A confounder controlled machine learning approach: Group analysis and classification of schizophrenia and Alzheimer's disease using resting-state functional network connectivity.

Resting-state functional magnetic resonance imaging (rs-fMRI) has increasingly been used to study both Alzheimer's disease (AD) and schizophrenia (SZ). While most rs-fMRI studies being conducted in AD and SZ compare patients to healthy controls, it is also of interest to directly compare AD and SZ patients with each other to identify potential biomarkers shared between the disorders. However, comparing patient groups collected in different studies can be challenging due to potential confounds, such as differences in the patient's age, scan protocols, etc. In this study, we compared and contrasted resting-state functional network connectivity (rs-FNC) of 162 patients with AD and late mild cognitive impairment (LMCI), 181 schizophrenia patients, and 315 cognitively normal (CN) subjects. We used confounder-controlled rs-FNC and applied machine learning algorithms (including support vector machine, logistic regression, random forest, and k-nearest neighbor) and deep learning models (i.e., fully-connected neural networks) to classify subjects in binary and three-class categories according to their diagnosis labels (e.g., AD, SZ, and CN). Our statistical analysis revealed that FNC between the following network pairs is stronger in AD compared to SZ: subcortical-cerebellum, subcortical-cognitive control, cognitive control-cerebellum, and visual-sensory motor networks. On the other hand, FNC is stronger in SZ than AD for the following network pairs: subcortical-visual, subcortical-auditory, subcortical-sensory motor, cerebellum-visual, sensory motor-cognitive control, and within the cerebellum networks. Furthermore, we observed that while AD and SZ disorders each have unique FNC abnormalities, they also share some common functional abnormalities that can be due to similar neurobiological mechanisms or genetic factors contributing to these disorders' development. Moreover, we achieved an accuracy of 85% in classifying subjects into AD and SZ where default mode, visual, and subcortical networks contributed the most to the classification and accuracy of 68% in classifying subjects into AD, SZ, and CN with the subcortical domain appearing as the most contributing features to the three-way classification. Finally, our findings indicated that for all classification tasks, except AD vs. SZ, males are more predictable than females.

Towards a multimodal neuroimaging-based risk score for mild cognitive impairment by combining clinical studies with a large (N>37000) population-based study.

Alzheimer's disease (AD) is the most common form of age-related dementia, leading to a decline in memory, reasoning, and social skills. While numerous studies have investigated the genetic risk factors associated with AD, less attention has been given to identifying a brain imaging-based measure of AD risk. This study introduces a novel approach to assess mild cognitive impairment MCI, as a stage before AD, risk using neuroimaging data, referred to as a brain-wide risk score (BRS), which incorporates multimodal brain imaging. To begin, we first categorized participants from the Open Access Series of Imaging Studies (OASIS)-3 cohort into two groups: controls (CN) and individuals with MCI. Next, we computed structure and functional imaging features from all the OASIS data as well as all the UK Biobank data. For resting functional magnetic resonance imaging (fMRI) data, we computed functional network connectivity (FNC) matrices using fully automated spatially constrained independent component analysis. For structural MRI data we computed gray matter (GM) segmentation maps. We then evaluated the similarity between each participant's neuroimaging features from the UK Biobank and the difference in the average of those features between CN individuals and those with MCI, which we refer to as the brain-wide risk score (BRS). Both GM and FNC features were utilized in determining the BRS. We first evaluated the differences in the distribution of the BRS for CN vs MCI within the OASIS-3 (using OASIS-3 as the reference group). Next, we evaluated the BRS in the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort (using OASIS-3 as the reference group), showing that the BRS can differentiate MCI from CN in an independent data set. Subsequently, using the sMRI BRS, we identified 10 distinct subgroups and similarly, we identified another set of 10 subgroups using the FNC BRS. For sMRI and FNC we observed results that mutually validate each other, with certain aspects being complementary. For the unimodal analysis, sMRI provides greater differentiation between MCI and CN individuals than the fMRI data, consistent with prior work. Additionally, by utilizing a multimodal BRS approach, which combines both GM and FNC assessments, we identified two groups of subjects using the multimodal BRS scores. One group exhibits high MCI risk with both negative GM and FNC BRS, while the other shows low MCI risk with both positive GM and FNC BRS. Moreover, in the UKBB we have 46 participants diagnosed with AD showed FNC and GM patterns similar to those in high-risk groups, defined in both unimodal and multimodal BRS. Finally, to ensure the reproducibility of our findings, we conducted a validation analysis using the ADNI as an additional reference dataset and repeated the above analysis. The results were consistently replicated across different reference groups, highlighting the potential of FNC and sMRI-based BRS in early Alzheimer's detection.

A Supervised Contrastive Learning-based Analysis of rs-tMRI Data Captures Gender Differences in Nonlinear Functional Network Coupling.

Many studies in neuroscience have focused on interpreting brain activity using functional connectivity (FC). The most widely used approach for measuring FC is based on linear correlation (e.g., the Pearson correlation), where the temporal cofluctuations between functional brain regions are computed. However, such approaches ignore nonlinear dependencies among regions that might carry distinctive information across groups of subjects. In this study, we offer a deep learning-based approach that also captures nonlinear temporal relationships between brain networks. Our approach consists of two main parts: an encoder that learns domain-specific embeddings of time courses estimated from independent component analysis (ICA) and a similarity metric that measures the similarities between the embeddings. We call such similarities as nonlinear functional relationships between networks. Our findings on a large dataset (including above 11k normal control subjects) suggest that male subjects exhibit stronger nonlinear network-network relationships than female subjects in most cases. Furthermore, we observe that, unlike FC, our approach could capture some intra-network relationships, especially between cognitive control and visual networks, which are significantly different between males and females, suggesting that our approach can provide a complementary interpretation of the functional brain activity to FC.

Individualized spatial network predictions using Siamese convolutional neural networks: A resting-state fMRI study of over 11,000 unaffected individuals.

Individuals can be characterized in a population according to their brain measurements and activity, given the inter-subject variability in brain anatomy, structure-function relationships, or life experience. Many neuroimaging studies have demonstrated the potential of functional network connectivity patterns estimated from resting functional magnetic resonance imaging (fMRI) to discriminate groups and predict information about individual subjects. However, the predictive signal present in the spatial heterogeneity of brain connectivity networks is yet to be extensively studied. In this study, we investigate, for the first time, the use of pairwise-relationships between resting-state independent spatial maps to characterize individuals. To do this, we develop a deep Siamese framework comprising three-dimensional convolution neural networks for contrastive learning based on individual-level spatial maps estimated via a fully automated fMRI independent component analysis approach. The proposed framework evaluates whether pairs of spatial networks (e.g., visual network and auditory network) are capable of subject identification and assesses the spatial variability in different network pairs' predictive power in an extensive whole-brain analysis. Our analysis on nearly 12,000 unaffected individuals from the UK Biobank study demonstrates that the proposed approach can discriminate subjects with an accuracy of up to 88% for a single network pair on the test set (best model, after several runs), and 82% average accuracy at the subcortical domain level, notably the highest average domain level accuracy attained. Further investigation of our network's learned features revealed a higher spatial variability in predictive accuracy among younger brains and significantly higher discriminative power among males. In sum, the relationship among spatial networks appears to be both informative and discriminative of individuals and should be studied further as putative brain-based biomarkers.

Deep learning in resting-state fMRI.

Modeling the rich, dynamic spatiotemporal variations captured by human brain functional magnetic resonance imaging (fMRI) data is a complicated task. Analysis at the brain's regional and connection levels provides more straightforward biological interpretation for fMRI data and has been instrumental in characterizing the brain thus far. Here we hypothesize that spatiotemporal learning directly in the four-dimensional (4D) fMRI voxel-time space could result in enhanced discriminative brain representations compared to widely used, pre-engineered fMRI temporal transformations, and brain regional and connection-level fMRI features. Motivated by this, we extend our recently reported structural MRI (sMRI) deep learning (DL) pipeline to additionally capture temporal variations, training the proposed 4D DL model end-to-end on preprocessed fMRI data. Results validate that the complex non-linear functions of the used deep spatiotemporal approach generate discriminative encodings for the studied learning task, outperforming both standard machine learning (SML) and DL methods on the widely used fMRI voxel/region/connection features, except the relatively simplistic measure of central tendency - the temporal mean of the fMRI data. Additionally, we identify the fMRI features for which DL significantly outperformed SML methods for voxel-level fMRI features. Overall, our results support the efficiency and potential of DL models trainable at the voxel level fMRI data and highlight the importance of developing auxiliary tools to facilitate interpretation of such flexible models.