DNA-Driven EEG monitoring for rapid seizure prediction in healthcare.

Incorporating global-local tissue changes to predict future breast cancer from longitudinal screening mammograms.

Neurobridge: Bridging functional and structural brain networks via neural coupling and consistency-Guided dynamic graph learning.

Time-frequency contrastive learning with context modeling for time series anomaly prediction.

Transfer learning from 2D natural images to 4D fMRI brain images via geometric mapping.

Traffic prediction based on real-world traffic data is a crucial task in Intelligent Transportation Systems (ITS). However, the issue of missing observations due to real-world disturbances undermines the robustness and accuracy of traffic prediction. This problem necessitates the development of a prediction model that integrates the imputation mechanism to be compatible with missing observations. This paper introduces ATTST, a self-imputation-assisted prediction model specifically designed to address the challenge of missing observations in the traffic prediction task. Unlike traditional approaches utilizing an additional supervised imputation model before prediction, the imputation unit in our model does not need the extra label for the missing observations. Our model employs a self-imputation unit to impute the missing observations by partially masking the observed data as the ground true labels. Thus, the self-imputation unit along with an encoder–decoder architecture and a graph evolving unit together directly predict future traffic data with multi-level missing observations. The effectiveness of ATTST is validated using several real-world traffic datasets, including speed and flow data, across various multi-step prediction scenarios with diverse missing observations. These validations demonstrate the model’s robustness and practical applicability in real-world traffic prediction tasks. The results show that ATTST can reliably predict traffic conditions even with incomplete data, making it a valuable tool for traffic management and planning. • The problem of imputation and prediction for traffic speed and flow data with missing observations is formulated within an end-to-end framework. • The proposed AttSt model addresses this problem by incorporating a self-imputation mechanism to effectively handle missing observations. • The model employs a graph network to capture and process spatiotemporal correlations within the traffic data. • Comprehensive evaluations on four real-world traffic speed and flow datasets validate the effectiveness of the proposed approach.

Explainable multimodal deep learning models for variable-length sequences in critically ill patients.

In bandwidth-limited and time-varying vehicle–road–cloud cooperative autonomous driving scenarios, real-time transmission and joint inference of high-dimensional multimodal perception data are simultaneously constrained by latency, reliability, and energy consumption. To address these challenges, this article proposes a task-oriented multimodal fusion framework named Multi-Agent Dynamic Diffusion Semantic Communication Network (MA-DDSCNet). On the vehicle side, we design a Task-Guided Multi-Modal Semantic Encoder (TG-MMSE) that performs spatio-temporal alignment, complementary memory gating, and differentiable discrete quantization to compress heterogeneous perception streams from cameras, Light Detection and Ranging (LiDAR), and vehicular state into task-weighted discrete token sequences. A hierarchical distillation scheme is further employed to maintain a unified semantic coordinate system across vehicles, Road Side Units (RSUs), and the cloud. On the communication side, a hierarchical controllable diffusion mechanism adaptively adjusts diffusion noise and time steps according to the importance of object detection, trajectory prediction, and motion planning tasks, as well as link-specific bandwidth budgets. A multi-agent deep scheduler enables collaborative utilization of communication resources among the cloud, RSUs, and vehicles, while an iterative joint semantic decoding and consistency calibration algorithm feeds residuals back into a global memory matrix to suppress semantic drift and yield isomorphic semantic representations at all three layers. Furthermore, we construct a learnable uncertainty-driven multi-objective loss function, combined with a gradient projection strategy, to achieve end-to-end joint optimization of detection, prediction, and planning within a single training loop. Simulation results demonstrate that, compared with baseline methods, MA-DDSCNet achieves average gains of 9.6–18.4% in mean Average Precision (mAP), Average Displacement Error (ADE), Final Displacement Error (FDE), Effective Bit Rate (EBR), and planning safety rate, while reducing the 95th-percentile end-to-end latency to 63 ms, indicating that the proposed framework can significantly enhance the overall performance of semantic perception tasks in complex vehicular networks.

Due to the significant non-stationarity and feature distribution discrepancies in surface electromyography (sEMG) signals during muscle fatigue monitoring, traditional fixed-parameter Transformer models often struggle to accurately capture the complex evolution of time-frequency characteristics across different fatigue stages. To address this limitation, this paper proposes a K-means clustering-guided neural architecture search method (CG-NAS) to achieve adaptive optimization of Transformer architectures based on data distribution characteristics. The method first classified input EMG features using the K-means clustering algorithm and constructed Gaussian distributions characterized by mean and variance to quantify the complexity of each cluster. These distribution priors then guided the neural architecture search process, enabling dynamic alignment between the architecture search space and data characteristics: for low-complexity data clusters with small mean and variance, lightweight Transformer architectures were selected, whereas for high-complexity clusters, architectures with greater width and depth were allocated. Experimental results demonstrated the superior performance of CG-NAS in muscle fatigue index prediction tasks, achieving a mean absolute error of 0.098 2 and a coefficient of determination of 0.957 3, significantly outperforming multiple benchmark models. The study shows that CG-NAS effectively aligns with the nonlinear evolution of time-frequency features during the fatigue process and provides an efficient and robust solution for fatigue monitoring.

Artificial intelligence (AI), as a socio-technical system, has been found to shape the socialization of human users by structuring how they acquire information, form judgments, and develop social values. Existing studies show that AI exhibits distinctive socio-cultural stances, yet its position within the Chinese value landscape and its potential influence on Chinese residents remain underexplored. In this paper, we draw on technical script theory from STS and conceptualize AI systems as carriers of implicit technical scripts, embodying underlying socio-cultural assumptions, imaginations, and expectations. Through two related studies, including an AI prediction task and one persuasion experiment, we show that AI agents, whether developed by U.S. or Chinese companies, tend to embody values that are more aligned with liberalism and postmaterialism than those held by most Chinese residents on topics of political freedom, gender equality, and sexual tolerance, as indicated by empirical data from the 2021 CGSS. These values shape how AI presents reality and interacts with users, potentially influencing users’ memories, cognition, and attitudes toward various moral claims. Our work contributes to technical script theory and the critical study of AI, while also pointing to a potential socio-cultural transformation driven by AI technologies in China.

Abstract Existing deep learning methods have achieved promising results in industrial time-series tasks such as remaining useful life (RUL) prediction, yet their performance still relies on repeated manual adjustment when degradation patterns vary across datasets and operating conditions. Although Neural Architecture Search (NAS) provides a natural way to automate architecture design, existing methods remain limited in industrial scenarios due to insufficient integration of heterogeneous temporal modules and inadequate support for cross-condition settings. To address this issue, this paper proposes Evo-NAS, a genetic algorithm-based neural architecture search framework that constructs candidate layers using temporal convolutional networks (TCN), bidirectional recurrent neural networks (BiLSTM/BiGRU), and multiple attention mechanisms, and adaptively searches for dataset-specific model configurations. On the four subsets of the NASA C-MAPSS benchmark, Evo-NAS achieves RMSE values of 11.93, 19.06, 11.78, and 18.35, corresponding to improvements of 4.71\%, 11.43\%, 14.26\%, and 16.30\% over the best fixed-architecture baselines, respectively. The method is further validated on a real-world aero-engine lubricating oil consumption prediction task, demonstrating its engineering applicability. Experimental results show that Evo-NAS reduces manual trial and error in model structure design within a controlled search space while maintaining competitive predictive performance.

Computational toxicity prediction has become a key component in modern drug discovery. Although machine learning or deep learning techniques have reformed this field in recent years, more in-depth studies on addressing data imbalance, missing labels, and lack of model interpretability are still needed. In this work, we develop a substructure-based deep graph learning architecture, by introducing various functional groups into the construction of molecular graphs and handling them through deep learning models. Our model, with several strategies adopted to deal with the missing labels and class imbalance in the datasets, performed well in toxicity prediction tasks. A functional group-based feature importance analysis provided further insights into different toxicity predictions and improved the interpretability of our model. It provides a solid foundation for the development of reliable toxicity prediction tools and supports rational decision-making in the drug development process.

Multimodal neural operators for real-time biomechanical modelling of traumatic brain injury.

Heart disease remains a leading cause of mortality worldwide, and timely and accurate diagnosis is crucial for improving patient outcomes. Medical imaging plays a pivotal role in this process, yet traditional diagnostic methods often suffer from limitations, including dependency on manual interpretation, susceptibility to observer variability, and inefficiency in handling large-scale data. Deep learning has emerged as an innovative technology in medical imaging, providing unparalleled advancements in feature extraction, segmentation, classification, and prediction tasks. Despite its proven potential, comprehensive reviews of deep learning methods specifically targeted at cardiac imaging remain scarce. This review paper seeks to bridge this gap by analyzing the state-of-the-art deep learning applications for heart disease diagnosis, covering the period from 2015 to 2025. Employing a well-structured methodology, this review categorizes and examines studies based on imaging modalities: Ultrasound (US), Magnetic Resonance Imaging (MRI), X-ray, Computed Tomography (CT), and Electrocardiography (ECG). For each modality, the analysis focuses on utilized datasets, processing techniques (e.g., extraction, segmentation and classification), and paradigms (e.g., transfer learning, federated learning, explainability, interpretability, and uncertainty quantification). Additionally, the types of heart disease addressed and prediction accuracy metrics are also scrutinized. These findings point toward future opportunities, including the study of data quality, optimization, transfer learning, uncertainty quantification and model explainability or interpretability. Furthermore, exploring advanced techniques such as recurrent expansion, transformers, and other architectures may unlock new pathways in cardiac imaging research. This review is a critical synthesis offering a roadmap for researchers and practitioners to advance the application of deep learning in heart disease diagnosis.

The online services have grown incredibly fast, and with such growth, also increases the fraudulent online services, such as phishing websites, email spoofing, internet domain, and over-the-phone frauds. These attacks take advantage of the vulnerability in the structure of Uniform Resource Locators (URLs), identity formats of the sender, domain registration system and numbering system in telecommunications. Such dynamic and emerging attacks are not usually that easily detected using traditional rule-based security mechanisms, which make use of fixed signatures and fixed patterns. The study suggests an Artificial Intelligence (AI)-based system to identify fraudulent web platforms based on structural analysis and behavioral analysis of phone numbers, email address, and URLs. The system combines telecommunication metadata analysis to detect suspicious phone numbers, Domain Name System (DNS) and Mail Exchange (MX) record authentication to determine sender integrity and a Random Forest machine learning classifier to profile URL and email-based threats. An interface based in a Flask allows meeting the task of real-time threat scanning and prediction. The model uses data preprocessing, feature engineering, model training, heuristic evaluation, and deployment. Results of trials performed on sample phishing data show high accuracy, precision, recall and F1-score, which are indicators of strong detection results. The framework is designed to be flexible to adapt to the new categories of cyber threats. The results indicate that the suggested system may be effectively used to reinforce traditional cybersecurity defenses through aid of the strengths revealed in automated detection and minimization of the use of rule-based approaches that are considered to be quite static.

Diagnosis and prognosis of lung cancer via PET/CT imaging have long been major clinical concerns. However, existing multimodal approaches often focus on feature aggregation rather than cross-modal interactive collaboration, failing to capture the structural-metabolic correlations and multi-scale synergy essential for characterizing complex lesions. Therefore, this study proposes TriFuse-Net, a tri-branch PET/CT fusion pyramid network (FPN) enhanced by lesion-guided structural-metabolic attention (LSMA) to improve both diagnosis and prognosis prediction tasks. The model is composed of two identical unimodal branches (PET/CT) and one pyramid branch with an interacting channel and spatial attention. The pyramid structure enables bidirectional multiscale feature extraction and fusion, capturing both local details and global semantic information of lesions. Comprehensive experiments validated the model's superiority across three clinical tasks. TriFuse-Net achieved a C-index of 0.747 for progression-free survival (PFS) prediction, showing improvements of 14.7% and 11.0% over ResNet-CT and ResNet-PET, respectively. Additionally, the clinical-integrated model (TriFuse-Net-Cli) achieved AUCs of 0.947 for differentiating lung cancer from tuberculosis and 0.937 for identifying lymph-node metastasis. Ablation studies further confirmed the essential contributions of both FPN and LSMA. In summary, the proposed framework demonstrates that integrating multi-scale structural-metabolic relationships significantly enhances diagnosis and prognosis in lung cancer.

Abstract In real-world applications, there is a fundamental problem: the data most critical to predict interesting events, anomalies, and high-stakes outliers are the rarest, while less interesting data is abundant. Although deep learning is deployed specifically for these difficult prediction tasks, data-driven models inevitably fail in underrepresented areas. This discrepancy between the empirical data- and the desired evaluation distribution is equivalent to a target distribution shift. The research field, termed Deep Imbalanced Regression (DIR), has emerged explicitly to address this challenge, which is particularly acute for continuous targets where most conventional classification-based methods are ill-suited. In this paper, we present the first comprehensive review of the DIR landscape, organized around a novel two-axis taxonomy that disentangles challenges along a Data Axis (target distribution shift, continuity, and density) and a Deep-Learning Axis (shared capacity, biased updates, and manifold distortion), where the latter captures a cascading failure mechanism through which deep models systematically neglect underrepresented targets. Within this framework, we systematically categorize and analyze 19 state-of-the-art methods spanning architectural, algorithm-level, and representation learning approaches, and empirically re-evaluate twelve of them with publicly available implementations under controlled, identical conditions. To stress-test generalization across the full target range, we introduce three novel targeted evaluation protocols, Balanced Extrapolation , Bimodal Interpolation , and Blind-Spot Isolation , that expose failure modes hidden by standard benchmarks ( https://github.com/noah-puetz/deconstructing_deep_imbalanced_regression ). Our study underscores the significant impact of imbalance on regression accuracy, offering a conceptual framework and practical benchmarks to catalyze further development of systems capable of capturing the rare as reliably as the common.

Artificial intelligence (AI)-driven molecular property prediction holds significant potential to accelerate drug discovery, yet the development of robust models is hindered by scarce, high-quality data and the diversity of prediction tasks. Although self-supervised learning (SSL), especially contrastive learning, has gained traction for molecular representation learning (MRL), the intrinsic structural integrity of molecules presents a unique challenge: it obstructs the straightforward creation of meaningful contrastive pairs. This often leads to suboptimal pretraining representations and, consequently, diminished downstream task performance. To overcome this limitation, we introduce a novel contrastive pair construction strategy based on molecular fragment contributions. Our method enables the learning of a higher-quality embedding space by utilizing information bottleneck theory to evaluate the importance of individual fragments for molecular properties─without relying on external prior knowledge. We implement a contrastive learning framework enhanced with an improved quadruplet loss that more effectively captures fine-grained molecular similarities. Empirical evaluations demonstrate that our approach achieves outstanding performance on the MoleculeNet benchmark and delivers promising results in predicting diverse pharmacokinetic (PK) and critical toxicity properties, highlighting its potential for real-world drug discovery applications.

In intensive care units, large-scale clinical time-series data are continuously accumulated through electronic medical records and bedside monitoring systems. However, direct utilization of such data for clinical decision-making remains challenging due to irregular sampling, pervasive missingness, unstructured diagnostic information, and incomplete ICD labeling. Automated ICD coding constitutes an extreme multi-class classification problem with thousands of long-tailed categories, while intervention prediction tasks, such as mechanical ventilation management, involve rare transition events and severe class imbalance. To address these challenges, we propose CAGE, a hierarchical Clinical Decision Support System framework that integrates diagnosis, time-series signals, and intervention prediction. The framework first infers admission-level diagnostic context using a partial-label Automated ICD Coding module that combines DCNv2 with an Adaptive CLPL loss, producing probability-weighted diagnostic embeddings. These embeddings are subsequently fused with ICU time-series tensors and processed by a multi-branch Temporal Convolutional Network equipped with an ICD-conditioned gating mechanism to predict future ventilation state transitions. The experimental results demonstrate that DCNv2 achieves consistent superiority across all hit@k and probability concentration metrics for ICD coding. For intervention prediction, the proposed method substantially outperforms existing baselines, achieving a Macro-AUC of 98.2, Macro-AUPRC of 77.4, and F1-score of 79.4. These findings indicate that reinjecting diagnostic context as a conditioning variable, together with imbalance-aware loss design, effectively enhances rare-event detection and improves the practical applicability of clinical decision support systems.

Drug repositioning (DR) identifies new therapeutic uses for approved drugs, reducing development burdens and offering safer treatment options for patients. While high-throughput technologies generate complex, large-scale multiomics data, existing DR tools struggle to comprehensively analyze the resulting biological networks. To address this challenge, we present DRHIN, an integrated, interactive web server for DR over heterogeneous information networks (HINs) using advanced deep learning techniques. DRHIN integrates transcriptomics, proteomics, and microbiome data, incorporating eight biological entities and 19 association types to build diverse HINs and elucidate the underlying molecular mechanisms. It includes 19 state-of-the-art graph representation algorithms, enabling flexible training, comparison, and evaluation of heterogeneous network data. The platform provides a code-free portal supporting three key predictive tasks: discovering drug-disease associations, repurposing existing drugs for new indications, and identifying potential therapies for specific diseases, making analyses accessible and reproducible. Leveraging high-performance computing, DRHIN efficiently processes million-scale networks, ensuring practical applicability in real-world scenarios. The web server is freely accessible at http://drhin.tianshanzw.cn.