Cross-subject electroencephalogram (EEG) emotion recognition is an important task in affective computing, which aims to learn discriminative and generalizable EEG features to reveal emotions across individuals. To achieve robust generalization, existing methods primarily focus on developing domain alignment constraints to learn features that remain consistent across all source domains, while ignoring the potential benefits of domain-specific information in enhancing model discriminability. As a result, these approaches struggle to utilize relevant source domain information to improve predictions for unseen domains. To address this limitation, we propose a novel prompt-guided domain generalization (PGDG) framework that extends the invariance perspective by incorporating domain-specific information through prompt learning. Specifically, the variational autoencoder is first used to extract domain-invariant features, and then the prompt of each source domain is learned by guiding the classifier in the optimal direction. Finally, a multi-head cross-attention mechanism adaptively integrates these domain prompts from multiple source domains to improve the model's generalization ability in target domains. Experimental results on the SEED and SEED-IV datasets show that PGDG outperforms state-of-the-art methods, demonstrating the potential of prompt-guided generalization in improving cross-subject EEG emotion recognition performance.

The motor imagery (MI) based brain–computer interface (BCI) holds broad application prospects in human–machine interaction. However, current MI recognition approaches primarily utilize complex attention modules for higher recognition accuracy, consequently hindering real-time BCI implementation. Furthermore, existing methods often overlook inter-subject variability, leading to inadequate generalization of model. Additionally, traditional BCI systems lack closed-loop feedback from the machine to the brain. To address these limitations, we develop a novel closed-loop motor imagery BCI system, which encompasses a spectral-temporal refined attention network via contrastive mutual learning (STRA-CML) and a brain-controlled perceived hand exoskeleton. Specifically, we first design a spectral temporal refined attention block to capture the most discriminative spectral and temporal features. Second, we investigate a contrastive mutual learning strategy incorporating supervised-contrastive learning to enhance the generalization of our STRA-CML. Finally, a brain–machine closed-loop interaction platform based on perceived hand exoskeleton is developed to validate the feasibility of the proposed STRA-CML and provide kinesthetic and visual feedback synchronized with MI. Competitive experimental results on two public datasets and a self-collected dataset demonstrate the effectiveness of our STRA-CML, indicating that our STRA-CML achieves superior classification performance of 83.89% on BCI IV 2a dataset, 86.93% on BCI IV 2b dataset, and 82.79% on self-collected dataset.

With the development of deep brain stimulation technique, single-channel intracranial electroencephalography (iEEG) based seizure prediction is a necessary and urgent needed tool for epilepsy closed-loop neuromodulation. However, previous prediction methods based on multi-channel scalp signals heavily relied on the spatial information, failing to fully exploit the interdependencies between temporal scales and spectral rhythms of single-channel iEEG. Additionally, current contrastive learning strategies can lead to model overfitting by excessively learning the feature distances in small samples, limiting the precision of seizure prediction. To tackle above issues, based on a single-channel iEEG, we propose a novel dual-cross tri-level routing transformer based metric learning network (DC-TRT-MLNet) for epileptic seizure prediction. First, a scale-rhythm dual-cross (DC) graph attention network is introduced to construct the dependent relationships across multi-scale temporal and multi-rhythm spectral features. Second, we design a tri-level routing transformer (TRT) network to comprehensively refine the most seizure-potential routing features while eliminating redundant information. Finally, a hard triplet optimization based metric learning (ML) strategy is developed to iteratively optimize the intra-class and inter-class distances of inter-ictal and pre-ictal routing features. Competitive experimental results on a private Xuanwu Single-Channel iEEG dataset validate the effectiveness of our proposed method, demonstrating the superior prediction performance of our DC-TRT-MLNet compared with the state-of-the-art methods. Our study may offer a new solution for intracranial single-channel seizure prediction.

Automatic anatomical localization is critical for radiology report generation. While many studies focus on lesion detection and segmentation, anatomical localization-accurately describing lesion positions in radiology reports-has received less attention. Conventional segmentation-based methods are limited to organ-level localization and often fail in severe disease cases due to low segmentation accuracy. To address these limitations, we reformulate anatomical localization as an image-to-text retrieval task. Specifically, we propose a CLIP-based framework that aligns lesion image patches with anatomically descriptive text embeddings in a shared multimodal space. By projecting lesion features into the semantic space and retrieving the most relevant anatomical descriptions in a coarse-to-fine manner, our method achieves fine-grained lesion localization with high accuracy across the entire body. Our main contributions are as follows: (1) hierarchical anatomical retrieval, which organizes 387 locations into a two-level hierarchy, by retrieving from the first level of 124 coarse categories to narrow down the search space and reduce localization complexity; (2) augmented location descriptions, which integrate domain-specific anatomical knowledge for enhancing semantic representation and improving visual-text alignment; and (3) semi-hard negative sample mining, which improves training stability and discriminative learning by avoiding selecting the overly similar negative samples that may introduce label noise or semantic ambiguity. We validate our method on two whole-body PET/CT datasets, achieving an 84.13% localization accuracy on the internal test set and 80.42% on the external test set, with a per-lesion inference time of 34 ms. The proposed framework also demonstrated superior robustness in complex clinical cases compared to segmentation-based approaches.

Depression detection using Electroencephalogram (EEG) signals obtained from wearable medical-assisted diagnostic systems has become a well-established approach in the field of affective disorders. However, despite recent advancements, on-board Artificial Intelligence (AI) models still demand substantial computational resources, presenting significant challenges for deployment on resource-constrained wearable medical devices. Embedded Multi-core Processors (MPs) offer a promising solution for accelerating these models. However, the limited computational capabilities of embedded MPs, combined with the structural diversity of AI models, complicate resource allocation and increase associated costs. To address these challenges, we propose a Memory-Aware Multi-Objective Iterative Local Search (MAMILS) algorithm to optimize task scheduling, thereby improving the efficiency of AI model deployment on wearable EEG devices. Experimental results across seven AI models demonstrate that, the MAMILS approach yields substantial improvements in key performance indicators: Total Energy Consumption ($\bm {TEC}$) with an average reduction of 47.57%,$\bm {Makespan}$with an average reduction of 48.75%, and$\bm {Throughput}$with an average increase of 198.37%, all while maintaining satisfactory classification performance for both Machine Learning (ML) and Deep Learning (DL) models. Especially, on-board deployment of EEGNeX achieves an accuracy of 93.4%, sensitivity of 91.6%, and specificity of 95.8%. Further analysis indicates that, when integrated with wearable EEG sensors and executable on-board AI models, the proposed MAMILS optimization strategy shows significant promise in facilitating the widespread adoption of low-power, real-time diagnostic systems for depression detection.

Vision Transformers (ViTs) have demonstrated impressive performance across various vision tasks. However, these models usually suffer from intensive computational costs and excessive memory requirements, making them impractical for deployment. In this paper, we propose a post-training pruning method specially tailored to ViTs. To achieve this goal, we underline two important facts: i) the importance of weights heavily relies on the attributes of the sample; ii) some weights show minor contributions in feed-forward calculation. Motivated by the first observation, we propose a weight selection and recovery mechanism, pruning the weights with the lowest score and then recovering the important weight via calibration samples for layer-wise tuning. Motivated by the second observation, we explore rank values with search space pruning to accelerate low-rank factorization for more effective compression. On top of these, we further conduct hardware optimization to support the pruned models. Promisingly, three designs could promote the effectiveness of each other and finally form an overall framework. Using the DeiT benchmarks, our framework achieves competitive acceleration and accuracy conditioned on using 1% data of ImageNet-1K.

Colorectal Cancer (CRC) is one of the leading causes of cancer-related deaths worldwide, and early screening plays a crucial role in improving patient outcomes. In this study, we present a novel AI-assisted CRC diagnostic system using Bowel Sound (BS) signals. We first develop two portable BS acquisition devices with distinct form factors for high-fidelity signal capture in both clinical and home-care scenarios. A total of 221 recordings were collected under expert-guided protocol, with 144 CRC recordings and 59 Non-CRC healthy controls using the developed device. To enable low-resource deployment, we design a lightweight deep learning model optimized for real-time, on-board inference. The model incorporates multiple training strategies, including transfer learning on a large-scale public BS dataset, self-supervised temporal feature learning, and a hybrid semi-and weakly-supervised approach that leverages both unlabeled and real-noise data. Furthermore, a Sound Event Detection (SED) attention mechanism and iterative consistency learning are introduced to enhance the model’s sensitivity to BS activity. The proposed model comprises only 264.7 K parameters and 253.2 M Floating-Point Operations (FLOPs), requiring 1.57 MB of RAM and 1.03 MB of FLASH when deployed on microcontroller. It performs inference in approximately 3.4 s with low power consumption, making it well-suited for low-resource environments. Despite its compact design, the model achieves 93.06% classification accuracy, 96.46% sensitivity, and 86.99% specificity for binary-classes in CRC diagnosis. These results demonstrate the system’s potential for accessible and cost-effective CRC screening in community, home, and rural healthcare scenarios.

Emotions regulation play an important role in human behavior, but exhibit considerable heterogeneity among individuals, which attenuates the generalization ability of emotion models. In this work, we aim to achieve robust emotion prediction through efficient disentanglement of affective semantic representations. In detail, the data generation mechanism behind observations from different perspectives is causally set, where latent variables that relate to emotion are explicitly separate into three parts: the intrinsic-related part, the extrinsic-related part, and the spurious-related part. Affective semantic features consist of the first two parts, with the understanding that spurious latent variables generate the inherent biases in the data. Furthermore, a variational autoencoder with a reformulated objective function is proposed to learn such disentangled latent variables, and only adopts semantic representations to perform the final classification task, avoiding the interference of spurious variables. In addition, for electroencephalography (EEG) data used in this article, a space-frequency mapping method is introduced to improve information utilization. Comprehensive experiments on popular emotion datasets show that the proposed method can achieve competitive intersubject generalization performance. Our results highlight the potential of efficient latent representation disentanglement in addressing the complexity challenges of emotion recognition.

Schizophrenia (SZ) is a severe mental disorder characterized by hallucinations, delusions, cognitive impairments, and social withdrawal. It leads to a series of brain abnormalities, particularly the deformation of the hippocampus and amygdala, which are highly associated with emotion, memory, and motivation. Most previous studies have used the hippocampal and amygdaloid volume, whereas surface-based morphometry reflects nuclear deformation more finely, but it is unclear the hippocampal and amygdaloid morphometry relates to schizophrenic pathology and its potential as a biomarker. In this study, we extracted individual multivariate morphometry statistics (MMS) of hippocampus and amygdala from MRI images and analyzed the morphometric differences between groups. After dictionary learning and max pooling, we obtain reduced dimensional features and use machine learning algorithms for individual diagnosis. The results showed that the hippocampus of the schizophrenia group was significantly atrophied bilaterally and the atrophied areas were symmetrical. Subregions of the amygdala are both atrophied and expanded, and in particular, the right amygdala shows a greater degree and extent of deformation. Using the random forest classifier, the accuracy of classification using hippocampal and amygdaloid morphometric features are 94.52% and 94.57%, respectively, and the accuracy of classification combining the two morphometric features reached 96.57%. Our study demonstrates the efficacy of MMS in identifying morphometric differences of the hippocampus and amygdala between healthy controls and schizophrenic, and these findings emphasize the potential of MMS as a reliable biomarker for the diagnosis of schizophrenia.

Heart sounds play a crucial role in assessing Coronary Artery Disease (CAD). The advancement of Artificial Intelligence (AI) technologies has given rise to Computer Audition (CA)-based methods for CAD detection. However, previous research has focused primarily on analyzing and modeling heart sound data, overlooking practical application scenarios. In this work, we design a pervasive heart sound collection device used for high-quality heart sound data acquisition. Moreover, we introduce an on-board executable lightweight network tailored for the designed portable device, referred to as TYKDModel. Further, heart sound data from 41 CAD patients and 22 non-CAD healthy controls are collected using the developed device. Experimental results show that the TYKDModel exhibits low-computational complexity, with 52.16 K parameters and 5.03 M Floating-Point Operations (FLOPs). When deployed on the board, it requires only 1.10 MB of Random Access Memory (RAM) and 236.27 KB of Read-Only Memory (ROM), and takes around 1.72 seconds to perform a classification. Despite the low computational and spatial complexity, the TYKDModel achieves a notable classification accuracy of 85.2%, specificity of 88.6%, and sensitivity of 82.8% on the board. These results indicate the promising potential of AI-assisted all-in-one integrated system for the diagnosis of heart sound-assisted CAD.

Depression detection via wearable Electroencephalogram (EEG) sensor-assisted diagnosis system demands computationally efficient models compatible with resource-constrained edge devices. Spiking Neural Networks (SNNs) offer inherent advantages for processing the spatio-temporal patterns of EEG through event-driven manner. In this study, we innovatively present LSNNet, a lightweight SNN model specifically designed for wearable EEG sensors. The model exhibits low computational complexity with 7.18 K parameters and 67.68 M Floating-Point Operations (FLOPs). It requires only 246.88 KB of Random Access Memory (RAM) and 57.33 KB of Read-Only Memory (ROM) for on-board execution, and has been validated on both the single-core STM32U535CET6 and the multi-core GAP8 microcontrollers. Despite its minimal computational and memory requirements, LSNNet achieves impressive performance metrics, with a classification accuracy of 89.2%, specificity of 92.4%, and sensitivity of 86.4% in independent tests conducted on EEG data collected from 73 depressed patients and 108 healthy controls using our three-lead EEG sensor. Especially, when running on the GAP8 microcontrollers, the LSNNet model has a low power consumption of 21.43 mW and a satisfactory inference time of 0.63 s while maintaining a classification accuracy of 87.5% (only with a reduction of 1.98%). These results underscore the potential of integrating wearable EEG sensors with the LSNNet model for depression detection in the Internet of Things (IoT) era.

Vigilance estimation is a critical task within the field of brain-computer interfaces, extensively applied in monitoring and optimizing user states during human-machine interaction using electroencephalography (EEG). However, most existing vigilance prediction frameworks are prone to spurious correlations stemming from inherent biases in collected data. These biases involve relevant but vigilance-independent information, which may lack robustness when applied to different data distributions, i.e., out-of-distribution (OOD) scenarios. The core idea of this study is to learn constraints that capture causal information from the input based on the assumed underlying data generating process. Leveraging the disentanglement and invariance principles behind the assumptions, we propose a constraint-driven causal representation learning (CCRL) to identify and separate spurious latent variables from biased training data for generalized vigilance estimation. The CCRL training process consists of two phases: self-supervised pretraining and constraint-driven causal information disentanglement. In the first phase, based on the masked autoencoder (MAE) architecture, unlabeled training data are used for reconstructing pretext tasks to capture the comprehensive and intrinsic contextual information from EEG data, which provides a powerful input for downstream disentanglement learning. In the second phase, we propose a novel disentanglement strategy to learn spurious-free latent representations causally related to the vigilance state driven by adversarial and invariance constraints. Comprehensive validation experiments conducted on two well-known public datasets demonstrate the effectiveness and superiority of the proposed framework. In general, this work has promising implications for addressing OOD challenges in vigilance estimation.

Functional connectivity (FC) network derived from resting-state functional magnetic resonance imaging (rs-fMRI) has been extensively employed in the automated identification of brain disorders. Conventional FC network modeling methods typically assign equal importance to different sampling points of rs-fMRI and brain regions of interest (ROIs). Nevertheless, considering temporal and regional heterogeneity, this assumption may not always be applicable. Moreover, sparse regression algorithms with regularization terms are commonly adopted to eliminate spurious FC. However, these conventional regularization terms impose a uniform sparse penalty on all FC without considering the prior knowledge that the brain tends to transmit information in an energetically efficient manner. To address the above problems, we propose an adaptive weight and Wasserstein distance constrained low-rank sparse representation (AW-WD-LSR) method to construct FC networks. Specifically, we employ adaptive weights to the reconstruction errors of rs-fMRI time series across various ROIs and sampling points, thereby amplifying the significance of crucial features in the joint representation, leading to a more robust FC network. Meanwhile, we incorporate a local constraint by introducing the optimal transport distance (i.e., Wasserstein distance) between ROIs to adjust the sparse penalty on the corresponding FC. The larger Wasserstein distance, the higher information transmission cost between two ROIs, resulting in a greater penalty imposed on the corresponding FC. The efficacy of the innovation modules is demonstrated in classification tasks involving major depressive disorder (MDD) with mixed features (MMF), MDD without mixed features (MDDnoMF), and healthy controls (HCs).

Currently, the integration of artificial intelligence (AI) techniques with multimodal physiological signals represents a pivotal approach to detect affective disorders (ADs). With the increasing complexity and diversity of physiological signal modalities, researchers have introduced various AI methods using multimodal physiological signals to improve model classification performance and explainability to increase trust and facilitate clinical adoption. Among these methods, spiking neural networks (SNNs) stand out as a promising avenue due to their alignment with the operating principles of the human brain, robust biological explainability, and adeptness in processing spatial–temporal information in an efficient event-driven manner with low power consumption. Furthermore, the emergence of neuromorphic computing (NC) chips based on SNNs has greatly bolstered the field of NC, enabling effective support for objective, pervasive, and wearable AI-assisted medical diagnostic devices for ADs and other diseases. This article presents a review of recent achievements in multimodal AD detection and points out the associated challenges in utilizing multimodal physiological signals and NC based on SNNs for AD detection. Building upon this foundation, we give perspectives on future work. The intended readership for this review consists of researchers in the fields of cognitive computing, computational psychophysiology, affective computing, NC, and brain-inspired computing. We hope that this survey not only garners increased attention from the scientific community but also serves as a valuable guide for future studies in this field.

Functional connectivity (FC) networks, built from analyses of resting-state magnetic resonance imaging (rs-fMRI), serve as efficacious biomarkers for identifying Autism Spectrum Disorders (ASD) patients. Given the neurobiological heterogeneity across individuals and the unique presentation of ASD symptoms, the fusion of individualized information into diagnosis becomes essential. However, this aspect is overlooked in most methods. Furthermore, the existing methods typically focus on studying direct pairwise connections between brain ROIs, while disregarding interactions between indirectly connected neighbors. To overcome above challenges, we build common FC and individualized FC by tangent pearson embedding (TP) and common orthogonal basis extraction (COBE) respectively, and present a novel multiview brain transformer (MBT) aimed at effectively fusing common and indivinformation of subjects. MBT is mainly constructed by transformer layers with diffusion kernel (DK), fusion quality-inspired weighting module (FQW), similarity loss and orthonormal clustering fusion readout module (OCFRead). DK transformer can incorporate higher-order random walk methods to capture wider interactions among indirectly connected brain regions. FQW promotes adaptive fusion of features between views, and similarity loss and OCFRead are placed on the last layer to accomplish the ultimate integration of information. In our method, TP, DK and FQW modules all help to model wider connectivity in the brain that make up for the shortcomings of traditional methods. We conducted experiments on the public ABIDE dataset based on AAL and CC200 respectively. Our framework has shown promising results, outperforming state-of-the-art methods on both templates. This suggests its potential as a valuable approach for clinical ASD diagnosis.

Conventional functional connectivity network (FCN) based on resting-state fMRI (rs-fMRI) can only reflect the relationship between pairwise brain regions. Thus, the hyper-connectivity network (HCN) has been widely used to reveal high-order interactions among multiple brain regions. However, existing HCN models are essentially spatial HCN, which reflect the spatial relevance of multiple brain regions, but ignore the temporal correlation among multiple time points. Furthermore, the majority of HCN construction and learning frameworks are limited to using a single template, while the multi-template carries richer information. To address these issues, we first employ multiple templates to parcellate the rs-fMRI into different brain regions. Then, based on the multi-template data, we propose a spatio-temporal weighted HCN (STW-HCN) to capture more comprehensive high-order temporal and spatial properties of brain activity. Next, a novel deep fusion model of multi-template called spatio-temporal weighted multi-hypergraph convolutional network (STW-MHGCN) is proposed to fuse the STW-HCN of multiple templates, which extracts the deep interrelation information between different templates. Finally, we evaluate our method on the ADNI-2 and ABIDE-I datasets for mild cognitive impairment (MCI) and autism spectrum disorder (ASD) analysis. Experimental results demonstrate that the proposed method is superior to the state-of-the-art approaches in MCI and ASD classification, and the abnormal spatio-temporal hyper-edges discovered by our method have significant significance for the brain abnormalities analysis of MCI and ASD.

Functional connectivity (FC) networks based on resting-state functional magnetic imaging (rs-fMRI) are reliable and sensitive for brain disorder diagnosis. However, most existing methods are limited by using a single template, which may be insufficient to reveal complex brain connectivities. Furthermore, these methods usually neglect the complementary information between static and dynamic brain networks, and the functional divergence among different brain regions, leading to suboptimal diagnosis performance. To address these limitations, we propose a novel multi-graph cross-attention based region-aware feature fusion network (MGCA-RAFFNet) by using multi-template for brain disorder diagnosis. Specifically, we first employ multi-template to parcellate the brain space into different regions of interest (ROIs). Then, a multi-graph cross-attention network (MGCAN), including static and dynamic graph convolutions, is developed to explore the deep features contained in multi-template data, which can effectively analyze complex interaction patterns of brain networks for each template, and further adopt a dual-view cross-attention (DVCA) to acquire complementary information. Finally, to efficiently fuse multiple static-dynamic features, we design a region-aware feature fusion network (RAFFNet), which is beneficial to improve the feature discrimination by considering the underlying relations among static-dynamic features in different brain regions. Our proposed method is evaluated on both public ADNI-2 and ABIDE-I datasets for diagnosing mild cognitive impairment (MCI) and autism spectrum disorder (ASD). Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art methods. Our source code is available at https://github.com/mylbuaa/MGCA-RAFFNet.

Clinically, retinal vessel segmentation is a significant step in the diagnosis of fundus diseases. However, recent methods generally neglect the difference of semantic information between deep and shallow features, which fail to capture the global and local characterizations in fundus images simultaneously, resulting in the limited segmentation performance for fine vessels. In this article, a global transformer (GT) and dual local attention (DLA) network via deep-shallow hierarchical feature fusion (GT-DLA-dsHFF) are investigated to solve the above limitations. First, the GT is developed to integrate the global information in the retinal image, which effectively captures the long-distance dependence between pixels, alleviating the discontinuity of blood vessels in the segmentation results. Second, DLA, which is constructed using dilated convolutions with varied dilation rates, unsupervised edge detection, and squeeze-excitation block, is proposed to extract local vessel information, consolidating the edge details in the segmentation result. Finally, a novel deep-shallow hierarchical feature fusion (dsHFF) algorithm is studied to fuse the features in different scales in the deep learning framework, respectively, which can mitigate the attenuation of valid information in the process of feature fusion. We verified the GT-DLA-dsHFF on four typical fundus image datasets. The experimental results demonstrate our GT-DLA-dsHFF achieves superior performance against the current methods and detailed discussions verify the efficacy of the proposed three modules. Segmentation results of diseased images show the robustness of our proposed GT-DLA-dsHFF. Implementation codes will be available on https://github.com/YangLibuaa/GT-DLA-dsHFF.

Dynamic functional connectivity (dFC) network inferred from resting-state fMRI reveals macroscopic dynamic neural activity patterns for brain disease identification. However, dFC methods ignore the causal influence between the brain regions. Furthermore, due to the complex non-Euclidean structure of brain networks, advanced deep neural networks are difficult to be applied for learning high-dimensional representations from brain networks. In this paper, a group constrained Kalman filter (gKF) algorithm is proposed to construct dynamic effective connectivity (dEC), where the gKF provides a more comprehensive understanding of the directional interaction within the dynamic brain networks than the dFC methods. Then, a novel virtual adversarial training convolutional neural network (VAT-CNN) is employed to extract the local features of dEC. The VAT strategy improves the robustness of the model to adversarial perturbations, and therefore avoids the overfitting problem effectively. Finally, we propose the high-order connectivity weight-guided graph attention networks (cwGAT) to aggregate features of dEC. By injecting the weight information of high-order connectivity into the attention mechanism, the cwGAT provides more effective high-level feature representations than the conventional GAT. The high-level features generated from the cwGAT are applied for binary classification and multiclass classification tasks of mild cognitive impairment (MCI). Experimental results indicate that the proposed framework achieves the classification accuracy of 90.9%, 89.8%, and 82.7% for normal control (NC) vs. early MCI (EMCI), EMCI vs. late MCI (LMCI), and NC vs. EMCI vs. LMCI classification respectively, outperforming the state-of-the-art methods significantly.

Dynamic functional connectivity (dFC) analysis using resting-state functional Magnetic Resonance Imaging (rs-fMRI) is currently an advanced technique for capturing the dynamic changes of neural activities in brain disease identification. Most existing dFC modeling methods extract dynamic interaction information by using the sliding window-based correlation, whose performance is very sensitive to window parameters. Because few studies can convincingly identify the optimal combination of window parameters, sliding window-based correlation may not be the optimal way to capture the temporal variability of brain activity. In this paper, we propose a novel adaptive dFC model, aided by a deep spatial-temporal feature fusion method, for mild cognitive impairment (MCI) identification. Specifically, we adopt an adaptive Ultra-weighted-lasso recursive least squares algorithm to estimate the adaptive dFC, which effectively alleviates the problem of parameter optimization. Then, we extract temporal and spatial features from the adaptive dFC. In order to generate coarser multi-domain representations for subsequent classification, the temporal and spatial features are further mapped into comprehensive fused features with a deep feature fusion method. Experimental results show that the classification accuracy of our proposed method is reached to 87.7%, which is at least 5.5% improvement than the state-of-the-art methods. These results elucidate the superiority of the proposed method for MCI classification, indicating its effectiveness in the early identification of brain abnormalities.

Mild cognitive impairment (MCI) detection is important, such that appropriate interventions can be imposed to delay or prevent its progression to severe stages, including Alzheimer's disease (AD). Brain connectivity network inferred from the functional magnetic resonance imaging data has been prevalently used to identify the individuals with MCI/AD from the normal controls. The capability to detect the causal or effective connectivity is highly desirable for understanding directed functional interactions between brain regions and further helping the detection of MCI. In this paper, we proposed a novel sparse constrained effective connectivity inference method and an elastic multilayer perceptron classifier for MCI identification. Specifically, a ultra-group constrained structure detection algorithm is first designed to identify the parsimonious topology of the effective connectivity network, in which the weak derivatives of the observable data are considered. Second, based on the identified topology structure, an effective connectivity network is then constructed by using an ultra-orthogonal forward regression algorithm to minimize the shrinking effect of the group constraint-based method. Finally, the effective connectivity network is validated in MCI identification using an elastic multilayer perceptron classifier, which extracts lower to higher level information from initial input features and hence improves the classification performance. Relatively high classification accuracy is achieved by the proposed method when compared with the state-of-the-art classification methods. Furthermore, the network analysis results demonstrate that MCI patients suffer a rich club effect loss and have decreased connectivity among several brain regions. These findings suggest that the proposed method not only improves the classification performance but also successfully discovers critical disease-related neuroimaging biomarkers.