ABSTRACT. This research paper presents a machine learning-based Arabic sentiment analysis of YouTube comments collected from the documentary that probes into the crucial events of October 7, 2023, showcasing the full and complete scenario of the aftermath and political ramifications of this operation, called "Al-Aqsa Flood." This pivotal incident brought about tremendous public engagement, especially in the Arab world, helping to choose 14,938 Arabic comments from 21,896 users' comments. ChatGPT 3.5 was employed using a zero-shot prompting approach to produce sentiment labels for the comments automatically as positive, negative, and neutral classification. It explores how large zero-shot language models can enhance sentiment classification in politically sensitive contexts. The labeled dataset was preprocessed and transformed using TF-IDF feature extraction before applying several machine learning classifiers, including Support Vector Machine (SVM), Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Multinomial Naïve Bayes (MNB), and K-Nearest Neighbors (K-NN). According to experimental data, SVM outperformed the other models, achieving the best balanced classification performance with 81% accuracy. It further emphasizes the value of machine learning models in enhancing the accuracy of classification and provides clues into the function of machine learning, especially in the analysis of digital text within contexts that are emotionally and politically charged.

ABSTRACT. Cloud computing indeed acts as the backbone of modern digital infrastructure; however, because of its inherently complex and dynamic nature, anoma-lies—with far-reaching consequences on perfor-mance and security—may arise. Hence, this paper attempts a literature review of current advances in real-time anomaly detection from cloud data streams. It systematically reviews the techniques, ranging from standard statistics and machine learn-ing algorithms to the more recent deep learning ar-chitectures, such as graph neural networks (GNNs) and transformers. Furthermore, it examines the main challenges in front of researchers and practi-tioners—such as the management of massive high-dimensional data, real-time detection demands, lim-ited labeled data, and very much explainable models. It also draws attention to some other intriguing fu-ture research directions, such as explainable artifi-cial intelligence (XAI), privacy-preserving federated learning, and generative AI. This review will provide some of the theoretical underpinnings for research-ers and practitioners in the arena and highlight criti-cal research gaps that may pave the way for future innovations toward smarter and more secure cloud monitoring systems.

ABSTRACT. Despite the recent emergence of neural text to speech (TTS) systems capable of near human speech quality, under resourced languages with high morphological richness still suffer from a considerable digital divide. This paper presents a approach to bridging this gap by building a monolithic multivariate model that synthesize three principal Amazigh variants: Tachelhit, Tamazight and Tarifit. By applying a stringent linguistic preprocessing pipeline, the proposed method eliminates orthographic and phonological variations through Tifinagh script normalization, morphological number transformation and adaptive prosodic segmentation. It uses the VITS inference model along with FreeVC module to avoid the speaker data training requirement through high zero shot speaker cloning and synthesized speech of high naturalness and quality. The presented model was trained on 8,139 audio samples covering the three variants, constituting a dataset of nearly 15 hours, both real and synthesized speech. To evaluate the system, a robust experimental design was used where the system was tested on 1000 never seen test samples in terms of the spoken content. The system achieved very competitive Mean Opinion Score (MOS) of 3.89 and 3.82 for speaker similarity and speech naturalness respectively. Meanwhile, objective measures (cosine similarity, F0 RMSE and MCD) and show the system achieves great performance with regard to voice quality. The research lays out a generalizable and linguistically informed solution for other under resourced languages and achieves a new benchmark in Amazigh TTS.

ABSTRACT. In this paper, the Project Management Office (PMO) is examined from the perspective of a viable moderating agent for the integration of local applications developed by courts into centralized information systems owned by the central administration of the Moroccan Ministry of Justice. The performance of centralized information systems is one of the main measurable outcomes of the successful integration of local applications into centralized information systems. This study used two methods in sequence. It was conducted with 100 respondents. It consists of two phases: First Phase is a qualitative thematic analysis taken from administration documents collected between 2018 and 2024, and second phase is a quantitative confirmatory analysis using PLS-SEM, confirms that there are four components that are good predictors of overall centralized IS Performance, namely "Local Application Integration"; Completeness, Technical Quality, Data Consistency and Synchronized Frequency and "PMO Maturity"; as a statistically significant Moderating Variable. This confirms the principles of IT Governance that the PMO uses as a strategic lever to generate improved performance from the Centralized IS in the Ministry of Justice.

ABSTRACT. Abstract—Digital transformation has reshaped work practices in the banking sector and created new questions about women’s inclusion in technology-based work environments, particularly in developing and conflict-affected contexts such as Yemen. Although supportive leadership is widely recognized as important for organizational adaptation, its role in women’s digital work integration in Hadhramout banks has received limited empirical attention. This study examines how four dimensions of supportive leadership—emotional intelligence, communication, participation in decision-making, and actual support—contribute to women’s integration into digital work environments. A quantitative descriptive-analytical approach was adopted, using a structured questionnaire distributed to employees in three banks operating in Hadhramout Governorate: the Central Bank of Yemen, Hadhramout Bank, and Yemen Commercial Bank. Of 191 questionnaires distributed, 154 valid responses were analyzed. The data were examined using descriptive statistics, multiple linear regression, and one-way analysis of variance. The findings indicate that supportive leadership positively contributes to women’s digital integration. Among the leadership dimensions, actual support and communication emerged as the strongest dimensions, suggesting that access to digital tools, technical assistance, training, and clear communication are central to women’s participation in digital banking work. The results also show that women’s involvement is stronger at the operational level than at the level of digital decision-making, where empowerment remains relatively limited. No statistically significant differences were found in respondents’ perceptions according to demographic and job-related variables. The study contributes empirical evidence from an underexplored banking context and highlights the need to strengthen inclusive leadership practices, digital training, and women’s participation in digital decision-making within banking institutions.

ABSTRACT. Abstract— Digital leadership has become an important concept in digital transformation research, yet its conceptual boundaries and measurement approaches remain fragmented. This paper presents a PRISMA-informed accessible-corpus review of how digital leadership has been conceptualised and measured in prior literature. The review focused on peer-reviewed studies published between 2010 and 2025 that provided definitions, dimensional frameworks, operational indicators, or measurement scales related to digital leadership and closely related constructs. From an accessible, peer-reviewed corpus, 23 records were identified; 4 duplicates were removed; 15 full-text reports were assessed; and 14 studies were included in the qualitative synthesis. The findings show that digital leadership has not yet converged around a single dominant factor structure. Instead, the literature can be organised into several measurement families, including short behavioural scales, competency-based frameworks, multidimensional capability models, adjacent construct-specific scales, and conceptual synthesis sources. Across these traditions, four recurring domains were identified: Digital Vision, Digital Competence, Digital Culture, and Digital Mindset. The proposed framework is not presented as a validated measurement model, but as a measurement-oriented synthesis that can support future instrument development and empirical validation, particularly in industrial and developing-country contexts.

ABSTRACT. Image security has become an important issue in modern communications, where increasing resistance against attacks usually leads to higher computational complexity and longer encryption time. This paper presents a lightweight chaos-based image encryption scheme that balances high security with low computational overhead. The proposed method combines Random Circular Shift (RCS) for row and column confusion with a Two-Dimensional Sine Tent Composite Map (2D-STCM) to generate a diffusion key. To enhance plaintext sensitivity and security, an MD5 hash of the plain image is utilized to derive initial keys. The scheme was rigorously evaluated using 8-bit grayscale and 24-bit color images from the USC-SIPI benchmark database across various dimensions. Experimental results demonstrate that the proposed algorithm provides high security and lower computational complexity compared with related methods. For a 1024 × 1024 image, the algorithm achieved 0.125 s encryption time, 33.4809% UACI, 99.6139% NPCR, and 7.9987 entropy in MATLAB 2020a.

ABSTRACT. Sustainability is now have been a global discussion and concern for long-term development. Fast fashion industry is one of industry that have been questioned and criticized by public due to its environmental, social, and economic impacts. This study examines how consumers perceive sustainability in fast fashion brands and how these perceptions influence brand trust and purchase satisfaction in e-commerce. Grounded in the Triple Bottom Line framework, sustainability is conceptualized as a multidimensional construct encompassing environmental, social, and economic dimensions. Many brands increasingly integrate artificial intelligence (AI) to support sustainability-related operations, Using a quantitative approach, 324 respondents’ data were collected from Indonesian consumers who have purchased fast fashion products and analyzed using Partial Least Squares Structural Equation Modeling (PLS-SEM). The findings highlight that sustainability significantly strengthens brand trust and enhances purchase satisfaction, positioning sustainability as a key trust-building mechanism in the fast fashion industry. This study contributes to the literature by empirically validating sustainability as a higher-order perceptual construct in the fast fashion context and demonstrating its role as a trust-building mechanism in purchase satisfaction. It further shows that technology-enabled practices such as AI influence consumer outcomes indirectly through sustainability perceptions rather than directly.

ABSTRACT. Artificial Intelligence (AI) and Learning Management Systems (LMS) are increasingly used in higher educational institutions to facilitate teaching and learning processes. However, it is also vital to explore the impact of these technologies on student academic performance. The purpose of this study is to examine the impact of Social Influence (SI), Learning Management System Usage (UL), Artificial Intelligence Usage (AI), and Sustainability Education (SE) on Academic Performance (AP). The methodology of this study is quantitative in nature, using survey methodology to collect data from 236 university students by using online questionnaires with five-point Likert scale items. Purposive sampling was used to select participants from university students who actively use LMS and AI in their learning activities. Data were analyzed by using Partial Least Squares-Structural Equation Modeling (PLS-SEM) with SmartPLS to explore relationships between five constructs. The results of this study show that social influence is very effective in motivating university students to use digital learning technologies. Moreover, LMS and AI are very effective in supporting university students' learning processes by providing access to resources and facilities for their academic activities. Sustainability education also plays an important role in shaping university students' use of digital technologies in their learning processes. Therefore, it is recommended that AI and LMS be used in the learning environment to improve student academic performance by providing positive social influence and sustainable digital learning practices.

ABSTRACT. This research investigates the influence of digital transformation (DT) on competitive advantage among industrial firms within the HSA Group in Taiz City, Yemen. Although digital transformation improves productivity, cost-effectiveness, and responsiveness to market demands, empirical studies regarding its impact on competitive advantage in this domain are still scarce. Utilizing a descriptive-correlational methodology, primary data was gathered by a Likert-scale questionnaire administered to 195 administrative officers from HSA Group enterprises in Taiz city, augmented by secondary sources. We used Structural Equation Modelling (SEM) with Diagonally Weighted Least Squares (DWLS) estimation for the analysis, as well as descriptive statistics and reliability testing. The results show that DT is being used at a high level, with process automation, data management, and digital infrastructure as the top sub-dimensions. Digital skills were the least important, Followed by data analytics and technology adoption. The study found that digital transformation has an impact on competitive advantage.

ABSTRACT. Phishing is a serious cybersecurity threat. Traditional detection methods, such as static blacklists and simple heuristic techniques, are becoming less effective as phishing strategies continuously evolve. To address this problem, this study proposes a phishing URL detection model based entirely on URL analysis. The proposed hybrid deep learning architecture combines Convolutional Neural Networks (CNNs) for local feature extraction with Long Short-Term Memory (LSTM) networks for learning sequential patterns. The model can automatically identify malicious textual and structural URL characteristics, including URL length, subdomain depth, HTTPS usage, and the frequency of special characters, without requiring webpage content analysis. Two datasets were used in this study: a primary dataset for training and testing, and an external dataset for validation. The proposed system operates independently without relying on third-party services, making it suitable for real-time phishing detection environments. Experimental evaluation on more than 450,000 URLs showed that the proposed model achieved 99.56% accuracy on the primary test set and 98.74% accuracy on the external validation dataset.

ABSTRACT. Telecommunication networks consist of several independent systems that operate to allow the exchange of information in a timely manner. Telecom networks are large, complicated systems that rely heavily on intelligent management to provide high levels of performance and reliability. Traditional optimization techniques are insufficient when it comes to accommodating the dynamic nature of these systems since they are unable to react quickly enough for real-time changes in the environment [1]. Within many undeveloped or war-torn countries, the power supply is unreliable, which will result in a decrease in telecommunication network performance. Additionally, power outages cause wasteful energy usage [2]. The purpose of this paper is to identify how implementing artificial intelligence-based optimization on unstable telecom networks can lead to lower operating costs and continue to maintain key performance indicators (KPI). The proposed methodology is to develop an approach for overcoming the power stability issue within current telecom models. Furthermore, the framework provides a model to optimize energy consumption with a focus on cost. Additionally, this model incorporates real-time data collection and monitoring, which allows for the dynamic management of energy consumption based on network performance/conditions as well as to meet the quality of service (QoS) expectations of customers. The simulation uses actual traffic data and a power outage model that is representative of actual grid instability in developing regions. Future work also involves validating the simulation with large-scale data sets. Previous studies on hybrid energy management systems indicated that optimization of diesel–battery systems through intelligent control could lead to a reduction in fuel consumption of up to 16.12% [3]. If the simulations are extended to incorporate deep learning or reinforcement learning to represent the intelligent communication networks, the results can demonstrate fuel savings of more than 16% for the diesel generator energy consumption with better battery power utilization via advanced energy-saving technologies. Simulation results show that the proposed AI-based optimization framework is capable of achieving a reduction in diesel generator energy consumption by 20-24% while ensuring that the KPIs are maintained.

ABSTRACT. Abstract— Examinations are considered among the most important evaluation tools. Still, several factors hinder the achievement of their objectives, including inefficient preparation, ineffective presentation methods, rising student numbers, and a weak educational infrastructure. Therefore, this study aimed to design a smart testing System based on artificial intelligence techniques to build examinations and analyze results. The System analyzes course content to extract questions aligned with Bloom's cognitive model (remembering, understanding, application) and provides detailed student reports. It prepares test examinations according to the specifications table and provides online and offline execution methods via personal phones using a special application, suitable for various educational infrastructures. The research and development methodology, based on the Analysis, Design, Development, Implementation, and Evaluation (ADDIE) model and the descriptive-analytical approach, was used to collect questionnaire data. The study applied a census sample of all (15) faculty members of the Computer Department during the first semester of the academic year 2025/2026. Results showed that testing System design procedures and construction mechanisms received a high degree of evaluation, and functions related to preparing and correcting examinations received very high appreciation. Using the testing System supports educational measurement and evaluation by organizing question banks, facilitating student monitoring, and providing analytical reports. Technically, the testing System received high evaluation regarding ease of use, interface clarity, and operating environment suitability. These results reinforce what some faculty members in the study sample did: applying to the testing System at other colleges outside the Computer Department at the College of Education, thereby reflecting the reliability of the study's findings.

ABSTRACT. Wireless Medical Sensor Networks (WMSNs) play a critical role in modern healthcare by enabling continuous monitoring of patients' vital signs. However, these networks face significant challenges in balancing security and resource constraints. This paper proposes a hybrid lightweight adaptive authentication protocol that combines the efficiency of hash-based mechanisms with the robust security of Elliptic Curve Cryptography (ECC). The protocol incorporates an adaptive system that dynamically adjusts security levels based on the sensitivity of medical data. The proposed solution was evaluated through extensive simulations across three medical scenarios: intensive care unit (ICU), home care, and emergency settings. Results indicate that the hybrid protocol achieves a good trade-off between security and performance. The proposed hybrid protocol achieved a 49.3% improvement in energy consumption compared to pure ECC, a 52.8% improvement in response time, and 96% of ECC's security strength. This demonstrates its ability to achieve an effective balance between security and efficiency across different medical environments.

ABSTRACT. Recently, artificial intelligence and machine learning have significantly advanced many fields, including computer vision. One of the most prominent applications of these technologies is gender recognition, which involves determining an individual's gender from facial images. Current model assessments are often low in accuracy. This research aims to improve the accuracy of gender recognition systems by applying advanced preprocessing techniques and feature extraction methods. We used the FERET dataset as a standard for training, testing, and performance evaluation. We implemented the Viola-Jones face detection algorithm, the COSFIRE segmentation algorithm based on the Gabor filter, and the spatial pyramid algorithm for feature extraction. The results showed a significant improvement in classification accuracy, reaching 97.9% when applying a deep learning-based classification design model for gender determination.

ABSTRACT. Intrusion Detection Systems (IDS) play an important role in the security of today’s networks against ever-evolving cyber threats. However, most of the machine learning-based intrusion detection systems use only one classifier to classify all the available data, which may not be efficient in handling large and varied network traffic. In this paper, the authors have proposed a framework that combines hierarchical binary SOM clustering with a classification pool architecture. The proposed approach is efficient in handling large datasets because it divides the available dataset into smaller and more homogeneous sets, allowing local classifiers to perform independently on smaller sets of data. The proposed approach has been evaluated using benchmark datasets like NSL-KDD and CICIDS2017. Experimental results show that the proposed approach has achieved 99.2% detection accuracy while reducing the dataset size up to 75%.

ABSTRACT. VANET has gained a lot of attention in recent years due to the rapid great service achievements in this field. As well as it provides us with one of the important developments to reach safest driving and smart self-driving vehicles in smart cities and intelligent/secure transportation. VANET is considered as one of Ad-hoc type dedicated vehicle-specific model. It is exactly differed from other Wireless technologies in Routing issues, because these nodes are attached into the vehicles that move in randomly manner, different in speeds and moving direction. Thus, one of the critical concerns in VANET to defeat routing critical concern and to find a mechanism with high level of routing efficiency. Hence, ensures speedy Vehicle’s communication and data transferring between nodes. This paper main contribution to address a new mechanism called One-Shot-Transmit (OST). It gives a high priority on scalable routing and fast data flow between moving Vehicles as well as with RSU station. SVROST indicates a specific mechanism by which vehicles within the VANET network can exchange data with RSUs. Our results and simulations in this main approach, demonstrate the effectiveness of SVROST in ensuring high efficiency and lower cost in both receiving / forwarding data to the trusted authority (TA).

ABSTRACT. Precision farming and sustainable agriculture both depend on effective water management. To provide intelligent irrigation control, this paper introduces a machine learning-based multi-zone smart irrigation system that uses Raspberry Pi and IoT cloud analytics. To automate irrigation decision-making across three agricultural zones, the recommended system combines soil moisture sensors, a DHT11 environmental sensor, relay-controlled solenoid valves, and ThingSpeak cloud monitoring. Using recently collected soil and environmental data, A Random Forest machine learning model is used to forecast irrigation needs. Using real-time sensor data acquired at regular intervals, experimental observations demonstrated effective irrigation prediction and successful multi-zone valve control using the Random Forest model. When compared to traditional threshold-based irrigation techniques. According to experimental findings, the proposed framework allows for effective real-time irrigation management with enhanced water-use efficiency, decreased manual intervention, and scalable agricultural automation across multiple zones. The results of the experiment show effective water management and the ability to make decisions in real time. Precision agriculture applications can benefit from the specified architecture’s increased irrigation efficiency and decreased water waste.

ABSTRACT. The rapid adoption of large language models (LLMs) has accelerated the development of intelligent agents capable of reasoning, tool invocation, and knowledge-grounded interaction. However, deploying such agents in real-world environments remains challenging due to fragmented data pipelines, limited configurability, and insufficient governance mechanisms. This paper presents Nonefinity Agent, a unified AI agent platform designed to support secure knowledge ingestion, structured data analytics, and dynamic tool orchestration within a production-ready architecture. The platform integrates semantic retrieval, SQL-based analytical querying, and external API tools into a configurable reasoning loop, enabling agents to perform multi-step tasks while maintaining transparency and data isolation autonomously. Experimental evaluation across four agent configurations on a 100-query benchmark shows that the full configuration achieves 85.3% reasoning accuracy, 95.4% tool success rate, and 17.6s average latency over five repeated runs, highlighting the platform's suitability for enterprise and research deployments.

ABSTRACT. This paper presents an investigation of the impact of performance expectancy (PE) and social influence on the intention to use mobile banking in Yemen. The study presents the results of a survey with 158 respondents and key informant interviews with banking professionals at six Yemeni banks. The results revealed that PE was a strong predictor of intention (average = 4.24) and that time-saving and daily-life convenience were significant drivers. On the other hand, SI showed a moderate positive impact, with an average of 3.53, mostly influenced by close friendships rather than by colleagues or formal relationships. The findings of this study supported hypotheses H1 and H2, confirming that PE and SI positively influence the intention to use mobile banking in Yemen, with PE as the main factor. This study provides new evidence for the UTAUT framework from conflict-affected and resource-limited settings, offering practical guidance for banks and policymakers seeking to promote mobile banking in unstable regions.

ABSTRACT. The technological disruption in the food and drink sector has changed everything through Food Delivery Apps (FDAs). The widespread acceptance at the beginning of the FDA is not surprising, however, understanding what drives users to continue to use FDAs over time is a major challenge. The purpose of this research was to investigate the determinants of the Continued Intention to use FDAs in Vietnam using an integrated theoretical model of SET (Social Exchange Theory), TAM (Technology Acceptance Model), TTF (Task-Technology Fit) and ECM (Expectation Confirmation Model). A quantitative methodology was used, collecting data from 346 people who responded to survey. Using PLS-SEM we analyzed the data with the aid of SmartPLS. Our results confirm that both Ease of Use and Performance Expectancy are significantly positively associated with both user Satisfaction and Continued Intention to use. We also found that although there is no direct relationship between Task-Technology Fit and Continued Intention (p = 0.078), it has a mediating influence on Satisfaction which is the primary determinant of continued usage. Satisfaction is the strongest determinant of long-term commitment, having a large path coefficient of 0.487. Therefore, satisfaction is the main "psychological gatekeeper" for converting functional utility into user loyalty. In terms of management implications for the FDA developers, they need to move their focus from small technical engineering towards improving the emotional user experience and UI/UX optimizations to prevent losing market share in a competitive marketplace.

ABSTRACT. The increasing adoption of digital learning platforms has transformed educational practices in higher education. However, improving students’ satisfaction with e-learning systems remains a critical issue for many institutions. This study investigates the relationships among E-Learning System Quality, Social Influence to Use E-Learning, Sustainability Education, and E-Learner Satisfaction. Data were collected from 403 university students in Indonesia through a questionnaire using a five-point Likert scale. The proposed model was analyzed using SmartPLS. The results show that Social Influence to Use E-Learning significantly influences both Sustainability Education and E-Learner Satisfaction. Furthermore, Sustainability Education positively affects learner satisfaction, while E-Learning System Quality significantly supports both sustainability education and satisfaction. These findings highlight the importance of integrating technological quality, social support, and sustainability-oriented learning content to enhance the effectiveness of e-learning systems and improve students’ satisfaction in higher education.

ABSTRACT. SphagnumPaludicultures on rewetted peatlands have intricate nutrient dynamics impacted by site-specific factors, time, and hydrology. In this study, a preliminary model is developed to assess whether machine learning algorithms can be used to forecast pH, ammonium (NH+ 4 ), phosphate (PO3− 4 ), potassium (K+), and sodium (Na+) porewater concentrations. We created prediction models based on characteristics including restoration site identities, historical NH+ 4 levels, and time since establishment using the Random Forest and XGBoost algorithms. While NH+ 4 and PO3− 4 exhibited low predictability, suggesting the effect of other unmeasured components, our data reveal that potassium and pH are the most predictable parameters (R2 up to 0.79). Nutrient prediction is similarly influenced by restoration age, delayed nutrient values, and site categorization, according to feature relevance analysis. Furthermore, correlation analysis showed a modest association between NH+ 4 and water table depth at the younger site and a substantial positive relationship between K+ and restoration age. These results support evidence-based management choices in climate-smart agriculture and highlight the need for data-driven methods for tracking nutrient dynamics in Sphagnum paludicultures.

ABSTRACT. Reconfigurable Intelligent Surfaces (RIS) are a promising technology for enhancing next-generation wireless systems. However, the existing works assume ideal hardware and perfect Channel State Information (CSI). We introduce RIPPLE framework to investigate robust RIS phase optimization for a downlink mmWave multi-user MISO system under conditions of hardware impairments (HWI) and imperfect CSI. We formulate the RIS design as a worst-case per-user mean square error (MSE) minimization problem to ensure robustness and fairness. Due to high-dimensional nature of the problem, RIPPLE framework is employed to learn efficient RIS configurations. The findings shows that the RIPPLE reduces worst-case MSE by up to 20–30\% compared to baseline Proximal Policy Optimization (PPO) and maintaining robust performance under varying system parameters and impairment levels. The findings highlight the effectiveness of reinforcement learning for RIS-assisted communication systems.

ABSTRACT. Clinical Decision Support Systems (CDSS) are used to assist healthcare workers in making real-time treatment decisions. One of the major drawbacks of current CDSS is that unstructured medical queries are not handled with low response time and high semantic accuracy. In this paper, an LSTM-based query processing framework is proposed, in which BioBERT embeddings are integrated for real-time processing of medical queries in CDSS. The system is trained on public datasets (MIMIC-III and PubMedQA), and it is tested at a rate of 45 queries per second, achieving an average latency of 230 milliseconds and a classification accuracy of 91.3%. The effectiveness of the LSTM model over conventional RNN and GRU models is validated through ablation results, and higher accuracy is achieved when domain-specific BioBERT embeddings are used instead of general-purpose Word2Vec embeddings. Certain limitations are identified, including the lack of clinical validation and limited support for multiple languages. The results are considered promising, demonstrating strong performance in benchmark evaluations; however, further clinical validation is required before deployment in real healthcare environments.

ABSTRACT. Clinical decision support systems (CDSS) systems are often used in healthcare to enhance the accuracy of making diagnoses and boost the success of treatments. Such systems work on top-quality data and advanced artificial intelligence to suggest helpful products. Sometimes, if a dataset does not have enough instances of some minority classes, it can hinder the models’ performance. Data augmentation approaches along with advanced recurrent neural networks (RNNs) are used in this study to deal with this matter. Earlier studies often faced difficulties mainly because rare conditions were handled poorly, it was hard to train vanilla RNNs and transformer models required much computing power. To solve these problems, SMOTE and Borderline SMOTE techniques are applied to the data to make it more balanced. Three types of architectures—Vanilla RNN, GRU and Transformer—are tested. In the suggested process, you preprocess clinical data, increase the number of minority classes and train the models using balanced datasets. The findings display that Borderline SMOTE performs well near decision borders, while the Transformer model reaches the highest level of accuracy (92-94%) and F1-score (89-91%). From the study, it is clear that using both data augmentation and advanced RNN models greatly boosts how a CDSS performs, including in the detection of uncommon medical conditions.

ABSTRACT. With the advancement of the Internet of Things (IoT) and Artificial Intelligence (AI), Predictive Maintenance (PdM) has become a key enabler of the factory of the future. This focused review examines twelve carefully selected and representative studies published between 2017 and 2025 that reveal the important role of IoT and AI in PdM of industrial equipment through fault data analysis. It discusses the various types of IoT-based sensors used for data acquisition, including temperature, vibration, and pressure sensors. Additionally, the study analyzes the performance of AI algorithms such as Artificial Neural Networks (ANN), Random Forest (RF), and Long Short-Term Memory (LSTM) in predicting equipment failures. Unlike prior review papers that address these dimensions in isolation, this review provides an integrated analysis covering Explainable AI (XAI), cybersecurity, edge computing, scalability challenges, and industrial system integration. Based on a comparative analysis of the selected studies, this paper identifies key research gaps related to data quality, model interpretability, and real-world validation. A conceptual framework and a hierarchical taxonomy are proposed to synthesize the main components of PdM systems. The findings indicate that while current approaches achieve high accuracy (83–98%), challenges such as missing data and the trade-off between accuracy and interpretability remain unresolved. This focused review provides a structured and practical perspective for researchers and practitioners, contributing to the advancement of smart manufacturing and Industry 4.0.

ABSTRACT. Inverse design of composite material microstructures used in the additive manufacturing process presents a significant challenge in the advanced materials engineering. The existing computational approaches have major shortcomings in searching the enormous multi-dimensional design space described by geometry, composition and processing parameters. The conventional topology optimization is computationally expensive to run quickly, whereas data-based generative models generate designs that are not physically plausible. In this work, a novel Physics-Informed Diffusion Model (PIDM) framework is presented to overcome these limitations. The generative capability of denoising diffusion probabilistic models is combined with the constraints of governing physical laws. The framework is trained on 12,000 simulated microstructure–property pairs for the isotropic case and 15,000 pairs for the anisotropic case. A physics-based loss is used to ensure that the denoised field remains physically admissible. The framework is used to design isotropic composites with a target elastic modulus of 8.5 GPa and a target Poisson’s ratio of 0.28, with a mean property error of 1.8% in silico. In addition, anisotropic composites with target thermal conductivities of 0.80, 0.30, and 0.30 W/m·K are designed, with a mean property error of 3.3% in silico. On-demand material design is demonstrated to be a robust, efficient and physically consistent paradigm as experimental validation through DLP 3D printing has fabricated properties within 6.7% of target values.

ABSTRACT. Boiler tube corrosion is considered one of the most significant issues in the operation of thermal power plants in ash-rich environments. In this study, a systematic experimental comparison is conducted between cerium oxide (CeO₂) and magnesium oxide (MgO) nanoparticles as corrosion inhibitors for ASTM A178 Grade A carbon steel in a simulated ash atmosphere at temperatures ranging from 600 °C to 900 °C, with an exposure time of 10 h for each condition using the dip-coating method. The corrosion rates are quantified using the weight loss method in accordance with ASTM G1, while the surface morphology and elemental composition are characterized by SEM, EDS, and XRD analyses. The experimental data are analyzed using a polynomial regression model, where the nanoparticle concentration (wt%) and exposure temperature (°C) are used as input features and the corrosion rate (mg/cm²·h) is used as the output variable. The model performance is validated by comparing the predicted corrosion rates with the experimental values, and good agreement is observed under all tested conditions. The results indicate that the lowest corrosion rate is obtained for CeO₂-coated specimens at 900 °C, with a value of 1.8 mg/cm²·h, whereas values of 3.9 mg/cm²·h and 8.4 mg/cm²·h are recorded for MgO-coated and uncoated specimens, respectively, at the same temperature. The superior performance of CeO₂ is attributed to the formation of dense and adherent oxide layers promoted by the reversible Ce³⁺/Ce⁴⁺ redox reaction. Furthermore, the developed predictive framework is demonstrated to be applicable for optimizing nanoparticle loading as a data-driven tool and is considered useful for supporting proactive maintenance scheduling in industrial boiler systems.

ABSTRACT. Breast cancer is one of the most prevalent and serious diseases worldwide, making early and accurate diagnosis essential for improving patient survival rates. With the exponential growth of medical data, the need has arisen for intelligent and automated systems capable of supporting the diagnostic process. However, the complexity and high dimensions of medical data pose a significant challenge to traditional classification methods. In this study, an efficient breast cancer classification model is proposed, based on an enhanced version of the Enhanced Grasshopper Optimization Algorithm (EGOA) to train a Multilayer Perceptron (MLP) network. The proposed model aims to improve classification performance by optimizing the neural network parameters and enhancing its generalization capabilities. To verify the efficiency of the proposed model, a series of experiments were conducted using breast cancer data, during which the performance of the EGOA-MLP model was evaluated. The experimental results showed that the proposed model achieved high accuracy, with a classification accuracy of 96%, reflecting the model's ability to effectively distinguish between benign and malignant cases. These results confirm that integrating the EGOA algorithm with the MLP network provides a robust and reliable framework for breast cancer classification, contributing to the advancement of early diagnosis systems and medical decision-making.

ABSTRACT. In recent years, we observed the emergence of Unmanned Aerial Vehicles (UAVs) in Search and Rescue operations (SAR). Offering advantages such as swift deployment and increased visibility on the ground, they revolutionized the field. They however remain relatively unknown and underutilized in the majority of Search and Rescue operations. In this paper, we provide insight for UAVs looking into their anatomy. In addition, we looked into how they affected the field of Search and Rescue by scaling existing methods or adding new technics. We discuss current limitations of the technology and propose potential solutions from current trends and avenues for future research, development and deployment of UAVs in the Search and Rescue field.

ABSTRACT. Corrosion represents a critical threat to metallic infrastructure integrity, imposing substantial economic and safety consequences globally. This study presents a comprehensive end-to-end deep learning pipeline for automated corrosion detection using the YOLOv8s object detection framework. A dataset of 506 real-world corrosion images was subjected to systematic exploratory data analysis, revealing characteristic warm chromatic bias consistent with iron oxide signatures. Offline augmentation via Albumentations expanded the training corpus from 354 to 1,380 images, enhancing model robustness against photometric and geometric variability. Supplementary feature extraction using ResNet-50 and handcrafted descriptors confirmed dominant morphological variance across the dataset. The YOLOv8s model was trained on an NVIDIA RTX 3070 GPU using the AdamW optimiser with early stopping, converging at epoch 94. Evaluation on the held-out validation set yielded a precision of 0.986, recall of 0.985, mAP@0.5 of 0.993, and mAP@0.5:0.95 of 0.981, with real-time inference at approximately 108 FPS. Results demonstrate that a lightweight single-stage detector, combined with rigorous data preprocessing, achieves near-perfect corrosion detection performance suitable for practical industrial deployment.

ABSTRACT. The increasing reliance on web-based digital services has intensified exposure to injection-related cybersecurity threats, particularly SQL injection, which continues to affect database-driven applications through unsafe input handling and insecure query construction. Although many studies have examined SQL injection and related defenses, the literature often remains fragmented across attack taxonomy, detection methods and prevention strategies. This paper employs a PRISMA-guided systematic review methodology to identify, evaluate and integrate pertinent studies on SQL injection and contemporary injection attacks in web applications. The review categorizes the literature into three primary perspectives: an overarching injection threat landscape, a comprehensive SQL injection attack taxonomy and contemporary detection and prevention methodologies. It also points out the real-world problems with current methods and finds areas where more research is needed on adaptive, hybrid, and cross-platform defense mechanisms

ABSTRACT. Recently, digital transformation has become the core of universities' planning, especially after the sudden change brought by the Coronavirus pandemic, which brought many existing and unaddressed challenges back to the forefront. To address these challenges, the Digital Transformation Strategies with Six Dimensions (DTS-6D) taxonomy is introduced as a framework to help higher education institutions build their own digital transformation strategies. A review of peer-reviewed literature from 2020 to 2025 reveals six strategic areas: Digital Infrastructure (DI), Cognitive Architecture (CA), Digital Infrastructure Governance (DIG), Culture and Capability Transformation (CCT), Community Integration (CI), and Sustainability (S). Rather than empirical testing of the framework, this study consolidates key concepts to address and fill existing gaps and ambiguities in the literature. The DTS-6D framework proposes an approach to future work on studies, policies, and leadership decisions. The framework offers a general approach, but it needs adaptation to emerging technologies, and potential biases in its application remain a challenge. Nevertheless, the DTS-6D taxonomy serves as a valuable guide for institutions that intend to handle the digital transformation and apply digital technologies fairly and effectively.

ABSTRACT. The need for safe ways to communicate data and information has grown in recent years due to the rapid expansion of computer networks and cyberspace. Intrusion Detection Systems (IDS) are crucial for improving network security because of the numerous cyber threats and malevolent assaults that have surfaced as a result of this rapid expansion. The creation of efficient detection systems is a constant problem since cyberattacks are dynamic and adaptable, making traditional detection methods less successful in recognising these threats. This paper proposes a multi-layered perception (MLP) model, a supervised learning approach utilising artificial neural networks (ANNs) for use in intrusion detection. The fundamental idea is to use the MSA algorithm, a nature-inspired optimisation technique, to optimise the neural network's training process. Accurately classifying network traffic into attacks and valid data is the goal of the suggested methodology. A validated standard dataset was used to assess the model's performance. According to experimental results, the MLP model optimised using MSA outperformed several other optimisation methods, achieving a high accuracy of up to 98.81%.

ABSTRACT. Wireless Sensor Networks (WSNs) are widely used in applications such as environmental monitoring, smart infrastructure, and Internet of Things (IoT) systems. However, the limited energy resources of sensor nodes make efficient topology control a critical challenge for extending network lifetime. Connected Dominating Set (CDS) based approaches have been widely adopted to construct virtual backbones that reduce redundant transmissions and improve routing efficiency. Nevertheless, most existing distributed CDS algorithms rely on static decision rules and often assume synchronous network operation, which limits their adaptability in dynamic and asynchronous environments. This paper proposes an artificial intelligence assisted distributed CDS construction algorithm for energy-efficient asynchronous wireless sensor networks. The proposed approach integrates a lightweight reinforcement learning mechanism into the classical Single Phase Multiple Initiator (SPMI) algorithm, enabling each node to autonomously decide whether to participate in the CDS based on local observations such as residual energy, neighborhood density, and connectivity contribution. Theoretical analysis demonstrates that the proposed algorithm guarantees CDS connectivity with linear message complexity. Simulation results show that the proposed method improves energy balancing, reduces CDS size, and significantly extends network lifetime compared to classical CDS-based approaches.