HCPTask Dataset: Human Connectome Project-Task

• HCPTask dataset is a large-scale fMRI collection from over 1,200 healthy adults performing seven cognitive tasks, providing dense temporal data for dynamic connectivity analysis.
• It employs standardized acquisition protocols and rigorous preprocessing pipelines, including motion correction, physiological noise removal, and brain parcellation with multiple atlases.
• The dataset supports advanced machine learning and graph-theoretic modeling techniques, facilitating accurate cognitive state decoding, biometric fingerprinting, and Bayesian activation inference.

The Human Connectome Project-Task (HCPTask) dataset comprises large-scale, multi-subject functional MRI (fMRI) recordings collected under standardized cognitive task paradigms, facilitating high-dimensional exploration of dynamic brain connectivity, cognitive state decoding, biometric fingerprinting, and activation signature modeling. Building on protocols from the HCP Young Adult release (S1200), HCPTask provides extensive coverage of healthy adults performing distinct tasks, with dense temporal sampling and rigorous preprocessing pipelines, enabling state-of-the-art machine learning and statistical inference on functional connectomes and activation patterns (Monti et al., 2015, Maji et al., 31 Dec 2025, Hannum et al., 2022, Miranda et al., 2021).

1. Dataset Structure and Acquisition Protocols

HCPTask encompasses fMRI data from over 1,200 healthy young adults (ages 22–35), each scanned using a Siemens 3T “Connectome Skyra” with a multiband gradient-echo EPI protocol (TR ≈ 720 ms, voxel size 2 mm isotropic, flip angle 52°, 32-channel head coil, multiband factor 8) (Hannum et al., 2022, Monti et al., 2015). Seven cognitive paradigms are included:

• Emotion processing
• Gambling
• Language comprehension
• Motor execution
• Relational processing
• Social cognition
• Working memory

Each subject typically completes two runs per task, with additional rest-state scans. Block timings and trial structure conform to HCP standards (e.g., 25 s blocks in working memory tasks), yielding over 7,443 task-evoked “runs” for recent GNN-based analyses (Maji et al., 31 Dec 2025).

2. Preprocessing Pipelines and Brain Parcellation

All workflows commence with HCP minimal preprocessing, followed by confound regression and tailored filtering steps:

• Motion correction via Friston’s 24-parameter model (rigid-body plus derivatives and squared terms)
• Removal of physiological noise (white matter and CSF mean signals)
• Temporal filtering (high-pass: 1/130 Hz (Monti et al., 2015), band-pass: 0.01–0.1 Hz (Maji et al., 31 Dec 2025), or 0.008–0.08 Hz (Hannum et al., 2022))
• Spatial normalization to MNI space
• Brain parcellation with atlases such as Desikan–Killiany (68 cortical, 16 subcortical; 84 total ROIs (Monti et al., 2015)), Schaefer–Yeo (100/200/400 cortical nodes (Hannum et al., 2022, Maji et al., 31 Dec 2025)), or Talairach atlas (298 ROIs for Bayesian voxel-level analyses (Miranda et al., 2021))

Node time series are extracted by averaging the BOLD signal within each ROI. Confound regression, z-scoring, detrending, and, where relevant, truncation of rest scans are applied. Variant pipelines (global signal regression, task regression, parcellation resolution, rest truncation) are explicitly tested for fingerprinting and decoding robustness (Hannum et al., 2022).

3. Functional Connectivity and Dynamic Network Estimation

Functional connectivity matrices are constructed by computing pairwise Pearson correlations between node time series: $$FC_{ij} = \frac{\sum_t (x_i(t) - \overline{x}_i)(x_j(t) - \overline{x}_j)}{\sqrt{\sum_t (x_i(t) - \overline{x}_i)^2} \cdot \sqrt{\sum_t (x_j(t) - \overline{x}_j)^2}}$$ with thresholds applied to enforce graph sparsity (e.g., adjacency $A_{ij}$) (Maji et al., 31 Dec 2025). For dynamic analyses, temporally local precision matrices are estimated with the Smooth Incremental Graphical Lasso Estimator (SINGLE): $$\min_{ \{ \Theta_i \} } \sum_{i=1}^n [ -\log\det\Theta_i + \text{trace}( \widehat{\Sigma}_i \Theta_i ) ] + \lambda_1 \sum_{i=1}^n \| \Theta_i \|_1 + \lambda_2 \sum_{i=2}^n \| \Theta_i - \Theta_{i-1} \|_1$$ where $\widehat{\Sigma}_i$ is the local covariance and $\lambda_1$, $\lambda_2$ balance sparsity and temporal homogeneity (Monti et al., 2015). This yields temporally-resolved connectivity snapshots, facilitating graph-theoretic and embedding analyses.

4. Machine Learning and Graph-Theoretic Modeling

Advanced classifiers and graph neural network architectures capitalize on HCPTask’s scale and granularity:

• Graph embeddings for dynamic connectivity: Linear projections (PCA, LDA) of Laplacians extracted from SINGLE-estimated networks reveal principal modes of variance (unsupervised) and discriminative patterns between cognitive loads (supervised), with low-dimensional embeddings directly interpretable in terms of connectivity subnetworks (Monti et al., 2015).
• Spectral graph neural networks: The SpectralBrainGNN model performs exact graph Fourier transforms (GFT) using the eigendecomposition of the normalized Laplacian. Spectral filtering is learned with channel mixing and per-eigenvalue attention, with final graph-level embeddings input to cross-entropy classifiers. Across 7,443 task graphs (400-node connectomes), SpectralBrainGNN achieves 96.25 ± 1.37% accuracy, outperforming conventional spatial GNNs (GCN, GAT, GraphSAGE) by >10% (Maji et al., 31 Dec 2025).
• Functional connectome fingerprinting and decoding: Linear discriminant analysis (LDA), support vector machines (SVM), neural networks (NN), and nearest-centroid classifiers robustly identify individual subjects (up to 99.7% accuracy) and decode cognitive states (up to 99.8% across eight states), with accuracy evaluated over multiple preprocessing strategies (Hannum et al., 2022).

Performance metrics include accuracy, precision, recall, F1-score, and confusion matrices to characterize misclassification, with statistical significance validated via permutation and t-tests (Maji et al., 31 Dec 2025).

Model Results on HCPTask (Mean ± Std over 30 runs from (Maji et al., 31 Dec 2025))

Model	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
GCN	86.29 ± 0.98	85.12 ± 1.05	86.45 ± 0.92	85.78 ± 0.97
GAT	85.60 ± 1.26	84.78 ± 1.31	85.92 ± 1.18	85.35 ± 1.24
ResGCN	93.75 ± 0.35	93.02 ± 0.41	93.89 ± 0.33	93.45 ± 0.38
BrainMAP	94.74 ± 0.07	94.12 ± 0.10	94.68 ± 0.06	94.40 ± 0.08
SpectralBrainGNN	96.25 ± 1.37	95.46 ± 1.42	94.32 ± 1.51	95.58 ± 1.39

5. Bayesian Activation Signature and Connectivity Inference

The HCPTask dataset supports fully Bayesian voxel-wise regression modeling of task fMRI responses and background connectivity. The hierarchical model is

$y_t(v) = \sum_{j=1}^p b_j(v) (s_j * h)(t) + e_t(v)$

with regression performed in tensor-basis space combining two-level spatial basis (within- and between-ROI PCA), and temporal decomposition via wavelet transforms (modeling long-memory covariance): $Y^* = \Theta Y \Upsilon$ where $\Upsilon = \Phi\Psi$ encodes local and global spatial components, and $\Theta = W$ applies multiscale Daubechies wavelets (Miranda et al., 2021). Bayesian inference yields posterior samples of activation coefficients, Simultaneous Credible Bands (SimBas), and estimates background covariance to reveal task-modulated subnetworks. CHSB methods detect activation clusters in occipital and cerebellar regions associated with working memory and visual processing, outperforming traditional GLM and spatially reduced approaches in sensitivity.

6. Robustness, Variability, and Best Practices

Extensive studies underscore several best-practice recommendations for HCPTask utilization:

• Dense temporal resolution (TR=0.72 s) enables fine-scale detection of network reconfigurations (Monti et al., 2015, Hannum et al., 2022).
• Rigorous nuisance regression and high-pass/band-pass filtering are critical for confound removal.
• Parcellation granularity (80–400 ROIs) balances anatomical fidelity and graph tractability; larger parcellations yield marginally higher classification accuracy (Hannum et al., 2022, Maji et al., 31 Dec 2025).
• For connectivity estimation, jointly enforcing sparsity and temporal smoothness prevents spurious and noisy edge inference.
• Linear and nonlinear embeddings should be mapped back to original edge or ROI dimensions for interpretable network visualizations.
• Leave-out and cross-run validation mitigates overfitting and confirms classifier generalization (e.g., training on LR, testing on RL phase-encoding).
• Bayesian spatial-temporal models improve sensitivity for activation pattern discovery and background network inference.

This suggests that future exploitation of HCPTask should prioritize multimodal preprocessing, robust classification paradigms, interpretable embedding architectures, and statistical inference models that leverage both spatial and temporal structure.

7. Significance and Research Applications

HCPTask has catalyzed advances across multiple domains:

• Cognitive state decoding: enabling real-time and cross-subject classification of task states from whole-brain connectomes (Maji et al., 31 Dec 2025, Hannum et al., 2022).
• Biometric subject fingerprinting: facilitating reliable identification based on FC signatures (Hannum et al., 2022).
• Graph-embedding approaches for dynamic network analysis: providing interpretable and discriminative views into functional reconfigurations (Monti et al., 2015).
• Bayesian modeling of activation and connectivity: supporting rigorous inference for both localized and distributed brain circuits under cognitive load (Miranda et al., 2021).

A plausible implication is that the combination of scale, standardization, and extensive metadata in HCPTask makes it a definitive benchmark for methodological development, validation, and cross-site replicability in neuroimaging-based connectivity research. The dataset has become foundational for brain network machine learning, biomarker development, and methodological innovation in cognitive connectomics.