This review provides detailed information on the current advancements in artificial intelligence in sports scenarios, including injury prediction, motion detection, and performance optimization. The AI diagnostic applications showcase ML and DL techniques, such as RFs, SVM, CNN, and ANN, which are all applied to forecasting injuries in various sports by integrating different data provided by individuals. Several sensor-based technologies emphasize the importance of inertial measurement units and sensor fusion in motion detection across a wide range of human joints. Meanwhile, in the sport of hammer throw, real-time biomechanical feedback is developed, aiming for more precise and accurate results. Generally, AI not only benefits athletes safely and thoroughly, but also promotes the sport medicine field to take a step forward by offering data and more research. The future of AI is optimistic and promising in the field of sports medicine, and the review has already identified ways to tackle several challenges that remain unresolved in the current situation.

In recent years, the popularity of smartphones and smartwatches, along with advancements in IoT and artificial intelligence (AI) technologies, has driven the widespread adoption of wearable sensors. Societally, the aging population has intensified the complexities of chronic diseases, home safety, and mental health issues. Wearable sensors enable home-based health monitoring and alleviate pressure on medical resources. This paper focuses on three sensing technologies: multimodal flexible sensor patches, graphene patches, and spectroscopic sensors. Multimodal flexible sensor patches monitor vital signs such as electrocardiograms (ECG), respiration, and skin temperature; graphene patches detect pulse waves via high-sensitivity pressure sensors to assess cardiovascular health; and spectroscopic sensors utilize optical signals for health monitoring and disease diagnosis. The paper also discusses data processing techniques, highlighting challenges such as sensor accuracy, stability, response time, autonomy, and follow-up response capabilities. Additionally, it explores the future feasibility of hybrid sensor systems combining wearable health sensors with robotics, AI, and smart home technologies.

This paper presents FedSC-MTL, a unified framework for multi-task federated learning (MTL) that incorporates dynamic regularization and dual control stabilization to address key challenges such as data heterogeneity, task divergence, and knowledge misalignment. While existing methods like FedProx and SCAFFOLD focus on data heterogeneity and client drift, they fail to manage task-level goals and knowledge transfer across clients effectively. To overcome this, FedSC-MTL introduces a novel dual control mechanism: the control variable 'c' stabilizes parameter updates and mitigates client drift, while the feature-level distillation control 'd' enables efficient knowledge transfer. By focusing on feature-based distillation, FedSC-MTL reduces communication costs compared to logits-level distillation methods like FedICT and FedDyn, without sacrificing task-level generalization. The effectiveness of this approach is evaluated on multi-task federated benchmarks, including EMNIST, CIFAR-10, and Fashion-MNIST datasets. Experimental results show that FedSC-MTL outperforms existing methods in terms of accuracy, convergence speed, and communication efficiency. Specifically, the feature-level distillation mechanism significantly improves cross-task generalization, and the dual control stabilization enhances the overall stability and convergence of the model. These findings demonstrate the potential of FedSC-MTL as an effective solution for multi-task federated learning in heterogeneous environments, offering both scalability and efficiency.

Deep learning-based autonomous vehicles (AVs) demonstrate significant potential in reducing traffic accident rates and enhancing transportation efficiency. However, the “black-box” nature of artificial intelligence systems such as deep neural networks (DNNs) has raised widespread concerns regarding the explainability, transparency, and safety of their decision-making processes. This paper focuses on the application of explainable artificial intelligence (XAI) in the autonomous driving domain as an effective approach to address current technological challenges. The paper first reviews two fundamental architectures of autonomous driving systems and provides a basic overview of X-AI technologies, explaining their significance within these systems. It then systematically outlines the classification framework of X-AI and its critical importance in high-risk domains. Through analysis of multiple cutting-edge research frameworks—including the SafeX framework for modular architectures, the XAI integration framework for end-to-end systems, and the XAI-ADS system for cybersecurity— This paper delves into how X-AI enhances the safety, regulatory compliance, and user trust of autonomous driving systems. Finally, it outlines future development directions for X-AI in autonomous driving technology and proposes research recommendations for building explainable, safe, and responsible autonomous driving systems.

This study presents a systematic empirical comparison of three leading large language models—DeepSeek, ChatGPT (GPT-4o), and Doubao—in multimodal emotion recognition tasks. Using a self-constructed dataset of 1,200 annotated text-image samples across three emotional scenarios (social gatherings, stress-induced tension, and anticipation-anxiety), the models were evaluated on overall performance, fine-grained emotion recognition, and context sensitivity. Results indicate that ChatGPT achieves the highest overall accuracy (89.5%) and demonstrates superior cross-modal reasoning and interpretability. Doubao excels in Chinese social contexts, with an F1 score of 91.5%, but shows limited cross-lingual generalization. DeepSeek performs stably in text-dominant tasks but lags in multimodal fusion scenarios. The findings highlight the context-dependent strengths of each model and provide practical guidance for model selection in real-world applications, such as global platforms, Chinese social media, and resource-constrained environments. This study addresses a critical gap in the comparative evaluation of multimodal LLMs and offers insights into future research in cross-cultural and lightweight multimodal emotion recognition.

With the development of intelligent transportation systems, traffic sign recognition plays a key role in autonomous driving and driver assistance systems. The performance of traffic sign recognition algorithms directly impacts the safety of vehicles during operation. Although traditional 32-bit floating-point convolutional neural networks (CNNs) offer the advantage of high recognition accuracy, their substantial number of parameters and computational intensity make them difficult to deploy directly on resource-constrained devices like in-vehicle FPGAs. Based on this, this paper proposes a lightweight CNN model for edge computing. At the algorithmic level, a VGG-like straight-line network structure with 5 convolutional and 4 pooling layers is designed. Employing 4-bit quantization-aware training (QAT), pseudo-quantization nodes are embedded into the forward training process, enabling the network to maintain discriminative capability within the 4-bit discrete space. At the toolchain level, an end-to-end pipeline from PyTorch → QAT → HLS → bitstream is constructed. It supports one-click weight/activation bit-width pruning and RTL generation. The final generated bitstream file can be directly applied in FPGA deployment. Experiments were conducted on the GTSRB dataset. The final quantized model achieved a test accuracy of 87%, with a model size of only 0.15 MB. The hardware was deployed on a Xilinx Zynq-7020 SoC, with resource utilization falling within an acceptable range for the chip. The results indicate that the proposed method achieves an excellent balance among accuracy, resource consumption, and power usage, offering an efficient, automated, and easily deployable hardware-friendly solution for traffic sign recognition at the edge.

As a hot field in current society, artificial intelligence text-to-image gen-eration has received extensive attention in recent years. Text-to-Image generation task refers to the process of converting natural language descriptions to corresponding visual content such as pictures and illus-trations, and has demonstrated a powerful influence in fields such as education, economic models, and artistic creation. Based on different technical frameworks, the current mainstream Text-to-Image large models can be divided into diffusion-based models, generative adver-sarial networks, variational autoencoder-based models and other methods. Different technical architectures have their own advantages and characteristics. Based on the above representative frameworks, this paper introduces some of the latest technological developments, ex-pounds its innovation direction and operation process, and analyzes its shortcomings. This paper introduces a few classic datasets such as LAION-5B and COCO, and analysis the performance of representative methods on these datasets. This paper summarizes the current prob-lems in the field of Text-to-Image, looks forward to the future develop-ment direction, and hopes to bring some inspiration to future researchers.

Bacteriophages (phages) are the “killers” and “regulators” of bacteria. In addition to being the foundation of life science research, phages are also an extremely promising alternative approach to solve the antibiotic crisis. The analysis of phage-specific structures is helpful for understanding host recognition and infection, and plays a key role in the study of viral evolution, new antibacterial applications, and biological tool development. However, the current bottlenecks in phage structure analysis are evident in two aspects: one is the gap between the static conformations extracted by structural biology methods and the dynamic functional changes of proteins in the natural infection environment; the other is that the strict requirements of advanced characterization technology for sample integrity and homogeneity, cumbersome procedures and high cost limit its characterization throughput and popularization in use. To overcome the dynamic functionality and throughput limitations of traditional structural analysis, people begin to explore computational methods: Basic Local Alignment Search Tool (BLAST) directly screens homologous proteins to infer functions; Convolutional Neural Networks (CNN) can directly predict structures and features of proteins based on their sequences or mass spectra; GNN is good at simulating protein interaction networks and their dynamic changes; and integrated platforms like DeepSP-GIN are the trend of future integration of multiple deep learning models to achieve high-throughput automatic analysis. This article compared and analyzed several commonly used data analysis methods in different situations, and summarized the applicable scenarios and effects of each method.

This article establishes a multi-stage mathematical model for the timing selection and fetal abnormality determination in NIPT testing, and comprehensively applies statistical analysis and machine learning methods to provide personalized testing strategies for different pregnant groups. Firstly, Spearman rank correlation analysis revealed a positive correlation (ρ=0.305) between Y chromosome concentration and gestational age, and a significant negative correlation (ρ=-0.534) with BMI. The conclusion was statistically significant after permutation test, Fisher z transform confidence interval estimation, and Benjamin Hochberg multiple test correction. Further validation through the Gradient Boosting Tree (GBDT) model revealed that the effects of proportion, gestational age, and BMI on Y concentration were most significant The model was cross validated with R²=01246. Although its generalization ability was limited, it provided a theoretical basis for subsequent modeling[1]. Based on K-means clustering, BMI was divided into three groups (low: 27.2-33.5, medium: 34.6-39.9, high: 41.0-44.2), and an optimization model was established with the goal of minimizing risk. Taking into account the early, middle, and late risk weights (1.0, 2.5, 5.0) and the re examination process, the optimal detection time points for each group were determined to be 11.0 weeks, 19.6 weeks, and 12.1 weeks, respectively, with an overall risk score of 119.80. By using the probability function and Monte Carlo simulation to analyze the detection error, the error tolerance of each group was determined to be±3.37 weeks,±4.02 weeks, and±3.75 weeks, respectively[2]. A follow-up strategy was proposed to control the error within±1.5 weeks, reducing the error by 55.5% -62.7%.

This paper focuses on optimising effective shielding duration within multi-target protection scenarios, proposing a solution that integrates geometric modelling with intelligent optimisation algorithms. By constructing a three-dimensional Cartesian coordinate system and establishing a geometric relationship model between targets and detection links, it precisely describes the parametric characteristics of each variable, thereby establishing distance calculation and indicator function mechanisms for shielding determination. For single-source deployment scenarios, a nonlinear programming model incorporating multiple decision variables is designed. Combined with genetic algorithms and local smoothing optimisation techniques, this achieves efficient solutions for shielding duration. For multi-source collaborative scenarios, a multi-agent collaborative optimisation framework is constructed using decision variable discretisation and exhaustive search strategies. This coordinates parameter configuration and temporal scheduling across deployment units to maximise global shielding effectiveness. Experimental results demonstrate that the proposed method significantly enhances effective masking duration, achieving an optimal single-source masking duration of 4.745 seconds and a total multi-source collaborative masking duration of approximately 23.7 seconds, thereby validating the scheme's effectiveness and engineering applicability.