Earticle

Voice pathology classification has become one of the primary objectives of research in biomedical engineering. This paper proposes PathoVoiceFAI, a technique that enhances the multiclass pathology classification by leveraging the power of attention layers and appropriate fusioning technique to fuse the multimodal inputs. The preliminary results show that use of mid-level fusion with attention layers improves the classification accuracy by 5% in comparison to the standard decision-level fusion technique. This highlights the effect of powerful feature extraction in enhancing the classification outcomes for application in clinical environment.

Bayesian Personalized Ranking (BPR) assigns ranks to a set of items to recommend them to users. This study proposes a novel approach to improve the performance of recommendation systems. The first proposed method for the enhanced recommendation system constructs positive preferences by utilizing only the items explicitly preferred by users rather than treating all interacted items (e.g., clicked, rated, or reviewed) as positive, as was traditionally done. The second method involves using explicitly non-preferred items as negative data, and the traditional approach of using only noninteracted items as negative. Message propagation is performed on subgraphs generated through meta-path design. The results of each subgraph are used to learn the representations of users and items through an attention mechanism and graph representation learning based on this data configuration. The system then predicts scores for user-item pairs. For evaluation, the performance of the recommendation system was assessed using not only traditional accuracy metrics but also by defining a pairwise ranking accuracy metric. Pairwise ranking accuracy assigns ranks to preferred and non-preferred items to determine if the model reflects user preferences. Experimental results showed improved performance in widely used evaluation metrics for recommendation systems, such as Hit Rate and Normalized Discounted Cumulative Gain, and higher performance in pairwise ranking accuracy.

This paper proposed a Feature-level fusion technique that combines facial expression and audio modalities for multimodal emotion recognition. The learning model utilizes a hybrid approach combining CNN and LSTM to learn the spatiotemporal characteristics of video and audio modalities effectively. Compared to a unimodal approach, speech emotion recognition achieved 74% accuracy, and facial emotion recognition achieved 83% accuracy, while the proposed multimodal approach achieved 93% accuracy, demonstrated that multimodal emotion recognition is more accurate than unimodal emotion recognition. Furthermore, in tests using the RAVDESS dataset, the proposed model achieved higher emotion recognition rates compared to related studies. This study demonstrated the possibility of multimodal emotion recognition and designed a model capable of recognizing emotions in various environments and situations. Through this, we aim to contribute to the advancement of emotion recognition technology.

The increasing demand for halal cosmetic products, especially in Muslim countries, has shown significant challenges for consumers seeking products that follow Islamic principles. Although various studies attempt to recommend halal status, mainly considering the discrete and specific relations within individual cosmetics, they ignore the high-order and complex relations between cosmetics and ingredients. To solve it, we propose a halal cosmetic recommendation framework that leverages a knowledge graph of cosmetics and their ingredients to recommend similar cosmetics and halal cosmetic predictions. Specifically, we construct a cosmetic knowledge graph representing the relations between various cosmetics, ingredients, and their properties. We then propose a pre-trained relational graph attention network model with residual connections identity mapping to learn the structural relation between entities in the knowledge graph. The pretrained model is then employed on downstream cosmetic data to recommend similar cosmetics and predict halal standards.

This study emphasizes the critical role that Non- Terrestrial Networks (NTN) will play in overcoming coverage limitations in remote, maritime, and aerial regions within 6G networks. NTN will maximize the potential of 6G by providing stable connectivity even in areas where traditional networks cannot reach. Both NR-NTN and IoT-NTN are essential for the expansion of the IoT (Internet of Things) and the realization of a hyper-connected society. The two payload architectures of NTN significantly impact its efficiency and service quality. Despite the remaining technical challenges in key technologies, NTN is crucial for meeting the high data rates, low latency, and enhanced mobility requirements of 6G.

Trusted Execution Environments (TEEs) are the most vital security features of contemporary computing, especially in a virtualized environment. Two popular hardware-based TEEs include Intel Software Guard Extensions (SGX) and AMD Secure Encrypted Virtualization (SEV), which respectively help protect sensitive computation from several forms of attacks. This paper investigates SGX and SEV very deeply, including their architecture, memory encryption mechanisms, and the security vulnerabilities they encounter. SGX adopts an enclave-based approach to application-level isolation, whereas SEV affords VM systemwide memory encryption. We discuss the implications of such designs in cloud computing environments and proffer recommendations that will help secure attacks emanating from side-channel and rollback.

As wireless communication demands continue to rise, Non-Orthogonal Multiple Access (NOMA) has emerged as a key technology for next-generation networks, including 5G and beyond. Unlike Orthogonal Multiple Access (OMA), NOMA enables multiple users to share the same resources by transmitting signals at different power levels, significantly improving spectral efficiency. This paper addresses the challenges of user pairing and dynamic NOMA/OMA switching in downlink NOMA systems, specifically within vehicular networks. We propose an optimization framework that jointly addresses Base Station- Vehicular User (BS-VU) association, channel assignment, and dynamic switching between OMA and NOMA to minimize power consumption. Simulation results demonstrate that our approach can effectively associate VUs to BS and dynamically switch between NOMA and OMA while meeting Quality of Service (QoS) requirements. These findings underline the viability of NOMA as a critical enabler of efficient and scalable communication in future networks.

Addressing traffic congestion in Vehicular Adhoc Networks (VANETs) is crucial for ensuring safety, social welfare, and economic progress. This study introduces a novel approach utilizing transfer learning in conjunction with the Gradient Boosting algorithm to optimize information transmission within VANETs. By leveraging pre-trained nodes as information sources, the proposed model effectively trains newly registered nodes, enhancing congestion control performance. Simulation results conducted in Python demonstrate the model's effectiveness, showcasing reduced execution times compared to traditional fuzzy logic-based methods. Integration of this model into existing congestion control systems promises real-time congestion screening capabilities. The study highlights the importance of further research collaboration to tackle realtime implementation challenges and advance traffic congestion management using AI-based techniques. Simulation results have indicated that the proposed system model achieves a performance of 95.43% accuracy. It also noted that the use of the proposed system in producing the HRA results is more accurate compared to the past methods.

This paper addresses the critical task of anomaly detection in river network sensor data, essential for accurate and continuous water quality monitoring. We propose M-MAD (Multi-Modal Anomaly Detection), a novel approach that integrates multi-modal features, including sensor data, weather information, and historical anomalies. M-MAD builds on the Graph Deviation Network (GDN) framework by introducing an improved anomaly threshold criterion derived from the learned graph structure. Our evaluation employs rigorous benchmarking simulations that mimic complex dependency structures and diverse anomalies, thoroughly assessing the strengths and weaknesses of M-MAD compared to existing methods. Results demonstrate M-MAD's superior performance in handling high-dimensional datasets and its enhanced interpretability, crucial for effective anomaly detection.

In semiconductor manufacturing, the design of devices and the selection of optimal materials are traditionally time-consuming and costly process. To address these challenges, machine learning techniques are being explored to improve simulation speed and efficiency without compromising accuracy. This study aims to optimize semiconductor manufacturing by identifying the optimal material based on slit design structure and transmittance. Traditionally, inverse design methods focused on developing slit designs from transmittance, requiring significant time and financial resources for material validation through simulations. We propose a multi-modal algorithm that combines slit design images and transmittance to predict optimal material. Additionally, we introduce a convolutional neural network that predicts transmittance from slit design image and materials. Our approach introduces a model that identifies optimal materials directly from transmittance and design structure, enhancing efficiency. This advancement allows for effective prediction and analysis of material properties in semiconductor devices through domain transformation.

The extensive use of technology and the internet in the modern digital age has enhanced our lives but also generated serious security risks, with phishing being one of the most common cybercrimes. Phishing attempts to get personal information by spoofing trustworthy websites and taking advantage of private information such as usernames, passwords, and account IDs. Researchers are using deep learning and machine learning approaches to tackle this problem. These methods are used in our study to identify phishing websites using a dataset of 48 characteristics and 10,000 occurrences, of which 5,000 are phishing and 5,000 are legal websites. We evaluated four deep learning models (ANN, LSTM, BiLSTM, and a hybrid ANN-LSTM model) and five machine learning models (Decision Tree, k-Nearest Neighbor, Naive Bayes, Logistic Regression, SVM) to assess their performance using evaluates for accuracy, F1 score, recall, and precision. Because of the drawbacks of adopting a small value, k-Nearest Neighbor fared the lowest, with 74% accuracy, while the hybrid ANNLSTM model outperformed the other models, with a maximum accuracy of 98%. Our results imply that deep learning models, especially hybrid ones, offer better phishing website detection capabilities.

Technological breakthroughs in Artificial Intelligence (AI) have led to the growth of human-like computers that can work independently and replicate a cognitive activity. The progression and interest among managers, researchers, and the general public have aroused interest in many industries, and significant corporate sectors are spending massively to profit gain through technology with the development of business interaction models. As information technology (IT) platforms become more advanced, and business activities become more autonomous, there is an increasing demand for business managers to better impact company operations and how they correspond with organizational goals. Machine learning and fuzzy logic design have lately been highlighted as recent innovations. Machine learning is an artificial intelligence approach that may enable smarter and more intelligent decision-making outcomes. In comparison, Fuzzy Logic Design (FLD) is a procedure that provides inferences or solutions from an ambiguous situation. In this research article, a fuzzy-inspired intelligent interaction model for the business sector is proposed, which utilizes a fuzzy logic design approach while enabling users to understand functions from a business standpoint and organize them related to the business targets, identify key indicators and carry out the necessary intelligent analysis on them to recognize causal factors of unforeseen metric values and enhance efficiency to improve business leadership.

Risk evaluation is a significant aspect of risk management; risk assessment, also known as risk description, determines its perspective and suitability by comparing risk to other hazards. Fuzzy logic is a kind of logic used in professional systems as well as additional Artificial Intelligence (AI) applications in which variables might get varying degrees of reliability or falsity denoted by a series of values among 1 (true) and 0 (false). Market, credit, insurance, and trading risk are all assessed using traditional risk models. On the other hand, fuzzy logic models are based on fuzzy set theory and fuzzy logic and are beneficial for analyzing hazards when there is insufficient knowledge or data. The primary motive of this research work is to analyze and propose a fuzzy logic-based risk assessment model to accurately assess the risks involved in establishing the timeline for a business sector development plan. The proposed model is to enhance classical risk assessment models based on "fuzzy logic" functioning on simple "If-Then" rules. The proposed model's simulation results provide effective results in a better way.

The duration of an inpatient stay affects hospital administration and improves hospital effectiveness in terms of controlling expenses and raising patient standards. It also assists in identifying the correlations among illnesses requiring hospitalization. For our study, we took 24,150 records from of the Open Data database pertaining to inpatient admissions in 2023. We used a number of methods, including Neural Networks, Deep Learning, Linear Regression, and Support Vector Machines, to predict the Length of Stay (LOS). We converted the data to numerical form for predictive purposes, dividing the dataset into 70% for training and 30% for testing. We assessed the model's performance using Root Mean Squared Error (RMSE) and split the forecast into four LOS categories: 0-2, 3-4, 5-7, and 8 days or more. The study also employed the Apriori algorithm to identify illness association rules that could impact LOS estimates. The results showed that identifying illness correlations is one element that might aid in enhancing the capacity to predict LOS.

Modern society operates on a data-centric basis, with the concept of data spaces becoming increasingly important. Data spaces are standardized digital infrastructures that enable trustworthy data exchange among various stakeholders, with data sovereignty as a core principle. The European Union is a leader in data spaces research and aims to efficiently manage and utilize health data in the medical field through the European Health Data Space (EHDS). EHDS serves as a cloud-based information source for patients, doctors, researchers, and businesses, enhancing the efficiency and accessibility of healthcare services and promoting medical technology research. To achieve this, the European Union is ensuring interoperability, establishing electronic identity verification and certification systems, and strengthening data security and governance. This paper analyzes the concept and development process of EHDS, the technical approaches for interoperability, and the legal and operational governance structures. It discusses how EHDS can contribute to innovation in healthcare systems and guarantee data sovereignty. Ultimately, EHDS is expected to provide efficient and personalized healthcare services based on data and promote the advancement of research and public health policies through the reuse of health data.

Detecting mass lesions not only helps reduce the cost of treating breast cancer but also enhances the lifespan of patients. Various computer-aided detection (CAD) systems have been developed to assist physicians in detecting mass in mammograms for early cancer screening. In this paper, a method for suspicious massive lesion segmentation in patches is proposed, which modified UNet with EfficientNet as the encoder. The proposed architectures are evaluated on publicly available dataset, namely the Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM). The quantitative results show that the proposed architecture can achieve mass segmentation with segmentation ac- curacy, Dice and IoU scores of 95.23%, 92.56% and 88.81% respectively in patches extracted from CBIS-DDSM.

Colorectal cancer is one of the most common cancers worldwide and poses a significant threat to humans, and it is preventable through early detection of polyps. Consequently, research is actively underway to improve the performance of early detection using artificial intelligence, particularly through segmentation tasks. This segmentation task requires expert annotated images of lesion areas and medical images, requiring large amounts of high-quality data. However, due to the limitations of medical data, there are constraints in obtaining high-quality data. Augmentation by image generation models has been actively researched to address this issue. In this study, we propose a diffusion model that generates diverse images by strongly applying different weights to the lesion area and the background area of the annotation, thereby improving the diverse generation capability in the background area. The evaluation involved assessing the quality of the generated images, augmenting the original dataset with this generated data, and applying it to the segmentation task to evaluate the performance improvement. This proves that assigning varying weights to the background can address the issue of data scarcity in medical imaging.

This paper introduces a system that uses Functional Electrical Stimulation (FES) for finger flexion control aimed rehabilitation for stroke patients. To address the variability in electrode between patients, Reinforcement Learning is applied together with a switching network that allows automatic electrode selection. This results in an adaptable system that does not require rigorous searching of the patient’s optimal stimulation points. Data that supports the differences in the stimulation location for individuals as well as the ability of the system to converge automatically to a stimulation point is presented.

The best defenses against AD are early detection and treatment. Recent developments in magnetic resonance imaging (MRI)-based computer-aided diagnostics have demonstrated promise in precisely classifying AD. To investigate these methods, this study makes use of the OASIS MRI dataset, which consists of 80,000 brain MRI pictures. Resampling was required in order to prevent bias towards the majority class when convolutional neural network (CNN) models were applied for classification due to the large imbalance of the dataset. The image slices obtained from 3D OASIS dataset's were examined for this study utilizing the EfficientNetV2B0 with customized classification layers and a condensed custom CNN model. There are four categories: nondemented, very mild demented, mild demented, and moderately demented. We covered multiclass as well as binary classification. Using various dataset sizes, the study evaluated two models: EfficientNetV2B0 and a customized sequential CNN model, with 98% and 96% accuracy, respectively. The findings address the disparity in different class sizes and demonstrate the promise of sophisticated CNN architectures for Alzheimer's disease classification and early detection.

Among the major reasons for death in humans, brain tumors are the most prevalent type and it affects humans of all ages. Brain tumors are treatable if detected in early stages. The classification of Tumors is being done by biopsy. On the Other hand, Magnetic Resonance Imaging (MRI) is a routine technique for humans to investigate this disease (Brain Tumors). In contrast, avoiding the need for a Radiologist, the detection and classification method proposed by using the Deep Learning Technique in this paper would benefit to all doctors globally. This work focused on a new Sequential base Convolutional Neutral Network (CNN) Architecture to classify the Brain Tumor types such as Glioma-Tumors, Meningioma tumors, No-tumors, and Pituitary tumors using MRI images. The proposed method gives better results for classifying Brain Images from a given dataset of Brain tumors with around 3264 MRI images. The purpose of our work is to use the Sequential base CNN model to detect brain cancers. The accuracy of our model's performance will be assessed. Consequently, we may infer that the Sequential base CNN model produces results that are very adequate and have an increased accuracy. Finally, the proposed method improves the accuracy up to 82.66%.

Traditional methods for plant phenotypic analysis, including attributes such as color, plant health status, and size, rely on expert manual analysis and judgment. Measurements are manually taken using tools, and all data is recorded by hand, which is highly inefficient. However, the development of artificial intelligence and deep learning has provided efficient solutions for plant phenotypic analysis. This study established a framework based on high-resolution images and utilized the SEGMENT ANYTHING MODEL (SAM) and Explainable Contrastive Language-Image Pretraining (ECLIP) for plant phenotypic analysis. Through the segmentation results of this model, the length and width of radishes, cucumbers, and pumpkins were measured. Since this method performed experiments without requiring model training or manual annotation of data, the framework demonstrated strong efficiency in the segmentation tasks prior to phenotypic analysis, significantly reducing manual labor costs. Moreover, the experimental results indicated that the mean absolute error (MAE) was below 0.05 for most test samples.

The metaverse is an expansive digital realm with endless potential for growth and has introduced unique safety challenges, especially when users are using head-mounted dis - plays (HMDs) that limit their awareness of real-world hazards. This paper emphasizes the urgent need for real-time risk detection in Unity-based augmented reality (AR) environments to improve user safety in the metaverse. We propose a new system architecture that integrates the YOLOv8 object detection model within a Unity environment to enable real-time processing of video feeds and to identify and respond to nearby risks. This system is designed to assist industrial workers in recognizing and addressing potential hazards on-site and enhance patient safety by identifying risk factors during medical procedures. As HMD usage extends to outdoor environments, assessing the safety implications of outdoor use becomes increasingly important. Our study aims to advance real-time risk detection capabilities in the metaverse by leveraging YOLOv8 to enhance both AR technology and risk management, with anticipated outcomes including significant contributions to AR development and improved safety protocols in various contexts.

In everyday communication, sign language is vital for those unable to speak. However, sign language education faces challenges such as mobility risks as well as limitations in solo practice. Recent attention has turned to Virtual Reality (VR) technology, which has shown positive effects in healthcare and rehabilitation. VR environments enhance learner immersion and support various physical activities, similar to traditional methods. Despite this, research on VR-based sign language education content is limited. This study proposes an immersive educational system that enables users to practice sign language in a virtual environment using a Head Mounted Display (HMD) and Leap Motion Controller(LMC) 2, with the aim of establishing a foundation for future VR sign language education content.

This paper proposes an algorithm to optimize a solar module optimizer's efficiency using the buck-boost functionality of a full-bridge converter. Given that the electrical characteristics of solar modules can vary significantly based on environmental conditions, it is necessary to have the capability to individually adjust each module's voltage and current. A fullbridge converter allows dynamic adjustment of input voltage to maintain optimal power output for each module and optimize the overall current of the system when modules are connected in series. The proposed algorithm was verified through MATLAB/Simulink simulations, demonstrating that it can increase total power production by up to 20%. This research contributes to enhancing the performance of solar power generation systems under various environmental conditions. Future studies will involve actual hardware implementation to further validate the algorithm and explore the application of different maximum power point tracking (MPPT) methods.

With global concerns mounting over the aging of infrastructure and roadways, and the growing emphasis on road maintenance, safety inspections have become a critical priority. Current crack detection models face challenges related to feature loss and performance degradation. In response, we propose an Attention U-Net model designed to minimize information loss and improve crack detection performance, particularly in low-quality images. Additionally, we focus on optimizing image capture and detection on low-spec devices, leading to the development of a lightweight model that performs effectively even in resource-constrained environments.

The advancement of deep learning has significantly improved image classification performance. However, the complexity of large-scale datasets continues to present challenges in terms of training time and resource consumption. One approach to address these issues is Self- Paced Curriculum Learning. Self-Paced Curriculum Learning begins by training on relatively easy data, allowing the model to autonomously select data and adjust difficulty levels as training progresses, gradually incorporating more complex data. This method improves the efficiency of the training process while minimizing performance degradation. In this study, we propose an approach that combines Self- Paced Curriculum Learning with the exclusion of low-noise data from the training process to further enhance training speed. The experimental results show that reducing the amount of data improves training speed. However, accuracy tends to decrease as the extent of data reduction increases.

As the metaverse evolves into a dynamic environment where users express their identity through avatars and fashion items, developing effective recommendation systems based on user interactions remains a significant challenge. To address this, we propose a novel technology that leverages Multi-Layer Perceptron (MLP)-based RGB and density values, processed using a Volume Rendering technique to convert them into a single-pixel representation. This approach enhances the accuracy of personalized fashion item recommendations by capturing visual and interactive data more precisely. Our model was trained on the publicly available H&M Personalized Fashion Recommendations dataset, achieving 79% similarity by measuring cosine similarity between item vectors. Additionally, we evaluated the system using data provided by a company that creates fashion items for real metaverse environments. Item IDs were used to define the source and target URLs, and the similarity between the items was measured to determine recommendations. This evaluation confirmed the model’s effectiveness in real-world scenarios.

Hospitals have accumulated large amounts of patient data, with each hospital’s data having unique characteristics and distributions. By leveraging this vast amount of data for machine learning, we can develop predictive models, such as those for predicting long-term outcomes in ischemic stroke patients and provide valuable information for treatment decisions. However, data privacy concerns prevent hospital data from being put together on a centralized server. This study investigates the applicability of federated learning for predicting long-term outcomes in ischemic stroke patients using data from Hallym University Sacred Heart Hospital in Pyeongchon and Hallym University Sacred Heart Hospital in Chuncheon. Patient outcomes are defined as favorable if the modified Rankin Scale (mRS) score is 0-2 and poor if the mRS score is 3-6. There are two tasks: one predicting patient outcomes at 3 months after stroke and the other predicting patient outcomes at 1 year after stroke. A simple deep neural networks model is used for implementation of the prediction model and the federated learning environment. In conclusion, the federated learning models using basic FedAVG and weighted averaging FedAVG achieved 99.4%-99.9% performance of traditional centralized learning models.

Although living organisms differ in shape and size, all are fundamentally structured by genetic sequences. Interpreting these sequences helps explain how organisms function. With the advancement of AI, significant breakthroughs have been made in protein sequencing and understanding protein function. However, there is still room for improvement, as data-intensive models require a substantial amount of protein sequences, many of which are not publicly available or lack quality. In this paper, we present a semi-supervised learning scheme to address the shortage of high-quality training data necessary for training protein language models. We demonstrate that this approach enhances the model's capability to classify toxic fungi protein sequences.

The rise of cryptocurrencies has also led to an increase in fraudulent activities, posing challenges in fraud detection on decentralized platforms like Ethereum. This issue is particularly pronounced in decentralized environments like Ethereum, where new transaction patterns continue to emerge. In this dynamic and changing environment, it is important to address the problem of concept drift, which refers to the continuous evolution of data patterns. To address the dynamic nature of these fraud patterns, we propose an automated hyperparameter optimization (HPO) approach using Proximal Policy Optimization (PPO). Unlike traditional HPO methods, PPO efficiently navigates the complex hyperparameter space, adapting to evolving fraud schemes with minimal human intervention. Our method enhances the adaptability and robustness of fraud detection models, effectively improving detection accuracy. Experimental results demonstrate that PPO outperforms existing HPO techniques, offering a more flexible and powerful tool for maintaining the performance of fraud detection systems in the rapidly changing cryptocurrency landscape.