The Human Connectome Project (HCP) has provided an unprecedented wealth of data for understanding the intricate networks of the human brain. These datasets, particularly from the Young-Adult (HCP-YA) and Aging (HCP-A) projects, offer invaluable insights into brain connectivity and its relationship to behavior and demographics. This article delves into the methodologies and findings derived from these datasets, focusing on how we can identify and interpret patterns – or what we might term “synchronised fashion” – in brain activity.
Human Connectome Project Datasets: A Rich Resource for Brain Network Analysis
The HCP-YA and HCP-A projects have meticulously collected comprehensive datasets, including behavioral data and functional MRI (fMRI) scans. Ethical approvals were secured for both projects, ensuring responsible data acquisition and analysis. These projects provide a foundation for exploring the complex interplay between brain structure, function, and individual characteristics.
HCP Young-Adult Dataset: Exploring Connectivity in the Prime of Life
The HCP-YA dataset, drawn from the S1200 release, encompasses resting-state fMRI (rs-fMRI) data from four sessions for each participant. The selection criteria prioritized subjects with complete data across all four sessions and for 25 predefined behavioral variables, resulting in a robust dataset of 771 individuals (384 female, 387 male). Ages in this cohort ranged from 22 to 37 years (mean age 28.41, standard deviation 3.74), representing a critical period of adulthood. The rs-fMRI sessions were conducted over two days, with variations in phase encoding directions (left-right [LR] and right-left [RL]) to enhance data quality and minimize artifacts. Scanning was performed using a 3T Siemens connectome-Skyra scanner with specific parameters optimized for capturing detailed brain activity: TE = 33.1 ms, TR = 720 ms, flip angle = 52°, 2.0 mm isotropic voxels, 72 slices, and a multiband factor of 8.
HCP Aging Dataset: Investigating Connectivity Across the Lifespan
To validate findings from the HCP-YA dataset and extend the investigation into older age groups, a subset of the HCP-A dataset was utilized. Similar to HCP-YA, HCP-A employed two rs-fMRI sessions across two days, using a 2D multiband (MB) gradient-recalled echo (GRE) echo-planar imaging (EPI) sequence (MB8, TR/TE = 800/37 ms, flip angle = 52°) with 2.0 mm isotropic voxels and 72 oblique-axial slices on a Siemens 3T Prisma scanner. Phase encoding polarity was varied (anterior-to-posterior [AP] and posterior-to-anterior [PA]) across runs within each session. Rigorous inclusion criteria ensured complete data for all four runs and selected behavioral and confounding variables, resulting in a sample of 558 subjects (316 female, 242 male) aged 36 to 100 years (mean age 59.87, standard deviation 15.03).
Both HCP-YA and HCP-A datasets benefited from the HCP’s minimal preprocessing pipeline. This crucial step involved motion correction, registration to standard space, and ICA-FIX (independent component analysis and FMRIB’s ICA-based X-noiseifier) to remove structured artifacts. Further noise reduction was achieved by regressing out 24 parameters, including rigid-body motion parameters, their temporal derivatives, and squared terms. This pre-processing ensured high-quality, denoised data ready for advanced analysis of brain connectivity.
Image Pre-processing: Refining the Raw Data for Meaningful Insights
Further pre-processing was applied to both datasets to refine the signal and isolate relevant brain activity. Confound regression, linear detrending, and bandpass filtering (0.008–0.08 Hz) were implemented using “nilearn.image.clean_img”. For the primary analysis, 12 parameters representing noise components were regressed out: mean time courses of white matter (WM), cerebrospinal fluid (CSF), global signal (GS), their squared terms, and temporal derivatives. Supplementary analyses explored results without global signal regression. A spike regressor addressed motion artifacts by flagging fMRI frames exceeding a motion threshold (framewise displacement [FD] > 0.25 mm). The Schaefer 200 parcellation was employed to define brain regions, with supplementary analyses using Schaefer 300 and 400 parcellations. In HCP-A data, time series were trimmed to ensure comparable bin sizes with HCP-YA, resulting in 440 volumes divided into 8 bins of 55 volumes each. This meticulous pre-processing is essential for extracting reliable functional connectivity measures.
Edge Time Series Construction and Functional Connectivity Estimation: Capturing Synchronised Brain Activity
To capture the dynamic nature of brain connectivity, edge time series were constructed. This method, detailed in prior research, involves z-scoring parcellated BOLD time series and then calculating the element-wise product between pairs of parcels. This product estimates co-fluctuation between regions over time. The magnitude of co-fluctuation was quantified using the root sum of squares (RSS) at each time point, generating a co-fluctuation time series for each subject. Time points in the BOLD time series were then ordered by their RSS magnitude (high to low).
To investigate the information captured by different levels of co-fluctuation, three sampling strategies were employed. Strategy (1), individual bins sampling, divided the ranked BOLD time series into bins (20 bins of 5% for HCP-YA, 8 bins of 12.5% for HCP-A) and sampled specific bins for analysis. Strategy (2), combined bins sampling, explored combinations of two bins to assess information gain from diverse co-fluctuation levels. Strategy (3), sequential sampling, used sequentially increasing thresholds to include varying numbers of time points based on co-fluctuation magnitude. Functional connectivity (FC) was estimated using pairwise Pearson’s correlation coefficients for each sampling strategy and co-fluctuation bin. FC matrices were averaged across phase encoding directions and further averaged to produce a single FC matrix per subject per co-fluctuation bin for prediction analyses. These strategies allowed for a detailed examination of how different patterns of co-fluctuation, or “synchronised fashion” moments, contribute to functional connectivity and its predictive power.
Structural Connectivity Extraction: Mapping the Brain’s Physical Network
In addition to functional connectivity, structural connectivity (SC) was also examined. Diffusion-weighted magnetic resonance imaging (dMRI) data, preprocessed using the HCP diffusion minimal preprocessing pipeline, was used to extract SC matrices. This pipeline corrected for distortions, motion, and gradient nonlinearities and registered diffusion data to structural T1w scans. An in-house workflow, using MRtrix3, was employed for probabilistic tractography. Ten million streamlines were generated via whole-brain probabilistic tractography (WBT), and fibre oriented distributions (FOD) were estimated using constrained spherical deconvolution. The Schaefer atlas with 200-area parcellation, based on Freesurfer cortical parcellation, was used to define regions. The tck2connectome function in MRtrix3 reconstructed SC matrices. Subjects with missing dMRI data or software errors during preprocessing were excluded, resulting in a final sample of 762 subjects for SC analyses. By analyzing structural connectivity alongside functional connectivity, a more complete picture of brain network organization emerges.
Subject Specificity: Differential Identifiability and Identification Accuracy
To quantify the uniqueness of individual brain connectivity profiles, both identification accuracy (I**acc) and differential identifiability (I**diff) were assessed. Identification accuracy measures the proportion of subjects correctly identified across different scanning sessions based on their FC profiles. Differential identifiability, calculated as the difference between mean within-subject correlations (I**self) and mean between-subject correlations (I**other), reflects the strength of an individual’s “connectome fingerprint”. Higher differential identifiability indicates a more distinct individual profile. These measures are crucial for understanding the reliability and individuality of functional connectivity patterns.
Prediction of Behavioural and Demographic Measures: Unveiling the Biological Meaning of Connectivity Patterns
Functional connectivity matrices, derived from the different sampling strategies, were used to predict behavioural and demographic measures. Phenotypes from categories like Cognition, In-scanner task performance, and Personality (25 targets in HCP-YA, four cognitive targets in HCP-A) were selected as prediction targets. Age at scan, sex at birth, and framewise displacement (FD) were controlled for as confounds in a cross-validation-consistent manner. In addition to behavioral targets, age and sex were also predicted.
Ridge regression with a Pearson kernel was the primary prediction model, benchmarked for its efficiency and performance. Support vector classifiers (SVC) were also used for sex classification to ensure robustness across different models. A nested cross-validation (CV) approach (10-fold for HCP-YA, 5-fold repeated 5 times for HCP-A) was implemented to assess out-of-sample prediction accuracy and select hyperparameters. Performance metrics included Pearson’s r, coefficient of determination (R2), mean absolute error (MAE) for regression, and accuracy/balanced accuracy for classification. These prediction analyses aim to determine how well “synchronised fashion” moments in brain connectivity can predict individual traits and characteristics, revealing the biological relevance of these patterns.
Bayesian Region-of-Practical-Equivalence (ROPE) Approach: Comparing Predictive Models
To statistically compare the predictive accuracy of different models (individual bins, combined bins, full FC), a Bayesian region-of-practical-equivalence (ROPE) approach was employed. This method defines practical equivalence between models if accuracy differences are within a pre-defined percentage (5% in this study). Bayesian estimation, starting with a normal prior distribution, was used to calculate posterior probabilities for three hypotheses: model y is better than x, models x and y are practically equivalent, and model x is better than y. The ROPE approach provides a robust statistical framework for comparing model performance by focusing on a range of practical equivalence rather than a point-wise null hypothesis.
Statistics and Reproducibility: Ensuring Rigor and Reliability
Statistical analyses were conducted with appropriate sample sizes for both HCP-YA and HCP-A datasets. Bootstrapping procedures were used to estimate variance and assess statistical significance in identification analyses. Wilcoxon-signed-rank tests, with Bonferroni correction for multiple comparisons, were used to compare HACF- and LACF-derived FC. Grouped cross-validation in HCP-YA accounted for family structure, while repeated cross-validation was used in HCP-A. The Bayesian ROPE approach was used for model comparison in the combined bins sampling strategy. Similarity between structural and functional connectivity was assessed using Pearson’s correlation. These rigorous statistical methods and transparent reporting ensure the reliability and reproducibility of the findings.
This comprehensive analysis of the HCP datasets, using advanced methodologies, provides valuable insights into the “synchronised fashion” of brain connectivity. By examining patterns of co-fluctuation and their predictive power, we gain a deeper understanding of the biological meaning of brain network organization and its relationship to individual differences in behavior and demographics.
