AROMA preprocessing (r = 0.10 ± 0.14; P = 0.0086), and using Schaefer300 (r = 0.12 ± 0.23; P = 0.044). Specificity: Prediction remained significant controlling for pre-probe RT variability (partial r = 0.10 ± 0.16; P = 0.040). SITUT-CPM showed minimal overlap and independence from SA-CPM and Cr-CPM; SA-CPM did not significantly predict SITUT (r = −0.073 ± 0.16; P = 0.093). - Network basis: Positive SITUT edges were dominated by DMN-FPCN and DMN-DMN contributions; negative edges by DMN–sensorimotor (SMN). SITUT was characterized by decreased DMNA–FPCNA anticorrelation and decreased DMNA–SMNA correlation. Univariate analyses highlighted DMN-FPCN as the only significantly SITUT-related pair after correction. CPM restricted to DMN-FPCN edges alone predicted SITUT (r = 0.14 ± 0.19; P = 0.015); lesioning these edges reduced performance (r = 0.079 ± 0.21; P = 0.15) within the derivation sample. - External validation in ADHD (MGH): Healthy-derived SITUT-CPM predicted within-individual SITUT in ADHD (r = 0.045 ± 0.086; P = 0.028). An ADHD-derived CPM also internally validated (r = 0.12 ± 0.15; P = 0.0032; MSE = 1.21 ± 0.22) and generalized to healthy adults (r = 0.072 ± 0.12; P = 0.015). The simplified DMN-FPCN-only model did not generalize to ADHD (r = 0.0053 ± 0.091; P = 0.74). Behaviorally, ADHD reported higher SITUT during gradCPT (55.3 ± 12.5 vs 34.8 ± 18.8; P = 9.6×10^-5) and lower meta-awareness (50.3 ± 15.3 vs 67.6 ± 16.7; P = 0.0016). Neural network strength was higher in ADHD (9.54 ± 8.98) than controls (4.98 ± 6.09; P = 0.039); SA-CPM strength did not differ (P = 0.37). - External validation from rs-fMRI (trait SITUT): Superstruct (n=911): SITUT-CPM predictions correlated with DDFS (Spearman ρ = 0.074; P = 0.025), remaining significant controlling for motion (partial ρ = 0.080; P = 0.016) and age (partial ρ = 0.072; P = 0.029). Among 67 phenotypes, correlation was strongest for daydreaming frequency; SA-CPM predictions were not associated with DDFS (ρ = 0.022; P = 0.50). Leipzig (n=144): Positive correlation with MW-D (ρ = 0.17; P = 0.043) and a trend for MW-S (ρ = 0.15; P = 0.068); effects persisted as partial correlations controlling for motion and age. - External validation in ADHD rs-fMRI (MIT, n=49): SITUT-CPM network strength was greater in high vs low MWQ subgroup (P = 0.043). Continuous association with MWQ trended positive (ρ ≈ 0.28; P = 0.055), remaining significant when controlling for motion (partial ρ ≈ 0.29; P = 0.042); SA-CPM did not relate to MWQ (P = 0.94). - State dynamics across rs-fMRI runs (Leipzig): Self-reported thoughts involving surroundings decreased after run 1 (interaction F1,502.74 = 4.68; P = 0.03; significant decrease run2<run1, PFDR = 0.016), indicating increased stimulus-independence over time. SITUT-CPM network strength increased across runs (interaction F1,419.73 = 14.56; P = 0.00016; significant increases run3>run1/run2 and run4>run1). Run-to-run network strength did not track the surroundings rating at this timescale (F1,669.65 = 0.22; P = 0.64).

The study demonstrates that a distributed functional connectivity pattern captures trial-by-trial SITUT within individuals and generalizes across datasets, contexts, and clinical status. Key mechanisms involve altered interactions between DMN and FPCN (reduced anticorrelation, especially DMNA–FPCNA), and increased antagonism between DMN and primary sensorimotor regions, consistent with perceptual decoupling during mind wandering. The SITUT-CPM was specific relative to sustained attention and creativity models, showing minimal overlap and unique behavioral associations. Generalization to ADHD supports clinical relevance; ADHD participants showed higher SITUT and higher expression of the SITUT network during tasks and at rest, though predictive edges beyond DMN-FPCN appeared important for ADHD generalization. From a biomarker perspective, the model linked state SITUT during tasks to trait SITUT measured at rest, suggesting that state mind-wandering contributes to rs-fMRI variability. The findings underscore the need to account for SITUT when interpreting rs-fMRI biomarkers and suggest that SITUT-related network dynamics operate across multiple timescales within individuals.

This work establishes a functional brain network-based marker (SITUT-CPM) that predicts intra-individual SITUT, generalizes to ADHD and to trait SITUT from rs-fMRI, and highlights specific DMN–FPCN and DMN–SMN interactions as core mechanisms. Although effect sizes are modest, they are consistent with large-sample functional connectivity predictions of self-report measures and robust across datasets and measures. The model offers a foundation for probing mind-wandering mechanisms, interpreting rs-fMRI variability due to mental state, and exploring clinical applications, such as identifying ADHD subgroups with high SITUT for targeted interventions. Future research should test the model in other disorders (e.g., depression, anxiety, Alzheimer’s), refine predictions with multimodal data (EEG, pupil, activation), examine different task contexts, and improve temporal resolution to better map SITUT episodes.

- Model derivation during the gradCPT may confound SITUT-related connectivity with task-specific processes (e.g., scene processing), potentially masking other SITUT interactions and limiting out-of-sample performance. - Connectivity was estimated in fixed 30-s pre-probe windows; this may not match the true timescale of SITUT episodes. - Retrospective categorization (SITUT vs external distractions vs task-related interferences) may be less precise than online multidimensional probing; distinct off-task categories can vary across trials and have different neural correlates. - Effects in rs-fMRI datasets were modest; task context and dataset heterogeneity (scanners, protocols, measures) may limit generalization. - The simplified DMN-FPCN-only model did not generalize to ADHD, suggesting potential overfitting and the need for broader network features in clinical populations.