Brain decoding of spontaneous thought: Predictive modeling of self-relevance and valence using personal narratives

Spontaneous thought is pervasive and often centers on personally significant, emotionally charged topics related to self-identity, goals, and memories. Its unconstrained nature and the Heisenberg-like effect of probing these experiences make assessment difficult. Motivated by evidence that personal relevance and valence are core dimensions summarizing spontaneous thought content, the study aimed to decode self-relevance (“is it relevant to me?”) and valence (“is it good or bad?”) directly from fMRI activity. The authors hypothesized that personally generated narratives would evoke brain representations resembling spontaneous thoughts and that predictive models trained on these stimuli could generalize to free-thinking and resting states, illuminating internal states relevant to mental health.

Prior work shows spontaneous thought is common, meaningful, and predictive of cognitive and affective traits, with disruptions linked to neuropsychiatric conditions. Narratives, rich in semantics and unfolding over time, have been effective in studying brain processes and may approximate the form of spontaneous thought. Self-referential processing engages distinct brain systems and is prominent in spontaneous cognition. Dimensionality reduction studies identify self-relevance and valence as central content dimensions. The default mode network (DMN), limbic, and ventral attention networks have been implicated in valuation, self-referential processing, and stimulus-independent thought. Previous work classified affect during rest using task-induced patterns (medial OFC), but did not build whole-brain regression models nor target self-relevance levels. This study extends those approaches using personalized narratives and multivariate decoding to model affective dimensions of spontaneous thought.

Participants: 57 healthy right-handed Korean participants consented; 49 (age 22.8 ± 2.4; 21 female) were included after excluding 8 for poor task performance or image quality. Design: Two-day protocol. Day 1: pre-scan questionnaires and one-on-one online interviews to generate personal stories; participants also read common stories to increase familiarity. Day 2: fMRI with five story-reading runs (~14 min/run) and two thought-sampling runs (~6 min/run), followed by a post-scan survey. Tasks: Story-reading: each participant read four personal and six common stories, with intermittent in-scanner valence ratings three times per story. Thought-sampling: participants freely thought and verbally reported words/phrases every 50.7 ± 5.6 s (5 s report windows). Post-scan survey: continuous ratings of self-relevance (0–1) and valence (−1 to 1) for stories; multidimensional ratings (including self-relevance and valence) for thought-sampling words. In-scanner vs. post-scan valence ratings correlated within-individuals (mean r = 0.844). MRI acquisition: 3T Siemens Prisma, 64-channel head coil. T1: TR = 2,400 ms, TE = 2.34 ms. EPI: TR = 460 ms, TE = 27.2 ms, multiband = 8, FoV = 220 mm, matrix 82×82, voxel 2.7 mm isotropic, 56 interleaved slices; 812 volumes for thought-sampling runs and 1,855 for story-reading runs. Predictive modeling: For feature-label disentanglement, story-reading data were binned into quintiles of self-relevance and valence, producing 25 averaged images per participant (5×5 grid). Models for self-relevance and valence were trained using principal component regression (PCR) with leave-one-subject-out cross-validation (LOSO-CV) and random-split CV (RS-CV). Additional validation included permutation tests and analyses of word-level representation (adjusted selection frequency based on valence, aFv). Model interpretation: Virtual isolation (using one network/region/searchlight at a time) and virtual lesion analyses examined feature importance across large-scale networks (visual, somatomotor, dorsal/ventral attention, limbic, frontoparietal, DMN) and ROIs (including mPFC). Searchlight-based maps identified spatially localized contributors. Generalization tests: Applied models to decode ratings during thought-sampling (n = 49) using a moving-window approach around report onset accounting for HRF delay (temporal Gaussian kernel, FWHM = 10 TRs). Compared with nine a priori maps (self, default, emotion, episodic, semantic, aversion meta-analytic maps; components from self-generated thought; PINES negative emotion signature). Applied models to two independent resting-state datasets (n = 90; n = 60); prediction evaluated using averaged pattern expression over windows near end of scan (e.g., last 31 TRs ≈ 14.3 s), with additional exploration of varying window sizes.

Model performance (story-reading): - Self-relevance: LOSO-CV mean r = 0.322, z = 9.204, P < 2.220e-16, mse = 0.148; RS-CV mean r = 0.332, mse = 0.144; permutation test P = 0.001. - Valence: LOSO-CV mean r = 0.205, z = 6.235, P = 4.511e-10, mse = 0.454; RS-CV mean r = 0.179, mse = 0.458; permutation test P = 0.001. - Vividness correlated with valence model performance across subjects (r = 0.370, P = 0.0090). - Weight pattern similarity between models: whole-brain r = 0.120; within mPFC r = 0.269. Both models showed vmPFC positive and dmPFC/sgACC negative weights (uncorrected P < 0.001 display). Validation analyses: - Personal vs. common stories: Self-relevance ratings higher for personal (0.75 ± 0.16) vs. common (0.42 ± 0.10), t₄₈ = 15.86, P < 2.220e-16; model responses higher for personal, t₄₈ = 10.18, P < 2.220e-16; forced-choice classification accuracy 93.8%, P = 6.980e-11. - Word-level aFv: correlation between aFv from actual ratings and valence model responses r = 0.2855, P = 1.110e-16. Feature importance: - Virtual isolation (networks/ROIs): default mode, ventral attention, and frontoparietal networks significantly predictive for both models; visual network predictive for self-relevance only; limbic network predictive for valence only; mPFC predictive for both. - Searchlight: aMCC, aINS, and visual areas were important for self-relevance; dmPFC, left TPJ, and temporal pole for valence; conjunction showed SMA, STG, IFG important to both. Generalization to free-thinking (thought-sampling): - Peak decoding near report onset: self-relevance mean r = 0.0518, P = 0.0141; valence mean r = 0.0495, P = 0.0095 (one-tailed, bootstrap). DMN was the only significant predictor for both within time-of-interest. - A priori maps did not predict ratings in this context. Resting-state decoding: - Dataset 1 (n = 90): Best window averaging last 31 TRs (14.3 s); self-relevance r = 0.189, P = 0.037; valence r = 0.300, P = 0.002 (one-tailed). Network importance: self-relevance—DMN only; valence—ventral attention, limbic networks, and brainstem. - Dataset 2 (n = 60): Predefined last-31-TR window—valence r = 0.320, P = 0.0063; self-relevance not significant (r = −0.094, P = 0.2375). Exploratory window of last 10 TRs (4.6 s) yielded self-relevance r = 0.218, P = 0.0470.

Personal narratives effectively evoked brain states resembling spontaneous thought, enabling decoding of self-relevance and valence directly from fMRI signals. The DMN, particularly mPFC, emerged as central for both dimensions, consistent with its role in stimulus-independent thought, memory, and self-related processing. Distinct systems also contributed uniquely: for self-relevance, salience network hubs (aINS, aMCC) and visual areas likely supported context-dependent salience detection when the self is the prevailing context; for valence, limbic components (e.g., OFC, temporal pole) and DMN’s dorsomedial subsystem (left TPJ, dmPFC) may support affective meaning and reflective, verbal imagination. The models generalized to free-thinking and resting states, suggesting that task-derived whole-brain patterns can decode aspects of ongoing internal cognition. Targeting affective dimensions complements semantic decoding by capturing idiosyncratic personal meanings, critical for assessing internal states and potential mental health markers.

The study introduces whole-brain, regression-based predictive models that decode self-relevance and valence—two central dimensions of spontaneous thought—from fMRI, trained on personalized narrative stimuli. It identifies key contributing networks and regions (DMN, salience, limbic, mPFC, aINS, aMCC, TPJ) and demonstrates generalization to free-thinking and resting-state contexts across independent datasets. These advances offer a pathway toward brain-based assessment of internal states during daydreaming and rest, with potential clinical utility when task-based fMRI is impractical. Future work should refine stimulus design (e.g., single story type with parametric self-relevance/valence), adopt personalized modeling to capture idiosyncratic representations, and evaluate functional connectivity-based approaches that may outperform activation-based models.

Personal versus common stories may differ qualitatively, potentially confounding self-relevance with attention/familiarity despite efforts to equalize exposure; personal stories still elicited higher concentration and familiarity. Decoding performances during free-thinking and rest were modest and did not survive multiple-comparison corrections, warranting cautious interpretation and raising risk of type I errors, though convergent time windows and replication for valence support generalizability. Neural representations of self-relevance and valence are interrelated (weak positive correlations in ratings and weight patterns), complicating disentanglement. The models were activation pattern-based; functional connectivity-based models may yield stronger performance in naturalistic contexts and should be tested.