Source-linked AI summary
Machine learning in resting-state fMRI analysis
Meenakshi Khosla, Keith Jamison, Gia H. Ngo, Amy Kuceyeski, Mert R. Sabuncu
TL;DR
rs-fMRI analysis must characterize rich, high-dimensional spontaneous activity while supporting both discovery of brain organization and individual-level prediction. This review surveys unsupervised and supervised machine-learning applications, organizing unsupervised methods across spatial, temporal, and population structure and examining supervised features and algorithms. It concludes by synthesizing the field’s current applications and identifying unresolved limitations, including the absence of a universal gold-standard atlas and constraints of commonly used methods.
Problem
rs-fMRI contains high-dimensional functional-connectivity data, while appropriate tools are needed to characterize its spatial organization, dynamics, and individual-level variation.
Method
The paper provides a comprehensive review and taxonomy of unsupervised and supervised machine-learning applications in rs-fMRI, organizing unsupervised methods by spatial, temporal, and population focus.
Results
The review identifies broad applications spanning functional-network mapping, dynamic connectivity-state discovery, population analysis, and supervised subject-level prediction.
Takeaways & Limitations
Machine learning provides a framework for extracting functional organization and dynamics from rs-fMRI and for studying individual differences and brain disorders.
Takeaways & Limitations
No unified framework or universal gold-standard atlas exists for evaluating the diverse brain parcellations produced by different methods and modalities.
Abstract
from arXiv · showhide
Machine learning techniques have gained prominence for the analysis of resting-state functional Magnetic Resonance Imaging (rs-fMRI) data. Here, we present an overview of various unsupervised and supervised machine learning applications to rs-fMRI. We present a methodical taxonomy of machine learning methods in resting-state fMRI. We identify three major divisions of unsupervised learning methods with regard to their applications to rs-fMRI, based on whether they discover principal modes of variation across space, time or population. Next, we survey the algorithms and rs-fMRI feature representations that have driven the success of supervised subject-level predictions. The goal is to provide a high-level overview of the burgeoning field of rs-fMRI from the perspective of machine learning applications.
1. Introduction
Resting-state fMRI measures spontaneous whole-brain BOLD fluctuations and supports investigation of functional organization, connectivity, and brain disorders. This review organizes machine-learning applications by unsupervised discovery across space, time, and populations, alongside supervised individual-level prediction.
- Background: Resting-state fMRI measures spontaneous BOLD fluctuations across the whole brain without a controlled experimental paradigm.
- Background: Correlated spontaneous fluctuations form resting-state networks, revealing temporally coherent activity among functionally related regions.
- Motivation: RSFC measures statistical dependence among distributed brain regions, and disruptions have been identified across neurological and psychiatric disorders.
- Review scope: Unsupervised methods such as decomposition and clustering discover spatial organization, temporal dynamics, or population structure, whereas supervised methods support individual-level predictions.
- Background: Traditional seed-based analysis correlates an a priori seed with all brain voxels but is limited by manual seed selection and one-system-at-a-time analysis.
- Temporal dynamics: Dynamic functional connectivity studies identify recurring connectivity states, whose dwell times differ between healthy controls and several patient populations.
2. Unsupervised Learning
Unsupervised learning in rs-fMRI extracts latent structure from high-dimensional, unlabeled data to characterize functional organization and dynamics. Applications include functional parcellation, clustering, and decomposition of spatial or temporal patterns.
- Overview: Unsupervised learning discovers latent representations and separates explanatory structure from noise without target outputs or labels.
- Clustering: K-means assigns observations to nearest-centroid Voronoi cells, while GMMs estimate Gaussian densities for probabilistic cluster assignments.
- Clustering: Hierarchical clustering creates nested partitions by iteratively merging clusters according to a linkage criterion.
- Clustering: Graph-based clustering partitions graph-structured data to minimize connections between distinct clusters.
- Functional organization: Functional parcellation segments the brain into discrete sub-units using functional data rather than cytoarchitectural or macroscopic anatomical features.
- Spatial organization: Decomposition expresses fMRI data through spatial patterns and associated time series, while clustering groups voxel time series or connectivity fingerprints into functional networks.
2.1. Discovering spatial patterns with coherent fluctuations
Machine-learning methods discover spatial organization in rs-fMRI by decomposing activity, learning sparse representations, or clustering functionally related regions. These approaches reveal reproducible and individualized network structure, but depend on modeling, graph-construction, distance, and parcellation choices.
- rs-fMRI and machine learning identify functionally distinct neuroanatomical boundaries and clusters of functionally coupled regions.
- Decomposition and dictionary learning: Decomposition methods represent 4D fMRI data as spatial modes with coherent temporal dynamics, including statistically independent components from ICA.ICA expresses brain activity as spatial maps with characteristic time courses and supports group-level inference through concatenation and back-projection or dual regression.
- Decomposition and dictionary learning: ICA-based group correspondence is limited because source separations can differ substantially across subjects, including through fragmentation.
- Decomposition and dictionary learning: Dictionary learning adds sparsity constraints while allowing subject-specific spatial maps to differ from a population-level atlas.This framework uses a factorization residual and a penalty on deviations of individual subject maps from the population representation.
- Clustering and parcellation: Clustering partitions brain surfaces or volumes into disjoint functional networks, with distances computed from voxel time series or connectivity profiles.The choice between these distance functions can produce different optimized parcellations, and the number of clusters generally requires a priori selection using validation or stability criteria.
- Clustering and parcellation: K-means and mixture models reveal intrinsic functional organization, including hierarchical networks and reproducible 7- and 17-network cortical parcellations.Yeo et al. estimated large-scale cortical organization from 1000 subjects; stability analysis selected the 7- and 17-network resolutions, which showed high replicability across samples.
- Limitations and comments: Spatial parcellation lacks a universal gold standard: clustering performance depends on graph construction, and no parcellation is consistently superior across evaluation metrics.Hierarchical clustering can also be biased by prior dimensionality reduction, greedy early partitions, noisy signals, and arbitrary distance-metric choices.
2.2. Discovering patterns of dynamic functional connectivity
Unsupervised methods characterize dynamic functional connectivity through recurring states, Markov transitions, and decomposed latent patterns across time. These approaches reveal subject-, behavior-, and disease-related variation while showing that dynamic connectivity can involve multiple overlapping patterns.
- Dynamic connectivity states: Clustering and generative models identify recurring connectivity states and estimate subject-level state occupancy or dwell time.The alternative hypotheses are cycling between discrete states or expressing connectivity as combinations of latent states.
- Clustering: K-means studies find recurring states whose functional connectivity departs from static FC and whose dwell times differ across several patient populations.Reported group differences include schizophrenia, bipolar disorder, and psychotic-like experience domains.
- Clustering: State estimates and summary measures, including mean dwell times and transition probabilities, were reproducible across independent population samples.Both standard and soft k-means were evaluated in the large-scale replicability study.
- Clustering: Dynamic FC states can correspond to internal high- and low-arousal states, linking RSFC fluctuations with behavioral state dependence.Two stable states identified by k-means showed correspondence with high- and low-arousal internal states.
- Markov modelling: HMMs model unobserved discrete states and reveal subject-specific occupancy, non-random transitions, hierarchical organization, and clinical discrimination of MCI.Relative occupancy was linked with behavioral traits and heredity, while modeled dynamics distinguished MCI patients from controls.
- Decomposition: Matrix decomposition assigns each dynamic FC matrix multiple latent factors, and reproducibility analysis indicates that dFC is better characterized by overlapping FC patterns.Decomposition has also revealed network-dynamic alterations across PTSD, multiple sclerosis, and developmental stages.
2.3. Disentangling latent factors of inter-subject FC variation
Unsupervised learning disentangles inter-subject FC variation by embedding connectomes in lower-dimensional spaces or grouping subjects by connectivity. These strategies address high dimensionality and support phenotype-related analyses, while autoencoder pre-training has improved some supervised classifications.
- Population-level variation: Population-level unsupervised learning either creates low-dimensional FC embeddings for later prediction or groups subjects by connectivity to differentiate phenotypes.These are the two applications identified for disentangling latent explanatory factors across populations.
- Dimensionality reduction: FC features grow as O(n2) with the number of regions, while typical sample sizes of tens or hundreds complicate learning generalizable patterns.This curse of dimensionality motivates dimensionality reduction for functional connectivity data.
- Dimensionality reduction: PCA, sparse dictionary learning, LLE, and autoencoders reduce or represent high-dimensional FC data for prediction, clustering, or feature compression.LLE preserves local distances, whereas autoencoders approximate inputs with minimal reconstruction loss.
- Dimensionality reduction: Autoencoder pre-training improved classification performance for autism and schizophrenia using RSFC.The pre-training stage directs supervised neural-network learning toward parameter spaces supporting generalization.
- Population grouping: Clustering groups subjects with similar FC and has been used to associate connectivity patterns with depression categories and identify depression subtypes.Maximum margin clustering separated depressed participants from controls, while hierarchical clustering identified clustered dysfunctional-connectivity patterns.
3. Supervised Learning
Supervised learning learns mappings from rs-fMRI features to labels or target predictions for unseen subjects. In rs-fMRI, applications include patient-control classification, disease prognosis and treatment guidance, and prediction of individual cognitive traits.
- Definition: Supervised learning learns a mapping from input features and corresponding labels to predict previously unseen data points.Autism prediction from rs-fMRI correlations is given as an example problem.
3.1. Deriving connectomic features
Supervised connectomic prediction commonly derives ROI time series and pairwise connectivity matrices before applying classification or regression. Feature construction must balance neurobiological representation with dimensionality and covariance-estimation constraints.
- Region definition: The common connectome pipeline begins by defining brain regions and extracting one time course per ROI, producing an N × T representation.Dense voxel-level connectomes are rarely used because of high dimensionality; functional, anatomical, or data-driven atlases reduce it.
- Region definition: Because whole-brain RSFC uses ROI pairs, the number of connectivity features grows as O(N 2) with the number of regions.This scaling is a central dimensionality constraint in connectomic feature extraction.
- Connectivity strength: The next pipeline step estimates covariance and converts it into connectivity strength, commonly using Pearson correlation, partial correlation, or tangent-based reparametrization.Shrinkage transformations can partially reduce covariance-estimation error caused by limited time points.
- Connectivity strength: Partial correlation estimates normalized association between two time series after removing effects of all other time series.It has yielded better network-connection estimates than Pearson correlation in simulated rs-fMRI data.
- Graph representations: Graph-based representations treat parcellated regions as nodes and functional connectivity as edge weights for deriving network-topology measures.Examples include modularity, clustering, and small-worldedness.
3.2. Algorithms
Supervised rs-fMRI studies primarily use discriminant learning to separate classes, with support vector machines widely applied and neural networks offering automated feature learning when sufficient data are available.
- Most supervised rs-fMRI methods are discriminant-based, estimating class boundaries without modeling likelihoods or posterior densities.
- Support vector machines maximize the margin between classes, with the decision boundary determined by training instances nearest to it.SVMs can also model non-linear separating boundaries.
- Neural networks learn multiple levels of feature abstraction directly from connectivity features and can approximate arbitrarily complex mappings with sufficient labelled data.Their adoption in rs-fMRI has increased alongside large-scale neuroimaging data repositories.
3.3. Applications of supervised learning
Supervised learning applies rs-fMRI connectivity to development, disease, cognition, vigilance, sleep vulnerability, heritability, and cross-modality prediction, producing individual-level biological and behavioral inferences.
- Brain development and aging: Machine learning models use RSFC to predict brain maturation and identify age-related changes in functional organization across development and aging.Reported changes include sensorimotor connectivity and increasingly distributed functional architecture, alongside applications to atypical neurodevelopment.
- Neurological and psychiatric disorders: Functional connectivity-based biomarkers have classified neurological diseases including Alzheimer’s disease, mild cognitive impairment, Parkinson’s disease, and amyotrophic lateral sclerosis.The passage reports promising accuracy but does not provide a single quantitative value.
- Neurological and psychiatric disorders: RSFC-based supervised models classify or predict symptom severity across psychiatric disorders, often using kernel-based SVMs and ROI-pair connectivity features.Applications include schizophrenia, depression, autism spectrum disorder, attention-deficit hyperactivity disorder, and several other conditions.
- Cognition and behavior: RSFC models predict individual differences in fluid intelligence, attention, memory, language, and personality traits across healthy and pathological populations.These applications use resting-state connectivity as a marker of inter-individual variability across multiple behavioral domains.
- Vigilance fluctuations and sleep studies: SVM classifiers detect sleep periods during rs-fMRI scans, revealing loss of wakefulness in one-third of subjects as early as 3 minutes into scanning.Vigilance classification can address confounds and help characterize functional reconfiguration during transitions into sleep.
- Vigilance fluctuations and sleep studies: Well-rested-state connectivity distinguishes people vulnerable to vigilance decline after sleep deprivation from more resilient subjects, revealing network differences between groups.
- Other neuroimaging modalities: Machine learning relates resting-state connectivity to genetic similarity, task-evoked responses, and structural connectivity measured with diffusion-weighted imaging.Joint structural-functional modeling also produced latent connectivity estimates that discriminated control and schizophrenic populations.
4. Discussion
The review identifies evaluation, feature-definition, multi-site heterogeneity, confounding, generalizability, interpretability, uncertainty, and differential diagnosis as major barriers to robust clinical use of rs-fMRI machine learning.
- Unsupervised learning: Unsupervised learning lacks a universal ground truth for evaluating brain parcellations, whose differing organizational scales complicate comparisons.
- Unsupervised learning: Dynamic connectivity states are difficult to interpret because resting-state scans lack behavioral probes and the repertoire of mental states may be effectively unbounded.Current approaches often use cluster statistics to fix the number of states, while infinite HMMs remain an unexplored alternative.
- Supervised prediction: Single-subject prediction varies with atlas choice, connectivity metric, and preprocessing strategy because rs-fMRI has no recognized feature-extraction or connectivity standard.Large prediction-performance deviations have been reported in relation to these choices.
- Supervised prediction: Multi-site acquisition protocols and scanner characteristics create heterogeneity, and larger multi-site samples have shown little to no prediction-accuracy improvement over single-site studies.
- Clinical translation: Diagnostic accuracies require caution because head motion can alter connectivity and differ systematically between diseased and healthy groups.
- Clinical translation: Small rs-fMRI samples can produce large cross-validation error bars, making data splits influential and motivating emphasis on generalizability, interpretability, and uncertainty estimation.These attributes are identified as critical for clinical translation.
- Future directions: Most studies classify one disease against controls, leaving differential diagnosis across multiple psychiatric disorders as an important need.Combining rs-fMRI with diffusion-weighted MRI is identified as a promising but challenging direction.
5. Conclusions
The paper provides a comprehensive overview of machine learning in rs-fMRI, organizing the literature by applications and techniques to help identify gaps in current practice.
- The review organizes the state of the art in machine learning for rs-fMRI by applications and techniques for researchers in neuroimaging and machine learning.