Introduction: Emotional regulation and regular physical activity play a critical role in shaping psychological well-being, stress management, and academic functioning among vocational college students. Despite growing recognition of these associations, empirical evidence derived from structured, intervention-based designs in vocational education settings remains limited and requires systematic investigation. Methods: A quasi-experimental pre–post study was conducted among 342 vocational college students in Jiangsu. Baseline comparisons were made between physically active and inactive students. Inactive participants then completed an 8-week supervised physical activity intervention. Data analytics techniques, including statistical modelling and mediation–moderation analysis, were applied to evaluate the impact of physical activity interventions on psychological well-being, contributing to data-driven performance improvement in healthcare-related educational settings. Results: Physically active students demonstrated significantly higher emotional regulation, greater life satisfaction, and lower stress compared to inactive peers (p <0.01). Following the intervention, previously inactive students showed significant reductions in stress (p <0.001) and significant changes in emotional regulation (p <0.001). Mediation analysis indicated that emotional regulation partially mediated the relationship between physical activity and psychological outcomes. Discussion: The findings indicate that physical activity improves psychological well-being by enhancing emotion regulation processes, with stronger effects among students experiencing greater academic pressure and stress. Conclusion: Structured physical activity interventions demonstrate potential to strengthen emotional regulation abilities and significantly reduce perceived stress among vocational college students. These findings support the integration of supervised exercise programs within higher education systems as evidence-based strategies to promote sustainable mental health and academic resilience.
Background: The exponential growth of textual data on social media and information networks poses a significant challenge to extracting valuable information. Text classification, a core task in Natural Language Processing (NLP), is essential for organizing and categorizing such data. Deep learning has emerged as an effective approach to overcoming the limitations of traditional machine learning methods in this field. Methods: This review describes the fundamental workflow of deep learning-based text classification, including text preprocessing, feature extraction, and classifier design. It also presents a detailed overview of key technologies and conducts a comparative analysis of mainstream models using commonly used datasets to evaluate their performance. Results: The analysis indicates that different model architectures generally exhibit scenariodependent characteristics. Pre-trained Language Models (PLMs) often show strong performance potential in short-text and context-sparse scenarios, whereas hybrid deep learning models may offer advantages in long-document modeling. Meanwhile, Graph Neural Networks (GNNs) demonstrate unique application value in processing structured text. Discussion: Despite considerable progress, challenges remain in domain adaptation, few-shot learning, and model interpretability. Dependence on large-scale labeled data, together with the computational cost of complex models, is also regarded as a limitation in practical deployment. Conclusion: Deep learning has profoundly impacted text classification. Future research should focus on transfer learning, data augmentation, and model compression to develop more efficient, adaptable, and robust classification systems.
Introduction/Objective: Artificial intelligence (AI) has revolutionized all aspects of life, including healthcare. Therefore, this study was conducted to evaluate the use of AI to analyze data and clarify its effectiveness in determining the predictive value of biomarkers for the diagnosis and management of psoriasis. Methodology: AI was applied using DeepSeek to analyze data from patients with psoriasis, myocardial infarction (MI), and a control group. A descriptive case-control study was conducted from June 1 to August 1, 2025. The data used were gathered from January 2013 to January 2014. The study population included individuals with psoriasis, those with myocardial infarction, or healthy subjects as matched controls, aged 18 to 50 years. The psoriasis group included 100 patients (46 female, 54 male) with a mean age of 43.56 years for females and 37.52 years for males. The myocardial infarction group included 72 patients (32 female, 40 male) with a mean age of 49.63 years for females and 54.40 years for males. The control group included 60 subjects (27 female, 33 male) with a mean age of 37.22 years for females and 42.63 years for males. Results: AI was effective in performing data analysis for patients with psoriasis and MI and provided predictive differentiation between patients and controls. A combination of biomarkers in the analysis was more predictive than a single biomarker for differentiating between groups, predicting cardiovascular disease (CVD) development in psoriatic patients, and monitoring treatment response in psoriasis. AI methods showed a 3–5% absolute improvement over logistic regression. The area under the curve (AUC) showed a 3–4% improvement in discriminative ability between psoriasis, MI, and control. AI showed a 4–6% improvement in detecting true positives, a 2–3% improvement in avoiding false positives, and reclassified 20–25% of cases more appropriately. For psoriasis biomarker analysis, random forest was used for the best balance of performance and interpretability, while logistic regression was used when regulatory approval or clinician trust was paramount, and XGBoost was used when pursuing maximum possible accuracy. Thus, the AI methods provide statistically significant and clinically meaningful improvements over conventional statistical methods, but the choice depends on the specific clinical context and constraints. Discussion: As previously reported using the same data, AI indicated that HDL, TNF-α, OPN, and IL-18 may be used to derive equations with the potential to aid in the diagnosis of psoriasis and MI; treatment monitoring; follow-up; and the assessment of the relative risk of CVD development in patients with psoriasis. Thus, metabolic and immunological abnormalities in patients with psoriasis may predispose them to an increased risk of other inflammatory diseases. Conclusion: AI was an effective tool for data analysis, contributing to increased accuracy in the diagnosis, treatment monitoring, follow-up, and assessment of comorbidity risk in patients with psoriasis. TNF-α demonstrated higher sensitivity and specificity for psoriasis, while IL-18 was more sensitive and specific for MI. HDL was beneficial as a marker for monitoring CVD risk in psoriasis. AI data analysis was superior to manual statistical analysis in providing deep insights.
Introduction/Objective: Glaucoma is a leading cause of visual impairment, characterized by progressive damage to the optic nerve. Early detection and timely intervention are critical for preventing irreversible vision loss. Recent advancements in Machine Learning (ML) and Deep Learning (DL) have significantly transformed medical imaging by enabling automated identification and diagnosis of glaucoma. The objective is to present a comprehensive analysis of ML and DL approaches for glaucoma detection using retinal fundus images and related ophthalmic data. Methods: This research analyzes state-of-the-art ML and DL models applied to glaucoma diagnosis. Convolutional Neural Networks (CNNs), transfer learning frameworks, and hybrid models are examined for detecting glaucoma using retinal fundus images, Optical Coherence Tomography (OCT), and visual field data. The study also reviews commonly used image segmentation techniques, evaluation metrics, and publicly available retinal image datasets used in recent studies. Results: We identified 1,111 records from Medline, Web of Science, and Embase; duplicate studies (i.e., 370) were removed, and we subsequently screened the remaining 741 unique studies. After assessing full-text articles, a total of 30 studies were included for qualitative synthesis based on their eligibility criteria. There was insufficient similarity among the studies regarding study design, architecture, and evaluation protocol to allow for a quantitative metaanalysis. All studies employed Convolutional Neural Networks (CNNs) that utilized transfer learning, achieving average accuracies of 85%–99% and AUCs of 0.90-0.99 across established benchmark datasets (RIM-ONE, DRISHTI-GS1, ORIGA, ACRIMA, LAG). The most prevalent imaging modality was fundus photography; however, combination imaging via multiple modalities (i.e., OCT and visual field) yielded improved performance. Emerging technologies, including Vision Transformers (ViTs) and Self-Supervised Learning (SSL), were identified, but there was insufficient standardized benchmarking in existing databases to allow for quantitative synthesis. Discussion: Despite promising results, several challenges remain, including data imbalance, limited availability of large annotated datasets, model interpretability, generalizability across diverse populations, and the translation of research outcomes into clinical practice. Conclusion: ML and DL techniques show strong potential for automated glaucoma detection from retinal imaging data. Future research should focus on larger datasets, improved model transparency, and clinically validated frameworks to enhance the reliability, robustness, and practical adoption of AI-based glaucoma diagnostic systems.
Introduction: In this study, a new method for comparing protein sequences, called the Minimal Moment Vector (MiniVec), is presented; it operates in the frequency domain using binary representations. Methods: Each protein sequence is broken down into 20 binary sequences, each representing one of the 20 standard amino acids. The Fast Fourier Transform (FFT) converts these binary vectors from a time-domain representation to a frequency-domain representation. The power spectrum for each component is normalized to a unit sum, and the descriptor is constructed as a 20-dimensional vector derived from the minimized second-order moments of these spectra. A Euclidean distance matrix is computed, and phylogenetic trees are generated using the UPGMA algorithm and visualized in iTOL. Result: The phylogenetic trees obtained from the MiniVec provide meaningful clustering across all datasets. Discussion: Qualitative analysis of the method suggests that species classified by MiniVec are as per biological classifications. Quantitative analysis shows that the proposed method performs better than other existing methods. Conclusion: The method’s accuracy is verified by testing it on four unique datasets: 9ND5 proteins, 50 Betaglobin, 19NADH, and 50 Coronavirus Spike Proteins. The qualitative evaluation and the Symmetric Distance (SD)-based quantitative comparisons demonstrate that the suggested method outperforms existing approaches and shows a high degree of consistency with existing biological classifications. The qualitative evaluation and Symmetric Distance (SD)- based quantitative comparisons demonstrate that the suggested method outperforms existing approaches and exhibits a high degree of consistency with established biological classifications.
Introduction/Objective: Emotions play a pivotal role in human experience, profoundly influencing behavior and decision-making. Identifying and understanding emotions is paramount in the field of neurological research. The objective of this study is to minimize the complexity of EEG-based emotion recognition by reducing the number of channels while achieving high recognition performance. Methods: In this study, we propose local mean decomposition (LMD)-based emotion recognition using single-channel electroencephalography (EEG). The EEG signals are decomposed into their product functions using LMD. Non-linear features are computed from each product function. The Kruskal-Wallis test is then conducted to identify significant features, ensuring the analysis focuses on the most informative elements. A decision tree classifier is employed to classify different emotion states. Results: The proposed analysis achieves a commendable average accuracy of 98.7% for EEG recordings from the F3 channel in the frontal region. Discussion: The accuracy results are further compared with other channels and reviewed studies, and it has been acknowledged that frontal neurons are best suited for emotion recognition. Conclusion: The experiment justifies the use of a focused EEG dataset acquired from the frontal region of the skull, and this study gives the scope and use of a single-channel EEG signal for emotion recognition.
Introduction: Chinese batik, as a significant intangible cultural heritage, faces critical challenges in digital preservation and innovative development. Traditional preservation methods suffer from limitations in detail capture, training instability, and scalability issues. Edge computing technologies provide unprecedented opportunities to develop distributed, realtime processing frameworks for efficient cultural heritage digitization while maintaining authenticity. Materials and Methods: This study proposes an edge computing-enabled deep learning framework integrating an improved Cycle-Consistent Generative Adversarial Network (CycleGAN) with distributed processing capabilities. The methodology incorporates: (1) U-Net generator with skip connections for fine-grained detail preservation; (2) Efficient Channel Attention (ECA) module for adaptive feature selection; (3) Wasserstein GAN loss function with VGG-19 perceptual loss for enhanced training stability. The model was trained on 344 floral images and 278 batik patterns at 256×256 pixel resolution. Results: The proposed edge computing-enhanced framework achieved: PSNR improvement of 73.95% over baseline CycleGAN, SSIM enhancement of 56.24%, and FID reduction of 62.75%. The framework successfully addresses batik digital preservation challenges while enabling real-time cultural heritage applications via edge computing integration. Discussion: Performance improvements stem from synergistic integration of architectural enhancements and edge computing optimization. The U-Net generator with skip connections maintains microscopic textures essential for authentic batik representation. The ECA module enables precise feature extraction while reducing computational overhead, suitable for edge deployment. Conclusion: This research demonstrates successful integration of advanced deep learning techniques with edge computing for preserving Chinese intangible cultural heritage batik. The framework bridges traditional craftsmanship with modern technology, enabling sustainable digital preservation and contemporary creative applications.
Introduction: Diabetic Retinopathy (DR) is a significant cause of avoidable blindness and requires effective deep learning image analysis techniques to screen and identify medical images for computer-aided diagnosis. The work introduces the GP-Net, an improved solution of the DR detection methodology combining Generalised Efficient Layer Aggregation (GELAN) and Programmable Gradient Information (PGI). Methods: The GP-Net architecture integrates gradient path planning into CSPNet and ELAN to improve efficiency and PGI; however, it includes an auxiliary reversible branch that provides the architecture with reliable gradients during training. The model was trained on the basis of 25 epochs on the Kaggle DR dataset, which was reduced to a binary classification problem, and its results were compared to those of YOLOv8x, YOLOv8s-seg, and Mask R-CNN. Results: Our proposed model was both 90.0 per cent precise and 88.0 per cent recall, and a balanced F1-score of 90.5. The result is better than the YOLOv8x baseline and than those of the segmentation-based models, such as YOLOv9s-seg and Mask R-CNN architectures. Moreover, it has a high efficiency of 308ms inference speed, thus it can be deployed in practice. Discussion: The integration of GELAN optimises feature extraction, and PGI minimises gradient loss. These results indicate that GP-Net can be successfully implemented in low-resource clinical areas, and it provides a scalable image of screening retinopathy in diabetics at an early stage. Conclusion: For DR detection, GP-Net provides an effective and validated architectural enhancement of YOLO. It demonstrates a superior balance of precision and recall, with a precision of 93.1% and efficient inference (308), making it suitable for medical practitioners.
Introduction: Nanotechnology and Artificial Intelligence (AI) have separately been revolutionizing the world of scientific and industrial innovation. The combination of both has resulted in fast material discovery, improved diagnostic abilities, intelligent therapeutic systems, and advanced computational systems. However, the current literature has presented a gap in structured evaluations of the impact of AI methodologies on nanoscale research and the role of nanoscale materials in the development of next-generation AI platforms Methods: The structured review approach was adopted in this study to identify peer-reviewed articles from prominent scientific databases (Scopus, Web of Science, IEEE Xplore, PubMed) between 2015 and 2025. A pre-defined search strategy and inclusion-exclusion criteria were used. The articles were screened to evaluate the comparison of AI models, nanomaterial applications, and performance trends. Results: The review shows that AI methods, such as Convolutional Neural Networks, Graph Neural Networks, and Transformers, greatly improve nanoscale design, prediction, and characterization. AI-assisted nanotechnology advances the accuracy of simulations, minimizes the number of experiments, and facilitates optimal drug delivery routes, biosensing, and material properties prediction. On the other hand, nanoscale materials advance energy-efficient AI hardware, neuromorphic computing, and high-performance computing infrastructure. Discussion: The results clearly show that there is a mutual technological front where AI advances nanotechnology and nanoscale materials advance energy-efficient AI hardware. The current limitations include a lack of data, a lack of interpretability of AI models, a lack of benchmarking standards, and the need for ethical and regulatory frameworks. Conclusion: The field of AI and nanotechnology is a rapidly growing interdisciplinary area with great potential for scientific and societal impact.
Introduction: Artificial intelligence (AI) is increasingly being explored in clinical decision support, creating a need for diagnostic models that are not only accurate but also transparent, reproducible, and clinically interpretable. This study presents UniBlendPredictor, a stability-aware explainable hybrid ensemble framework for breast cancer diagnosis using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset. Methods: The proposed framework extends the hybrid UniBlend architecture by combining stacking, soft voting, and hard voting within a unified ensemble design. Explainability was incorporated using SHAP and LIME to provide global and local interpretation of model predictions. In addition, bootstrap-based confidence intervals and stability-focused explanation analysis were used to assess the reproducibility of model outputs. A deterministic rule-based clinical decision-support layer was included to translate model predictions into structured recommendations, while a prototype Gemini-enabled explanation layer was implemented to provide clinician-oriented natural-language explanations without participating in model training or predictive inference. Results: The optimal UniBlendPredictor achieved an accuracy of 99.12% and an AUC of 0.9977 on the WDBC test set, outperforming the individual ensemble components while maintaining strong explanatory consistency. The framework identified clinically relevant morphological features as dominant predictors of malignancy and demonstrated stable agreement between SHAP- and LIME-based explanations. Discussion: The findings indicate that integrating explainability, uncertainty awareness, and deterministic clinical decision-support logic can improve the transparency and interpretability of hybrid ensemble predictions without compromising diagnostic performance. Although Uni- BlendPredictor showed strong performance, direct comparison with previously published models should be interpreted cautiously, as differences in preprocessing, feature engineering, validation design, and experimental setup can influence reported results even when the same dataset is used. Conclusion: UniBlendPredictor shows that a hybrid ensemble framework can support highperformance breast cancer diagnosis while also improving interpretability, reproducibility, and decision transparency. The study provides a foundation for future validation on broader datasets and for further development of clinician-oriented explanation workflows.
Introduction: The electoral system ensures that the administrative machinery functions smoothly at all levels, including the panchayat, municipal, state, and central levels. However, the conventional electoral system suffers from security limitations, including electoral fraud, vote manipulation, and ballot stuffing. The major objective of this research is to address issues with the conventional electoral system and to provide a secure voting environment for the voting community. Methods: This study proposes a Cloud Computing-based Voting System (CVS) with an integrated interface. The primary focus of the proposed system is voter authentication, which is an important component. The authors have implemented a hybrid security system for voter authentication that leverages multimodal biometrics, image processing, steganography, and encryption techniques across three biometric image types: palmprint, face, and fingerprint. Results: The research findings suggest that the biometric image quality improved after applying hybrid security techniques. This is indicated by the Peak Signal-to-Noise Ratio (PSNR) values of 51.1769 dB for the palmprint image, 12.1418 dB for the facial image, and 4.2350 dB for the fingerprint image. For an m×m image with n-bit pixels, the total time complexity is O(m2n). Discussion: The proposed hybrid security method improves overall biometric image quality and provides a secure way to store and communicate biometric image data. Conclusion: For secure biometric data transmission and to provide an integrated interface for voting, a secure CVS has been suggested. This enhances the overall electoral process and supports efficient electronic governance service delivery to citizens.
Introduction: The combination of Artificial Intelligence (AI) and data analytics is used to transform the sphere of healthcare by moving away from the approach of reactive risk management to proactive risk management. Predictive models have the potential to detect risks earlier, maximize interventions, and enhance clinical outcomes with the increasing access to Electronic Health Records, EHRs, and wearable devices (WD), and real-time patient monitoring. Nevertheless, issues like data privacy, algorithmic bias, and interoperability continue to be a serious obstacle to adoption. This paper explores the creation and use of computational models and machine learning algorithms of Logistic Regression, Random Forest, and Neural Networks as predictive healthcare analytics. The objective is to assess their usefulness in predicting chronic diseases, reducing costs, and early intervention, and to address ethical and technical issues. Methods: Heterogeneous data, such as EHRs, wearable sensors, and medical imaging, were preprocessed by standard cleaning and feature engineering methods and anonymization. Crossvalidation and performance measures, including accuracy, sensitivity, and specificity, were used to predict the development and validation of the predictive models. A comparative study of Logistic Regression, Random Forest, and Neural Networks was conducted to evaluate predictive performance. The concept of deployment was taken into consideration concerning healthcare IoT and communication systems. Results: Neural Networks were identified to be the most accurate predictor of chronic diseases (92%), compared to Random Forest (90%) and Logistic Regression (85%). The application of AI-based predictive analytics has led to a decline in hospital readmission rates by 25 percent, patient care expenses by 18 percent, and a 40 percent rise in the rate of early interventions. Optimization of resources also led to a decrease in the average hospital stay by 22.6 and the minimization of medication errors by 66.7. These findings indicate the huge potential of AI in improving healthcare outcomes and efficiency. Discussion: However, the findings suggest that AI and predictive analytics are used to change the potential to shift the healthcare model with proactive care. Successful implementation requires addressing challenges like data interoperability, algorithmic bias, and ethical governance. Neural Networks offer high accuracy but lack interpretability, while Logistic Regression (LR) and Random Forest (RF) strike a balance between interpretability and effectiveness. The paper highlights the significance of hybrid solutions between explainability and performance to be effectively deployed into clinical practice. Conclusion: Artificial intelligence and data analytics have the power to make healthcare a proactive and patient-centered ecosystem, which facilitates early diagnosis, saves money, and streamlines resource distribution. Logistic Regression is interpretable, Random Forest is robust, and Neural Networks are maximally accurate. Although the results are promising, to achieve successful adoption, it is necessary to resolve the problem of data privacy, interoperability, and algorithmic fairness. This paper shows that AI-based predictive modeling, in conjunction with healthcare communication systems, can enhance risk management and clinical decision-making to a large extent.
Background: Cloud computing is now the most popular way for businesses to run workflow apps. But scheduling tasks and allocating resources efficiently remains hard due to factors such as task dependencies, system heterogeneity, and high computational demands. This paper proposes a scheduling framework, Federated Learning-enabled Deep Q-Learning (FedDQL), to address these problems. The goal is to improve performance, use less energy, and ensure data is handled safely while making the most of available resources. A real-world dataset used for a comparative analysis shows that the proposed method outperforms existing methods. Objective: To develop a secure, energy-efficient, and high-performing task scheduling framework for cloud environments that improves Quality of Service (QoS) through optimized resource allocation and monitoring. Methods: The FedDQL framework is based on the idea of combining Federated Learning (FL) and Deep Q-Learning. People can learn this way without having to give up their data. The Enhanced Multi-Verse Optimization (EMVO) algorithm finds the optimal scheduling settings. For effective resource monitoring, the framework incorporates the Jordan Normal Form-Deep Kronecker Neural Network (JNF-DKNN). Additionally, the Chebyshev Distance-based Fuzzy Self-Defense Algorithm (CD-FSDA) is employed for dynamic virtual machine selection and monitoring. This holistic approach ensures secure, intelligent, and adaptive scheduling within the cloud environment. Results: Experimental evaluation shows that the proposed FedDQL framework outperforms existing approaches, including DQ-HEFT, DDQNEC, DRQL, ANN, and DQL, in terms of execution time, energy consumption, and overall system performance. Discussion: This study presents a DQL-based scheduling framework augmented with federated learning to ensure secure and energy-efficient workflow scheduling in cloud computing. The system enhances resource allocation, minimizes expenses, makespan, and energy consumption, while improving accuracy by incorporating algorithms such as WHA, FedDQLEMVO, JNF-DKNN, and CD-FSDA. Performance assessment using empirical datasets demonstrates exceptional outcomes. Conclusion: This research introduces a FedDQL-based energy-efficient and secure workflow scheduling framework integrating federated learning for privacy protection and multialgorithm optimization (WHA, FedDQL-EMVO, JNF-DKNN, CD-FSDA). The system enhances task scheduling accuracy, reduces energy use, cost, and imbalance, and outperforms existing methods. However, its reliance on simulations, predefined workflows, and simplified energy modeling limits real-world applicability. Future work will emphasize real-world validation, adaptive optimization, and thermal- and SLA-aware scheduling in heterogeneous cloud environments.
Introduction: Predicting high-frequency intraday stock prices remains challenging due to volatility, noise, and nonstationary market behavior. To address these issues, this study proposes a hybrid framework integrating K-Means clustering with a dual-stage Long ShortTerm Memory (LSTM) network for regime-specific learning and adaptive forecasting. The proposed Cluster-Based Deep Sequential LSTM (CB-DS-LSTM) model segments market regimes dynamically and trains specialized LSTM models for each cluster, enhancing prediction accuracy, adaptability, and robustness under varying market conditions. Methods: The proposed Cluster-Based Deep Sequential LSTM (CB-DS-LSTM) framework segments historical stock data into distinct clusters representing different market regimes using K-means clustering. For each cluster, dedicated dual-stage LSTM models are trained to capture both short-term fluctuations and long-term trends. The model leverages 5-minute interval intraday stock data from the Indian stock market, incorporating multiple technical indicators such as moving averages, Relative Strength Index (RSI), and MACD to enrich the input features. This architecture enhances the model’s temporal pattern recognition capabilities and allows adaptation to varying market conditions. Results: Empirical evaluation shows that CB-DS-LSTM significantly outperforms conventional machine learning and standard deep learning models. It achieves a Root Mean Square Error (RMSE) of 0.6148 and an R² score of 0.9975, demonstrating superior predictive accuracy. The clustering-based segmentation enables specialized modelling of market regimes, leading to improved robustness under volatile market conditions. Discussion: This hybrid approach effectively addresses the limitations of generic LSTM models by tailoring predictions to cluster-specific market behaviors. The dual-stage LSTM captures multi-scale temporal dependencies, while integration of technical indicators provides enriched contextual information. Conclusion: CB-DS-LSTM offers a robust, adaptive framework for high-frequency stock price prediction, with significant potential to optimize trading strategies in dynamic financial markets.
Introduction: In recent years, new medical image segmentation methods have been constantly emerging, and their segmentation effects have been significantly improved compared with traditional methods. Medical image segmentation methods based on deep learning have also achieved many excellent advancements. However, they still face problems such as difficulties in multi-scale feature fusion, blurred organ boundaries, and the absence of small targets. Methods: This paper proposes an MDFNet model based on multi-scale dynamic fusion and a deformable cross-attention mechanism. This model takes PVTv2 as the backbone network and effectively overcomes the limitations of traditional one-way feature fusion by introducing deformable convolution and bidirectional cross-attention mechanisms, achieving the collaborative expression of local details and global semantics. results: The experimental results show that the model has a good segmentation effect on seven medical datasets such as ClinicDB. The Dice coefficient on Kvasir-SEG reached 91.8%, which was 10% higher than that of U-Net, and the number of parameters was only 21.5 million. Results: The experimental results show that the model has a good segmentation effect on five medical datasets, such as ClinicDB. The Dice coefficient of Kvasir-SEG reaches 91.8%, which is 10 percentage points higher than that of U-Net, and the number of parameters is only 27 million. Discussion: To verify the effectiveness and accuracy, experiments on five popularly used datasets are carefully designed and implemented. Meanwhile, the performance of the proposed method is compared with typical segmentation models, such as the other seven models, like UNet and SANet. conclusion: This paper proposes a new medical image segmentation model, MDFNet. This model effectively solves the scale sensitivity problem in medical image segmentation through multi-level dynamic feature interaction and spatio-temporal attention mechanism, providing an effective solution for clinical tasks such as digestive system lesion detection and tumor edge delineation. Conclusion: This paper proposes a new medical image segmentation model, MDFNet. This model effectively solves the scale sensitivity problem in medical image segmentation through multi-level dynamic feature interaction and a spatio-temporal attention mechanism, providing an effective solution for clinical tasks such as tumor edge delineation.
Introduction: Diabetic Retinopathy (DR) is a severe outcome of diabetes that may lead to vision impairment. Manual screening of DR is time-demanding, subjective, and less accessible in resource-limited areas. Existing deep learning (DL) models outperform in classifying between the No-DR and Severe-DR classes. However, several existing methods report reduced sensitivity in detecting mild and moderate DR stages due to biased training using publicly available imbalanced datasets. Methods: This study introduces DReXDetect, a hybrid computer-aided diagnostic (CAD) framework driven by Explainable Artificial Intelligence (XAI). The proposed model exclusively hybridizes established transfer learning (TRL) models, ResNet-50 and Inception-V3, for robust DR severity classification. The incorporation of XAI analysis, including Gradientweighted Class Activation Mapping (Grad-CAM), Local Interpretable Model-Agnostic Explanations (LIME), and Shapley Additive Explanations (SHAP), offers visual and feature-level insights into model decision-making, potentially fostering trust in clinical environments. Results: Experimental results contribute towards the development of a clinically trustworthy XAI-assisted CAD system, with particular emphasis on grade-wise reliability and interpretability across different DR severity stages. Discussion: The performance of DReXDetect is evaluated against various cutting-edge TRL architectures, including standalone DenseNet-121, Inception-V3, and Xception. ResNet-50 is considered the baseline model because of its reliable residual learning. DenseNet-121 and Xception are excluded from the fusion process to prevent feature redundancy and to enhance discriminatory information; instead, ResNet-50 is hybridized with Inception-V3 to extract multi- scale, diverse, and discriminative deep features. Conclusion: This paper presents a systematic, interpretable, and hybrid feature-learning architecture within a unified, transparent evaluation framework for five-stage DR classification, thereby emphasizing robustness, transparency, and class-wise reliability.
Introduction: Clinical decision support systems backed by machine learning approaches are widely deployed at primary care centers for the prediction of diseases. New findings for the ongoing study in the field of clinical informatics, incorporating biomarker data, have recently focused on studies that validate the conclusions made from the biomarker data. The intersection of biomarkers and asthma control is a rapidly advancing field that holds promise for improving personalized care. Materials and Methods: The work presented in the paper proposes the implementation of a machine learning framework based on regularization techniques for the assessment of various asthma control levels by discovering the distinctive data via an embedded approach. The approach is seen to effectively identify the best of the features that mark the cytokine and cellular profile of subjects to identify the predictive biomarkers. results: The performance evaluation results of the various classifiers using the predominant cytokine combination identified in the reduced set yielded a classification accuracy of 80% when compared to 64% classification accuracy that was obtained with when the entire cytokine profile was used without involving any feature selection technique Results: The performance evaluation results of the various classifiers using the predominant cytokine combination identified in the reduced set yielded a classification accuracy of 80%, compared to 64% obtained when the entire cytokine profile was used without any feature selection technique. Discussion: A differential classification model to effectively differentiate the two categories of controlled and uncontrolled asthmatics was implemented and validated by identifying the predominant biomarkers from the available data panel. Conclusion: The performance evaluation results of the various classifiers using the predominant cytokine combination identified in the reduced set were carried out. An optimal mixing of both L1 and L2 regularizations was chosen by an empirical means for the elastic net. The ReliefF- based regularization technique used for predicting asthma control levels performed well with almost all regularization methods and performed extremely well with elastic net, yielding a lower error rate.
Introduction: In the context of digital education, integrating heterogeneous learning- related data sources remains a significant challenge due to schema variability, semantic mismatches, and platform diversity. This study aims to systematically evaluate current research on data integration within e-learning environments, focusing on identifying challenges, effective techniques, performance metrics, and future directions. Methods: A Systematic Literature Review (SLR) was conducted using Kitchenham’s guidelines, targeting empirical studies published between 2015 and 2024. The review applied rigorous inclusion, exclusion, and quality assessment criteria, resulting in 142 high-quality studies. Extracted data covered integration techniques, evaluation metrics, and educational contexts such as Learning Management Systems (LMS) and Student Information Systems (SIS). Results: Semantic modeling, ontology-based integration, and service-oriented architectures (SOAs) emerged as dominant techniques. Performance evaluations frequently relied on metrics such as accuracy, scalability, interoperability, and semantic completeness. Persistent challenges include real-time data handling, adaptability, and achieving semantic interoperability in largescale educational systems. Discussion: The review highlights a growing shift toward AI-driven and semantic approaches, reflecting the need for scalable and intelligent integration. However, the lack of standardized evaluation metrics and limited real-world deployment restrict generalizability. There is also a need for solutions addressing both technical and user-centric factors, including personalization and quality of experience. Conclusion: This review synthesizes current research on educational data integration, outlining key methods, challenges, and research gaps. Future efforts should prioritize real-time, adaptive frameworks supported by standardized evaluation protocols and validated through practical implementation in diverse educational settings.
Introduction: The working condition identification of an industrial system (WCIIS) is important for optimizing system operation in real time, thereby enhancing production efficiency. However, existing WCIIS methods face challenges such as cross-domain adaptation and data missing. Methods: To address these issues, we propose a novel domain adversarial transfer network (DAITN). Specifically, we propose a self-attention generation-adversarial multi-discriminator dual-imputation module (SGMDM) to address the problem of data missing in WCIIS. Furthermore, two asymmetric encoder networks are designed to learn hierarchical representations from both the source and target domains. The network parameters learned from the source domain are utilized to initialize the parameters trained by the target domain. Additionally, domain adversarial training with a loss function is employed to handle distribution drift between the source and target domains. Results: The performance of SGMDM is evaluated using the anode current signal (ACS), the Tennessee Eastman Process (TEP), and the Continuous Stirred Tank Heater (CSTH). The validation of the DAITN is conducted on ACS and TEP, and the accuracies on ACS and TEP are 99.17+0.69, 99.11±0.76, respectively. Discussion: The DAITN model works well because it combines random forests and a selfattention GAN to fill in loss data accurately, and a domain-adversarial network with two CNNs that share knowledge from a source to a target domain, helping the model generalize better across different datasets. Conclusion: Experimental results demonstrate that the proposed method effectively addresses cross-domain WCIIS challenges in the presence of data missing.
In the originally published article [1], certain phrases and expressions were unclear in the abstract and introduction section, which may have affected readability. These have now been revised to improve clarity and ensure that the intended meaning is accurately conveyed. The corrections do not affect the results, interpretations, or conclusions of the article. The authors apologize for any inconvenience caused to the readers. The original article can be found online at: https://www.eurekaselect.com/public/article/144170 Details of the error and correction are provided here. Original: Abstract: Wireless networks are essential communication technologies that prevent cable installation prices and burdens. Because of this technology's pervasive usage, wireless network safety is a significant problem. INTRODUCTION This could be shared in some groups as network safety to protect a network of computers against intruders; app safety, which has a target in protecting software and devices against threats; info safety guarantees operational security, confidentiality as well and data integrity that contains processes and decisions associated with data processing and protection as well others like disaster recovery and user training [8]. Corrected: Abstract: Wireless networks are essential communication technologies that prevent cable installation prices and burdens. Because of this technology's pervasive usage, wireless network security is a significant problem. INTRODUCTION This could be shared in some groups as network security to protect a network of computers against intruders; app safety, which has a target in protecting software and devices against threats; info safety guarantees operational security, confidentiality as well and data integrity that contains processes and decisions associated with data processing and protection as well others like disaster recovery and user training [8].
