Multiscale spatiotemporal neural network with multi-attention mechanism using brain partitioning for motor imagery recognition.

SigNet-10: a dataset and CNN-based benchmark for signal-level network traffic classification

The binary paradigm of modern electronics imposes intrinsic limits on information density and energy efficiency. Zero differential transconductance (ZDT)-based multi-valued logic (MVL) offers a promising alternative, providing a compelling platform for MVL architectures. However, achieving stable and tunable ZDT operation remains challenging due to the difficulty of precisely controlling ZDT behavior. Here, we present the plateau transistors (ptTs), which enables highly tunable, intrinsic ternary operation through a distinctive polaron-mediated charge transport mechanism. Polaronic behavior in cobalt ferrite (CoFe2O4) was predicted through simulations and subsequently confirmed using a designed charge-smearing near-infrared spectroscopy technique. Incorporating a CoFe2O4 gate dielectric yields stable near-ZDT operation within a standard field-effect transistor architecture, enabling distinct electrical separation of multiple logic states. We further demonstrated that this approach is broadly applicable to a wide range of 2D materials, including MoS2, graphene, and WSe2. The demonstrated ternary operation provides a viable device-level platform for multi-level signal representation, offering new opportunities for energy-efficient electronic devices.

Unregulated accessibility to the latest deepfake technologies presents escalating, unprecedented threats to personal security, public trust, and democratic integrity, owing to the ever-increasing sophistication and realism of these forgeries. The biggest challenge is the inability of human verification to ascertain the original from the forgeries. Therefore, this research aims to establish an initial framework of detection and verification. This research presents a completely new way of detecting manipulation by looking for second-order spatiotemporal inconsistencies in chromatic energy distributions, as opposed to existing deepfake detection methods that rely on complicated multi-stream architectures or first-order pixel-level features. The theoretical importance comes from the fact that generative models can convincingly copy static visual features, but they always fail to keep colour and texture changes that make sense in both space and time. The Chromatic Gradient Anomaly Network (CrGAN) is an architecture that will be built and tested to capture changes of the various components of a video over time in order to reveal patterns of inconsistency between the spatiotemporal levels of a video and the changes of its chromatic components. This method is useful in two ways: first, it gets state-of-the-art detection accuracy without needing complicated multi-modal fusion; second, and more importantly, it lets forensic analysts see exactly where and how a video was changed at the pixel level, which is very important for legal and investigative purposes. One of the most important contributions of this research is the analysis of the second-order derivatives (in this case, the Chromatic Gradient Fields) of the Spatiotemporal Chromatic Energy Distributions, revealing the synthesis boundary of temporally sparse flickers and the physically implausible discontinuities of the blend. The results for CrGAN demonstrate the highest level of diagnostic confidence, reporting a detection rate of 97.9%, and most importantly a level of pixel-wise localized mapping of the detected region that is statistically differentiated from other detection models, achieving state-of-the-art performance while maintaining architectural simplicity. This is a big change in how deepfake detection works: it moves complexity from model architecture to forensic signal representation, which makes the solution more elegant, easier to understand, and more generalizable. In conclusion, this study validates how targeting second-order spatiotemporal inconsistencies using chromatic gradients not only acts as an efficient detection mechanism but also as an interpretable tool in the combat against digital deception by identifying the how and where of video forgery.

DeepSoundVisionNet: A new approach to urban sound classification using visual representations of audio signals

Activity of sensory neurons is influenced not only by external stimuli but also by the animal's behavioral state. It is well documented that behavior influences the general properties of neural activity, such as response gain. However, it is not known whether it could affect the sensory tuning of individual neurons in a more refined way and what the functional benefit of such nuanced modulation might be. Here, we investigate this in the mouse visual cortex using the data made available by the Allen Brain Observatory. Our analysis indicates that locomotion can modulate not only the gain of the entire neuronal response, but also more selectively control responses to specific stimuli. This modulation results in changes of neuronal tuning in different behavioral states. Using numerical simulations, we demonstrate that such patterns of gain modulation can multiplex behavioral information in sensory populations without compromising the accuracy of sensory coding. In that way, the visual cortex could instantiate an accurate, joint representation of sensory and movement-related signals and support computations that simultaneously require both types of information.

Noninvasive neuromonitoring provides valuable insights into cerebral physiology and function, with the potential to support individualized care for patients who are critically ill. As precision medicine continues to advance within the intensive care unit, the integration of multiple noninvasive monitoring techniques may offer a more comprehensive, real-time understanding of dynamic cerebral pathophysiology while avoiding the risks of invasive procedures. This study aimed to identify the implementation challenges of combining multiple noninvasive neuromonitoring modalities. Six modalities were selected: blind transcranial Doppler ultrasonography, optic nerve sheath diameter measurement, multichannel continuous electroencephalography (cEEG), automated pupillometry, near-infrared spectroscopy (NIRS), and skull extensometer (Brain4Care). To assess the readiness of each modality for integration into a multimodal noninvasive monitoring setup, a group of experts propose the Noninvasive Neuromonitoring Multimodal Readiness (nnMR) score. This scoring system aims to provide a structured evaluation of each modality's multimodal suitability based on criteria such as data continuity and processing, physiological signal representation, spatial and temporal resolution, and software integration in the intensive care unit setting. nnMR scores showed that cEEG and NIRS enable continuous monitoring without manual processing. Blind transcranial Doppler and NIRS offer a good temporal resolution; optic nerve sheath diameter provides good spatial resolution; cEEG and skull extensometer combine both. Only cEEG demonstrates anatomical clarity. Software integration remains limited across all modalities. Key challenges to multimodal implementation include the need for skilled personnel, artifact susceptibility, mismatched temporal and spatial resolutions, and limited long-term signal stability. Integration into unified analytical platforms is further constrained by a lack of standardization, requiring separate manual data input, processing, and quantification. The nnMR score was developed to assess the current readiness of each modality for multimodal, continuous, and long-term clinical application. To validate the proposed nnMR score, engagement with the broader neurocritical care community is essential. Realizing the full potential of multimodal neuromonitoring will require advances in signal validation, software integration, resolution alignment, artifact detection, fixation techniques, and standardization, necessitating close collaboration among researchers, engineers, manufacturers, and clinicians.

Multi-component LFM signal representation method under impulsive noise: Principle, method and application

This paper presents a Load-Dependent Multimodal Vibration Signal Enhancement and Fusion Framework (LD-MVSEFF) for load-specific condition monitoring, building on the Customised Load Adaptive Framework (CLAF). The proposed approach enhances the classification of CLAF load-dependent subclasses, namely, Healthy, Mild, Moderate, and Severe, by integrating complementary information from raw vibration signals and encoded signal representations. Three input channels are employed, combining time–frequency domain features with Continuous Wavelet Transform (CWT) and Gramian Angular Difference Field (GADF) image encodings, with each channel independently trained and evaluated to identify its most effective classifiers. To address the reduced separability of the Mild and Moderate fault subclasses under varying load conditions, a weighted decision-fusion strategy is introduced, assigning classifier contributions according to their class-specific strengths. Experimental evaluation over five runs demonstrates high and stable performance, with the best configuration achieving an overall accuracy of 99.04% ± 0.22% and an average training time of 18 min and 30 s. The results confirm the effectiveness of LD-MVSEFF as a robust multimodal methodology for load-specific condition monitoring.

Falls represent a major health risk for older adults living independently, motivating the development of unobtrusive and privacy-preserving monitoring solutions. This study investigates whether Bluetooth Low Energy (BLE) 6.0 Channel Sounding (CS) can support device-free fall detection using low-complexity signal representations suitable for residential deployment. The proposed system employs two BLE nodes performing periodic channel sounding, from which only scalar distance estimates are extracted. Time-domain and temporal-dynamic features are computed from sliding windows of the distance signal and used for supervised classification. Three widely used classifiers-Support Vector Machine with radial basis function kernel, Random Forest, and gradient-boosted decision trees (XGBoost)-are evaluated under both a default operating point and a sensitivity-first regime achieved through validation-based decision threshold adjustment, reflecting the higher cost of missed fall detections in assisted living scenarios. Experiments conducted in a furnished indoor environment with six participants performing realistic fall and non-fall scenarios demonstrate strong window-level sensitivity under subject-independent evaluation, with XGBoost providing the most favorable sensitivity-specificity balance. Under sensitivity-first operation, very high recall is achieved at the expense of increased false alarms. Given the limited dataset and single-environment setting, the reported results should be interpreted as a proof-of-concept demonstration of feasibility rather than definitive large-scale performance. The findings suggest that BLE CS captures motion-relevant signal variations that may support practical fall detection while maintaining low deployment complexity and privacy preservation.

A Conserved Systems-Level Constraint in Biological Acoustic Signaling.

This study proposes a supervised machine learning framework for vibration-based fault diagnosis of rotating machinery using integrated data-driven and physics-informed feature sets. A dataset acquired under variable load and multiple operating conditions was used for model training. Parallel signal processing techniques were applied to capture fault-related information across multiple frequency bands including time-domain analysis, frequency-domain analysis, baseband analysis, and envelope analysis. From the corresponding signal representations, statistical, spectral, and physics-based features associated with characteristic fault frequencies were extracted and combined into integrated feature sets. The diagnostic performance of models trained using purely data-driven features was systematically compared with models incorporating integrated data-driven and physics-informed features. Support Vector Machine, Random Forests, Gradient Boosting, and an ensemble classifier were evaluated using accuracy, precision, recall, and F1-score metrics. The proposed framework employs a two-layer classification strategy, where the first layer performs multiclass fault identification, while the second layer evaluates the presence of imbalance as a coexisting fault. In addition, the influence of different feature groups as well as individual measurement axes and their combinations on diagnostic performance were analyzed. Validation using a new dataset measured in laboratory conditions confirmed the robustness and generalization capability of the proposed diagnostic framework.

Advances in artificial intelligence (AI) and multi-omics integration are reshaping precision oncology by enabling deeper mechanistic understanding, improved characterization of tumor heterogeneity, and the accelerated discovery of targeted therapeutics. Antimicrobial peptides (AMPs) have emerged as promising candidates for cancer treatment due to their selective cytotoxicity, immunomodulatory properties, and ability to alter the tumor microenvironment. However, their accurate computational identification remains challenging because existing models struggle to capture the complex structural and functional determinants of AMP activity. In this study, we propose GAC-BiTCNN-AMP, a hybrid generative and explainable deep learning framework designed to advance peptide discovery for precision oncology. The architecture integrates a Generative Adversarial Network to enhance data diversity, Capsule Networks to model hierarchical molecular dependencies, and a Bidirectional Temporal Convolutional Neural Network for capturing contextual sequence information. To strengthen biological signal representation, the model incorporates embeddings from advanced protein language model including ProtTrans-T5, UniRep, and ESM-2 alongside a novel PsePSSM-DCT evolutionary descriptor. A wrapper-based XGBoost Forward Feature Selection strategy further refines the feature space by identifying the most discriminative sequence patterns. GAC-BiTCNN-AMP delivers strong predictive performance, achieving 97.42% accuracy and 0.923 MCC in cross-validation, and 95.32% accuracy with 0.914 MCC on the same independent test set. SHapley Additive exPlanations (SHAP) analysis highlights key contributions from the fused latent representations to peptide activity, demonstrating the framework's interpretability at the representation level. By integrating generative modeling, deep representation learning, and explainable AI, this study provides a scalable computational pipeline supporting therapeutic peptide discovery for targeted, immune-modulatory, and precision cancer applications.

Knowledge of how vocal communication signals are represented in the auditory system is crucial for understanding the perceptual basis of vocal communication. Using male and female zebra finches, we identified differentially expressed molecular markers that helped define distinct (caudal, rostral, dorsal, and ventral) domains within the caudomedial nidopallium (NCM), a high-order cortical auditory area known for its song-selective responses. Using expression analysis of the activity-inducible gene zenk, we found that the number of activated neurons is more stimulus dependent in NCM than in the auditory midbrain or the caudomedial mesopallium and that information on the density and spatial distribution of responsive neurons in NCM is sufficient to discriminate responses to conspecific song from other stimuli. We observed stronger activation of dorsal NCM, higher selectivity of caudal NCM toward conspecific song, and strong activation of the inhibitory network of rostral NCM by nonconspecific song stimuli. The spatial organization of responsive cells was particularly sensitive to both spectral and temporal components of song. We also obtained evidence of broadly distributed song-selective neuronal ensembles and that individual NCM neurons participate in the representation of different conspecific songs, implying independent activation and molecular induction responses. We conclude that some basic aspects of the cortical response to complex auditory stimuli are topographically organized, a finding that has been elusive in other systems. These findings advance our knowledge of the functional organization of a key song-processing cortical area, providing novel insights into the auditory representation of vocal communication signals.

According to data provided by the World Health Organization (WHO), falls are one of the major reasons for unintentional deaths or injuries in older adults. Although many fall detection methods and algorithms exist, there is no efficient artificial intelligence strategy for fall detection. Various studies have stated that Fall Detection among Elderly Persons (FDEP) provides the possibility of developing an efficient and cost-effective way to tackle this problem. This study generated a signal-based image dataset, SimgFall, from the existing accelerometer or gyroscope-based sensor data of the SiSFall dataset for the early detection of falls to accelerate the medical assistance process. The SimgFall dataset was used to train and evaluate the FallCNN model, a novel deep Convolutional Neural Network (CNN) architecture comprising multiple CNN folds to effectively learn discriminative features from the transformed signal representations. These models utilize depth-wise convolution with varying dilation rates to efficiently extract diversified features from the SimgFall dataset. The dataset contained 1992 signal-based images, of which 498 were samples collected for fall, jump, stumble, and walk for the 4 classes. The initial architecture, referred to as FallCNN_1, with two basic convolutional layers and max-pooling, which is simple and efficient in feature extraction and dimensionality reduction, resulted in 94% accuracy for detecting the 4 classes. The incorporation of the average pooling and dropout layers in FallCNN_2 reduced overfitting and improved feature extraction, thereby enhancing the accuracy to 95%. Expanding the feature dimensions in FallCNN_3 further refined the capacity of the model to capture intricate patterns, achieving a notable accuracy of 97%. Finally, FallCNN_4 with three convolutional blocks and additional intermediate layers achieved the highest accuracy of 98%, demonstrating cumulative performance improvements through architectural enhancements. Furthermore, the performance of the generated dataset using different pretrained and custom models was evaluated based on the loss and accuracy curves. The experimental results showed that the highest classification accuracy was 98%, with a loss of 0.0833, using categorical cross-entropy as the loss function.

Electrocardiogram (ECG) analysis is a fundamental tool for diagnosing various cardiac conditions; however, accurately distinguishing between normal and abnormal ECG signals remains challenging due to high inter-individual variability and the inherent complexity of ECG waveforms. In this study, We propose a novel Sparse Temporal Autoencoder (STAE) for unsupervised ECG anomaly detection that leverages Temporal Convolutional Networks (TCNs) to extract hierarchical features from both time-domain and frequency-domain representations of ECG signals. Unlike traditional approaches requiring annotated abnormal samples, the proposed model is trained exclusively on normal ECG data, making it well-suited for real-world deployment. A STAE integrates a masked signal reconstruction strategy and a hybrid sparse attention mechanism combining sparse block and sparse strided attention to capture critical temporal and spectral patterns efficiently. The proposed method is evaluated on the PTB-XL dataset, where it achieves the highest ROC-AUC of 0.872 among compared unsupervised methods while maintaining a low inference time of 0.009 s, demonstrating that STAE achieves state-of-the-art performance in ECG anomaly detection, highlighting its potential as a powerful tool for automated and intelligent ECG analysis.

This study aims to support reliable ECG signal interpretation by reducing human-dependent variability through computer-aided analysis methods. Machine learning and deep learning methods were employed to examine 2D ECG representations and Synthetic Minority Over-Sampling Technique (SMOTE)-based balancing in ECG classification. Unlike existing ECG classification studies that typically address signal representation and class imbalance separately, this study jointly investigates the interaction between two-dimensional QRS representation and SMOTE-based data balancing within a unified experimental framework, thereby providing a systematic analysis of their combined impact on classification performance. Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), and K-Nearest Neighbors (KNN) algorithms were implemented and comparatively analyzed. ECG beats from record 108 of the MIT-BIH Arrhythmia dataset were represented in a vision-based form for classification. To address severe class imbalance, SMOTE was applied only to the training data, and its effect on two-dimensional ECG representations was explicitly examined. Normal and Abnormal heartbeats were classified using a stratified 5-fold cross-validation strategy. Experimental results demonstrated that the CNN model achieved the most successful performance after applying SMOTE, reaching a weighted average F1-score of 99.82% ± 0.002, highlighting the combined effectiveness of two-dimensional QRS representation and data balancing in improving automated ECG classification.

Sparse representation of high-density retinal signal in time-frequency domain to support diagnosis in psychiatric disorders

An adaptive decoupling learning system informed by the brain functional structure for EEG decoding.

This paper proposes a new technique for computer modeling linear filters based on the spectral form of mathematical description of linear systems. It assumes the representation of input and output signals of the filter as orthogonal expansions, while filters themselves are described by two-dimensional non-stationary transfer functions. This technique allows one to model the output signal in continuous time, and it is successfully tested on the Butterworth, Linkwitz-Riley, and Chebyshev filters with different orders.