Heterogenous brain activations across individuals localize to a common network

Over three decades, task-based fMRI has advanced understanding of human brain function, but individual-level analyses suffer from low reliability and reproducibility due to poor temporal SNR and short scan durations. A key manifestation is inconsistent activation locations across people performing the same cognitive task, motivating group averaging to gain statistical power. However, brain functions may localize better to distributed, connected networks than to isolated regions. Lesion network mapping (LNM) shows that heterogeneous lesion locations causing similar symptoms often converge on common functional networks. By analogy, the authors hypothesize that heterogeneous task-evoked activations across individuals may likewise localize to a common network, and that reproducibility should be defined in network terms rather than by discrete activation loci. The study focuses on working memory (WM), a well-studied domain with evidence for network-based organization and substantial inter-individual behavioral variability, enabling brain–behavior prediction. The authors employ activation network mapping (ANM), which uses resting-state functional connectivity to identify networks functionally connected to task-evoked activation sites. They test three aims using Human Connectome Project (HCP) data: (1) whether heterogeneous WM activations localize to a common network, (2) whether the identified WM network is specific relative to other tasks, and (3) whether individual WM networks predict individual differences in WM performance.

Data and participants: The study used publicly available HCP Young Adult (WU-Minn) 3T data. From 850 healthy participants with complete imaging, 100 subjects were randomly selected. Each had 14 task runs (2 per 7 tasks: WM, emotion, social, gambling, motor, relational, language) and four resting-state runs (two LR and two RL). All analyses were conducted in voxel space and thresholded at voxelwise FWE-corrected P < 0.05.

WM task paradigm: An N-back task with 0-back and 2-back conditions using blocks of pictures (places, faces, body parts, tools). Each scan had 8 task blocks (10 trials of 2.5 s; 25 s) and 4 fixation blocks (15 s). Stimuli were presented for 2 s with 500 ms ITI. The contrast of interest for activation was 2-back > 0-back. Behavioral WM performance was mean accuracy across stimuli in 2-back blocks.

Control tasks: Six HCP tasks served as controls: gambling (reward vs baseline), language (story vs math), motor (movement vs baseline), emotion (faces vs shapes), social cognition (social vs random), and relational processing (relational vs match), using HCP standard paradigms.

MRI acquisition and preprocessing: fMRI acquired on a 3T Siemens Skyra with multiband EPI: TR 720 ms, TE 33.1 ms, flip angle 52°, 2 mm isotropic resolution, multiband factor 8, FOV 280 × 180 mm. Left-to-right and right-to-left phase-encoding runs were collected. Task fMRI preprocessing used HCP “fMRIVolume” pipeline: gradient unwarping, motion correction, EPI distortion correction, registration to T1, nonlinear warp to standard space, intensity normalization, and spatial smoothing (4 mm FWHM). Resting-state preprocessing followed prior lab methods.

Activation maps (WM and control tasks): For each subject, GLM (FSL FILM) modeled eight predictors (four for 0-back stimulus types; four for 2-back). Temporal derivatives were included as confounds; high-pass filter (200 s) and prewhitening (film_gls) applied. Fixed-effects combined the two runs. Subject-level z-maps for 2-back > 0-back were thresholded voxelwise FWE P < 0.05 and binarized to yield activation maps. Control task activations were derived analogously for their contrasts.

Activation Network Mapping (ANM): For each subject’s binarized activation map (activation seed), functional connectivity to all brain voxels was computed using a resting-state normative connectome of 1000 subjects. For each normative subject, the mean seed time course was correlated voxelwise to obtain r-maps, transformed to Fisher z. A voxelwise one-sample t-test across the 1000 z-maps produced a network t-map per study subject. Each network t-map was thresholded at voxelwise FWE P < 0.05 and binarized to yield the subject-level activation network map. As a complementary analysis, each subject’s own resting-state data (four concatenated runs) was used to compute seed-based RSFC from the activation seed; r-maps were thresholded against zero at voxelwise FWE P < 0.05 and binarized to generate own-connectome activation network maps.

Interindividual consistency metrics: Binarized activation maps and activation network maps were overlapped across subjects. Overlap maps highlighted voxels present in a high percentage of subjects (e.g., >80%). Pairwise Dice index (DI = 2 V_overlap / (V1 + V2)) quantified between-subject similarity for activation vs activation network maps.

Robustness analyses: (1) Matched sparsity: To control for differences in map extent, both activation and activation network maps were thresholded to retain the top 5% of voxels before binarization and consistency/Dice computations. (2) Peak-seed ANM: To test whether network patterns were driven by distributed seeds, 8-mm spheres centered on each subject’s activation peak (max of the largest cluster) were used as seeds in ANM; resulting network overlap and similarity to distributed-seed ANM were assessed.

Specificity of WM network: Binarized activation network t-maps for WM (N=100) were compared to pooled control task activation network maps (N=600) using a nonparametric Liebermeister test in NiiStat, including only voxels present in >10% of maps. Voxelwise FWE P < 0.05 with 5000 permutations determined specificity.

Behavioral prediction (ridge regression): Nested leave-one-out cross-validation (LOOCV) ridge regression predicted 2-back accuracy from imaging features. Three predictor types were used separately: (a) binarized activation maps, (b) binarized ANM maps based on the normative connectome, and (c) binarized ANM maps based on each subject’s own connectome. An additional model combined predictors. Prior to modeling, whole-brain voxel features (2-mm³ voxels; 238,955) were standardized and reduced via PCA retaining components explaining 95% variance. Inner LOOCV tuned λ in [1e-5, 1e5] (log scale) to maximize R²; outer LOOCV assessed generalization. Performance was quantified by coefficient of determination (R²). Significance of R² was assessed via 10,000-label permutation tests. Model comparison used paired t-tests on squared prediction errors. Consensus feature weights were averaged across outer folds, tested against permutation-derived nulls with FWE correction, and projected back to voxel space for visualization (top 25% absolute weights). A control prediction analysis repeated models using top-5% thresholded activation and ANM maps to match sparsity.

• Heterogeneity of activations: Single-subject WM activations (2-back > 0-back) were highly heterogeneous. Only 50% of subjects showed activation in the most convergent region; 25–50% showed activation in bilateral dorsal/ventral prefrontal cortex, dorsal anterior cingulate cortex, and dorsal parietal cortex.
• Convergence at the network level: Despite heterogeneous activation loci, >90% of subjects’ activation seeds were functionally connected (via ANM) to a common network including bilateral lateral prefrontal cortex, parietal cortex, cingulate gyrus, inferior temporal gyrus, thalamus, and basal ganglia. Similar overlap was observed using subjects’ own connectomes.
• Quantitative similarity: Pairwise Dice indices were significantly higher for activation network maps than for activation maps: • Normative connectome ANM vs activation: t9898 = −248.55, p < 0.0001. • Own connectome ANM vs activation: t9898 = −245.26, p < 0.0001.
• Matched-sparsity robustness: When both map types were thresholded to top 5% voxels, ANM maps still showed greater interindividual consistency: • Normative ANM vs activation: t9898 = −100.21, p < 0.0001. • Own ANM vs activation: t9898 = −91.59, p < 0.0001.
• Peak-seed robustness: Using only activation peaks as seeds, locations remained heterogeneous (only 8/100 peaks in the most convergent region), but >80% of peaks were connected to common brain regions; the peak-seed network overlapped highly with distributed-seed ANM (spatial correlation r = 0.89, unthresholded).
• Specificity: The WM ANM network was significantly more connected than control task ANM networks to key regions; WM overlap (>80%) strongly intersected with the WM-specificity map (Liebermeister test, FWE-corrected).
• Behavioral prediction of WM performance: • Activation maps: R² = 0.06. • ANM maps (normative connectome): R² = 0.27; significantly outperformed activation maps (paired t98 = 2.11, p = 0.037). Unthresholded ANM maps also predicted well (R² = 0.24, p < 0.0001). • ANM maps (own connectome): R² = 0.18. • Combined predictors (activation + ANM types): R² = 0.07 (no improvement). • Control task ANM maps failed to predict WM; five of six had negative R² values; only WM ANM yielded significant prediction (p < 0.0001).
• Weight maps: For activation prediction, positive weights localized to commonly activated regions (lateral PFC, somatosensory association cortex, TPJ); negative weights in primary sensorimotor, thalamus, occipital/temporal. For ANM (normative), positive weights in lateral PFC, primary sensorimotor, occipital/inferior temporal; negative in medial/dorsal frontal and temporal regions; similar patterns for own-connectome ANM.

The study demonstrates that while WM task-evoked activations are spatially heterogeneous across individuals, these activations converge on a reproducible functional network when examined via connectivity. The WM network identified by ANM includes bilateral lateral PFC, parietal cortex, cingulate, inferior temporal cortex, thalamus, and basal ganglia—overlapping with the frontoparietal control network implicated in WM. This network has convergent support from lesion, neuromodulation, training, and connectivity studies linking its integrity to WM capacity. The low reproducibility of individual activation loci likely reflects limited temporal SNR and short task scan durations typical of task fMRI, which bias detection toward the strongest nodes while missing weaker network components. ANM acts as an indirect solution by leveraging a high-powered normative resting-state connectome to recover distributed, functionally connected regions associated with the task-defined seeds. The findings align with highly sampled individual studies that show alignment between task-evoked responses and resting-state networks. Importantly, ANM-derived individual WM networks predicted behavioral performance substantially better than raw activation maps, establishing practical utility for individualized brain–behavior mapping. In contrast to recent reports that LNM-derived networks poorly predict behavioral deficits, ANM performed well—likely because activation seeds are gray-matter–focused and task-specific, while lesions in LNM are large, often span white matter, and may mix multiple networks. The results also provide the first direct empirical evidence that a large normative connectome can effectively substitute for individual connectomes in network localization, yielding even higher cross-subject consistency and predictive performance, presumably due to improved SNR from large sample size. Finally, the most predictive regions in ANM maps were not necessarily the most commonly connected regions, consistent with the use of binarized features and reduced variance in universally connected voxels. Overall, network localization reframes reproducibility in task fMRI, improving interpretability and predictive linkage between cognitive processes and neuroanatomy.

This study shows that heterogeneous WM activations across individuals converge on a common, highly reproducible functional network when assessed via activation network mapping. ANM identifies individualized WM networks that not only exhibit greater intersubject consistency than discrete activation maps but also predict individual WM performance substantially better. The WM network is specific relative to networks derived from other tasks. The work provides empirical support for using a large normative connectome to approximate individual connectivity in both ANM and related network localization approaches. These findings suggest that network-based definitions can reconcile apparent low reproducibility of task activations and enhance brain–behavior mapping at the individual level. Future research should generalize the ANM framework to other cognitive domains, perform causal validation (e.g., lesion datasets, neuromodulation), and examine how individualized and common network components jointly shape behavior.

• Thresholding and binarization: Main analyses used binarized, FWE-thresholded activation and ANM maps, potentially discarding informative subthreshold variation; however, unthresholded ANM maps showed similar predictive power.
• Correlational nature of fMRI: Functional imaging is correlational; causal validation of the ANM-identified WM network is needed (e.g., prediction of deficits in lesion cohorts, neuromodulation studies).
• Individual variability: While a common network was identified, individual-specific WM network differences likely coexist; the study emphasizes commonality with minor intersubject variations.
• Domain focus: The investigation centered on WM; generalization of the principle that heterogeneous activations map to common networks requires testing across other cognitive processes.