Resting-state functional magnetic resonance imaging (rs-fMRI) is a noninvasive technique pivotal for understanding human neural mechanisms of intricate cognitive processes. Most rs-fMRI studies compute a single static functional connectivity matrix across brain regions of interest, or dynamic functional connectivity matrices with a sliding window approach. These approaches are at risk of oversimplifying brain dynamics and lack proper consideration of the goal at hand. While deep learning has gained substantial popularity for modeling complex relational data, its application to uncovering the spatiotemporal dynamics of the brain is still limited. In this study we propose a novel interpretable deep learning framework that learns goal-specific functional connectivity matrix directly from time series and employs a specialized graph neural network for the final classification. Our model, DSAM , leverages temporal causal convolutional networks to capture the temporal dynamics in both low- and high-level feature representations, a temporal attention unit to identify important time points, a self-attention unit to construct the goal-specific connectivity matrix, and a novel variant of graph neural network to capture the spatial dynamics for downstream classification. To validate our approach, we conducted experiments on the Human Connectome Project dataset with 1075 samples to build and interpret the model for the classification of sex group, and the Adolescent Brain Cognitive Development Dataset with 8520 samples for independent testing. Compared our proposed framework with other state-of-art models, results suggested this novel approach goes beyond the assumption of a fixed connectivity matrix, and provides evidence of goal-specific brain connectivity patterns, which opens up potential to gain deeper insights into how the human brain adapts its functional connectivity specific to the task at hand. Our implementation can be found on https://github.com/bishalth01/DSAM . • Multilevel Temporal Feature Extraction: • Utilizes a novel multilevel Temporal Convolutional Network (TCN) adaptation to directly extract temporal features from raw brain activity data. • Shared Temporal Attention for Key Time Points: • Implements shared temporal attention mechanisms to selectively focus on the most informative time points, enhancing model efficiency. • Node-Node Self-Attention for Dynamic Brain Connectivity: • Leverages self-attention mechanisms to dynamically construct a goal-specific brain connectivity matrix, capturing complex inter-node interactions. • ROI-Aware Graph Neural Networks (GNNs): • Introduces ROI-aware GNNs to model spatial brain dynamics, ensuring region-specific contextual learning and improving interpretability.

Schizophrenia (SZ) is a neuropsychiatric disorder that affects millions globally. Current diagnosis of SZ is symptom-based, which poses difficulty due to the variability of symptoms across patients. To this end, many recent studies have developed deep learning methods for automated diagnosis of SZ, especially using raw EEG, which provides high temporal precision. For such methods to be productionized, they must be both explainable and robust. Explainable models are essential to identify biomarkers of SZ, and robust models are critical to learn generalizable patterns, especially amidst changes in the implementation environment. One common example is channel loss during EEG recording, which could be detrimental to classifier performance. In this study, we developed a novel channel dropout (CD) approach to increase the robustness of explainable deep learning models trained on EEG data for SZ diagnosis to channel loss. We developed a baseline convolutional neural network (CNN) architecture and implement our approach as a CD layer added to the baseline (CNN-CD). We then applied two explainability approaches to both models for insight into learned spatial and spectral features and show that the application of CD decreases model sensitivity to channel loss. The CNN and CNN-CD achieved accuracies of 81.9% and 80.9% on testing data, respectively. Furthermore, our models heavily prioritized the parietal electrodes and the a-band, which is supported by existing literature. It is our hope that this study motivates the further development of explainable and robust models and bridges the transition from research to application in a clinical decision support role.

Deep learning methods are increasingly being applied to raw electroencephalography (EEG) data. Relative to traditional machine learning methods, deep learning methods can increase model performance through automated feature extraction. Nevertheless, they are also less explainable. Existing raw EEG explainability methods identify relative feature importance but do not identify how features relate to model predictions. In this study, we combine well-characterized first layer filters with a novel post hoc statistical analysis of the filter activations that linearly relates properties of model activations with predictions. We implement our approach within the context of automated sleep stage classification, finding that the model uncovers positive relationships between waveforms resembling sleep spindles and NREM2, between high frequency filters and Awake samples, and between δ activity and NREM3. Our approach represents a significant step forward for raw EEG explainability and has the potential to provide many insights relating learned features and model predictions in future studies.

As the field of deep learning has grown in recent years, its application to the domain of raw resting-state electroencephalography (EEG) has also increased. Relative to traditional machine learning methods or deep learning methods applied to manually engineered features, there are fewer methods for developing deep learning models on small raw EEG datasets. One potential approach for enhancing deep learning performance, in this case, is the use of transfer learning. While a number of studies have presented transfer learning approaches for manually engineered EEG features, relatively few approaches have been developed for raw resting-state EEG. In this study, we propose a novel EEG transfer learning approach wherein we first train a model on a large publicly available single-channel sleep stage classification dataset. We then use the learned representations to develop a classifier for automated major depressive disorder diagnosis with raw multichannel EEG. Statistical testing reveals that our approach significantly improves the performance of our model (p < 0.05), and we also find that the performance of our approach exceeds that of many previous studies using both engineered features and raw EEG. We further examine how transfer learning affected the representations learned by the model through a pair of explainability analyses, identifying key frequency bands and channels utilized across models. Our proposed approach represents a significant step forward for the domain of raw resting-state EEG classification and has broader implications for use with other electrophysiology and time-series modalities. Importantly, it has the potential to expand the use of deep learning methods across a greater variety of raw EEG datasets and lead to the development of more reliable EEG classifiers.

In recent years, many neuroimaging studies have begun to integrate gradient-based explainability methods to provide insight into key features. However, existing explainability approaches typically generate a point estimate of importance and do not provide insight into the degree of uncertainty associated with explanations. In this study, we present a novel approach for estimating explanation uncertainty for convolutional neural networks (CNN) trained on neuroimaging data. We train a CNN for classification of individuals with schizophrenia (SZs) and controls (HCs) using resting state functional magnetic resonance imaging (rs-fMRI) dynamic functional network connectivity (dFNC) data. We apply Monte Carlo batch normalization (MCBN) and generate an explanation following each iteration using layer-wise relevance propagation (LRP). We then examine whether the resulting distribution of explanations differs between SZs and HCs and examine the relationship between MCBN-based LRP explanations and regular LRP explanations. We find a number of significant differences in LRP relevance for SZs and HCs and find that traditional LRP values frequently diverge from the MCBN relevance distribution. This study provides a novel approach for obtaining insight into the level of uncertainty associated with gradient-based explanations in neuroimaging and represents a significant step towards increasing reliability of explainable deep learning methods within a clinical setting.

Schizophrenia (SZ) is a mental disorder that affects millions of people globally. At this time, diagnosis of SZ is based upon symptoms, which can vary from patient to patient and create difficulty with diagnosis. To address this issue, researchers have begun to look for neurological biomarkers of SZ and develop methods for automated diagnosis. In recent years, several studies have applied deep learning to raw EEG for automated SZ diagnosis. However, the use of raw time-series data makes explainability more difficult than it is for traditional machine learning algorithms trained on manually engineered features. As such, none of these studies have sought to explain their models, which is problematic within a healthcare context where explainability is a critical component. In this study, we apply perturbation-based explainability approaches to gain insight into the spectral and spatial features learned by two distinct deep learning models trained on raw EEG for SZ diagnosis for the first time. We develop convolutional neural network (CNN) and CNN long short-term memory network (CNN-LSTM) architectures. Results show that both models prioritize the T8 and C3 electrodes and the δ- and y-bands, which agrees with previous literature and supports the overall utility of our models. This study represents a step forward in the implementation of deep learning models for clinical SZ diagnosis, and it is our hope that it will inspire the more widespread application of explainability methods for insight into deep learning models trained for SZ diagnosis in the future.

Schizophrenia (SZ) is a neuropsychiatric disorder that adversely effects millions of individuals globally. Current diagnostic efforts are symptom based and hampered due to the variability in symptom presentation across individuals and overlap of symptoms with other neuropsychiatric disorders. This spawns the need for (1) biomarkers to aid with empirical SZ diagnosis and (2) the development of automated diagnostic approaches that will eventually serve in a clinical decision support role. In this study, we train random forest (RF) and support vector machine (SVM) models to differentiate between individuals with schizophrenia and healthy controls using spectral features extracted from resting state EEG data. We then perform two explainability analyses to gain insight into key frequency bands and channels. In our explainability analyses, we examine the reproducibility of SZ biomarkers across models with the goal of identifying those that have potential clinical implications. Our model performance results are well above chance level indicating the broader utility of spectral information for SZ diagnosis. Additionally, we find that the RF prioritizes the upper$\gamma$-band and is robust to loss of information from individual electrodes, while the SVM prioritizes the$\alpha$and$\theta$-bands and P4 and T8 electrodes. It is our hope that our findings will inform future efforts towards the empirical diagnosis of SZ and towards the development of clinical decision support systems for SZ diagnosis.

The application of clustering algorithms to fMRI functional network connectivity (FNC) data has been extensively studied over the past decade. When applied to FNC, these analyses assign samples to an optimal number of groups without a priori assumptions. Through these groupings, studies have provided insights into the dynamics of network connectivity through the identification of different brain states and have identified subgroups of individuals with unique brain activity. However, the manner in which underlying brain networks influence the identified groups is yet to be fully understood. In this study, we apply k-means clustering to resting-state fMRI-based static FNC data collected from 37,784 healthy individuals. We identified 2 groups of individuals with statistically significant differences in cognitive performance in several test metrics. Then, by applying two different versions of G2PC, a global permutation feature importance approach, and logistic regression with elastic net regularization, we were able to identify the relative importance of brain network pairs and their underlying features to the resulting groups. Through these approaches, together with the visualization of centroids' connectivity matrices, we were able to explain the observed differences in cognition in terms of specific key brain networks. We expect that our results will shed further light upon the effect of underlying brain networks on encountered cognitive differences between groups with unique brain activity.

Spectral analysis of electroencephalography (EEG) data has developed as an important area of EEG research. EEG spectra have been analyzed with explainable machine learning and deep learning methods. However, as deep learning has developed, many studies have used raw EEG data, which is poorly suited for traditional explainability methods. Several studies have introduced methods for spectral insight into classifiers trained on raw EEG data. These studies have provided global insight into the frequency bands important to a classifier but not local insight into the frequency bands important to the classification of individual samples. Local explainability could be particularly helpful for EEG domains like sleep stage classification that feature multiple evolving states. We present a novel local spectral explainability approach and use it to explain a convolutional neural network trained for automated sleep stage classification. We use our approach to show how the importance of frequency bands varies over time and even within the same sleep stages. Also, to better understand how our approach compares to existing methods, we compare a global estimate of spectral importance generated from our local results with an existing global spectral importance approach. We find that the δ band is most important for most sleep stages, though ß is most important for the non-rapid eye movement 2 (NREM2) sleep stage. Additionally, 0 is particularly important for identifying Awake and NREM1 samples. Our study represents the first approach developed for local spectral insight into deep learning classifiers trained on raw EEG time series.

In recent years, more biomedical studies have begun to use multimodal data to improve model performance. Many studies have used ablation for explainability, which requires the modification of input data. This can create out-of-distribution samples and lead to incorrect explanations. To avoid this problem, we propose using a gradient-based feature attribution approach, called layer-wise relevance propagation (LRP), to explain the importance of modalities both locally and globally for the first time. We demonstrate the feasibility of the approach with sleep stage classification as our use-case and train a 1-D convolutional neural network with electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG) data. We also analyze the relationship of our local explainability results with clinical and demographic variables to determine whether they affect our classifier. Across all samples, EEG is the most important modality, followed by EOG and EMG. For individual sleep stages, EEG and EOG have higher relevance for awake and non-rapid eye movement 1 (NREM1). EOG is most important for REM, and EEG is most relevant for NREM2-NREM3. Also, LRP gives consistent levels of importance to each modality for the correctly classified samples across folds but inconsistent levels of importance for incorrectly classified samples. Our statistical analyses suggest that medication has a significant effect upon patterns learned for EEG and EOG NREM2 and that subject sex and age significantly affects the EEG and EOG patterns learned, respectively. Our results demonstrate the viability of gradient-based approaches for explaining multimodal electrophysiology classifiers and suggest their generalizability for other multimodal classification domains.

With the growing use of multimodal data for deep learning classification in healthcare research, more studies are presenting explainability methods for insight into multimodal classifiers. Among these studies, few utilize local explainability methods, which can provide (1) insight into the classification of samples over time and (2) better understanding of the effects of demographic and clinical variables upon patterns learned by classifiers. To the best of our knowledge, we present the first local explainability approach for insight into the importance of each modality to the classification of samples over time. Our approach uses ablation, and we demonstrate how it can show the importance of each modality to the correct classification of each class. We further present a novel analysis that explores the effects of demographic and clinical variables upon the multimodal patterns learned by the classifier. As a use-case, we train a convolutional neural network for automated sleep staging with electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG) data. We find that EEG is the most important modality across most stages, though EOG is particularly important for non-rapid eye movement stage 1. Further, we identify significant relationships between the local explanations and subject age, sex, and state of medication which suggest that the classifier learned features associated with these variables across multiple modalities and correctly classified samples. Our novel explainability approach has implications for many fields involving multimodal classification. Moreover, our examination of the degree to which demographic and clinical variables may affect classifiers could provide direction for future studies in automated biomarker discovery.

Spectral analysis remains a hallmark approach for gaining insight into electrophysiology modalities like electroencephalography (EEG). As the field of deep learning has progressed, more studies have begun to train deep learning classifiers on raw EEG data, which presents unique problems for explainability. A growing number of studies have presented explainability approaches that provide insight into the spectral features learned by deep learning classifiers. However, existing approaches only attribute importance to different frequency bands. Most of the methods cannot provide insight into the actual spectral values or the relationship between spectral features that models have learned. Here, we present a novel adaptation of activation maximization for electrophysiology time-series that generates samples that indicate the features learned by classifiers by optimizing their spectral content. We evaluate our approach within the context of EEG sleep stage classification with a convolutional neural network, and we find that our approach is able to identify spectral patterns known to be associated with each sleep stage. We also find surprising results suggesting that our classifier may have prioritized the use of eye and motion artifact when identifying Awake samples. Our approach is the first adaptation of activation maximization to the domain of raw electrophysiology classification. Additionally, our approach has implications for explaining any classifier trained on highly dynamic, long time-series.

Spectral analysis of brain activation in different regions, either in the form of network time-courses or regions of interest (ROI) time-series, has been a topic of interest in recent studies. Such studies hypothesize that observed brain fluctuations are due to different underlying sources of neurophysiological activation. Among these studies, brain fluctuations during the resting-state, as an unconstrained condition, have been a subject of interest. Some clinical studies have employed spectral analysis to locate differences between diagnostic groups such as schizophrenia and bipolar disorder. Other studies have argued that resting-state brain fluctuations are in fact dynamic, and that activation and connectivity of brain regions develops and evolves spontaneously. In this study, we combine both approaches and focus on capturing dynamics of the spectral properties of network time-courses estimated from independent components analysis (ICA) and categorizing spontaneous frequency profiles of network time-courses into three major profiles, which we call "frequency modes". We show that brain networks have distinct time-varying frequency domain characteristics, differing from one another in their occupancy rates of the frequency modes. Additionally, we identify some networks in which the occurrence rates of the different modes are significantly different based on the gender of the subjects.

Patterns of resting state fMRI functional network connectivity in schizophrenia patients have been shown to differ markedly from those of healthy controls. While some studies have explored connectivity within fixed frequency bands, the question of network phase synchrony across disparate frequency bands, or cross-frequency connectivity , has remained surprisingly underexplored. Computational modeling at the neuronal scale however has long acknowledged the existence of coupled fast and slow subsystems. Here, we present preliminary evidence that cross-frequency coupling exists at the network level, that it patterns in meaningful ways over functional domains, and that this patterning differs between the healthy population and individuals with diagnosed schizophrenia.