Brain-age models based on functional connectivity have shown promise for characterizing large-scale age-related network changes and advancing understanding of neural aging mechanisms. However, existing approaches still face challenges in achieving both robust predictive performance and stable identification of connectivity signatures. To address this, we developed a brain-age modeling framework that integrates SHAP-guided PCA back-projection to link model predictions to stable, interpretable connectivity signatures and aging trajectories. Resting-state fMRI data from 599 adults (age range 36–95 years) were used to compute whole-brain functional connectivity. An Orthogonal Auxiliary Guidance Convolutional Autoencoder (OAG-CAE) was developed to learn age-relevant latent representations while separating age-residual variation, followed by a regressor for brain-age prediction. To achieve stable connection-level interpretability, we designed a SHAP-guided PCA back-projection module that applies SHAP to PCA-reduced FC features, propagates attribution through the latent representation and predictive pathway, and projects PCA-level attributions back to connectivity space to localize age-relevant circuits. The brain-age model achieved robust accuracy across mid- to late adulthood (MAE = 6.308 ± 0.566 years; R^2 = 0.702 ± 0.043). SHAP analysis revealed systematic patterns of functional reorganization with aging. Negative SHAP-identified connectivity was concentrated within and between the somatomotor, ventral attention, and default mode networks, as well as in basal ganglia–thalamus–cortical pathways. Notably, a substantial proportion of these connections exhibited an inverted U trajectory peaking around age 70, including within-network effects in the somatomotor and default mode networks and cross-network thalamo-cortical and somatomotor–ventral attention links. In contrast, positive SHAP-identified connectivity emerged mainly in cerebellar links with the default mode, frontoparietal, and dorsal attention networks, as well as default mode–visual and dorsal attention–thalamic pathways. By integrating OAG-CAE-based latent representation learning with SHAP-guided PCA back-projection, this study establishes an interpretable brain-age modeling framework that links prediction to specific connectivity circuits. The identified signatures reveal stage-dependent reorganization across large-scale networks, including non-linear transitions during mid- to late adulthood. These findings offer new insight into large-scale functional brain aging and provide a structured basis for future investigations in pathological aging and independent cohorts.
Autism spectrum disorder (ASD) is a group of neurodevelopmental disorders characterized by deficits in social communication and restricted, repetitive behaviors or interests. The development of ASD involves complex interactions between genetic and environmental factors. This study aims to investigate potential molecular links between target genes of chlorpyrifos and ASD, thereby elucidating potential mechanisms by which environmental exposures contribute to ASD development and exploring candidate biomarkers. We analyzed the ASD dataset GSE18123, autism-related genes (ARGs), and chlorpyrifos target-related genes (CRGs). A total of 206 autism and chlorpyrifos target-related genes (ACRGs) were identified. Through protein-protein interaction (PPI) network analysis, six hub genes (MYC, HIF1A, SP1, CXCL8, IFNG, and IL10) were screened. Functional enrichment analyses indicated that these ACRGs are primarily involved in energy metabolism and immune/inflammatory regulation. Using least absolute shrinkage and selection operator (LASSO) penalized logistic regression, an exploratory four-gene signature (CXCL8, IFNG, MYC, and SP1) was identified, which showed discriminatory ability in the current dataset. Moreover, immune cell infiltration analysis via single-sample gene set enrichment analysis (ssGSEA) indicated significant differences in 10 immune cell types, which were correlated with key genes. Molecular docking and dynamics simulations suggested potential binding interactions of MYC with methylprednisolone and SP1 with terameprocol, providing preliminary computational clues for further experimental validation. Our findings revealed an exploratory four-gene signature linked to chlorpyrifos exposure and ASD, and suggested that MYC and SP1 may represent candidate molecules for future validation in therapeutic studies. These results provide new insights into the interaction between environmental exposures and genetic factors in ASD, offering novel clues for future research in precision diagnosis and potential therapeutic exploration.
Deep learning has further accelerated progress in automated Magnetic Resonance Imaging (MRI) based classification of brain tumors, whereas prior studies have often reflected critical limitations such as narrow classification scope, insufficient Explainable Artificial Intelligence, weak statistical validation, and incomplete deployment metrics. These ultimately impede clinical trust and practical deployment. This study presents the CAE-BrainNet, a statistically validated Class-Adaptive Attention Ensemble model, which integrates representations of EfficientNetV2-M, DenseNet201, and ConvNeXt-Base dynamically, with complementary inductive biases based on different tumor morphology. Unlike uniform ensembles, a learned class-adaptive attention mechanism, parameterised by a 4× 3 weight matrix with a trainable temperature scalar, adaptively leverages model-specific morphological strengths, serving as the primary performance driver confirmed by ablation analysis. A hybrid optimization strategy based on grid search refinement via gradient-based optimization is also employed to estimate optimal hyperparameters. In addition, McNemar’s and Cochran’s Q tests were conducted, proving that the accuracy gains are statistically significant and not stochastic in nature. Experimental results show that the proposed CAE-BrainNet achieves 99.39
Major Depressive Disorder (MDD) is a prevalent mental illness and a leading cause of suicide, making early detection critical. The electroencephalogram (EEG) provides a non-invasive method for recording brain activity; however, current diagnostic methods rely on manual evaluation by neurologists, which can be subjective, labor-intensive, and time-consuming. While Machine Learning (ML) and Deep Learning (DL) have shown promise in medical diagnosis, many existing models for MDD detection lack generalizability due to a large number of learnable parameters, small datasets, and data leakage from Patient-Independent evaluation protocols. In this work, we propose an ensemble classifier of lightweight convolutional neural networks (CNNs) for MDD detection using EEG signals, designed with minimal learnable parameters to ensure efficiency and generalization. The method was evaluated on the MUMTAZ benchmark dataset using Leave-Some-Subject-Out Cross Validation (LSSOCV) to ensure robust performance, achieving an accuracy of 94.1
This study investigates the topological differences between functional networks in the left and right hemispheric regions during visual word processing using task-based fMRI data from 45 subjects in the OpenfMRI project. The analysis employs Automated Anatomical Labeling (AAL) atlases to construct functional networks under two experimental conditions: word processing and consonant string processing. Four key topological metrics are systematically compared between hemispheres: degree centrality, global efficiency, clustering coefficient, and local efficiency. Significant hemispheric differences in local efficiency occurred between 13–19 ≤ 1 ≤ 1 ≤ 4
Hotline text data, characterized by their colloquial, fragmented, and noisy nature, present significant challenges for traditional event extraction methods based on formal written text. To address these challenges, this study proposes EENE, a Graph-Aware Transformer model that reformulates multi-stage hotline event extraction as a graph- structured decoding task centered on trigger words to achieve efficient event extraction from fragmented and noisy hotline texts. By integrating BERT-based semantic encoding with Transformer-based dependency modeling, EENE effectively captures long-range contextual dependencies and fine-grained semantic associations for accurately extracting token–token association from fragmented and noisy texts. Furthermore, these connections are utilized to construct a token–token association matrix in which triggers serve as central nodes and arguments as edge nodes, thereby enabling event extraction by structured graph reasoning. Multi-stage event extraction is thus transformed into a single-step association extraction task, effectively avoiding cascading errors and improving the accuracy of event extraction. Experimental results on two self-built Chinese hotline datasets covering urban flooding, public complaints, and help requests demonstrate that EENE achieves state-of-the-art performance, surpassing representative baselines such as EDEE and CASEE by 5.15
For any machine learning or deep learning model, explainability is of paramount importance. As data complexity increases, deep learning models renowned for their outstanding performance have been applied to a wide range of tasks. However, these models remain typical “black boxes” whose predictions are difficult to interpret. To address this limitation, this study proposes a progressive framework that integrates feature engineering with knowledge distillation for interpreting deep learning-based hypertension prediction models. Statistical relevance testing are performed to select target-associated variables, and numeric variables are first partitioned with a supervised decision tree to obtain interval-based, clinically legible features with interpretable thresholds. On this basis, FT-Transformer is adopted as the teacher to generate soft labels that capture nonlinear structure for distillation. Finally, a globally interpretable linear model is trained on the transformed features and soft labels, preserving accuracy while delivering quantitative, interval-level risk attribution. Experiments on the public MIMIC-III demonstrate that our method significantly outperforms traditional interpretable models and machine learning baselines. Although AUC and accuracy slightly lag behind deep learning models, our approach delivers intuitive feature interpretations while maintaining competitive performance.
Alzheimer’s disease (AD) is a progressive neurodegenerative disorder marked by structural brain changes detectable through neuroimaging, particularly Magnetic Resonance Imaging (MRI), which can reveal the extent of atrophy in cortical and subcortical regions. We propose an ensemble of linear models for AD classification, combining regression and classification techniques for a novel methodology to identify potential biomarkers from MRI data, termed Apparent Brain Features (ABF). These biomarkers represent morphological brain regions automatically selected to optimise classification accuracy while preserving interpretability. Unlike deep learning or other nonlinear methods, our approach maintains the anatomical semantics of the input space. A key innovation is a feature score that quantifies the influence of each selected morphological region on classification, enabling both diagnostic utility and neuroscientific insights. We validate our approach on MRI scans from 1990 subjects gathered from four publicly available repositories: ADNI, AIBL, PPMI, and IXI. The results show that our ensemble methodology achieves high classification accuracy while offering an interpretable framework for assessing the role of brain morphology in AD. A systematic selection and evaluation of brain regions can provide a transparent and clinically relevant tool, supporting both computational neuroscience research and practical diagnostic applications.
Brain-Computer Interfaces (BCIs) represent a rapidly advancing field that enables communication between the brain and external devices. Motor Imagery (MI) and Motor Execution (ME) paradigms are particularly valuable for BCI applications, especially in neurorehabilitation contexts. This study introduces a novel hybrid TCN-Transformer architecture designed to automatically classify electroencephalography (EEG) signals from MI and ME tasks. The proposed model combines a Temporal Convolutional Network (TCN) for feature extraction with a Transformer encoder for modeling global temporal dependencies. We validated our approach through comprehensive experiments on three datasets: Shuqfa-103 (103 subjects), Brunner-9 (9 subjects), and Kodera-29 (29 subjects). The model demonstrated strong performance across multiple classification scenarios, achieving accuracies of 84.84
This paper proposes a deep learning-based cross-modality fusion framework that integrates neuroimaging data (MRI and DaTSCAN) with handwriting-derived image data to capture distinctive patterns of Parkinson disease. While neuroimaging modalities provide valuable insights into structural and functional brain abnormalities, handwriting-based assessments capture subtle motor dysfunctions that are also highly relevant to disease characterization. By combining structural brain features with quantitative handwriting characteristics, the proposed approach demonstrates that complementary information across modalities can be effectively leveraged. Following preprocessing and enhancement procedures applied to the images, important features were extracted using a ResNet50 model fine-tuned with the multimodal dataset. The obtained feature set was then used to train a stacking classifier. This ensemble-based strategy allows the model to leverage complementary strengths of different classifiers, providing a more robust decision mechanism. Experimental results demonstrate that this cross-modality integration is feasible and yields promising results, highlighting the potential of generating multimodal datasets for Parkinson research. Such datasets could enable the design of task-specific models in future studies, ultimately contributing to more comprehensive and reliable assessment.
Transcranial direct current stimulation (tDCS) is a non-invasive neuromodulation technique that has shown promise for enhancing cognitive functions such as attention and working memory, yet its effects remain inconsistent across studies. This study introduces a modified phase-based analysis to capture the temporal dynamics of tDCS-induced changes in both behavior and brain activity. Nineteen healthy adults performed a visual working memory task under active and sham stimulation targeting the left and right posterior parietal cortex (lPPC, rPPC) and right dorsolateral prefrontal cortex (rDLPFC). Trials were chronologically divided into Early, Middle, and Late phases to examine evolving cognitive states over time. Behavioral outcomes-hit rate and reaction time-did not differ significantly between stimulation and sham, suggesting stable performance across conditions. However, event-related potential (ERP) analyses revealed phase-dependent neural modulations: rDLPFC stimulation showed a trend toward reduced Contralateral Delay Activity (CDA) in the Early phase and enhanced P3 amplitude in the Late phase, indicating improved memory efficiency and consolidation, while lPPC stimulation tended to increase N2PC amplitude during the Early phase, reflecting enhanced attentional allocation. These findings demonstrate that even without overt behavioral effects, tDCS can induce temporally specific neural changes. The modified phase-based analysis provides a sensitive framework for linking behavioral stability with dynamic ERP modulations, offering new insights into how tDCS influences attention and working memory across time.
We present EEGScope, a novel framework for real time EEG acquisition, processing, and dynamic feature visualization. Unlike traditional offline EEG toolkits, EEGScope offers a modular and user-friendly system that integrates live filtering, cognitive metric computation, and auditory sonification. Application scenarios include attentional monitoring, inter subject synchrony, and emotional tracking. A real-time experiment on Alpha Frontal Asymmetry highlights the framework’s ability to reveal rhythm modulations through immediate Power Spectral Density (PSD) shifts. EEGScope provides a lightweight, extensible platform to enhance exploratory neuroscience research, neurofeedback training, and adaptive BCI development.
Aiming at the problems that early screening of chronic diseases relies too much on doctors’ experience, the utilization of patients’ health data is not comprehensive, and the traditional deep learning model lacks clinical interpretability and traceability, this paper explores the application of large language model and retrieval enhancement generation technology in the classification of chronic diseases. Time alignment and semantic links are performed on multimodal data such as patient lifestyle, health records, electronic medical records, laboratory indicators, and medical images through data preprocessing. Based on clinical guidelines and medical literature, a domain vector knowledge base is constructed, and prompt word engineering and RAG technology are integrated to dynamically retrieve authoritative medical literature, providing context for large language model reasoning, realizing collaborative optimization of multimodal data fusion and dynamic knowledge incremental updating, and breaking through the limitations of traditional large models resulted from poor domain data quality and static knowledge storage. The experiments show that the accuracy of chronic diseases classification can be improved to more than 91
Understanding how localized anatomical variations influence brain age predictions is critical for clinical interpretability. However, existing explainability methods primarily capture associations and do not support controlled interventions to assess how specific morphometric changes affect model output. These approaches lack the ability to simulate coherent “what-if” scenarios at the regional level. We present a simulation framework that enables localized morphometric interventions with biologically consistent propagation of effects. When a single feature (e.g., gray matter volume) is perturbed, the framework automatically adjusts the remaining features within the same brain region to preserve statistical dependencies, estimated from a reference population. This produces realistic, covarying input profiles that any pretrained brain age model can evaluate without retraining. The method is tested on the publicly available OpenBHB dataset, focusing on two regions of interest selected for their contrasting correlation with age. Controlled perturbations of 1
This study explores the optimization of classical Brain-Computer Interface (BCI) paradigms—Motor Imagery (MI) and Steady-State Visual Evoked Potentials (SSVEP)—using a portable, dry-electrode EEG system in operating conditions. By adjusting key parameters, such as stimulation frequency, window length, and classifier type, we aim to bridge the gap between controlled laboratory settings and practical BCI applications. Experiments were conducted across varied environments and times to capture ecological variability. Our results reveal strong inter- and intra-subject performance variability, emphasizing the need for personalized adaptation. In MI, we evaluated three classification methods, with TSLDA offering the best trade-off for real-time use. For SSVEP, we assessed target properties and high-frequency stimuli to enhance comfort without sacrificing signal quality. The dry EEG device demonstrated performance comparable to wet systems, with certain participants achieving high classification accuracy and Information Transfer Rates in real-time conditions. These findings highlight the feasibility of deploying dry EEG-based BCIs outside laboratory settings, provided careful tuning of paradigm-specific parameters. This work lays a foundation for developing adaptive BCI systems designed for operational deployment.
Estimating music perceived by individuals from electroencephalography (EEG) signals holds considerable promise for both medical and engineering applications. In this study, we propose an EEG-driven music generation framework based on latent discrete diffusion models, designed to generate music conditioned on EEG recordings. The proposed method comprises two primary stages. First, a variational autoencoder (VAE) is employed to extract informative latent representations from EEG signals, yielding a conditional vector that encapsulates neural activity patterns associated with auditory perception. Second, a discrete diffusion model operates in the latent space, leveraging these EEG-derived feature vectors to generate music data that reflect the underlying neural responses to music stimuli. We evaluate the proposed approach using real EEG data recorded from human subjects during music listening. Given the inherently noisy nature of EEG signals, ensuring robustness to such noise is a critical challenge. Experimental results demonstrate that our model effectively generates music while maintaining resilience against noise, highlighting its potential as a novel machine learning-based approach for EEG-driven music generation and broader cross-modal neural-audio synthesis.
This research project aims to represent the development and emergence of the severe neuropsychiatric condition named Cotard’s delusion by implementing a six-level higher adaptive network model capturing biological, cognitive, and epigenetic levels. The model emphasizes adaptive processes via dynamic interactions between environmental triggers, stressors, epigenetic mechanisms, neural plasticity, and most importantly the relationship between the aforementioned and beliefs. More specifically, it portrays how disturbances originating from both the upmost epigenetic level and the baseline environmental factors propagate through subsequent states result in persistent maladaptive cognitive states belonging to Cotard’s syndrome. Furthermore, the research aims to model targeted epigenetic therapeutic interventions, leveraging the adaptive structure to identify ideal points to mitigate the nihilistic delusions. Based on the limited research due to its sub parent nature as a disorder, another major aim of this paper is to shed light on the rare characteristics of patients suffering from this, who are usually grouped and diagnosed into other major, far more common disorders such as MDD (Major Depressive Disorder) and psychosis.
In this study, electroencephalography (EEG) is used to investigate the mental states involved in music imagery. A subject was exposed to eight different songs, each lasting four seconds, which they hear, repeat, and later recall and check. The paper focuses on using temporal generalization techniques to analyze how imagined and real musical experiences are reflected in brain activity. This study shows it is crucial to analyze how EEG signals change over time, specifically in response to musically salient events within the imagined song. By focusing on these dynamic shifts, rather than treating the mental state as constant, we could leverage EEG’s high temporal resolution to decode imagined music with greater accuracy. These findings have implications for neurology, cognitive neuroscience, and brain-computer interfaces.
Recent studies in decoding neural signals for speech-related applications have shown considerable promise for advanced brain-computer interfaces (BCIs). However, most studies have focused on speech production, while auditory speech reconstruction remains a challenging task. This paper introduces NeuroLingua, a lightweight and interpretable machine learning framework for bilingual auditory speech reconstruction from stereotactic electroencephalography (sEEG) signals. While high-frequency sEEG features are often used exclusively, we propose to integrate both low- and high-frequency neural features that complement one another, and employ an extreme gradient boosting (XGBoost) regression model paired with Shapley additive explanations (SHAP) for enhanced interpretability. To evaluate NeuroLingua, we collected and analyzed a bilingual sEEG-audio dataset from epilepsy patients undergoing intracranial monitoring. We show that the proposed framework consistently outperforms conventional single-band approaches in speech reconstruction. Furthermore, the model allows us to identify the most informative neural channels for bilingual speech reconstruction tasks. This study advances the neural speech decoding studies that support the development of next-generation BCIs for assistive communication and rehabilitation in multilingual populations. Code is publicly available ( https://github.com/seegdecoding/NeuroLingua ).
Visual evoked potentials (VEPs) are widely used in EEG-based brain-computer interface (BCI) systems. Among these, code-modulated VEPs (c-VEPs) offer high communication speeds, support a large number of stimulus options and provide robust classification performance through the use of m-sequences. Although convolutional neural networks such as EEGNet have been successfully applied to frequency-based steady-state VEPs (SSVEPs), their performance on c-VEPs remains underexplored. In this study, an ablation analysis is conducted to evaluate architectural and training modifications aimed at improving EEGNet’s performance in c-VEP tasks. Moreover, a simple yet effective multi-cycle c-VEP classification method based on logit aggregation is proposed. Additionally, a calibration method that leverages artificially shifted m-sequences is introduced to construct multiclass training sets from a single sequence, thereby reducing data requirements. The methods are evaluated on three public datasets and one private dataset in a cross-subject setting. Results show that combining layer normalization, mixup augmentation, weight decay, and exponential moving average (EMA) improves EEGNet’s classification accuracy by up to 28.3
