Constructing large-scale cortical brain networks from scalp EEG with Bayesian nonnegative matrix factorization

Time-frequency-space transformer EEG decoding for spinal cord injury

Transformer neural networks based on multi-head self-attention are effective in several fields. To capture brain activity on electroencephalographic (EEG) signals and construct an effective pattern recognition model, this paper explores the multi-channel deep feature decoding method utilizing the self-attention mechanism. By integrating inter-channel features with intra-channel features, the self-attention mechanism generates a deep feature vector that encompasses information from all brain activities. In this paper, a time-frequency-spatial domain analysis of motor imagery (MI) based EEG signals from spinal cord injury patients is performed to construct a transformer neural network-based MI classification model. The proposed algorithm is named time-frequency-spatial transformer. The time-frequency and spatial domain feature vectors extracted from the EEG signals are input into the transformer neural network for multiple self-attention depth feature encoding, a peak classification accuracy of 93.56% is attained through the fully connected layer. By constructing the attention matrix brain network, it can be inferred that the channel connections constructed by the attention heads have similarities to the brain networks constructed by the EEG raw signals. The experimental results reveal that the self-attention coefficient brain network holds significant potential for brain activity analysis. The self-attention coefficient brain network can better illustrate correlated connections and show sample differences. Attention coefficient brain networks can provide a more discriminative approach for analyzing brain activity in clinical settings.

Unveiling perinatal depression: A dual-network EEG analysis for diagnosis and severity assessment

Perinatal depression (PD), which affects about 10-20 percent of women, often goes unnoticed because related symptoms frequently overlap with those commonly experienced during pregnancy. Moreover, identifying PD currently depends heavily on the use of questionnaires, and objective biological indicators for diagnosis has yet to be identified. This research proposes a safe and non-invasive method for diagnosing PD and aims to delve deeper into its underlying mechanism. Considering the non-invasiveness and clinical convenience of electroencephalogram (EEG) for mothers-to-be and fetuses, we collected the resting-state scalp EEG of pregnant women (with PD/healthy) at the 38th week of gestation. To compensate for the low spatial resolution of scalp EEG, source analysis was first applied to project the scalp EEG to the cortical-space. Afterwards, cortical-space networks and large-scale networks were constructed to investigate the mechanism of PD from two different level. Herein, differences in the two distinct types of networks between PD patients and healthy mothers-to-be were explored, respectively. We found that the PD patients illustrated decreased network connectivity in the cortical-space, while the large-scale networks revealed weaker connections at cerebellar area. Further, related spatial topological features derived from the two different networks were combined to promote the recognition of pregnant women with PD from those healthy ones. Meanwhile, the depression severity at patient level was effectively predicted based on the combined spatial topological features as well. These findings consistently validated that the two kinds of networks indeed played off each other, which thus helped explore the underlying mechanism of PD; and further verified the superiority of the combination strategy, revealing its reliability and potential in diagnosis and depression severity evaluation.

Large-Scale Cortical Network Analysis and Classification of MI-BCI Tasks Based on Bayesian Nonnegative Matrix Factorization

Motor imagery (MI) is a high-level cognitive process that has been widely applied to clinical rehabilitation and brain-computer interfaces (BCIs). However, the decoding of MI tasks still faces challenges, and the neural mechanisms underlying its application are unclear, which seriously hinders the development of MI-based clinical applications and BCIs. Here, we combined EEG source reconstruction and Bayesian nonnegative matrix factorization (NMF) methods to construct large-scale cortical networks of left-hand and right-hand MI tasks. Compared to right-hand MI, the results showed that the significantly increased functional network connectivities (FNCs) mainly located among the visual network (VN), sensorimotor network (SMN), right temporal network, right central executive network, and right parietal network in the left-hand MI at the β (13-30Hz) and all (8-30Hz) frequency bands. For the network properties analysis, we found that the clustering coefficient, global efficiency, and local efficiency were significantly increased and characteristic path length was significantly decreased in left-hand MI compared to right-hand MI at the β and all frequency bands. These network pattern differences indicated that the left-hand MI may need more modulation of multiple large-scale networks (i.e., VN and SMN) mainly located in the right hemisphere. Finally, based on the spatial pattern network of FNC and network properties, we propose a classification model. The proposed model achieves a top classification accuracy of 78.2% in cross-subject two-class MI-BCI tasks. Overall, our findings provide new insights into the neural mechanisms of MI and a potential network biomarker to identify MI-BCI tasks.

The long-term effect of modulated acoustic stimulation on alteration in EEG brain network of chronic tinnitus patients: An exploratory study

Acoustic stimulation is one of the most influential techniques for distressing tinnitus, while how it functions to reverse neural changes associated with tinnitus remains undisclosed. In this study, our objective is to investigate alterations in brain networks to shed light on the enigma of acoustic intervention for tinnitus. We designed a 75-day long-term acoustic intervention experiment, during which chronic tinnitus patients received daily modulated acoustic stimulation with each session lasting 15 days. Every 15 days, professional tinnitus assessments were conducted, collecting both electroencephalogram (EEG) and tinnitus handicap inventory (THI) data from the patients. Thereafter, we investigated the changes in EEG network organizations during continuous acoustic stimulation and their progressive evolution throughout long-term therapy, alongside exploring the associations between the evolving changes of the network alterations and THI. Our current study findings reveal reorganization in alpha/beta long-range frontal-parietal-occipital connections as well as local frontal and parietal-occipital regions induced by acoustic stimulation. Furthermore, we observed a decrease in modulation effects as therapy sessions progressed. These alterations in brain networks reflect the reversal of tinnitus-related neural activities, particularly distress and perception; thus contributing to tinnitus rehabilitation through long-term modulation effects. This study provides unique insights into how long-term acoustic intervention affects the network organizations of tinnitus patients and deepens our understanding of the pathophysiological mechanisms underlying tinnitus rehabilitation.

Study subnetwork developing pattern of autism children by non-negative matrix factorization

BACKGROUND

As a developmental disorder, the brain networks of autism children show abnormal patterns compared with that of typically developing. The differences between them are not stable due to the developing progress of children. It has become a choice to study the differences of developing trajectories between autistic and typically developing children by investigating the change of each group respectively. Related researches studied the developing of brain network by analyzing the relationship between network indices of the entire or sub brain networks and the cognitive developing scores.

METHODS

As a matrix decomposition algorithm, non-negative matrix factorization (NMF) was applied to decompose the association matrices of brain networks. By NMF, we can obtain subnetworks in an unsupervised way. The association matrices of autism and control children were estimated by their magnetoencephalography data. NMF was applied to decompose the matrices to obtain common subnetworks of both groups. Then we calculated the expression of each subnetwork in each child's brain network by two indices, energy and entropy. The relationship between the expression and the cognitive and development indices were investigated.

RESULTS

We found a subnetwork with left lateralization pattern in α band showed different expression tendency in two groups. The expression indices of two groups were correlated with cognitive indices in autism and control group in an opposite way. In γ band, a subnetwork with strong connections on right hemisphere of brain showed a negative correlation between the expression indices and development indices in autism group.

CONCLUSION

NMF algorithm can effectively decompose brain network to meaningful subnetworks. The finding of α band subnetworks confirms the results of abnormal lateralization of autistic children mentioned in relevant studies. We assume the results of decrease of expression of the subnetwork may relate to the dysfunction of mirror neuron. The decrease expression of γ subnetwork of autism may be related to the weaken process of high-frequency neurons in the neurotrophic competition.

A Novel Method for Constructing EEG Large-Scale Cortical Dynamical Functional Network Connectivity (dFNC): WTCS

As a kind of biological network, the brain network conduces to understanding the mystery of high-efficiency information processing in the brain, which will provide instructions to develop efficient brain-like neural networks. Large-scale dynamical functional network connectivity (dFNC) provides a more context-sensitive, dynamical, and straightforward sight at a higher network level. Nevertheless, dFNC analysis needs good enough resolution in both temporal and spatial domains, and the construction of dFNC needs to capture the time-varying correlations between two multivariate time series with unmatched spatial dimensions. Effective methods still lack. With well-developed source imaging techniques, electroencephalogram (EEG) has the potential to possess both high temporal and spatial resolutions. Therefore, we proposed to construct the EEG large-scale cortical dFNC based on brain atlas to probe the subtle dynamic activities in the brain and developed a novel method, that is, wavelet coherence-S estimator (WTCS), to assess the dynamic couplings among functional subnetworks with different spatial dimensions. The simulation study demonstrated its robustness and availability of applying to dFNC. The application in real EEG data revealed the appealing "Primary peak" and "P3-like peak" in dFNC network properties and meaningful evolutions in dFNC network topology for P300. Our study brings new insights for probing brain activities at a more dynamical and higher hierarchical level and pushing forward the development of brain-inspired artificial neural networks. The proposed WTCS not only benefits the dFNC studies but also gives a new solution to capture the time-varying couplings between the multivariate time series that is often encountered in signal processing disciplines.

Predicting individual muscle fatigue tolerance by resting-state EEG brain network

OBJECTIVE

Exercise-induced muscle fatigue is a complex physiological phenomenon involving the central and peripheral nervous systems, and fatigue tolerance varies across individuals. Various studies have emphasized the close relationships between muscle fatigue and the brain. However, the relationships between the resting-state electroencephalogram (rsEEG) brain network and individual muscle fatigue tolerance remain unexplored.

APPROACH

Eighteen elite water polo athletes took part in our experiment. Five-minute before- and after-fatigue-exercise rsEEG and fatiguing task (i.e., elbow flexion and extension) electromyography (EMG) data were recorded. Based on the graph theory, we constructed the before- and after-task rsEEG coherence network and compared the network differences between them. Then, the correlation between the before-fatigue rsEEG network properties and the EMG fatigue indexes when a subject cannot keep on exercising anymore was profiled. Finally, a prediction model based on the before-fatigue rsEEG network properties was established to predict fatigue tolerance.

MAIN RESULTS

Results of this study revealed the significant differences between the before- and after-exercise rsEEG brain network and found significant high correlations between before-exercise rsEEG network properties in the beta band and individual muscle fatigue tolerance. Finally, an efficient support vector regression (SVR) model based on the before-exercise rsEEG network properties in the beta band was constructed and achieved the accurate prediction of individual fatigue tolerance. Similar results were also revealed on another thirty-subject swimmer data set further demonstrating the reliability of predicting fatigue tolerance based on the rsEEG network.

SIGNIFICANCE

Our study investigates the relationship between the rsEEG brain network and individual muscle fatigue tolerance and provides a potential objective physiological biomarker for tolerance prediction and the regulation of muscle fatigue.

Constructing Time-varying Directed EEG network by Multivariate Nonparametric Dynamical Granger Causality

Time-varying directed electroencephalography (EEG) network is the potential tool for studying the dynamical causality among brain areas at a millisecond level; which conduces to understanding how our brain effectively adapts to information processing, giving inspiration to causality- and brain-inspired machine learning. Currently, its construction still mainly relies on the parametric approach such as multivariate adaptive autoregressive (MVAAR), represented by the most widely used adaptive directed transfer function (ADTF). Restricted by the model assumption, the corresponding performance largely depends on the MVAAR modeling which usually encounters difficulty in fitting complex spectral features. In this study, we proposed to construct EEG directed network with multivariate nonparametric dynamical Granger causality (mndGC) method that infers the causality of a network, instead, in a data-driven way directly and therefore avoids the trap in the model-dependent parametric approach. Comparisons between mndGC and ADTF were conducted both with simulation and real data application. Simulation study demonstrated the superiority of mndGC both in noise resistance and capturing the instantaneous directed network changes. When applying to the real motor imagery (MI) data set, distinguishable network characters between left- and right-hand MI during different MI stages were better revealed by mndGC. Our study extends the nonparametric causality exploration and provides practical suggestions for the time-varying directed EEG network analysis.

L1-norm based time-varying brain neural network and its application to dynamic analysis for motor imagery

OBJECTIVE

EEG-based motor imagery (MI) brain-computer interface offers a promising way to improve the efficiency of motor rehabilitation and motor skill learning. In recent years, the power of dynamic network analysis for MI classification has been proved. In fact, its usability mainly depends on the accurate estimation of brain connection. However, traditional dynamic network estimation strategies such as adaptive directed transfer function (ADTF) are designed in the L2-norm. Usually, they estimate a series of pseudo connections caused by outliers, which results in biased features and further limits its online application. Thus, how to accurately infer dynamic causal relationship under outlier influence is urgent.

APPROACH

In this work, we proposed a novel ADTF, which solves the dynamic system in the L1-norm space (L1-ADTF), so as to restrict the outlier influence. To enhance its convergence, we designed an iteration strategy with the alternating direction method of multipliers (ADMM), which could be used for the solution of the dynamic state-space model restricted in the L1-norm space. Furthermore, we compared L1-ADTF to traditional ADTF and its dual extension across both simulation and real EEG experiments.

MAIN RESULTS

A quantitative comparison between L1-ADTF and other ADTFs in simulation studies demonstrates that fewer bias errors and more desirable dynamic state transformation patterns can be captured by the L1-ADTF. Application to real MI EEG datasets seriously noised by ocular artifacts also reveals the efficiency of the proposed L1-ADTF approach to extract the time-varying brain neural network patterns, even when more complex noises are involved.

SIGNIFICANCE

The L1-ADTF may not only be capable of tracking time-varying brain network state drifts robustly but may also be useful in solving a wide range of dynamic systems such as trajectory tracking problems and dynamic neural networks.

Impaired Memory Awareness and Loss Integration in Self-Referential Network Across the Progression of Alzheimer's Disease Spectrum

BACKGROUND

Anosognosia, or unawareness of memory deficits, is a common manifestation of Alzheimer's disease (AD), but greatly variable in subjective cognitive decline (SCD) and amnestic mild cognitive impairment (aMCI) subjects. Self-referential network (SRN) is responsible for self-referential processing and considered to be related to AD progression.

OBJECTIVE

Our aim is to explore connectivity changes of SRN and its interaction with memory-related network and primary sensorimotor network (SMN) in the AD spectrum.

METHODS

About 444 Alzheimer's Disease Neuroimaging Initiative subjects (86 cognitively normal [CN]; 156 SCD; 146 aMCI; 56 AD) were enrolled in our study. The independent component analysis (ICA) method was used to extract the SRN, SMN, and memory-related network from all subjects. The alteration of functional connectivity (FC) within SRN and its connectivity with memory-related network/SMN were compared among four groups and further correlation analysis between altered FC and memory awareness index as well as episodic memory score were performed.

RESULTS

Compared with CN group, individuals with SCD exhibited hyperconnectivity within SRN, while aMCI and AD patients showed hypoconnectivity. Furthermore, aMCI patients and AD patients both showed the interruption of the FC between the SRN and memory-related network compared to CN group. Pearson correlation analysis showed that disruptive FC within SRN and its interaction with memory-related network were related to memory awareness index and episodic memory scores.

CONCLUSION

In conclusion, impaired memory awareness and episodic memory in the AD spectrum are correlated to the disconnection within SRN and its interaction with memory-related network.

Blind Visualization of Task-Related Networks From Visual Oddball Simultaneous EEG-fMRI Data: Spectral or Spatiospectral Model?

Various disease conditions can alter EEG event-related responses and fMRI-BOLD signals. We hypothesized that event-related responses and their clinical alterations are imprinted in the EEG spectral domain as event-related (spatio)spectral patterns (ERSPat). We tested four EEG-fMRI fusion models utilizing EEG power spectra fluctuations (i.e., absolute spectral model - ASM; relative spectral model - RSM; absolute spatiospectral model - ASSM; and relative spatiospectral model - RSSM) for fully automated and blind visualization of task-related neural networks. Two (spatio)spectral patterns (high δ 4 band and low β 1 band) demonstrated significant negative linear relationship (p FWE < 0.05) to the frequent stimulus and three patterns (two low δ 2 and δ 3 bands, and narrow θ 1 band) demonstrated significant positive relationship (p < 0.05) to the target stimulus. These patterns were identified as ERSPats. EEG-fMRI F-map of each δ 4 model showed strong engagement of insula, cuneus, precuneus, basal ganglia, sensory-motor, motor and dorsal part of fronto-parietal control (FPCN) networks with fast HRF peak and noticeable trough. ASM and RSSM emphasized spatial statistics, and the relative power amplified the relationship to the frequent stimulus. For the δ 4 model, we detected a reduced HRF peak amplitude and a magnified HRF trough amplitude in the frontal part of the FPCN, default mode network (DMN) and in the frontal white matter. The frequent-related β 1 patterns visualized less significant and distinct suprathreshold spatial associations. Each θ 1 model showed strong involvement of lateralized left-sided sensory-motor and motor networks with simultaneous basal ganglia co-activations and reduced HRF peak and amplified HRF trough in the frontal part of the FPCN and DMN. The ASM θ 1 model preserved target-related EEG-fMRI associations in the dorsal part of the FPCN. For δ 4, β 1, and θ 1 bands, all models provided high local F-statistics in expected regions. The most robust EEG-fMRI associations were observed for ASM and RSSM.

AOAR: an automatic ocular artifact removal approach for multi-channel electroencephalogram data based on non-negative matrix factorization and empirical mode decomposition

OBJECTIVE

Electroencephalogram (EEG) signals suffer inevitable interference from artifacts during the acquisition process. These artifacts make the analysis and interpretation of EEG data difficult. A major source of artifacts in EEGs is ocular activity. Therefore, it is important to remove ocular artifacts before further processing the EEG data.

APPROACH

In this study, an automatic ocular artifact removal (AOAR) method for EEG signals is proposed based on non-negative matrix factorization (NMF) and empirical mode decomposition (EMD). First, the amplitude of EEG data was normalized in order to ensure its non-negativity. Then, the normalized EEG data were decomposed into a set of components using NMF. The components containing ocular artifacts were extracted automatically through the fractal dimension. Subsequently, the temporal activities of these components were adaptively decomposed into some intrinsic mode functions (IMFs) by EMD. The IMFs corresponding to ocular artifacts were removed. Finally, the de-noised EEG data were reconstructed.

MAIN RESULTS

The proposed method was tested against seven other methods. In order to assess the effectiveness and reliability of the AOAR method in processing EEG data, experiments on ocular artifact removal were performed using simulated EEG data. Experimental results indicated that the proposed method was superior to the other methods in terms of root mean square error, signal-to-noise ratio (SNR) and correlation coefficient, especially in cases with a lower SNR. To further evaluate the potential applications of the proposed method in real life, the proposed method and others were applied to preprocess real EEG data recorded from children with and without attention-deficit/hyperactivity disorder (ADHD). After artifact rejection, the event-related potential feature was extracted for classification. The AOAR method was best at distinguishing the children with ADHD from the others.

SIGNIFICANCE

These results indicate that the proposed AOAR method has excellent prospects for removing ocular artifacts from EEG data.

KD-KLNMF: Identification of lncRNAs subcellular localization with multiple features and nonnegative matrix factorization

Long non-coding RNAs (lncRNAs) refer to functional RNA molecules with a length more than 200 nucleotides and have minimal or no function to encode proteins. In recent years, more studies show that lncRNAs subcellular localization has valuable clues for their biological functions. So it is count for much to identify lncRNAs subcellular localization. In this paper, a novel statistical model named KD-KLNMF is constructed to predict lncRNAs subcellular localization. Firstly, k-mer and dinucleotide-based spatial autocorrelation are incorporated as the feature vector. Then, Synthetic Minority Over-sampling Technique is used to deal with the imbalance dataset. Next, Kullback-Leibler divergence-based nonnegative matrix factorization is applied to select optimal features. And then we utilize support vector machine as the classifier after comparing with other classifiers. Finally, the jackknife test is performed to evaluate the model. The overall accuracies reach 97.24% and 92.86% on training dataset and independent dataset, respectively. The results are better than the previous methods, which indicate that our model will be a useful and feasible tool to identify lncRNAs subcellular localization. The datasets and source code are freely available at https://github.com/HuijuanQiao/KD-KLNMF.

Assessment of nonnegative matrix factorization algorithms for electroencephalography spectral analysis

BACKGROUND

Nonnegative matrix factorization (NMF) has been successfully used for electroencephalography (EEG) spectral analysis. Since NMF was proposed in the 1990s, many adaptive algorithms have been developed. However, the performance of their use in EEG data analysis has not been fully compared. Here, we provide a comparison of four NMF algorithms in terms of accuracy of estimation, stability (repeatability of the results) and time complexity of algorithms with simulated data. In the practical application of NMF algorithms, stability plays an important role, which was an emphasis in the comparison. A Hierarchical clustering algorithm was implemented to evaluate the stability of NMF algorithms.

RESULTS

In simulation-based comprehensive analysis of fit, stability, accuracy of estimation and time complexity, hierarchical alternating least squares (HALS) low-rank NMF algorithm (lraNMF_HALS) outperformed the other three NMF algorithms. In the application of lraNMF_HALS for real resting-state EEG data analysis, stable and interpretable features were extracted.

CONCLUSION

Based on the results of assessment, our recommendation is to use lraNMF_HALS, providing the most accurate and robust estimation.