Deep neural networks and kernel regression achieve comparable accuracies for functional connectivity prediction of behavior and demographics

Exploring the potential of representation and transfer learning for anatomical neuroimaging: Application to psychiatry

The perspective of personalized medicine for brain disorders requires efficient learning models for anatomical neuroimaging-based prediction of clinical conditions. There is now a consensus on the benefit of deep learning (DL) in addressing many medical imaging tasks, such as image segmentation. However, for single-subject prediction problems, recent studies yielded contradictory results when comparing DL with Standard Machine Learning (SML) on top of classical feature extraction. Most existing comparative studies were limited in predicting phenotypes of little clinical interest, such as sex and age, and using a single dataset. Moreover, they conducted a limited analysis of the employed image pre-processing and feature selection strategies. This paper extensively compares DL and SML prediction capacity on five multi-site problems, including three increasingly complex clinical applications in psychiatry namely schizophrenia, bipolar disorder, and Autism Spectrum Disorder (ASD) diagnosis. To compensate for the relative scarcity of neuroimaging data on these clinical datasets, we also evaluate three pre-training strategies for transfer learning from brain imaging of the general healthy population: self-supervised learning, generative modeling and supervised learning with age. Overall, we find similar performance between randomly initialized DL and SML for the three clinical tasks and a similar scaling trend for sex prediction. This was replicated on an external dataset. We also show highly correlated discriminative brain regions between DL and linear ML models in all problems. Nonetheless, we demonstrate that self-supervised pre-training on large-scale healthy population imaging datasets (N≈10k), along with Deep Ensemble, allows DL to learn robust and transferable representations to smaller-scale clinical datasets (N≤1k). It largely outperforms SML on 2 out of 3 clinical tasks both in internal and external test sets. These findings suggest that the improvement of DL over SML in anatomical neuroimaging mainly comes from its capacity to learn meaningful and useful abstract representations of the brain anatomy, and it sheds light on the potential of transfer learning for personalized medicine in psychiatry.

Individual characteristics outperform resting-state fMRI for the prediction of behavioral phenotypes

In this study, we aimed to compare imaging-based features of brain function, measured by resting-state fMRI (rsfMRI), with individual characteristics such as age, gender, and total intracranial volume to predict behavioral measures. We developed a machine learning framework based on rsfMRI features in a dataset of 20,000 healthy individuals from the UK Biobank, focusing on temporal complexity and functional connectivity measures. Our analysis across four behavioral phenotypes revealed that both temporal complexity and functional connectivity measures provide comparable predictive performance. However, individual characteristics consistently outperformed rsfMRI features in predictive accuracy, particularly in analyses involving smaller sample sizes. Integrating rsfMRI features with demographic data sometimes enhanced predictive outcomes. The efficacy of different predictive modeling techniques and the choice of brain parcellation atlas were also examined, showing no significant influence on the results. To summarize, while individual characteristics are superior to rsfMRI in predicting behavioral phenotypes, rsfMRI still conveys additional predictive value in the context of machine learning, such as investigating the role of specific brain regions in behavioral phenotypes.

Brain age monotonicity and functional connectivity differences of healthy subjects

Alterations in the brain's connectivity or the interactions among brain regions have been studied with the aid of resting state (rs)fMRI data attained from large numbers of healthy subjects of various demographics. This has been instrumental in providing insight into how a phenotype as fundamental as age affects the brain. Although machine learning (ML) techniques have already been deployed in such studies, novel questions are investigated in this work. We study whether young brains develop properties that progressively resemble those of aged brains, and if the aging dynamics of older brains provide information about the aging trajectory in young subjects. The degree of a prospective monotonic relationship will be quantified, and hypotheses of brain aging trajectories will be tested via ML. Furthermore, the degree of functional connectivity across the age spectrum of three datasets will be compared at a population level and across sexes. The findings scrutinize similarities and differences among the male and female subjects at greater detail than previously performed.

Computational limits to the legibility of the imaged human brain

Our knowledge of the organisation of the human brain at the population-level is yet to translate into power to predict functional differences at the individual-level, limiting clinical applications and casting doubt on the generalisability of inferred mechanisms. It remains unknown whether the difficulty arises from the absence of individuating biological patterns within the brain, or from limited power to access them with the models and compute at our disposal. Here we comprehensively investigate the resolvability of such patterns with data and compute at unprecedented scale. Across 23 810 unique participants from UK Biobank, we systematically evaluate the predictability of 25 individual biological characteristics, from all available combinations of structural and functional neuroimaging data. Over 4526 GPU*hours of computation, we train, optimize, and evaluate out-of-sample 700 individual predictive models, including fully-connected feed-forward neural networks of demographic, psychological, serological, chronic disease, and functional connectivity characteristics, and both uni- and multi-modal 3D convolutional neural network models of macro- and micro-structural brain imaging. We find a marked discrepancy between the high predictability of sex (balanced accuracy 99.7%), age (mean absolute error 2.048 years, R2 0.859), and weight (mean absolute error 2.609Kg, R2 0.625), for which we set new state-of-the-art performance, and the surprisingly low predictability of other characteristics. Neither structural nor functional imaging predicted an individual's psychology better than the coincidence of common chronic disease (p < 0.05). Serology predicted chronic disease (p < 0.05) and was best predicted by it (p < 0.001), followed by structural neuroimaging (p < 0.05). Our findings suggest either more informative imaging or more powerful models will be needed to decipher individual level characteristics from the human brain. We make our models and code openly available.

Reliability and predictability of phenotype information from functional connectivity in large imaging datasets

One of the central objectives of contemporary neuroimaging research is to create predictive models that can disentangle the connection between patterns of functional connectivity across the entire brain and various behavioral traits. Previous studies have shown that models trained to predict behavioral features from the individual's functional connectivity have modest to poor performance. In this study, we trained models that predict observable individual traits (phenotypes) and their corresponding singular value decomposition (SVD) representations - herein referred to as latent phenotypes from resting state functional connectivity. For this task, we predicted phenotypes in two large neuroimaging datasets: the Human Connectome Project (HCP) and the Philadelphia Neurodevelopmental Cohort (PNC). We illustrate the importance of regressing out confounds, which could significantly influence phenotype prediction. Our findings reveal that both phenotypes and their corresponding latent phenotypes yield similar predictive performance. Interestingly, only the first five latent phenotypes were reliably identified, and using just these reliable phenotypes for predicting phenotypes yielded a similar performance to using all latent phenotypes. This suggests that the predictable information is present in the first latent phenotypes, allowing the remainder to be filtered out without any harm in performance. This study sheds light on the intricate relationship between functional connectivity and the predictability and reliability of phenotypic information, with potential implications for enhancing predictive modeling in the realm of neuroimaging research.

Beyond network connectivity: A classification approach to brain age prediction with resting-state fMRI

The brain is a complex, dynamic organ that shows differences in the same subject at various periods. Understanding how brain activity changes across age as a function of the brain networks has been greatly abetted by fMRI. Canonical analysis consists of determining how alterations in connectivity patterns (CPs) of certain regions are affected. An alternative approach is taken here by not considering connectivity but rather features computed from recordings at the regions of interest (ROIs). Using machine learning (ML) we assess how neural signals are altered by and prospectively predictive of age and sex via a methodology that is novel in drawing upon pairwise classification across six decades of subjects' chronological ages. ML is used to answer the equally important questions of what properties of the computed features are most predictive as well as which brain networks are most affected by aging. It was found that there is decreased differentiation among the neural signals of older subjects that are separated in age by the same number of years as younger subjects. Furthermore, the burstiness of the signals change at different rates between males and females. The findings provide insight into brain aging via an ROI-based analysis, the consideration of several feature groups, and a novel classification-based ML pipeline. There is also a contribution to understanding the effects of data aggregated from different recording centers on the conclusions of fMRI studies.

Whole-brain functional connectivity predicts groupwise and sex-specific tau PET in preclincal Alzheimer's disease

Preclinical Alzheimer's disease, characterized by the initial accumulation of amyloid and tau pathologies without symptoms, presents a critical opportunity for early intervention. Yet, the interplay between these pathological markers and the functional connectome during this window remains understudied. We therefore set out to elucidate the relationship between the functional connectome and amyloid and tau, as assessed by PET imaging, in individuals with preclinical AD using connectome-based predictive modeling (CPM). We found that functional connectivity predicts tau PET, outperforming amyloid PET models. These models were predominantly governed by linear relationships between functional connectivity and tau. Tau models demonstrated a stronger correlation to global connectivity than underlying tau PET. Furthermore, we identify sex-based differences in the ability to predict regional tau, without any underlying differences in tau PET or global connectivity. Taken together, these results suggest tau is more closely coupled to functional connectivity than amyloid in preclinical disease, and that multimodal predictive modeling approaches stand to identify unique relationships that any one modality may be insufficient to discern.

Lightweight neural network for Alzheimer's disease classification using multi-slice sMRI

Alzheimer's disease (AD) is a progressive neurodegenerative disease. Early detection and intervention are crucial in preventing the progression of AD. To achieve efficient and scalable AD auto-detection based on structural Magnetic Resonance Imaging (sMRI), a lightweight neural network using multi-slice sMRI is proposed in this paper. The backbone for feature extraction is based on ShuffleNet V1 architecture, which is effective for overcoming the limitations posed by limited sMRI data and resource-restricted devices. In addition, we incorporate Efficient Channel Attention (ECA) to capture cross-channel interaction information, enabling us to effectively enhance features of disease associated brain regions. To optimize the model, we employ both cross entropy loss and triplet loss functions to constrain the predicted probabilities to the ground-truth labels, and to ensure appropriate representation of distances between different classes in the learned features. Experimental results show that the classification accuracies of our method for AD vs. CN, AD vs. MCI, and MCI vs. CN classification tasks are 95.00%, 87.50%, and 85.62% respectively. Our method utilizes only 3.42 M parameters and 6.08G FLOPs, while maintaining a comparable level of performance compared to the other 5 latest lightweight methods. This model design is computationally efficient, allowing it to process large amounts of data quickly and accurately in a timely manner. Additionally, it has the potential to advance the intelligent detection of Alzheimer's disease on devices with limited computing capabilities.

Brain Networks and Intelligence: A Graph Neural Network Based Approach to Resting State fMRI Data

Resting-state functional magnetic resonance imaging (rsfMRI) is a powerful tool for investigating the relationship between brain function and cognitive processes as it allows for the functional organization of the brain to be captured without relying on a specific task or stimuli. In this paper, we present a novel modeling architecture called BrainRGIN for predicting intelligence (fluid, crystallized and total intelligence) using graph neural networks on rsfMRI derived static functional network connectivity matrices. Extending from the existing graph convolution networks, our approach incorporates a clustering-based embedding and graph isomorphism network in the graph convolutional layer to reflect the nature of the brain sub-network organization and efficient network expression, in combination with TopK pooling and attention-based readout functions. We evaluated our proposed architecture on a large dataset, specifically the Adolescent Brain Cognitive Development Dataset, and demonstrated its effectiveness in predicting individual differences in intelligence. Our model achieved lower mean squared errors, and higher correlation scores than existing relevant graph architectures and other traditional machine learning models for all of the intelligence prediction tasks. The middle frontal gyrus exhibited a significant contribution to both fluid and crystallized intelligence, suggesting their pivotal role in these cognitive processes. Total composite scores identified a diverse set of brain regions to be relevant which underscores the complex nature of total intelligence.

Few-Shot Domain-Adaptive Anomaly Detection for Cross-Site Brain Images

Early screening is essential for effective intervention and treatment of individuals with mental disorders. Functional magnetic resonance imaging (fMRI) is a noninvasive tool for depicting neural activity and has demonstrated strong potential as a technique for identifying mental disorders. Due to the difficulty in data collection and diagnosis, imaging data from patients are rare at a single site, whereas abundant healthy control data are available from public datasets. However, joint use of these data from multiple sites for classification model training is hindered by cross-domain distribution discrepancy and diverse label spaces. Herein, we propose few-shot domain-adaptive anomaly detection (FAAD) to achieve cross-site anomaly detection of brain images based on only a few labeled samples. We introduce domain adaptation to mitigate cross-domain distribution discrepancy and jointly align the general and conditional feature distributions of imaging data across multiple sites. We utilize fMRI data of healthy subjects in the Human Connectome Project (HCP) as the source domain and fMRI images from six independent sites, including patients with mental disorders and demographically matched healthy controls, as target domains. Experiments showed the superiority of the proposed method compared with binary classification, traditional anomaly detection methods, and several recognized domain adaptation methods.

MRI economics: Balancing sample size and scan duration in brain wide association studies

A pervasive dilemma in neuroimaging is whether to prioritize sample size or scan duration given fixed resources. Here, we systematically investigate this trade-off in the context of brain-wide association studies (BWAS) using resting-state functional magnetic resonance imaging (fMRI). We find that total scan duration (sample size × scan duration per participant) robustly explains individual-level phenotypic prediction accuracy via a logarithmic model, suggesting that sample size and scan duration are broadly interchangeable. The returns of scan duration eventually diminish relative to sample size, which we explain with principled theoretical derivations. When accounting for fixed costs associated with each participant (e.g., recruitment, non-imaging measures), we find that prediction accuracy in small-scale BWAS might benefit from much longer scan durations (>50 min) than typically assumed. Most existing large-scale studies might also have benefited from smaller sample sizes with longer scan durations. Both logarithmic and theoretical models of the relationships among sample size, scan duration and prediction accuracy explain well-predicted phenotypes better than poorly-predicted phenotypes. The logarithmic and theoretical models are also undermined by individual differences in brain states. These results replicate across phenotypic domains (e.g., cognition and mental health) from two large-scale datasets with different algorithms and metrics. Overall, our study emphasizes the importance of scan time, which is ignored in standard power calculations. Standard power calculations inevitably maximize sample size at the expense of scan duration. The resulting prediction accuracies are likely lower than would be produced with alternate designs, thus impeding scientific discovery. Our empirically informed reference is available for future study design: WEB_APPLICATION_LINK.

Prediction of cognitive performance differences in older age from multimodal neuroimaging data

Differences in brain structure and functional and structural network architecture have been found to partly explain cognitive performance differences in older ages. Thus, they may serve as potential markers for these differences. Initial unimodal studies, however, have reported mixed prediction results of selective cognitive variables based on these brain features using machine learning (ML). Thus, the aim of the current study was to investigate the general validity of cognitive performance prediction from imaging data in healthy older adults. In particular, the focus was with examining whether (1) multimodal information, i.e., region-wise grey matter volume (GMV), resting-state functional connectivity (RSFC), and structural connectivity (SC) estimates, may improve predictability of cognitive targets, (2) predictability differences arise for global cognition and distinct cognitive profiles, and (3) results generalize across different ML approaches in 594 healthy older adults (age range: 55-85 years) from the 1000BRAINS study. Prediction potential was examined for each modality and all multimodal combinations, with and without confound (i.e., age, education, and sex) regression across different analytic options, i.e., variations in algorithms, feature sets, and multimodal approaches (i.e., concatenation vs. stacking). Results showed that prediction performance differed considerably between deconfounding strategies. In the absence of demographic confounder control, successful prediction of cognitive performance could be observed across analytic choices. Combination of different modalities tended to marginally improve predictability of cognitive performance compared to single modalities. Importantly, all previously described effects vanished in the strict confounder control condition. Despite a small trend for a multimodal benefit, developing a biomarker for cognitive aging remains challenging.

Finger motor representation supports the autonomy in arithmetic: neuroimaging evidence from abacus training

Researches have reported the close association between fingers and arithmetic. However, it remains unclear whether and how finger training can benefit arithmetic. To address this issue, we used the abacus-based mental calculation (AMC), which combines finger training and mental arithmetic learning, to explore the neural correlates underlying finger-related arithmetic training. A total of 147 Chinese children (75 M/72 F, mean age, 6.89 ± 0.46) were recruited and randomly assigned into AMC and control groups at primary school entry. The AMC group received 5 years of AMC training, and arithmetic abilities and resting-state functional magnetic resonance images data were collected from both groups at year 1/3/5. The connectome-based predictive modeling was used to find the arithmetic-related networks of each group. Compared to controls, the AMC's positively arithmetic-related network was less located in the control module, and the inter-module connections between somatomotor-default and somatomotor-control modules shifted to somatomotor-visual and somatomotor-dorsal attention modules. Furthermore, the positive network of the AMC group exhibited a segregated connectivity pattern, with more intra-module connections than the control group. Overall, our results suggested that finger motor representation with motor module involvement facilitated arithmetic-related network segregation, reflecting increased autonomy of AMC, thus reducing the dependency of arithmetic on higher-order cognitive functions.

Performance reserves in brain-imaging-based phenotype prediction

This study examines the impact of sample size on predicting cognitive and mental health phenotypes from brain imaging via machine learning. Our analysis shows a 3- to 9-fold improvement in prediction performance when sample size increases from 1,000 to 1 M participants. However, despite this increase, the data suggest that prediction accuracy remains worryingly low and far from fully exploiting the predictive potential of brain imaging data. Additionally, we find that integrating multiple imaging modalities boosts prediction accuracy, often equivalent to doubling the sample size. Interestingly, the most informative imaging modality often varied with increasing sample size, emphasizing the need to consider multiple modalities. Despite significant performance reserves for phenotype prediction, achieving substantial improvements may necessitate prohibitively large sample sizes, thus casting doubt on the practical or clinical utility of machine learning in some areas of neuroimaging.

Macroscopic resting-state brain dynamics are best described by linear models

It is typically assumed that large networks of neurons exhibit a large repertoire of nonlinear behaviours. Here we challenge this assumption by leveraging mathematical models derived from measurements of local field potentials via intracranial electroencephalography and of whole-brain blood-oxygen-level-dependent brain activity via functional magnetic resonance imaging. We used state-of-the-art linear and nonlinear families of models to describe spontaneous resting-state activity of 700 participants in the Human Connectome Project and 122 participants in the Restoring Active Memory project. We found that linear autoregressive models provide the best fit across both data types and three performance metrics: predictive power, computational complexity and the extent of the residual dynamics unexplained by the model. To explain this observation, we show that microscopic nonlinear dynamics can be counteracted or masked by four factors associated with macroscopic dynamics: averaging over space and over time, which are inherent to aggregated macroscopic brain activity, and observation noise and limited data samples, which stem from technological limitations. We therefore argue that easier-to-interpret linear models can faithfully describe macroscopic brain dynamics during resting-state conditions.

Multilayer meta-matching: translating phenotypic prediction models from multiple datasets to small data

Resting-state functional connectivity (RSFC) is widely used to predict phenotypic traits in individuals. Large sample sizes can significantly improve prediction accuracies. However, for studies of certain clinical populations or focused neuroscience inquiries, small-scale datasets often remain a necessity. We have previously proposed a "meta-matching" approach to translate prediction models from large datasets to predict new phenotypes in small datasets. We demonstrated large improvement of meta-matching over classical kernel ridge regression (KRR) when translating models from a single source dataset (UK Biobank) to the Human Connectome Project Young Adults (HCP-YA) dataset. In the current study, we propose two meta-matching variants ("meta-matching with dataset stacking" and "multilayer meta-matching") to translate models from multiple source datasets across disparate sample sizes to predict new phenotypes in small target datasets. We evaluate both approaches by translating models trained from five source datasets (with sample sizes ranging from 862 participants to 36,834 participants) to predict phenotypes in the HCP-YA and HCP-Aging datasets. We find that multilayer meta-matching modestly outperforms meta-matching with dataset stacking. Both meta-matching variants perform better than the original "meta-matching with stacking" approach trained only on the UK Biobank. All meta-matching variants outperform classical KRR and transfer learning by a large margin. In fact, KRR is better than classical transfer learning when less than 50 participants are available for finetuning, suggesting the difficulty of classical transfer learning in the very small sample regime. The multilayer meta-matching model is publicly available at GITHUB_LINK.

Multi-level and joint attention networks on brain functional connectivity for cross-cognitive prediction

Deep learning on resting-state functional MRI (rs-fMRI) has shown great success in predicting a single cognition or mental disease. Nevertheless, cognitive functions or mental diseases may share neural mechanisms that can benefit their prediction/classification. We propose a multi-level and joint attention (ML-Joint-Att) network to learn high-order representations of brain functional connectivities that are specific and shared across multiple tasks. We design the ML-Joint-Att network with edge and node convolutional operators, an adaptive inception module, and three attention modules, including network-wise, region-wise, and region-wise joint attention modules. The adaptive inception learns brain functional connectivity at multiple spatial scales. The network-wise and region-wise attention modules take the multi-scale functional connectivities as input and learn features at the network and regional levels for individual tasks. Moreover, the joint attention module is designed as region-wise joint attention to learn shared brain features that contribute to and compensate for the prediction of multiple tasks. We employed the Adolescent Brain Cognitive Development (ABCD) dataset (n =9092) to evaluate the ML-Joint-Att network for the prediction of cognitive flexibility and inhibition. Our experiments demonstrated the usefulness of the three attention modules and identified brain functional connectivities and regions specific and common between cognitive flexibility and inhibition. In particular, the joint attention module can significantly improve the prediction of both cognitive functions. Moreover, leave-one-site cross-validation showed that the ML-Joint-Att network is robust to independent samples obtained from different sites of the ABCD study. Our network outperformed existing machine learning techniques, including Brain Bias Set (BBS), spatio-temporal graph convolution network (ST-GCN), and BrainNetCNN. We demonstrated the generalization of our method to other applications, such as the prediction of fluid intelligence and crystallized intelligence, which also outperformed the ST-GCN and BrainNetCNN.

Gaussian Process-based prediction of memory performance and biomarker status in ageing and Alzheimer's disease-A systematic model evaluation

Neuroimaging markers based on Magnetic Resonance Imaging (MRI) combined with various other measures (such as genetic covariates, biomarkers, vascular risk factors, neuropsychological tests etc.) might provide useful predictions of clinical outcomes during the progression towards Alzheimer's disease (AD). The use of multiple features in predictive frameworks for clinical outcomes has become increasingly prevalent in AD research. However, many studies do not focus on systematically and accurately evaluating combinations of multiple input features. Hence, the aim of the present work is to explore and assess optimal combinations of various features for MR-based prediction of (1) cognitive status and (2) biomarker positivity with a multi-kernel learning Gaussian process framework. The explored features and parameters included (A) combinations of brain tissues, modulation, smoothing, and image resolution; (B) incorporating demographics & clinical covariates; (C) the impact of the size of the training data set; (D) the influence of dimensionality reduction and the choice of kernel types. The approach was tested in a large German cohort including 959 subjects from the multicentric longitudinal study of cognitive impairment and dementia (DELCODE). Our evaluation suggests the best prediction of memory performance was obtained for a combination of neuroimaging markers, demographics, genetic information (ApoE4) and CSF biomarkers explaining 57% of outcome variance in out-of-sample predictions. The highest performance for Aβ42/40 status classification was achieved for a combination of demographics, ApoE4, and a memory score while usage of structural MRI further improved the classification of individual patient's pTau status.

Functional brain networks are associated with both sex and gender in children

Sex and gender are associated with human behavior throughout the lifespan and across health and disease, but whether they are associated with similar or distinct neural phenotypes is unknown. Here, we demonstrate that, in children, sex and gender are uniquely reflected in the intrinsic functional connectivity of the brain. Unimodal networks are more strongly associated with sex while heteromodal networks are more strongly associated with gender. These results suggest sex and gender are irreducible to one another not only in society but also in biology.

A Siamese Network With Node Convolution for Individualized Predictions Based on Connectivity Maps Extracted From Resting-State fMRI Data

Deep learning has demonstrated great potential for objective diagnosis of neuropsychiatric disorders based on neuroimaging data, which includes the promising resting-state functional magnetic resonance imaging (RS-fMRI). However, the insufficient sample size has long been a bottleneck for deep model training for the purpose. In this study, we proposed a Siamese network with node convolution (SNNC) for individualized predictions based on RS-fMRI data. With the involvement of Siamese network, which uses sample pair (rather than a single sample) as input, the problem of insufficient sample size can largely be alleviated. To adapt to connectivity maps extracted from RS-fMRI data, we applied node convolution to each of the two branches of the Siamese network. For regression purposes, we replaced the contrastive loss in classic Siamese network with the mean square error loss and thus enabled Siamese network to quantitatively predict label differences. The label of a test sample can be predicted based on any of the training samples, by adding the label of the training sample to the predicted label difference between them. The final prediction for a test sample in this study was made by averaging the predictions based on each of the training samples. The performance of the proposed SNNC was evaluated with age and IQ predictions based on a public dataset (Cam-CAN). The results indicated that SNNC can make effective predictions even with a sample size of as small as 40, and SNNC achieved state-of-the-art accuracy among a variety of deep models and standard machine learning approaches.

Connectome-based predictive modeling of fluid intelligence: evidence for a global system of functionally integrated brain networks

Cognitive neuroscience continues to advance our understanding of the neural foundations of human intelligence, with significant progress elucidating the role of the frontoparietal network in cognitive control mechanisms for flexible, intelligent behavior. Recent evidence in network neuroscience further suggests that this finding may represent the tip of the iceberg and that fluid intelligence may depend on the collective interaction of multiple brain networks. However, the global brain mechanisms underlying fluid intelligence and the nature of multi-network interactions remain to be well established. We therefore conducted a large-scale Connectome-based Predictive Modeling study, administering resting-state fMRI to 159 healthy college students and examining the contributions of seven intrinsic connectivity networks to the prediction of fluid intelligence, as measured by a state-of-the-art cognitive task (the Bochum Matrices Test). Specifically, we aimed to: (i) identify whether fluid intelligence relies on a primary brain network or instead engages multiple brain networks; and (ii) elucidate the nature of brain network interactions by assessing network allegiance (within- versus between-network connections) and network topology (strong versus weak connections) in the prediction of fluid intelligence. Our results demonstrate that whole-brain predictive models account for a large and significant proportion of variance in fluid intelligence (18%) and illustrate that the contribution of individual networks is relatively modest by comparison. In addition, we provide novel evidence that the global architecture of fluid intelligence prioritizes between-network connections and flexibility through weak ties. Our findings support a network neuroscience approach to understanding the collective role of brain networks in fluid intelligence and elucidate the system-wide network mechanisms from which flexible, adaptive behavior is constructed.

Brain-Based Predictions of Psychiatric Illness-Linked Behaviors Across the Sexes

BACKGROUND

Individual differences in functional brain connectivity can be used to predict both the presence of psychiatric illness and variability in associated behaviors. However, despite evidence for sex differences in functional network connectivity and in the prevalence, presentation, and trajectory of psychiatric illnesses, the extent to which disorder-relevant aspects of network connectivity are shared or unique across the sexes remains to be determined.

METHODS

In this work, we used predictive modeling approaches to evaluate whether shared or unique functional connectivity correlates underlie the expression of psychiatric illness-linked behaviors in males and females in data from the Adolescent Brain Cognitive Development Study (N = 5260; 2571 females).

RESULTS

We demonstrate that functional connectivity profiles predict individual differences in externalizing behaviors in males and females but predict internalizing behaviors only in females. Furthermore, models trained to predict externalizing behaviors in males generalize to predict internalizing behaviors in females, and models trained to predict internalizing behaviors in females generalize to predict externalizing behaviors in males. Finally, the neurobiological correlates of many behaviors are largely shared within and across sexes: functional connections within and between heteromodal association networks, including default, limbic, control, and dorsal attention networks, are associated with internalizing and externalizing behaviors.

CONCLUSIONS

Taken together, these findings suggest that shared neurobiological patterns may manifest as distinct behaviors across the sexes. Based on these results, we recommend that both clinicians and researchers carefully consider how sex may influence the presentation of psychiatric illnesses, especially those along the internalizing-externalizing spectrum.

An atlas of trait associations with resting-state and task-evoked human brain functional organizations in the UK Biobank

Functional magnetic resonance imaging (fMRI) has been widely used to identify brain regions linked to critical functions, such as language and vision, and to detect tumors, strokes, brain injuries, and diseases. It is now known that large sample sizes are necessary for fMRI studies to detect small effect sizes and produce reproducible results. Here we report a systematic association analysis of 647 traits with imaging features extracted from resting-state and task-evoked fMRI data of more than 40,000 UK Biobank participants. We used a parcellation-based approach to generate 64,620 functional connectivity measures to reveal fine-grained details about cerebral cortex functional organizations. The difference between functional organizations at rest and during task was examined, and we have prioritized important brain regions and networks associated with a variety of human traits and clinical outcomes. For example, depression was most strongly associated with decreased connectivity in the somatomotor network. We have made our results publicly available and developed a browser framework to facilitate the exploration of brain function-trait association results (http://fmriatlas.org/).

Accounting for temporal variability in functional magnetic resonance imaging improves prediction of intelligence

Neuroimaging-based prediction methods for intelligence have seen a rapid development. Among different neuroimaging modalities, prediction using functional connectivity (FC) has shown great promise. Most literature has focused on prediction using static FC, with limited investigations on the merits of such analysis compared to prediction using dynamic FC or region-level functional magnetic resonance imaging (fMRI) times series that encode temporal variability. To account for the temporal dynamics in fMRI, we propose a bi-directional long short-term memory (bi-LSTM) approach that incorporates feature selection mechanism. The proposed pipeline is implemented via an efficient algorithm and applied for predicting intelligence using region-level time series and dynamic FC. We compare the prediction performance using different fMRI features acquired from the Adolescent Brain Cognitive Development (ABCD) study involving nearly 7000 individuals. Our detailed analysis illustrates the consistently inferior performance of static FC compared to region-level time series or dynamic FC for single and combined rest and task fMRI experiments. The joint analysis of task and rest fMRI leads to improved intelligence prediction under all models compared to using fMRI from only one experiment. In addition, the proposed bi-LSTM pipeline based on region-level time series identifies several shared and differential important brain regions across fMRI experiments that drive intelligence prediction. A test-retest analysis of the selected regions shows strong reliability across cross-validation folds. Given the large sample size of ABCD study, our results provide strong evidence that superior prediction of intelligence can be achieved by accounting for temporal variations in fMRI.

A meta-analysis and systematic review of single vs. multimodal neuroimaging techniques in the classification of psychosis

BACKGROUND

Psychotic disorders are characterized by structural and functional abnormalities in brain networks. Neuroimaging techniques map and characterize such abnormalities using unique features (e.g., structural integrity, coactivation). However, it is unclear if a specific method, or a combination of modalities, is particularly effective in identifying differences in brain networks of someone with a psychotic disorder.

METHODS

A systematic meta-analysis evaluated machine learning classification of schizophrenia spectrum disorders in comparison to healthy control participants using various neuroimaging modalities (i.e., T1-weighted imaging (T1), diffusion tensor imaging (DTI), resting state functional connectivity (rs-FC), or some combination (multimodal)). Criteria for manuscript inclusion included whole-brain analyses and cross-validation to provide a complete picture regarding the predictive ability of large-scale brain systems in psychosis. For this meta-analysis, we searched Ovid MEDLINE, PubMed, PsychInfo, Google Scholar, and Web of Science published between inception and March 13th 2023. Prediction results were averaged for studies using the same dataset, but parallel analyses were run that included studies with pooled sample across many datasets. We assessed bias through funnel plot asymmetry. A bivariate regression model determined whether differences in imaging modality, demographics, and preprocessing methods moderated classification. Separate models were run for studies with internal prediction (via cross-validation) and external prediction.

RESULTS

93 studies were identified for quantitative review (30 T1, 9 DTI, 40 rs-FC, and 14 multimodal). As a whole, all modalities reliably differentiated those with schizophrenia spectrum disorders from controls (OR = 2.64 (95%CI = 2.33 to 2.95)). However, classification was relatively similar across modalities: no differences were seen across modalities in the classification of independent internal data, and a small advantage was seen for rs-FC studies relative to T1 studies in classification in external datasets. We found large amounts of heterogeneity across results resulting in significant signs of bias in funnel plots and Egger's tests. Results remained similar, however, when studies were restricted to those with less heterogeneity, with continued small advantages for rs-FC relative to structural measures. Notably, in all cases, no significant differences were seen between multimodal and unimodal approaches, with rs-FC and unimodal studies reporting largely overlapping classification performance. Differences in demographics and analysis or denoising were not associated with changes in classification scores.

CONCLUSIONS

The results of this study suggest that neuroimaging approaches have promise in the classification of psychosis. Interestingly, at present most modalities perform similarly in the classification of psychosis, with slight advantages for rs-FC relative to structural modalities in some specific cases. Notably, results differed substantially across studies, with suggestions of biased effect sizes, particularly highlighting the need for more studies using external prediction and large sample sizes. Adopting more rigorous and systematized standards will add significant value toward understanding and treating this critical population.

The challenges and prospects of brain-based prediction of behaviour

Relating individual brain patterns to behaviour is fundamental in system neuroscience. Recently, the predictive modelling approach has become increasingly popular, largely due to the recent availability of large open datasets and access to computational resources. This means that we can use machine learning models and interindividual differences at the brain level represented by neuroimaging features to predict interindividual differences in behavioural measures. By doing so, we could identify biomarkers and neural correlates in a data-driven fashion. Nevertheless, this budding field of neuroimaging-based predictive modelling is facing issues that may limit its potential applications. Here we review these existing challenges, as well as those that we anticipate as the field develops. We focus on the impacts of these challenges on brain-based predictions. We suggest potential solutions to address the resolvable challenges, while keeping in mind that some general and conceptual limitations may also underlie the predictive modelling approach.

Automatic brain MRI motion artifact detection based on end-to-end deep learning is similarly effective as traditional machine learning trained on image quality metrics

Head motion artifacts in magnetic resonance imaging (MRI) are an important confounding factor concerning brain research as well as clinical practice. For this reason, several machine learning-based methods have been developed for the automatic quality control of structural MRI scans. Deep learning offers a promising solution to this problem, however, given its data-hungry nature and the scarcity of expert-annotated datasets, its advantage over traditional machine learning methods in identifying motion-corrupted brain scans is yet to be determined. In the present study, we investigated the relative advantage of the two methods in structural MRI quality control. To this end, we collected publicly available T1-weighted images and scanned subjects in our own lab under conventional and active head motion conditions. The quality of the images was rated by a team of radiologists from the point of view of clinical diagnostic use. We present a relatively simple, lightweight 3D convolutional neural network trained in an end-to-end manner that achieved a test set (N = 411) balanced accuracy of 94.41% in classifying brain scans into clinically usable or unusable categories. A support vector machine trained on image quality metrics achieved a balanced accuracy of 88.44% on the same test set. Statistical comparison of the two models yielded no significant difference in terms of confusion matrices, error rates, or receiver operating characteristic curves. Our results suggest that these machine learning methods are similarly effective in identifying severe motion artifacts in brain MRI scans, and underline the efficacy of end-to-end deep learning-based systems in brain MRI quality control, allowing the rapid evaluation of diagnostic utility without the need for elaborate image pre-processing.

Intermediately synchronised brain states optimise trade-off between subject specificity and predictive capacity

Functional connectivity (FC) refers to the statistical dependencies between activity of distinct brain areas. To study temporal fluctuations in FC within the duration of a functional magnetic resonance imaging (fMRI) scanning session, researchers have proposed the computation of an edge time series (ETS) and their derivatives. Evidence suggests that FC is driven by a few time points of high-amplitude co-fluctuation (HACF) in the ETS, which may also contribute disproportionately to interindividual differences. However, it remains unclear to what degree different time points actually contribute to brain-behaviour associations. Here, we systematically evaluate this question by assessing the predictive utility of FC estimates at different levels of co-fluctuation using machine learning (ML) approaches. We demonstrate that time points of lower and intermediate co-fluctuation levels provide overall highest subject specificity as well as highest predictive capacity of individual-level phenotypes.

Prediction of cognitive scores by joint use of movie-watching fMRI connectivity and eye tracking via Attention-CensNet

Background

Brain functional connectivity under the naturalistic paradigm has been shown to be better at predicting individual behaviors than other brain states, such as rest and doing tasks. Nevertheless, the state-of-the-art methods have found it difficult to achieve desirable results from movie-watching paradigm functional magnetic resonance imaging (mfMRI) -induced brain functional connectivity, especially when there are fewer datasets. Incorporating other physical measurements into the prediction method may enhance accuracy. Eye tracking, becoming popular due to its portability and lower expense, can provide abundant behavioral features related to the output of human's cognition, and thus might supplement the mfMRI in observing participants' subconscious behaviors. However, there are very few studies on how to effectively integrate the multimodal information to strengthen the performance by a unified framework.

Objective

A fusion approach with mfMRI and eye tracking, based on convolution with edge-node switching in graph neural networks (CensNet), is proposed in this article.

Methods

In this graph model, participants are designated as nodes, mfMRI derived functional connectivity as node features, and different eye-tracking features are used to compute similarity between participants to construct heterogeneous graph edges. By taking multiple graphs as different channels, we introduce squeeze-and-excitation attention module to CensNet (A-CensNet) to integrate graph embeddings from multiple channels into one.

Results

The proposed model outperforms those using a single modality and single channel, and state-of-the-art methods.

Conclusions

The results indicate that brain functional activities and eye behaviors might complement each other in interpreting trait-like phenotypes.

Few temporally distributed brain connectivity states predict human cognitive abilities

Human functional brain connectivity can be temporally decomposed into states of high and low cofluctuation, defined as coactivation of brain regions over time. Rare states of particularly high cofluctuation have been shown to reflect fundamentals of intrinsic functional network architecture and to be highly subject-specific. However, it is unclear whether such network-defining states also contribute to individual variations in cognitive abilities - which strongly rely on the interactions among distributed brain regions. By introducing CMEP, a new eigenvector-based prediction framework, we show that as few as 16 temporally separated time frames (< 1.5% of 10min resting-state fMRI) can significantly predict individual differences in intelligence (N = 263, p < .001). Against previous expectations, individual's network-defining time frames of particularly high cofluctuation do not predict intelligence. Multiple functional brain networks contribute to the prediction, and all results replicate in an independent sample (N = 831). Our results suggest that although fundamentals of person-specific functional connectomes can be derived from few time frames of highest connectivity, temporally distributed information is necessary to extract information about cognitive abilities. This information is not restricted to specific connectivity states, like network-defining high-cofluctuation states, but rather reflected across the entire length of the brain connectivity time series.

Distinct brain network features predict internalizing and externalizing traits in children and adults

Internalizing and externalizing traits are two distinct classes of behaviors in psychiatry. However, whether shared or unique brain network features predict internalizing and externalizing behaviors in children and adults remain poorly understood. Using a sample of 2262 children from the Adolescent Brain Cognitive Development (ABCD) study and 752 adults from the Human Connectome Project (HCP), we show that network features predicting internalizing and externalizing behavior are, at least in part, dissociable in children, but not in adults. In ABCD children, traits within internalizing and externalizing behavioral categories are predicted by more similar network features concatenated across task and resting states than those between different categories. We did not observe this pattern in HCP adults. Distinct network features predict internalizing and externalizing behaviors in ABCD children and HCP adults. These data reveal shared and unique brain network features accounting for individual variation within broad internalizing and externalizing categories across developmental stages.

Unsupervised representation learning of spontaneous MEG data with Nonlinear ICA

Resting-state magnetoencephalography (MEG) data show complex but structured spatiotemporal patterns. However, the neurophysiological basis of these signal patterns is not fully known and the underlying signal sources are mixed in MEG measurements. Here, we developed a method based on the nonlinear independent component analysis (ICA), a generative model trainable with unsupervised learning, to learn representations from resting-state MEG data. After being trained with a large dataset from the Cam-CAN repository, the model has learned to represent and generate patterns of spontaneous cortical activity using latent nonlinear components, which reflects principal cortical patterns with specific spectral modes. When applied to the downstream classification task of audio-visual MEG, the nonlinear ICA model achieves competitive performance with deep neural networks despite limited access to labels. We further validate the generalizability of the model across different datasets by applying it to an independent neurofeedback dataset for decoding the subject's attentional states, providing a real-time feature extraction and decoding mindfulness and thought-inducing tasks with an accuracy of around 70% at the individual level, which is much higher than obtained by linear ICA or other baseline methods. Our results demonstrate that nonlinear ICA is a valuable addition to existing tools, particularly suited for unsupervised representation learning of spontaneous MEG activity which can then be applied to specific goals or tasks when labelled data are scarce.

Relationship Between Prediction Accuracy and Feature Importance Reliability: an Empirical and Theoretical Study

There is significant interest in using neuroimaging data to predict behavior. The predictive models are often interpreted by the computation of feature importance, which quantifies the predictive relevance of an imaging feature. Tian and Zalesky (2021) suggest that feature importance estimates exhibit low split-half reliability, as well as a trade-off between prediction accuracy and feature importance reliability across parcellation resolutions. However, it is unclear whether the trade-off between prediction accuracy and feature importance reliability is universal. Here, we demonstrate that, with a sufficient sample size, feature importance (operationalized as Haufe-transformed weights) can achieve fair to excellent split-half reliability. With a sample size of 2600 participants, Haufe-transformed weights achieve average intra-class correlation coefficients of 0.75, 0.57 and 0.53 for cognitive, personality and mental health measures respectively. Haufe-transformed weights are much more reliable than original regression weights and univariate FC-behavior correlations. Original regression weights are not reliable even with 2600 participants. Intriguingly, feature importance reliability is strongly positively correlated with prediction accuracy across phenotypes. Within a particular behavioral domain, there is no clear relationship between prediction performance and feature importance reliability across regression models. Furthermore, we show mathematically that feature importance reliability is necessary, but not sufficient, for low feature importance error. In the case of linear models, lower feature importance error is mathematically related to lower prediction error. Therefore, higher feature importance reliability might yield lower feature importance error and higher prediction accuracy. Finally, we discuss how our theoretical results relate with the reliability of imaging features and behavioral measures. Overall, the current study provides empirical and theoretical insights into the relationship between prediction accuracy and feature importance reliability.

Brain-age prediction: A systematic comparison of machine learning workflows

The difference between age predicted using anatomical brain scans and chronological age, i.e., the brain-age delta, provides a proxy for atypical aging. Various data representations and machine learning (ML) algorithms have been used for brain-age estimation. However, how these choices compare on performance criteria important for real-world applications, such as; (1) within-dataset accuracy, (2) cross-dataset generalization, (3) test-retest reliability, and (4) longitudinal consistency, remains uncharacterized. We evaluated 128 workflows consisting of 16 feature representations derived from gray matter (GM) images and eight ML algorithms with diverse inductive biases. Using four large neuroimaging databases covering the adult lifespan (total N = 2953, 18-88 years), we followed a systematic model selection procedure by sequentially applying stringent criteria. The 128 workflows showed a within-dataset mean absolute error (MAE) between 4.73-8.38 years, from which 32 broadly sampled workflows showed a cross-dataset MAE between 5.23-8.98 years. The test-retest reliability and longitudinal consistency of the top 10 workflows were comparable. The choice of feature representation and the ML algorithm both affected the performance. Specifically, voxel-wise feature spaces (smoothed and resampled), with and without principal components analysis, with non-linear and kernel-based ML algorithms performed well. Strikingly, the correlation of brain-age delta with behavioral measures disagreed between within-dataset and cross-dataset predictions. Application of the best-performing workflow on the ADNI sample showed a significantly higher brain-age delta in Alzheimer's and mild cognitive impairment patients compared to healthy controls. However, in the presence of age bias, the delta estimates in the patients varied depending on the sample used for bias correction. Taken together, brain-age shows promise, but further evaluation and improvements are needed for its real-world application.

One Size Does Not Fit All: Methodological Considerations for Brain-Based Predictive Modeling in Psychiatry

Psychiatric illnesses are heterogeneous in nature. No illness manifests in the same way across individuals, and no two patients with a shared diagnosis exhibit identical symptom profiles. Over the last several decades, group-level analyses of in vivo neuroimaging data have led to fundamental advances in our understanding of the neurobiology of psychiatric illnesses. More recently, access to computational resources and large, publicly available datasets alongside the rise of predictive modeling and precision medicine approaches have facilitated the study of psychiatric illnesses at an individual level. Data-driven machine learning analyses can be applied to identify disease-relevant biological subtypes, predict individual symptom profiles, and recommend personalized therapeutic interventions. However, when developing these predictive models, methodological choices must be carefully considered to ensure accurate, robust, and interpretable results. Choices pertaining to algorithms, neuroimaging modalities and states, data transformation, phenotypes, parcellations, sample sizes, and populations we are specifically studying can influence model performance. Here, we review applications of neuroimaging-based machine learning models to study psychiatric illnesses and discuss the effects of different methodological choices on model performance. An understanding of these effects is crucial for the proper implementation of predictive models in psychiatry and will facilitate more accurate diagnoses, prognoses, and therapeutics.

Predicting sex, age, general cognition and mental health with machine learning on brain structural connectomes

There is an increasing expectation that advanced, computationally expensive machine learning (ML) techniques, when applied to large population-wide neuroimaging datasets, will help to uncover key differences in the human brain in health and disease. We take a comprehensive approach to explore how multiple aspects of brain structural connectivity can predict sex, age, general cognitive function and general psychopathology, testing different ML algorithms from deep learning (DL) model (BrainNetCNN) to classical ML methods. We modelled N = 8183 structural connectomes from UK Biobank using six different structural network weightings obtained from diffusion MRI. Streamline count generally provided the highest prediction accuracies in all prediction tasks. DL did not improve on prediction accuracies from simpler linear models. Further, high correlations between gradient attribution coefficients from DL and model coefficients from linear models suggested the models ranked the importance of features in similar ways, which indirectly suggested the similarity in models' strategies for making predictive decision to some extent. This highlights that model complexity is unlikely to improve detection of associations between structural connectomes and complex phenotypes with the current sample size.

Predicting Future Depressive Episodes from Resting-State fMRI with Generative Embedding

After a first episode of major depressive disorder (MDD), there is substantial risk for a long-term remitting-relapsing course. Prevention and early interventions are thus critically important. Various studies have examined the feasibility of detecting at-risk individuals based on out-of-sample predictions about the future occurrence of depression. However, functional magnetic resonance imaging (MRI) has received very little attention for this purpose so far. Here, we explored the utility of generative models (i.e. different dynamic causal models, DCMs) as well as functional connectivity (FC) for predicting future episodes of depression in never-depressed adults, using a large dataset (N=906) of task-free ("resting state") fMRI data from the UK Biobank. Connectivity analyses were conducted using timeseries from pre-computed spatially independent components of different dimensionalities. Over a three year period, 50% of participants showed indications of at least one depressive episode, while the other 50% did not. Using nested cross-validation for training and a held-out test set (80/20 split), we systematically examined the combination of 8 connectivity feature sets and 17 classifiers. We found that a generative embedding procedure based on combining regression DCM (rDCM) with a support vector machine (SVM) enabled the best predictions, both on the training set (0.63 accuracy, 0.66 area under the curve, AUC) and the test set (0.62 accuracy, 0.64 AUC; p<0.001). However, on the test set, rDCM was only slightly superior to predictions based on FC (0.59 accuracy, 0.61 AUC). Interpreting model predictions based on SHAP (SHapley Additive exPlanations) values suggested that the most predictive connections were widely distributed and not confined to specific networks. Overall, our analyses suggest (i) ways of improving future fMRI-based generative embedding approaches for the early detection of individuals at-risk for depression and that (ii) achieving accuracies of clinical utility may require combination of fMRI with other data modalities.

Comparison Between Gradients and Parcellations for Functional Connectivity Prediction of Behavior

Resting-state functional connectivity (RSFC) is widely used to predict behavioral measures. To predict behavioral measures, representing RSFC with parcellations and gradients are the two most popular approaches. Here, we compare parcellation and gradient approaches for RSFC-based prediction of a broad range of behavioral measures in the Human Connectome Project (HCP) and Adolescent Brain Cognitive Development (ABCD) datasets. Among the parcellation approaches, we consider group-average "hard" parcellations (Schaefer et al., 2018), individual-specific "hard" parcellations (Kong et al., 2021a), and an individual-specific "soft" parcellation (spatial independent component analysis with dual regression; Beckmann et al., 2009). For gradient approaches, we consider the well-known principal gradients (Margulies et al., 2016) and the local gradient approach that detects local RSFC changes (Laumann et al., 2015). Across two regression algorithms, individual-specific hard-parcellation performs the best in the HCP dataset, while the principal gradients, spatial independent component analysis and group-average "hard" parcellations exhibit similar performance. On the other hand, principal gradients and all parcellation approaches perform similarly in the ABCD dataset. Across both datasets, local gradients perform the worst. Finally, we find that the principal gradient approach requires at least 40 to 60 gradients to perform as well as parcellation approaches. While most principal gradient studies utilize a single gradient, our results suggest that incorporating higher order gradients can provide significant behaviorally relevant information. Future work will consider the inclusion of additional parcellation and gradient approaches for comparison.

Supervised Phenotype Discovery From Multimodal Brain Imaging

Data-driven discovery of image-derived phenotypes (IDPs) from large-scale multimodal brain imaging data has enormous potential for neuroscientific and clinical research by linking IDPs to subjects' demographic, behavioural, clinical and cognitive measures (i.e., non-imaging derived phenotypes or nIDPs). However, current approaches are primarily based on unsupervised approaches, without the use of information in nIDPs. In this paper, we proposed a semi-supervised, multimodal, and multi-task fusion approach, termed SuperBigFLICA, for IDP discovery, which simultaneously integrates information from multiple imaging modalities as well as multiple nIDPs. SuperBigFLICA is computationally efficient and largely avoids the need for parameter tuning. Using the UK Biobank brain imaging dataset with around 40,000 subjects and 47 modalities, along with more than 17,000 nIDPs, we showed that SuperBigFLICA enhances the prediction power of nIDPs, benchmarked against IDPs derived by conventional expert-knowledge and unsupervised-learning approaches (with average nIDP prediction accuracy improvements of up to 46%). It also enables the learning of generic imaging features that can predict new nIDPs. Further empirical analysis of the SuperBigFLICA algorithm demonstrates its robustness in different prediction tasks and the ability to derive biologically meaningful IDPs in predicting health outcomes and cognitive nIDPs, such as fluid intelligence and hypertension.

Exploring the latent structure of behavior using the Human Connectome Project's data

How behavior arises from brain physiology has been one central topic of investigation in neuroscience. Considering the recent interest in predicting behavior from brain imaging using open datasets, there is the need for a principled approach to the categorization of behavioral variables. However, this is not trivial, as the definitions of psychological constructs and their relationships-their ontology-are not always clear. Here, we propose to use exploratory factor analysis (EFA) as a data-driven approach to find robust and interpretable domains of behavior in the Human Connectome Project (HCP) dataset. Additionally, we explore the clustering of behavioral variables using consensus clustering. We find that four and five factors offer the best description of the data, a result corroborated by the consensus clustering. In the four-factor solution, factors for Mental Health, Cognition, Processing Speed, and Substance Use arise. With five factors, Mental Health splits into Well-Being and Internalizing. Clustering results show a similar pattern, with clusters for Cognition, Processing Speed, Positive Affect, Negative Affect, and Substance Use. The factor structure is replicated in an independent dataset using confirmatory factor analysis (CFA). We discuss how the content of the factors fits with previous conceptualizations of general behavioral domains.

Classification and prediction of cognitive performance differences in older age based on brain network patterns using a machine learning approach

Age-related cognitive decline varies greatly in healthy older adults, which may partly be explained by differences in the functional architecture of brain networks. Resting-state functional connectivity (RSFC) derived network parameters as widely used markers describing this architecture have even been successfully used to support diagnosis of neurodegenerative diseases. The current study aimed at examining whether these parameters may also be useful in classifying and predicting cognitive performance differences in the normally aging brain by using machine learning (ML). Classifiability and predictability of global and domain-specific cognitive performance differences from nodal and network-level RSFC strength measures were examined in healthy older adults from the 1000BRAINS study (age range: 55-85 years). ML performance was systematically evaluated across different analytic choices in a robust cross-validation scheme. Across these analyses, classification performance did not exceed 60% accuracy for global and domain-specific cognition. Prediction performance was equally low with high mean absolute errors (MAEs ≥ 0.75) and low to none explained variance (R2 ≤ 0.07) for different cognitive targets, feature sets, and pipeline configurations. Current results highlight limited potential of functional network parameters to serve as sole biomarker for cognitive aging and emphasize that predicting cognition from functional network patterns may be challenging.

Challenges for machine learning in clinical translation of big data imaging studies

Combining deep learning image analysis methods and large-scale imaging datasets offers many opportunities to neuroscience imaging and epidemiology. However, despite these opportunities and the success of deep learning when applied to a range of neuroimaging tasks and domains, significant barriers continue to limit the impact of large-scale datasets and analysis tools. Here, we examine the main challenges and the approaches that have been explored to overcome them. We focus on issues relating to data availability, interpretability, evaluation, and logistical challenges and discuss the problems that still need to be tackled to enable the success of "big data" deep learning approaches beyond research.

Cross-cohort replicability and generalizability of connectivity-based psychometric prediction patterns

An increasing number of studies have investigated the relationships between inter-individual variability in brain regions' connectivity and behavioral phenotypes, making use of large population neuroimaging datasets. However, the replicability of brain-behavior associations identified by these approaches remains an open question. In this study, we examined the cross-dataset replicability of brain-behavior association patterns for fluid cognition and openness predictions using a previously developed region-wise approach, as well as using a standard whole-brain approach. Overall, we found moderate similarity in patterns for fluid cognition predictions across cohorts, especially in the Human Connectome Project Young Adult, Human Connectome Project Aging, and Enhanced Nathan Kline Institute Rockland Sample cohorts, but low similarity in patterns for openness predictions. In addition, we assessed the generalizability of prediction models in cross-dataset predictions, by training the model in one dataset and testing in another. Making use of the region-wise prediction approach, we showed that first, a moderate extent of generalizability could be achieved with fluid cognition prediction, and that, second, a set of common brain regions related to fluid cognition across cohorts could be identified. Nevertheless, the moderate replicability and generalizability could only be achieved in specific contexts. Thus, we argue that replicability and generalizability in connectivity-based prediction remain limited and deserve greater attention in future studies.

A multicohort geometric deep learning study of age dependent cortical and subcortical morphologic interactions for fluid intelligence prediction

The relationship of human brain structure to cognitive function is complex, and how this relationship differs between childhood and adulthood is poorly understood. One strong hypothesis suggests the cognitive function of Fluid Intelligence (Gf) is dependent on prefrontal cortex and parietal cortex. In this work, we developed a novel graph convolutional neural networks (gCNNs) for the analysis of localized anatomic shape and prediction of Gf. Morphologic information of the cortical ribbons and subcortical structures was extracted from T1-weighted MRIs within two independent cohorts, the Adolescent Brain Cognitive Development Study (ABCD; age: 9.93 ± 0.62 years) of children and the Human Connectome Project (HCP; age: 28.81 ± 3.70 years). Prediction combining cortical and subcortical surfaces together yielded the highest accuracy of Gf for both ABCD (R = 0.314) and HCP datasets (R = 0.454), outperforming the state-of-the-art prediction of Gf from any other brain measures in the literature. Across both datasets, the morphology of the amygdala, hippocampus, and nucleus accumbens, along with temporal, parietal and cingulate cortex consistently drove the prediction of Gf, suggesting a significant reframing of the relationship between brain morphology and Gf to include systems involved with reward/aversion processing, judgment and decision-making, motivation, and emotion.

Proportional intracranial volume correction differentially biases behavioral predictions across neuroanatomical features, sexes, and development

Individual differences in brain anatomy can be used to predict variations in cognitive ability. Most studies to date have focused on broad population-level trends, but the extent to which the observed predictive features are shared across sexes and age groups remains to be established. While it is standard practice to account for intracranial volume (ICV) using proportion correction in both regional and whole-brain morphometric analyses, in the context of brain-behavior predictions the possible differential impact of ICV correction on anatomical features and subgroups within the population has yet to be systematically investigated. In this work, we evaluate the effect of proportional ICV correction on sex-independent and sex-specific predictive models of individual cognitive abilities across multiple anatomical properties (surface area, gray matter volume, and cortical thickness) in healthy young adults (Human Connectome Project; n = 1013, 548 females) and typically developing children (Adolescent Brain Cognitive Development study; n = 1823, 979 females). We demonstrate that ICV correction generally reduces predictive accuracies derived from surface area and gray matter volume, while increasing predictive accuracies based on cortical thickness in both adults and children. Furthermore, the extent to which predictive models generalize across sexes and age groups depends on ICV correction: models based on surface area and gray matter volume are more generalizable without ICV correction, while models based on cortical thickness are more generalizable with ICV correction. Finally, the observed neuroanatomical features predictive of cognitive abilities are unique across age groups regardless of ICV correction, but whether they are shared or unique across sexes (within age groups) depends on ICV correction. These findings highlight the importance of considering individual differences in ICV, and show that proportional ICV correction does not remove the effects of cranial volume from anatomical measurements and can introduce ICV bias where previously there was none. ICV correction choices affect not just the strength of the relationships captured, but also the conclusions drawn regarding the neuroanatomical features that underlie those relationships.

Comparison of individualized behavioral predictions across anatomical, diffusion and functional connectivity MRI

A fundamental goal across the neurosciences is the characterization of relationships linking brain anatomy, functioning, and behavior. Although various MRI modalities have been developed to probe these relationships, direct comparisons of their ability to predict behavior have been lacking. Here, we compared the ability of anatomical T1, diffusion and functional MRI (fMRI) to predict behavior at an individual level. Cortical thickness, area and volume were extracted from anatomical T1 images. Diffusion Tensor Imaging (DTI) and approximate Neurite Orientation Dispersion and Density Imaging (NODDI) models were fitted to the diffusion images. The resulting metrics were projected to the Tract-Based Spatial Statistics (TBSS) skeleton. We also ran probabilistic tractography for the diffusion images, from which we extracted the stream count, average stream length, and the average of each DTI and NODDI metric across tracts connecting each pair of brain regions. Functional connectivity (FC) was extracted from both task and resting-state fMRI. Individualized prediction of a wide range of behavioral measures were performed using kernel ridge regression, linear ridge regression and elastic net regression. Consistency of the results were investigated with the Human Connectome Project (HCP) and Adolescent Brain Cognitive Development (ABCD) datasets. In both datasets, FC-based models gave the best prediction performance, regardless of regression model or behavioral measure. This was especially true for the cognitive component. Furthermore, all modalities were able to predict cognition better than other behavioral components. Combining all modalities improved prediction of cognition, but not other behavioral components. Finally, across all behaviors, combining resting and task FC yielded prediction performance similar to combining all modalities. Overall, our study suggests that in the case of healthy children and young adults, behaviorally-relevant information in T1 and diffusion features might reflect a subset of the variance captured by FC.

Statistical quantification of confounding bias in machine learning models

BACKGROUND

The lack of nonparametric statistical tests for confounding bias significantly hampers the development of robust, valid, and generalizable predictive models in many fields of research. Here I propose the partial confounder test, which, for a given confounder variable, probes the null hypotheses of the model being unconfounded.

RESULTS

The test provides a strict control for type I errors and high statistical power, even for nonnormally and nonlinearly dependent predictions, often seen in machine learning. Applying the proposed test on models trained on large-scale functional brain connectivity data (N= 1,865) (i) reveals previously unreported confounders and (ii) shows that state-of-the-art confound mitigation approaches may fail preventing confounder bias in several cases.

CONCLUSIONS

The proposed test (implemented in the package mlconfound; https://mlconfound.readthedocs.io) can aid the assessment and improvement of the generalizability and validity of predictive models and, thereby, fosters the development of clinically useful machine learning biomarkers.

Decoding working memory-related information from repeated psychophysiological EEG experiments using convolutional and contrastive neural networks

OBJECTIVE

Extracting reliable information from EEG signals is difficult because the low signal-to-noise ratio and significant intersubject variability seriously hinder statistical analyses. However, recent advances in explainable machine learning open a new strategy to address this problem.

APPROACH

The current study evaluates this approach using results from the classification and decoding of electrical brain activity associated with information retention. We designed four neural network models differing in architecture, training strategies, and input representation to classify single experimental trials of a working memory task.

MAIN RESULTS

Our best models achieved an accuracy of 65.29$±0.76 and Matthews correlation coefficient of 0.288±0.018, outperforming the reference model trained on the same data. The highest correlation between classification score and behavioral performance was 0.36 (p=0.0007). Using analysis of input perturbation, we estimated the importance of EEG channels and frequency bands in the task at hand. The set of essential features identified for each network varies. We identified a subset of features common to all models that identified brain regions and frequency bands consistent with current neurophysiological knowledge of the processes critical to attention and working memory. Finally, we proposed sanity checks to examine further the robustness of each model's set of features.

SIGNIFICANCE

Our results indicate that explainable deep learning is a powerful tool for decoding information from EEG signals. It is crucial to train and analyze a range of models to identify stable and reliable features. Our results highlight the need for explainable modeling as the model with the highest accuracy appeared to use residual artifactual activity.