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 of BOLD data, leading researchers to favor group averaging. A key manifestation is inconsistent activation locations across individuals for the same cognitive process. Emerging evidence suggests brain functions localize to connected networks rather than isolated regions. Lesion network mapping (LNM) shows that heterogeneous lesion locations causing similar symptoms map to a common functional network. The authors hypothesize that heterogeneous task-evoked activations across individuals likewise localize to a common network and that reproducibility should be defined in terms of connectivity. They test this using working memory (WM), selected because it is well-studied, is better localized to networks than isolated regions, and exhibits broad individual differences enabling brain–behavior prediction. They propose activation network mapping (ANM) to identify, at the individual level, brain regions functionally connected to each person’s activation during WM, and evaluate network reproducibility, specificity versus other tasks, and behavioral prediction.

Prior work highlights limited reliability of individual task-fMRI and greater correspondence of cognitive functions to networks. LNM has mapped diverse symptoms (e.g., prosopagnosia, amnesia, movement disorders) to common networks despite spatially heterogeneous lesions, suggesting network-level localization. WM is strongly associated with the frontoparietal control network, with converging evidence from fMRI/EEG studies linking stronger functional and structural connectivity to higher WM capacity, TMS to dorsolateral prefrontal cortex impacting WM, and training increasing frontoparietal connectivity. Resting-state architecture reflects co-activation patterns during tasks, motivating ANM’s use of a normative resting-state connectome to infer networks from task activations. Although ANM has been used for meta-analytic integration across studies, it has not previously been used to identify individual-level networks of cognitive processes. The study also addresses debates over using large normative connectomes versus individual connectomes for mapping, providing empirical tests of their equivalence and performance.

Data: Public Human Connectome Project (HCP) Young Adult 3T dataset (WU-Minn; 1200 release). From ~850 fully scanned healthy subjects, 100 were randomly selected who completed seven tasks (two runs each: WM, emotion, social, gambling, motor, relational, language) and four resting-state runs. Ethics approvals were in place. Task paradigm: HCP WM N-back task with 0-back and 2-back blocks across four stimulus types (places, faces, body parts, tools). Behavioral measure: mean accuracy in 2-back blocks. Acquisition: 3T Siemens Skyra, 32-channel coil; multiband EPI: TR=720 ms, TE=33.1 ms, flip angle=52°, 2 mm isotropic, multiband factor=8, FOV 280×180 mm²; LR and RL phase encoding. HCP minimal preprocessing for task data (gradient unwarping, motion correction, distortion correction, registration to T1, nonlinear warp to standard space, intensity normalization), then 4 mm FWHM smoothing. Resting-state preprocessing per prior lab protocol. Activation maps: For each subject and task, GLM (FSL FILM) with eight predictors (four 0-back, four 2-back), including temporal derivatives; high-pass filter 200 s; prewhitening (film_gls). Contrast: 2-back > 0-back for WM; task-specific contrasts for control tasks. Fixed-effects across runs. Z-maps thresholded voxelwise FWE-corrected P<0.05 and binarized. Activation Network Mapping (ANM): Each subject’s binarized activation map served as a seed. Using a normative resting-state connectome (n=1000), seed time series were averaged and correlated voxelwise, converted to Fisher z, and subjected to one-sample voxelwise t-tests across the 1000 subjects, yielding a network t-map per subject. Thresholding at voxelwise FWE P<0.05 with binarization produced individual activation network maps. For comparison, subject-specific ANM was also computed using each subject’s concatenated four resting-state runs to derive seed-based correlation maps, thresholded vs zero at voxelwise FWE P<0.05 and binarized. Reproducibility metrics: Overlap maps computed across subjects; overlap threshold >80% to define commonly connected regions. Pairwise Dice Index (DI) calculated between binarized maps for activations and networks to quantify interindividual similarity. Robustness analyses: (1) Matched suprathreshold voxel proportion by applying a top 5% threshold to both activation and network maps before overlap and DI analyses. (2) Peak-seed ANM: 8-mm sphere at each subject’s peak activation (max of largest cluster) used as seed to repeat ANM and compare network patterns with those derived from distributed activation seeds (spatial correlation across unthresholded overlap maps). Specificity: Compared WM ANM maps (N=100) to pooled ANM maps from six control tasks (emotion, social, gambling, motor, relational, language; N=600) using nonparametric Liebermeister tests in NiiStat with 5000 permutations, including only voxels present in >10% of maps; voxelwise FWE P<0.05. Behavioral prediction: Ridge regression with nested leave-one-out cross-validation (LOOCV). Inputs: whole-brain voxel features (238,955 2-mm voxels) from (i) activation maps, (ii) ANM maps using normative connectome, (iii) ANM maps using subject’s own connectome; both binarized, with additional analysis on unthresholded ANM. Standardization and PCA retained components explaining 95% variance before ridge; inner LOOCV optimized λ (10⁻⁵ to 10² log-spaced) maximizing R²; outer LOOCV evaluated generalization. Model performance: coefficient of determination (R²). Statistical significance via 10,000-label permutation tests. Model comparisons via paired t-tests on squared prediction errors. Consensus weight maps averaged across outer folds; significance assessed against permutation-derived null and FWE-corrected; weights projected back to brain; visualization emphasized top 25% absolute weights. Additional model combining activation and ANM predictors evaluated.

• Single-subject WM activations were spatially 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, and dorsal parietal cortex.
• Despite heterogeneous activation locations, ANM revealed convergence to a common network: >90% of subjects had activation seeds connected to bilateral lateral prefrontal cortex, parietal cortex, cingulate gyrus, inferior temporal gyrus, thalamus, and basal ganglia. Similar patterns emerged using each subject’s own connectome.
• Interindividual similarity (Dice Index) was significantly higher for activation network maps than for activation maps, both with normative connectome (t9898 = −248.55, p < 0.0001) and with subject’s own connectome (t9898 = −245.26, p < 0.0001).
• When matching the percentage of suprathreshold voxels (top 5%) between activation and network maps, network maps still showed greater between-subject consistency (normative: t9898 = −100.21, p < 0.0001; own: t9898 = −91.59, p < 0.0001).
• Peak-seed ANM produced networks highly similar to distributed-seed ANM (spatial correlation r = 0.89 for unthresholded network overlap). Although only 8/100 activation peaks fell in the most convergent location, >80% connected to common network regions.
• Specificity: Regions commonly connected in WM ANM overlap significantly overlapped with voxels identified as specific to WM versus six control tasks (emotion, social, gambling, motor, relational, language) via Liebermeister testing (voxelwise FWE P<0.05).
• Behavioral prediction: Activation maps explained 6% of variance in WM accuracy (R² = 0.06). ANM maps (normative connectome) explained 27% (R² = 0.27) and performed significantly better than activation maps (t98 = 2.11, p = 0.037). Unthresholded ANM (normative) explained R² = 0.24. ANM maps using the subject’s own connectome explained 18% (R² = 0.18). Combining activation and ANM predictors did not improve performance (R² = 0.07). Control-task ANM maps failed to predict WM (five of six had negative R²), supporting specificity. Contribution weight maps highlighted lateral prefrontal, sensorimotor, occipital/inferior temporal (positive) and medial/dorsal frontal and temporal (negative) regions for ANM predictors.

Findings support the hypothesis that heterogeneous individual WM activations localize to a common, reproducible functional network when analyzed via connectivity. Although voxelwise activations show low overlap across subjects, ANM reveals a consistent network encompassing bilateral lateral prefrontal, parietal, cingulate, inferior temporal, thalamic and basal ganglia regions—overlapping the frontoparietal control network implicated in WM. The higher interindividual consistency of network maps versus activations persisted after matching suprathreshold extents and when using activation peaks as seeds, indicating robustness and that many heterogeneous peaks reside within different nodes of the same network. Behaviorally, individual ANM-derived networks predicted WM performance substantially better than activation maps, with normative connectome-derived ANM outperforming subject-specific ANM, likely due to improved signal-to-noise from the large normative sample. Specificity analyses showed the WM ANM network is distinct from networks derived from other tasks and uniquely predictive of WM behavior. The results reconcile individual activation heterogeneity with reliable network-level organization, providing a framework to map individual cognitive function to neuroanatomy. They also offer empirical support for using normative connectomes in network mapping, complementing prior evidence that resting-state networks reflect task co-activation. Compared with LNM, ANM may better support behavioral prediction because activation seeds are focal, gray-matter–based and task-specific, avoiding confounds from large, heterogeneous lesions that can span multiple networks. The identified network aligns with causal evidence from lesion and neuromodulation studies (e.g., dorsolateral prefrontal TMS effects; basal ganglia gating) and training-induced connectivity changes, underscoring its functional relevance.

This study demonstrates that seemingly low reproducibility of individual WM activations masks high reproducibility at the network level: heterogeneous activation locations across individuals converge on a common WM network revealed by ANM. Individual ANM networks robustly predict WM performance and outperform predictions based on activation maps, with normative connectome–based ANM showing the best generalization. The WM network localized via ANM is specific relative to networks from other cognitive tasks, and results support the practical use of normative connectomes for individual network mapping. Future directions include: (1) testing generalization of the “heterogeneous activations → common network” principle to other cognitive domains; (2) causal validation of ANM-derived networks using lesion cohorts, neuromodulation, or longitudinal interventions; (3) refining models using richer features beyond binarization, multimodal data, and extended individual scan times; and (4) exploring individual-specific deviations within the common network to understand trait variability and clinical translation.

• Binarization of activation and network maps may discard informative subthreshold and amplitude information, though unthresholded ANM showed similar predictive power in supplementary analyses.
• fMRI provides correlational evidence; causal involvement of identified regions in WM requires quantitative validation (e.g., lesion datasets, neuromodulation).
• While a common network emerged, individual differences in WM networks likely coexist, and minor interindividual variations were observed.
• The study focused on WM; other cognitive processes may show different patterns of network convergence and will require separate validation.
• Normative connectome use cannot capture individual-specific factors (age, comorbidities) in connectivity, although here it performed as well or better than individual connectomes, likely due to higher SNR.
• Short task scan durations inherent to standard fMRI may limit detection of weakly activated nodes; ANM indirectly addresses this via normative data rather than increased per-subject scan time.