Network Neuroscience

Linking structural connectivity (SC) to functional connectivity (FC) through mechanistic models remains challenging in network neuroscience. In this study, empirical data of diffusion magnetic resonance imaging (dMRI) and resting-state functional MRI (rs-fMRI) were used to reconstruct SC and FC of a swine connectome. We evaluated a structurally constrained Kuramoto phase-oscillator framework to reproduce resting-state FC and then assessed the models sensitivity to traumatic brain injury (TBI) and its longitudinal progression post-TBI. A joint tuning procedure was implemented to calibrate data-informed natural frequencies and global coupling strength. The tuned Kuramoto model was then used to evolve oscillator phases constrained by the SC, followed by a Balloon-Windkessel hemodynamic model. The optimized model produced significant edge-wise correspondence between averaged simulated FC and the empirical FC (r = 0.61, p < 0.001). Graph-theoretical analysis across network densities (30-50%) showed strong agreement for global efficiency, characteristic path length, and clustering coefficient, while modularity and small-worldness exhibited deviations. Longitudinal analysis of the swine TBI dataset revealed modest reductions in structure-function coupling over time but no significant differences across injury severities. These results demonstrate that optimized Kuramoto models can reproduce key functional network features while preserving inter-subject variability.

The connectivity of brain networks is extremely sparse due to metabolic, physical and spatial constraints. Although wiring sparsity can confer computational advantages for biological and artificial neural networks, sparse networks require fine parameter tuning and exhibit strong sensitivity to perturbations. How brains achieve their efficiency and robustness is unclear. Here we addressed this by analysing the dynamical properties of Echo State Networks with wiring based on the Drosophila melanogaster fruit fly connectome, compared with sparsity-matched random-wiring networks. We evaluated these networks on a set of eight cognitive tasks, and found that connectome-based neural networks (CoNNs) typically showed narrowly distributed task engagement across their neurons. The importance of a neuron for task performance correlated with its node degree, local clustering, and selfrecurrency, and these correlations were stronger in CoNNs than in random networks. CoNNs were more robust to neuronal loss, retaining their task performance and beneficial dynamical properties such as criticality and spectral radius better than random networks. Similarly, CoNNs were more robust to hyperparameter variations in both input and recurrent weight scaling. Using theoretical arguments and numerical simulations, we show that excess CoNN node self-recurrency is sufficient to explain this enhanced robustness. Overall, these results identify non-random features of connectome wiring that allow brains to reconcile extreme sparsity with reliable computation. SignificanceBrain networks support robust computation even though they operate under extreme wiring sparsity due to metabolic and spatial constraints. While sparse networks typically require fine-tuning and are sensitive to perturbations, we show that biological connectomes support specialised, efficient task engagement and remain robust to neuron loss and parameter variation. We identify excess neuronal selfrecurrency as a key structural feature underlying this stability. These results reveal how non-random connectivity stabilises computation in extremely sparse networks, providing principles for understanding brain function and designing robust, efficient artificial neural systems.

Functional brain networks are conventionally constructed using measures of direct temporal synchrony between neural signals, implicitly restricting connectivity to scale-specific interactions. Here, we introduce an alternative framework in which interregional similarity is defined through correlations between multiscale entropy (MSE) profiles, enabling network construction based on scale-dependent dynamical structure rather than instantaneous alignment. Using resting-state fMRI data from the Human Connectome Project (N = 1003), we systematically compare MSE-based networks with conventional time-series-based networks across conventional/spectral graph-theoretical, and information-theoretic measures. We show that MSE-based networks exhibit stronger modular organization, enhanced local segregation, and distinct global integration patterns, reflecting a reorganization of functional architecture when multiscale dynamics are taken into account. Importantly, MSE-based networks demonstrate substantially greater sensitivity to biologically meaningful variability, revealing robust and reproducible sex differences across multiple network measures, in contrast to the limited and inconsistent effects observed in conventional networks. These findings suggest that multiscale representations provide a more informative and biologically grounded basis for functional brain network construction, capturing aspects of neural organization that are not accessible through direct synchrony alone.

White matter structural connectivity constrains large-scale brain communication, yet most network models do not account for biologically meaningful differences between connections. Although axonal diameter and myelination influence neural signaling at the microscale, how these features shape systems-level functional connectivity remains unclear. Here, we test whether structural connectomes weighted by white matter microstructure give rise to distinct communication regimes that differentially predict multimodal functional connectivity. Combining quantitative MRI and advanced diffusion modeling, we constructed whole-brain networks weighted by tract caliber and multiple myelin-sensitive measures. To these, we applied routing- and diffusion-based communication models and used the resulting communication metrics to predict haemodynamic and frequency-resolved electromagnetic connectivity. Myelin-weighted networks preferentially enhanced long-range communication efficiency and redistributed spectral energy toward globally integrative topological eigenmodes. In contrast, caliber-weighted networks emphasized mesoscale organization and short-range communication. Across nested regression models controlling for geometric embedding and network topology, myelin-sensitive communication explained unique variance in functional connectivity with effects varying systematically across cortical systems and frequency bands. The strongest coupling was observed for alpha-band connectivity in association and attentional networks, consistent with a role for myelin-dependent communication delays in supporting long-range alpha synchrony. These findings demonstrate how distinct white matter microstructural features give rise to heterogeneous large-scale communication regimes: tract caliber and myelin bias communication toward locally specialized and globally integrative architectures, respectively. By integrating biologically informed connectomics with communication modeling and multimodal functional data, this work advances a mechanistic account of how white matter microstructure shapes macroscale brain dynamics.

Brain decoding from fMRI data using artificial neural networks traditionally operates at the regional level, identifying which brain areas activate during tasks but ignoring how these regions interact through structural networks. While Graph Neural Networks can capture connectivity, they require prohibitively large datasets for typical neuroscience studies. We introduce a message-passing mechanism that allows a shallow neural network to incorporate structural connectivity, enabling network-level interpretation from limited data. Using motor task data from 30 Human Connectome Project subjects, we evaluate seven structural connectivity matrices derived from deterministic and probabilistic tractography. Our approach achieves 83.0% classification accuracy while revealing functional network organization. We demonstrate that sparser, anatomy-driven connectivity matrices outperform dense alternatives, and that normalizing for network size improves model performance. Critically, our method is capable of exposing structural pathways contributing towards classification, distinguishing between complete network recruitment and selective regional activation. This approach bridges the gap between high-performance brain decoding and biological fidelity of the model, enhancing neuroscientific understanding, with implications for analyzing network dysfunctions in neurological disorders such as Alzheimers disease (AD), attention deficit hyperactivity disorder (ADHD), autism spectrum disorder (ASD), bipolar disorder, mild cognitive impairment (MCI), and schizophrenia.

Neurodegenerative diseases such as Mild Cognitive Impairment (MCI), Multiple Sclerosis (MS), Parkinson s Disease (PD), and Amyotrophic Lateral Sclerosis (ALS) are becoming more prevalent. Each of these diseases, despite its specific pathophysiological mechanisms, leads to widespread reorganization of brain activity. However, the corresponding neurophysiological signatures of these changes have been elusive. As a consequence, to date, it is not possible to effectively distinguish these diseases from neurophysiological data alone. This work uses Magnetoencephalography (MEG) resting-state data, combined with interpretable machine learning techniques, to support differential diagnosis. We expand on previous work and design a Riemannian geometry-based classification pipeline. The pipeline is fed with typical connectivity metrics, such as covariance or correlation matrices. To maintain interpretability while reducing feature dimensionality, we introduce a classifier-independent feature selection procedure that uses effect sizes derived from the Kruskal-Wallis test. The ensemble classification pipeline, called REDDI, achieved a mean balanced accuracy of 0.81 (+/-0.04) across five folds, representing a 13% improvement over the state-of-the-art, while remaining clinically transparent. As such, our approach achieves reliable, interpretable, data-driven, operator-independent decision-support tools in Neurology.

Resting-state functional MRI has been widely used to study brain connectivity, yet the test-retest reliability of commonly used metrics remains a concern. To improve reliability, extended scan lengths and larger subject cohorts are often recommended. However, these solutions can be impractical and pose challenges, particularly in studies of clinical populations. Here, we systematically assess the reliability of two main types of functional connectivity measures: node-based connectome metrics (edge-level intraclass correlation coefficient [ICC], connectome-level ICC, functional connectivity fingerprinting, and discriminability); and voxel-based resting-state networks (RSNs) (spatial similarity of independent component analysis [ICA]-derived RSN maps quantified using the Dice coefficient). Using data from the Human Connectome Project, we evaluated the effects of scan length (3.6, 7.2, 10.8, and 14.4 minutes) and number of participants (n = 10, 20, 50, and 100), on both within-session and between-session reliability. We found that multivariate connectome metrics demonstrated greater reliability than edge-level measures, and that scan length had stronger influence on test-retest reliability than the number of participants. For connectome metrics, 14 minutes of scanning and a cohort of approximately 20 participants were sufficient to achieve reliable estimates. In contrast, RSN measures benefited from larger cohort sizes. Our findings provide practical guidelines for designing resting-state fMRI studies in terms of scan length and number of participants, balancing reliability and feasibility. Ultimately, protocol choices should be guided by the specific study objectives and the functional connectivity metric of interest.

Transcendent thinking (TT) is an enduring affective and cognitive process characterized by abstract meaning-making, moral reflection, self-referential integration, and strong emotional engagement. Despite growing interest in its developmental and affective significance, the intrinsic neural dynamics that predict individual differences in disposition to TT remain poorly understood. Most prior work has relied on linear functional connectivity measures, which may be insufficient to capture the nonlinear and multiscale nature of brain dynamics underlying higher-order affective dispositions like TT. Here, we introduce a nonlinear functional brain network (FBN) framework based on multiscale entropy (MSE) to investigate whether intrinsic resting-state nonlinear brain dynamics predict disposition to TT in adolescents. Functional connectivity was defined as inter-regional similarity in MSE profiles derived from resting-state fMRI, yielding weighted networks that capture scale-dependent dynamical correspondence rather than linear synchrony. Graph-theoretical, spectral, and information-theoretic measures were computed and evaluated against signal-level and network-level null models. Predictive performance was assessed using machine-learning models and compared with conventional time series-based FBNs. Global intelligence (IQ) was examined as a control cognitive variable. MSE-based network features, particularly spectral energy and Shannon entropy, showed significant associations with TT and enabled reliable prediction of individual differences, whereas time series-based network measures failed to predict TT. No network measures reliably predicted IQ. Overall, these results indicate that intrinsic nonlinear brain dynamics carry predictive information about affective dispositions, rather than domainspecific or network-localized cognitive abilities such as IQ. This work demonstrates that nonlinear, multiscale network representations of resting-state brain activity provide a principled and predictive framework for modeling individual differences in enduring affective dispositions.

Cognitive processes such as decision-making, working memory, and motor planning operate across a hierarchy of timescales, manifesting as rapid neural transients alongside slower physiological mechanisms like short-term plasticity. Conventional Dynamic Causal Modelling (DCM) limits our ability to study these dynamics by assuming stationary parameters, whilst recent time-varying approaches often rely on segmenting data into epochs. This segmentation artificially resets neural states between windows, fundamentally obscuring the continuous hysteresis essential to sequential processing. To address this limitation, we introduce DCM for Sequential Responses (DCM-SR), a generative framework that embeds parameter evolution directly within the first-level model whilst employing a continuous state-space formulation that removes the requirement for epoching. This approach generalises non-stationarity to all neural mass parameters, including synaptic gains and time constants, modelling them as piecewise smooth trajectories that evolve alongside continuous neural states. Consequently, the model explicitly captures two distinct forms of temporal memory: transient history dependence, where responses are shaped by the carryover effect of recent perturbations, and path dependence, where the systems trajectory through parameter space determines its responsiveness. The framework accommodates both exogenous, stimulus-locked transitions and endogenous, autonomous state changes, permitting inference on both external perturbations and internal drivers of network evolution. Simulations establish the models face validity, demonstrating robust parameter recovery and conservative model selection that accurately discriminates between genuine parameter evolution and spurious complexity. We applied the framework to empirical data from an auditory go/no-go task, modelling a full sequence of cognitive phases from initial cue processing and anticipation through to motor preparation and execution. This analysis established construct validity by resolving the biophysical generators of the contingent negative variation, attributing this slow potential to sustained thalamocortical drive and deep-layer hyperpolarisation rather than superficial-layer activity. Furthermore, the model captured trial-specific modulations of the hyperdirect pathway during motor inhibition, tracking the dynamic interplay between prefrontal executive control and basal ganglia gating. DCM-SR offers the first principled approach to decomposing compound signals such as slow cortical potentials into evolving synaptic mechanisms and continuous state trajectories, and provides a necessary bridge for investigating the biophysical implementation of extended cognitive phenomena including evidence accumulation and physiological hysteresis.

Understanding how brain regions influence one another over time is a central goal of neuroscience. However, existing approaches to effective connectivity often involve tradeoffs among mechanistic interpretability, computational scalability, and time-resolved estimation. Here, we introduce information channel capacity, a model-based measure of directed information transfer between brain regions, and combine it with a sliding-window framework to estimate time-varying directional interactions. We validate channel capacity across multimodal neuroimaging datasets in humans and rodents because this breadth is needed to test three complementary properties that no single dataset can establish alone: sensitivity to evoked information transfer, specificity against false-positive directional effects, and the ability to capture meaningful temporal variability in directed brain-network interactions. Human motor-task fMRI tests sensitivity, showing that channel capacity detects task-related increases in directed interactions and stronger directional effects during task than during rest in motor-related regions. Concurrent rat local field potential (LFP)-fMRI tests specificity, showing minimal spurious directional asymmetry relative to scan-to-scan variability and consistent temporal dynamics across neural and BOLD signals. Mouse Ca2+-fMRI tests temporal variability, showing that channel-capacity patterns identify reproducible connectivity states and transitions over time. Together, these results establish channel capacity as a physiologically grounded framework for measuring dynamic directional interactions across species and neuroimaging modalities.

Mesoscale neuronal networks represent an intermediate organizational level linking single-neuron activity to large-scale brain networks. Here, we used in vivo two-photon calcium imaging and graph-theoretical analysis to characterize functional network topology in the primary motor cortex across behavioral states. Motion networks exhibited the largest functional connectivity architectures, whereas anesthesia networks showed reduced network scales together with stronger modular segregation and more pronounced small-world topology. Network sign further shaped topology, with negative associations associated with reduced modularity and weakened small-world structure. Hub analyses revealed additional state-dependent differences: anesthesia networks exhibited stronger hub connectivity despite reduced neuronal activity, whereas motion networks showed higher hub activity with weaker connectivity structure. These findings demonstrate that mesoscale neuronal networks exhibit structured and state-dependent organization and provide a framework for studying cortical network dynamics in normal brain function and brain disorders.

The relationship between neonatal brain activity patterns and later cognitive development has become a central topic in developmental neuroscience. Addressing this question requires whole-brain analytical approaches capable of identifying which large-scale functional systems carry stable and generalizable predictive signals. However, most existing studies remain focused on specific brain regions or localized functional circuits, such as thalamocortical pathways and amygdala-centered emotional networks. While these region-specific investigations have provided important insights, they are inherently limited in terms of robustness and cross-sample generalizability. As a result, systematic evidence identifying which large-scale functional systems reliably support stable and generalizable predictive signals remains scarce. Overcoming the methodological constraints of conventional whole-brain analytical paradigms has therefore become a key bottleneck in advancing our understanding of how early brain activity patterns relate to subsequent cognitive development. Here, using data from 402 infants in the developing Human Connectome Project (278 term-born; 124 preterm-born), we introduce a region-of-interest (ROI)-constrained variant of Connectome-Based Predictive Modeling (CPM) that incorporates ROI-degree-guided feature selection to predict 18-month Bayley-III cognitive, language, and motor outcomes. Model performance declined as progressively lower-degree regions were included, indicating that conventional whole-connectome CPM may obscure robust predictive signals by incorporating low signal-to-noise (SNR) features. Our models robustly predicted cognitive, language, and motor outcomes at 18 months of age. Cohort-specific connectivity patterns emerged. In term-born infants, dominant predictive features were concentrated in visual-auditory interactions, as well as connections between visual and auditory networks and other cortical regions. Interhemispheric and intrahemispheric connections contributed in roughly equal proportions. In contrast, among preterm infants, predictive features were primarily concentrated in connectivity involving auditory and temporoparietal networks, with interhemispheric connections comprising approximately twice the number of intrahemispheric connections. The whole-cohort model (term + preterm) reflected the combined contributions of both term- and preterm-associated connectivity patterns. Predictions generalized across Bayley composite and subscale scores and were supported by permutation testing and held-out validation. These findings identify early sensory hubs--particularly visual and auditory regions--as promising early biomarkers for later neurodevelopmental outcomes. Furthermore, they demonstrate that ROI-constrained CPM can reveal meaningful predictive signals that may be obscured by conventional connectome-wide approaches.

The early development of the prefrontal cortex is crucial for higher cognitive functions. However, current research presents inconsistent findings regarding whether intra-prefrontal connectivity increases or decreases in infants younger than six months. Do dynamic changes in connection strength across different states over time carry information about prefrontal maturation? This study used functional near-infrared spectroscopy (fNIRS) to record prefrontal brain activity in 48 healthy infants aged 1-8 months during natural sleep and auditory stimulation. By analyzing the fluctuations in frequency-domain characteristics of functional connectivity (FC) and various brain network properties, we found that: under auditory stimulation, the intensity of FC fluctuations in the ultra-low frequency range was positively correlated with age; while in the resting state, the fluctuation intensity of network properties in relatively higher frequency bands decreased with age. Furthermore, auditory stimulation reconfigured the energy distribution of network fluctuations, shifting it towards higher frequency bands. These results suggest that the early development of the infant prefrontal internal network is characterized by state-dependent optimization of its dynamic fluctuation properties, shedding light on the developmental tuning of functional network dynamics in infancy.

Lesion network mapping (LNM) is increasingly used to link focal brain lesions to distributed functional networks. Recent work has raised concerns that LNM results may be spatially biased by dominant features of the normative connectome. If this were the case, three testable predictions would follow: (i) a consistent spatial pattern of false positives across LNM studies, (ii) that this pattern can be consistently explained by intrinsic connectome organization, and (iii) that symptom-associated LNM findings preferentially occur in regions with high spatial bias. We tested these predictions across three independent LNM datasets (n = 49/101/200), evaluating each prediction in all cohorts. Spatial bias maps derived from 4,000,000 random permutations under the null hypothesis showed minimal correspondence across cohorts (R2 = 0.4-0.8%), indicating strong cohort specificity. Moreover, dominant connectome features--captured by the first 10 principal components of connectivity profiles from 1,000 atlas regions--did not systematically explain these bias maps. Finally, symptom-associated results showed no enrichment in high-bias regions. Together, these findings provide strong evidence that spatial bias in LNM is not driven by dominant connectome features. With appropriate inferential statistics and rigorous study design, LNM remains a valid approach for mapping symptom-related brain networks.

Because many neurobehavioural functions rely on distributed brain networks, anatomically diverse brain lesions can cause the same neurobehavioural deficit. Lesion network mapping (LNM) builds on this principle to better understand the functional anatomy of the brain, by mapping focal lesions onto a normative functional connectome. Yet, recent work raised concerns about anatomical specificity of LNM, showing that commonly used LNM procedures converge on nonspecific connectome properties. Here, we show that anatomically specific LNM is possible with the right statistical approach using symptom-label permutation as a null model. We demonstrate this in a multicenter dataset of 2,950 stroke patients across 12 cohorts, comparing patients with and without impairment in 6 cognitive domains. First, we showed that permutation-based LNM yielded distinct and biologically plausible network maps with modest cross-cognitive domain similarity. Second, we replicated the previously raised concern of nonspecific connectome-driven convergence when using parametric statistics. Third, we assessed specificity of our approach through simulation analyses across 10,000 null studies which confirmed that the permutation framework maintained valid type I error control. These findings demonstrate that permutation-based null models preserve anatomical specificity in LNM, enabling the identification of brain networks that are genuinely linked to distinct neurobehavioural functions. This approach may thus allow researchers to more reliably map the network basis of neurobehavioural deficits from focal brain lesions.

The brains functional organization has been extensively studied through pairwise connectivity analyses. While these approaches have provided important insights into brain network organization, they fall short in capturing the complexity of high-order functional interactions (HOI). Particularly relevant is the investigation of redundancy and synergy patterns -not addressable with pairwise interactions-, revealing fundamental mechanisms of brain integration and information processing across various cognitive functions and clinical conditions. Conventional neuroimaging software packages are primarily designed for classical (general linear model-like) analyses but lack native support for HOI metrics. To address this gap, this study introduces a novel framework that bridges high-order information theory with conventional neuroimaging analysis pipelines and is subsequently applied to resting-state functional MRI to demonstrate its practical utility. By representing HOI into voxel-level metrics, our approach allows standard neuroimaging analyses to probe complex multivariate dependencies. Moreover, voxel-level group-comparison analyses show age differences linked with reduced redundancy in default mode network interactions. These findings advance our understanding of the complex relationship between multivariate functional interactions, voxel-level neuroimaging, and behavior, highlighting novel analytic strategies to study high-order information processing underlying cognitive function and its alterations in pathological conditions.

Sequential experimental paradigms are fundamental to cognitive neuroscience, yet standard event-related response analysis struggles with the temporal variability inherent to these designs. Conventional epoching treats each event within a sequence as an independent response, discarding the temporal dependencies between successive events and obscuring systematic changes in neural state that accumulate across the sequence. In order to generate responses that capture the entire sequence, it necessitates alignment across trials to correct for the inherent temporal jitter that would otherwise blur averaged responses and obscure the true sequential dynamics. Existing temporal alignment methods warp observed signals directly, making them vulnerable to correlated noise and potentially disrupting multichannel temporal relationships essential for connectivity and causal analyses. Event-Related Warping (ERW) addresses these limitations by aligning template functions encoding experimental event structure rather than neural signals themselves. Templates constructed from event onsets and durations undergo smooth monotonic warping via gradient-based optimisation, then estimated trajectories are applied uniformly across all channels, preserving inter-channel timing relationships and causal structure. This design-level alignment exploits experimentally observable jitter whilst maintaining signal integrity. Simulations with known ground truth incorporating Gaussian jitter, skewed latencies, amplitude-latency coupling, and multi-parameter dependencies yielded standardised root-mean-square errors (sRMSE) of 0.27-0.38. Distance-weighted averaging, emphasising temporally consistent trials, provided 5-13% improvement when jitter exceeded 100 ms, with maximal benefit ({approx}13% reduction) under quadratic amplitude-latency coupling. Empirical validation using an auditory go/no-go dataset with cue-to-target intervals of 1.5-4.1 seconds demonstrated that ERW recovers jittered target-locked responses with comparable fidelity (sRMSE 0.24-0.51) to conventional epoching of time-locked events, whilst preserving inter-channel lag relationships (cross-covariance sRMSE 0.63-0.82). ERW thus extends standard trial averaging to scenarios where temporal variability would otherwise preclude coherent response recovery, supporting investigation of temporally extended processing in ecologically valid paradigms whilst maintaining compatibility with established ERP frameworks and downstream connectivity analyses.

PurposeWe investigate whether bilingual versus monolingual language environments in early infancy are associated with differences in intrinsic functional organization measured from resting-state fNIRS connectivity. ApproachUsing the RS4 infant resting-state fNIRS cohort (HbO), we studied two complementary subject-level representations of resting-state connectivity: correlation-based symmetric positive definite (SPD) operators and learned-graph Laplacian operators. Correlation matrices were estimated over fixed non-overlapping temporal windows, regularized by shrinkage, and aggregated at the subject level using a Jensen- Bregman LogDet (JBLD/Stein) barycentric mean. Dominant eigenspaces were used as compact descriptors of functional organization and compared across subjects through principal angles augmented with spectral jump features. In parallel, learned functional graphs provided a complementary Laplacian-based representation of network structure. All analyses followed a strict leave-one-subject-out protocol on a common subject set (N = 94), with all templates and model parameters estimated from the training fold only. ResultsThe strongest individual branch was the correlation-based spectral-subspace representation (CORR-ANGLES: ROC-AUC = 0.811), while the learned-graph spectral branch also showed clear above-chance performance (LAP-ANGLES: ROC-AUC = 0.785). Fusion improved performance both within representation families and across them. Within-family fusion yielded ROC-AUC = 0.836 for the correlation branch and ROC-AUC = 0.805 for the Laplacian branch, whereas fusion of the two spectral branches reached ROC-AUC = 0.883, supporting the view that covariance-based and learned-graph representations capture complementary aspects of infant functional connectivity. The best overall performance was achieved by the main reported hierarchical four-branch fusion, with balanced accuracy = 0.826, F1 score = 0.781, and ROC-AUC = 0.900. ConclusionsResting-state infant fNIRS contains subtle spectral-geometric structure associated with bilingual exposure. Correlation-based and learned-graph representations provide complementary information, and their hierarchical fusion improves separability under strict cross-subject evaluation.

Dynamic functional connectivity (dFC) models have become increasingly popular over the past decade for characterising time-varying interactions between brain regions. However, assessing and comparing dFC models remains challenging. Here, we introduce bi-cross-validation as a general framework for evaluating dFC models and selecting key hyperparameters, such as the number of states. By jointly partitioning the data across subjects and brain regions, bi-cross-validation enables out-of-sample evaluation without re-estimating latent states on the same data used for testing, thereby avoiding circularity. Using simulated data with known ground-truth dynamics, we show that bi-cross-validation favours models that accurately capture the underlying state structure. Applying the framework to real resting-state fMRI data, we demonstrate that bi-cross-validation naturally balances goodness-of-fit against model complexity, with performance improving and then declining as model complexity increases. Finally, we use bi-cross-validation to directly compare static and dynamic FC models, showing that dynamic models underperform static models at low spatial dimensionality, but outperform static models at sufficiently high dimensionality. Together, these results establish bi-cross-validation as a principled tool for dFC model selection, evaluation, and comparison.

Modeling individual brain dynamics from resting-state fMRI (rs-fMRI) remains challenging due to substantial inter-subject variability, measurement noise, and limited data length per subject. Here, we systematically evaluate a hierarchical dynamical systems framework based on shallow piecewise-linear recurrent neural networks (shPLRNNs) for individualized modeling of rs-fMRI data, with a particular focus on reproducing subject-specific functional connectivity (FC). We applied the framework to 1,423 rs-fMRI samples from healthy participants of the Marburg-Munster Affective Disorders Cohort Study (MACS). Simulated rs-fMRI data robustly reproduced empirical FC patterns, with comparable reconstruction accuracy on training and independent validation sets. Generalization to unseen individuals was heterogeneous and strongly depended on how typical a subjects connectivity pattern was relative to the training cohort, with template similarity explaining 37% of variance in reconstruction accuracy. Learned subject-specific parameters exhibited significant test-retest stability and higher within-subject than between-subject similarity on longitudinal data from two different timepoints, supporting their interpretation as individualized dynamical markers. Associations between individual parameters and demographic or cognitive variables were statistically significant but modest in effect size, and predictive performance remained below that obtained using empirical rs-fMRI features directly. Together, these results demonstrate that hierarchical shPLRNNs can extract meaningful and stable individual-specific dynamical structure from rs-fMRI data, while highlighting current limitations in capturing fine-grained individual differences. The findings delineate key trade-offs between model expressivity, generalization and subject specificity, and point to directions for future methodological refinement in individualized brain modeling.