In-hospital cardiac arrest in intensive care remains frequent (often cited incidence roughly 0.5%-7.8% of admissions), while causes differ in what staff and equipment must be ready. We ask whether vital-sign trajectories from a standard EHR can classify which of three cardiac-related mechanisms is most salient arrhythmia, acute myocardial infarction (AMI), or respiratory failure or hypoxia so ICU resources can be aligned with risk. Using MIMIC-IV, we extracted diagnoses and charted vitals in the 12 hours before the index event, applied cleaning, aggregation, label encoding, sequence padding, and class balancing (3,000 cases per class), then trained and compared eXtreme gradient boosting (XGBoost), random forest (RF), support vector machine (SVM), and logistic regression (LR) with 5-fold cross-validation on an 80/20 split. XGBoost performed best (about 93% accuracy; sensitivity 89.15%; specificity 90.43%; AUC-ROC 0.94). Feature importance highlighted heart rate, oxygen saturation, and blood pressure patterns consistent with bedside monitoring practice. The study supports mechanism-oriented triage labels derived from widely recorded vitals, as a complement to generic early warning scores, for prioritizing telemetry, respiratory support, and cardiology pathways. External validation and prospective evaluation are needed before deployment.
Student academic success is influenced by various factors, both academic and non-academic. This study aims to examine the correlation between key student attributes and final grade point average (GPA), as well as to develop a machine learning model to predict academic success. The correlation analysis involved academic variables such as admission scores (Mathematics, English, Indonesian, and academic aptitude test/TPA), English ability test (EAT), spiritual formation (SF), and first-year GPA (GPA_1). The results indicate that GPA_1 has the highest correlation with final GPA (0.63), followed by SF (0.44), while other variables exhibit lower correlations. To enhance prediction accuracy, a machine learning approach using three primary models was employed: Naïve Bayes, support vector machine (SVM), and an ensemble learning method based on a stacking classifier that combines SVM and Naïve Bayes. The evaluation used five train-test split ratios and performance metrics, including accuracy, precision, recall, and F1-score. Experimental results reveal that the SVM model achieves the highest accuracy at 88.40%, followed by the ensemble model combining SVM and Naïve Bayes (88.00%) and the Naïve Bayes model (87.10%). These findings confirm that the machine learning approaches, could effectively predict student academic success, providing a foundation for academic decision-making and educational intervention strategies.
Research on halal fashion has largely focused on consumer purchase decisions, with limited attention to how halal fashion brands interact with consumers and manage their brand image on social media platforms such as Instagram. This study addresses this gap by examining brand interaction patterns and image management strategies among leading halal fashion brands in Indonesia. Using a netnographic approach, it analyzed 1,321 Instagram posts from six halal fashion brands over six months (July–December 2022), applying content and image‑management codes to classify post types (photos and videos) and representation strategies (personalized, contextual, and celebrity use). The findings show a slightly higher proportion of photo posts (51%, 674 posts) than video posts (49%, 647 posts), with hijab fashion brands more active than Muslim and sports fashion brands in producing content. Across all brands, image management relied predominantly on personal context and non‑celebrity representation, while professional context and celebrity‑based posts were used less frequently. These results suggest that halal fashion brands strategically emphasize relatable, personalized, and non‑celebrity content to build brand image and engagement on Instagram, offering practical guidance for brand managers in designing effective social media strategies and contributing novel empirical evidence on brand interaction and image management in the halal fashion sector.
The continued scaling of semiconductor devices has exposed the limitations of traditional two-dimensional (2D) integrated circuit architectures. To address performance bottlenecks and interconnect constraints, the industry is increasingly adopting three-dimensional (3D) integration technologies. through-silicon vias (TSVs) are a fundamental enabler of this advancement, facilitating vertical signal transmission between stacked silicon layers. Despite their benefits, TSVs face critical challenges related to crosstalk, power dissipation, and signal delay issues that are especially pronounced in dense via arrays. This research explores the use of multi-walled carbon nanotube (MWCNT) based TSVs insulated with different dielectric liners, including silicon dioxide (SiO₂), PPC, polyimide, and benzocyclobutene (BCB). HSPICE simulations are used to evaluate crosstalk noise, power dissipation, power delay product (PDP), and energy delay product (EDP) across varying TSV pitches. Among the materials studied, BCB demonstrates the most promising results. Specifically, MWCNT TSVs with BCB at a 10,000 μm pitch achieve up to 58% reduction in functional crosstalk, 75% in dynamic crosstalk, 78% in power dissipation, and a 52% improvement in PDP compared to single-walled CNT (SWCNT) based TSVs. These findings confirm the suitability of combining MWCNT cores with low-k BCB liners for enhancing performance, energy efficiency, and signal reliability in advanced 3D integrated circuits.
The timely identification and diagnosis of leaf diseases is crucial for crop productivity and health. This study proposes a robust approach to this issue by combining beetle swarm optimization (BSO) with other ML models. Four different datasets were used to train our model: apple leaf, grape leaf, plant village leaf, and tomato leaf for disease detection. The process begins with preparing the leaf images, involving contrast enhancement and noise reduction. Through color-based segmentation, we can distinguish healthy regions from diseased ones, aiding in the classification process. Our research demonstrates the effectiveness of the BSO-convolutional neural networks (CNN) method in recognizing and categorizing plant diseases with high accuracy rates. Leveraging the power of BSO to adjust the model’s parameters and incorporating color-based segmentation enhances the model’s robustness and accuracy. The results of this study highlight the potential of automated disease management systems for agriculture, providing agronomists and farmers with the necessary tools to address and monitor emerging threats to their crops effectively.
As digital data rapidly grows, content-based image retrieval (CBIR) has become important for optimizing collections of visual data. This work proposes a retrieval framework which operates in two stages and improves accuracy by using systematic fusion of features. In the first stage, first-stage wide-scope descriptors called bag-of-visual-words (BoVW), scattering wavelet transform (SWT), discrete cosine transform (DCT), and principal component analysis (PCA) retrieve initial candidate images. The second stage undertakes detailed re-ordering of candidate images by implementing the local binary pattern (LBP), histogram of oriented gradients (HOG), and singular value decomposition (SVD) descriptors to re-evaluate similarity scores. Each individual descriptor returned results for mean average precision for the top 10 retrieved images (mAP, top-10) of between 0.63 and 0.79 and the fused framework achieved 0.88, which is evidence of the viability of complementary feature integration. These findings support the hypothesis that while multiple descriptors performed well and delivered high retrieval accuracy, hierarchical fusion of multiple handcrafted descriptors does not involve the computational costs associated with deep learning methods.
Tracking the origins of viral content in social media is crucial for identifying misinformation, ensuring accountability, and protecting intellectual property. By tracing content back to its source, platforms can curb the spread of false narratives, hold malicious actors responsible, and safeguard creators’ rights. In this work, “Effective blockchain based system for tracking viral content origins in social media-(BCTVCO)” is proposed. The BCTVCO is a blockchain-based content authentication platform that leverages interplanetary file system (IPFS), smart contracts, and cryptographic hashing to verify digital assets and detect unauthorized reuse. This decentralized application addresses the challenges of content authenticity, ownership verification, and intellectual property protection in the digital space. The system integrates Ethereum smart contracts (Solidity) to store immutable content records and uses SHA-256 hashing for secure file integrity verification. Content is uploaded to IPFS via Web3. Storage, ensuring distributed and tamper-resistant storage. The React-based frontend with MetaMask authentication allows users to seamlessly register, upload, and track their content. In implementation, BCTVCO outperformed existing methods and proved to be a scalable, transparent, and secure blockchain-based content verification system. Future enhancements of BCTVCO involve multi-chain support and AI-powered content analysis to strengthen security and usability.
Cultural heritage preservation increasingly relies on digital forensics to ensure authenticity and consistency in heritage documentation. This study presents a digital forensic framework based on machine learning for classifying multi-device images of the historic Surabaya City Hall. The dataset was collected from nine smartphone devices and preprocessed through standardization, 360° rotational augmentation, and three filtering methods: gaussian, median, and laplacian. Three supervised algorithms (support vector machine (SVM), K-nearest neighbor (KNN), and logistic regression (LR)) were evaluated using accuracy, macro average, and weighted average of precision, recall, and F1-score. The results indicate that image preprocessing substantially affects model performance, with the gaussian-filtered KNN achieving the best result, reaching 92% accuracy, and balanced macro and weighted F1-scores of 0.92-0.93. Confusion-matrix analysis revealed minor misclassifications among iPhone models with similar sensor characteristics, while other devices were accurately identified. The findings confirm that gaussian filtering improves feature consistency and that KNN’s distance-based classification exhibits robustness across heterogeneous image sources. However, the study is limited to a single heritage object and a restricted number of devices, which may affect generalizability. The proposed framework provides a reproducible and interpretable method that supports digital authenticity verification and aligns with UNESCO’s vision for open, transparent cultural heritage preservation.
The volatile nature of financial markets requires sophisticated tools that integrate advanced analytics with accessible interfaces to facilitate informed investment decisions. This research introduces Insight Invest, an intelligent investment assistant that combines sentiment analysis with time-series forecasting to deliver comprehensive stock market insights. The platform introduces the emotional quotient (EQ), a novel metric derived from the sentiment analysis of financial news, to quantify market sentiment and align it with historical stock price data. Leveraging long short-term memory (LSTM) models, the system provides precise predictions of future stock trends. Automated data collection and processing are achieved through a Flask-based backend, while an OpenAI-powered chatbot delivers intuitive interpretations of predictions and EQ values. The user-centric design, implemented using Next.js, ensures a seamless and responsive experience. By integrating state-of-the-art machine learning techniques with intuitive interfaces, Insight Invest bridges the gap between complex predictive analytics and practical usability, offering a robust framework for informed investment strategies.
Automated student attentiveness estimation is a fundamental component of intelligent e-learning systems and adaptive classroom analytics. Traditional convolutional and recurrent architectures often struggle to model long-range temporal dependencies and complex inter-modal relationships inherent in engagement behavior. To address these limitations, this paper proposes a cross-modal attention fusion framework built upon a vision transformer (ViT) backbone for robust student attentiveness estimation. The proposed architecture leverages patch-based visual encoding through a ViT to capture global spatial dependencies, while behavioral cues such as gaze direction, head pose, and blink dynamics are embedded into a shared latent representation space. A cross-modal multi-head attention mechanism is introduced to dynamically learn interactions between visual and behavioral modalities, replacing static weighted fusion strategies. Temporal dynamics are modeled using a Transformer encoder, enabling effective long-range sequence modeling without recurrent dependencies. Experimental evaluation on a benchmark attentiveness dataset demonstrates superior performance compared to CNN–LSTM-based models, achieving improved accuracy, F1 score, and robustness under challenging lighting and occlusion conditions. Ablation studies validate the contribution of cross-modal attention and transformer-based temporal modeling. The proposed framework maintains real-time feasibility while significantly enhancing discriminative capability.
Parking demand continues to rise as private vehicle use increases, making timely information about available spaces essential for efficient parking management. Many existing monitoring approaches still rely on fixed slot sensors or visual detectors that report accuracy without examining how confidence settings affect the final availability decision. This work investigates Faster region-based convolutional network (Faster R-CNN) with a ResNet-50 backbone for image-based parking availability detection using a public parking-lot dataset annotated in Pascal visual object classes (VOC) format. The experiment evaluates several confidence thresholds to determine how each setting changes the balance among accuracy, precision, recall, and F1-score. The most balanced setting was obtained at a threshold of 0.5, where the model achieved 95% accuracy and 97.3% for precision, recall, and F1-score. These results show that threshold configuration is an important factor in reducing missed detections and false alarms, although validation using real campus CCTV data and direct comparison with lightweight detectors remain necessary before practical deployment.
Image super-resolution (SR) is essential in applications such as surveillance, medical imaging, and remote sensing, but existing deep learning (DL) models often require high computational resources and struggle to recover fine details in lightweight architectures. Although feedback and attention based methods have shown improvements, they still lack an effective combination of efficient feature refinement, edge enhancement, and low parameter complexity. To address this gap, we propose a lightweight parallel feedback network (LPFN) that combines three key components: a feedback block for repeated feature refinement, a dispersion-aware attention residual block (DARB) for highlighting important spatial and channel details, and EdgeNet for edge sharpening for sharper boundaries. These components are supported by curriculum reinforcement learning (CRL), an adaptive training strategy that gradually improves the model’s learning behavior. Instead of relying on a fixed loss function, LPFN uses a dynamically learned global feedback loss to refine reconstruction quality at each stage. Experiments on DIV2K and Flickr2K show that LPFN achieves higher PSNR and SSIMscores while keeping the model lightweight and efficient. This study emphasizes an effective lightweight feedback framework, an enhanced attention and edge-refinement mechanism, and an adaptive learning strategy that improves both accuracy and stability under different degradation conditions.
Moore’s Law has driven the development of very large-scale integration (VLSI) technology, allowing continuous transistor scaling to increase speed, density, and performance. However, as two-dimensional (2D) integrated circuits (ICs) near their physical and performance boundaries, and 2.5D ICs still face interconnect delay and power issues, three-dimensional (3D) integration has become a practical solution. In 3D ICs, multiple active layers are vertically stacked and connected via through-silicon vias (TSVs), providing short, high-bandwidth interconnects between layers. Electrical TSVs are essential for signal transmission, but also cause noise coupling between adjacent TSVs, where an aggressive TSV can induce interference in a nearby TSV. This coupling can impair signal integrity, increasing delay and power consumption. To mitigate this, low-dielectric-constant (low-k) materials are used to reduce capacitive coupling. In this study, materials such as benzocyclobutene (BCB), Perylene-N, and Teflon AF 1600 are compared with conventional SiO₂. Generally, TSVs are two structures — single-liner and stacked-liner — which are analysed at 10 GHz and 1 THz frequencies. At 10 GHz, the single-liner structure incorporating SiO₂ exhibits a noise reduction of about 6.56 dB, whereas the stacked-liner configuration using Teflon AF 1600 provides a noticeably greater reduction of 8.40 dB. As the operating frequency increases to 1 THz, the advantage of the low-k dielectric becomes more evident, yielding 9.63 dB noise reduction for the single-liner and 12.04 dB for the stacked-liner structure. These results indicate that low-k materials effectively suppress capacitive coupling and mitigate high-frequency interference in 3D ICs. The stacked-liner design contributes additional isolation by creating a secondary dielectric barrier, which further minimizes electric field interaction between neighboring interconnects. Thus, the integration of low-k dielectrics with optimized liner architectures significantly enhances signal integrity and overall electromagnetic performance in advanced high-frequency 3D IC systems.
Nowadays, smart homes have become quite complicated systems. Thus, an appropriate technique for controlling all those devices is necessary, especially considering that certain nodes are likely to be broken. In that connection, we have proposed two algorithms related to Gaussian mixture models (GMM): GMM equal and GMM unequal. They were compared with graph neural network (GNN) equal, GNN unequal, and the LucasWheel algorithms. The peculiarity of the GMM equal algorithm consists in the fact that all clusters should have similar sizes and shapes, which is quite useful for routing and balancing purposes, while the clusters in the GMM unequal algorithm can have various sizes and shapes depending on the data distribution. All five models were analyzed using 843 nodes, where failure rates ranged from zero to fifty percent. The surprising outcome of this analysis is that GMM equal performed better than the other four models in every aspect. Efficiency was steady and steadily increased in accordance with the rising failure rate. The Wiener index gradually fell from its initial value to nearly zero, suggesting a dense connection among the nodes and an evenly spread-out network. Furthermore, GMM equal attained the highest modularity among the five models at every failure level. In combination, these results indicate that GMM equal is the most balanced topology, with the best balance between reliability, efficient communication, and scalability when applied to the internet of things and smart homes.
Visual question answering (VQA) is a challenging research area that enables machines to answer natural language questions based on visual content by jointly understanding images and text. Conventional VQA systems typically produce a single answer for each image–question pair. However, many real world visual questions are ambiguous or complex, allowing multiple valid answers to exist. This systematic literature review (SLR) focuses on multi answer VQA systems and the use of object detection, following the PRISMA 2020 guidelines. We analyzed 58 peer-reviewed journal articles retrieved from the Scopus database published between 2020 and 2025. Ten of these studies clearly stated that generating multiple answers was their main goal. Forty-eight others indirectly supported answer variability by using object-based or multi-instance reasoning. Through this review, we examine the current methodologies for supporting multi-answer generation, including model architecture, datasets, and evaluation metrics. Most multi answer generation approaches utilize attention mechanisms, graph neural networks, and transformer-based models. Additionally, we propose a taxonomy of multi-answer VQA organized along four dimensions. Limitations are identified in datasets and evaluation metrics (i.e., answer ambiguity/subjectivity). Future research should focus on improving model interpretability and designing an evaluation framework that incorporates subjective and context-sensitive responses.
Water quality monitoring is vital for protecting aquatic ecosystems and ensuring sustainable water resource management. Traditional manual sampling methods are often costly, time-consuming, and unsuitable for real-time assessment. This study presents a newly designed solar-powered IoT-based water quality monitoring system for remote and continuous data collection. The system utilizes an ESP32 microcontroller integrated with pH, temperature, and total dissolved solids (TDS) sensors, powered by a 10W solar panel. Data is transmitted to a cloud-based platform wirelessly, enabling remote access and visualization via a mobile app. Performance evaluation included descriptive statistics and one-way ANOVA across four sampling sites. ANOVA results showed statistically significant differences (p < 0.05) in water quality parameters among locations, confirming the system’s sensitivity. Sensor accuracy was validated against standard meters, revealing mean relative errors below 5% for pH and TDS. The system reliably provides real-time, accurate data, supporting proactive water quality management. Integrating IoT with renewable energy offers a cost-effective, scalable, and energy-efficient solution for environmental monitoring in remote or resource-limited areas.
The fast development of internet of things (IoT) networks has led to an increased probability of cyberattacks. Intrusion detection systems (IDS) are needed for identifying unauthorised access and malicious activities in such dynamic environments. However, existing machine learning (ML) models failed to handle the complexity and variability of modern cyber threats. In this work, a hybrid deep learning (DL)-based anomaly detection model is presented for IoT cybersecurity. The model combines three types of features: (i) supervised feature extraction using linear discriminant analysis (LDA) to extract the most discriminative features, (ii) unsupervised feature learning through autoencoders to capture latent representations of the input data, and (iii) statistical features such as mean, variance, skewness, and kurtosis to learn input characteristics. The fused feature matrix is fed into a learning based echo state network (LBESN) for final detection. The parameters of the LBESN model are tuned using black eagle optimizer (BEO). Experimental results on standard intrusion detection datasets such as UNSW-NB15, KDD99, and InSDN show that the proposed model achieves superior performance in terms of accuracy, precision, recall, and F1-score compared to conventional DL techniques.
Strawberries are severely affected by the main fungal diseases such as anthracnose fruit rot, grey mould, and powdery mildew, directly reducing commercial yield and post-harvest quality. In this paper, we propose an enhanced hybrid deep-learning method by combining graph neural networks (GNNs) and multi-layer perceptrons (MLPs) for effective strawberry disease detection in real environments of fields. This dataset contains images acquired from both public domain repositories as well as farm operational settings, encompassing three classes of diseases with varying lighting conditions, background noise, and occlusion. A structured pre-processing workflow comprising contrast enhancement, denoising, and synthetic hyperspectral simulation is used to enhance subtle lesion features and stabilize the subsequent feature extraction. The hybrid GNN–MLP framework possesses the merits of spatially local lesion topology and two-point contextual information, which can improve disease classification compared with common CNN-based structures. 5-fold cross-validation shows that the model holds 93.59% accuracy and a macro F1-score of 90.39%, showing good generalization even with heterogeneous input regimes. Transparent models use local interpretable model-agnostic explanations (LIME) and gradient-weighted class activation mapping (Grad-CAM) in combination, highlighting disease-relevant areas that give an interpretable rationale for each prediction. To conclude, the developed system offers an accurate and explainable solution that has a computationally efficient commitment for real-time monitoring of disease in smart agriculture settings, particularly on low-cost hardware assets.
Cognitive radio (CR) technology is an adaptive, intelligent radio and network technology that can automatically detect available channels in a wireless spectrum. Spectrum sensing is the most important component in CR due to its ability to sense and recognize parameters related to the radio channel characteristics. However, there are some spectrums that are not used known as spectrum holes. It is challenging to accurately identify these spectrum holes, especially when employing traditional energy detection techniques, which suffer from incorrect threshold selection at low signal-to noise ratio (SNR) levels. This work suggests a wavelet-based spectrum sensing technique in conjunction with an enhanced thresholding method to improve detection accuracy and decrease noise to overcome this constraint. MATLAB simulations are used for evaluating three threshold functions: hard, soft, and improved. The results indicate that the improved threshold achieves superior denoising performance and a higher detection probability compared to the traditional energy detection method. In this study, the energy detection technique was also implemented for comparison with the wavelet-based approach. The findings reveal that wavelet-based sensing consistently provides a higher detection probability (𝑃𝐷𝐸𝑇), demonstrating its effectiveness and reliability for cognitive radio (CR) application.
Originating in India, Ayurveda is an ancient medical system focused on holistic healing that considers the mind, body, and spirit. This study utilizes knowledge graph (KG) technology to develop a KG model for an Ayurveda question-and-answer system. The system includes modules for knowledge extraction from चरकसंहिता, कायहचहकत्सा, भैषज्यरत्नावली and द्रव्यगुण संग्रि, construction of KG from this extracted knowledge and construction of AI supported Question answer system. In the methodology, domain-specific KG is constructed in Neo4j. Entities such as diseases (Vyadhi व्याधी), symptoms (Lakshana लक्षण), doshas (दोष), herbs, and treatments are incorporated. Advanced Sanskrit natural language processing (NLP) pipelines using ByT5-Sanskrit, SanskritBERT, and fine-tuned BioBERT facilitate named entity recognition (NER) and relation extraction. Graph-based reasoning models such as graph attention networks (GAT) and graph reasoning enhanced language models (GREASELM) enhance multi-hop reasoning across Ayurvedic concepts. Evaluation was conducted using a gold-standard annotated dataset of Charak Samhita verses mapped to disease–symptom–treatment relationships. Performance metrics included precision, recall, F1-score, mean reciprocal rank (MRR), and overlap coefficient. Superior accuracy can be seen in the proposed model as compared to baseline BERT-QA and subgraph QA approaches. This research has integrated Sanskrit computational linguistics and KG science. The approach mentioned in this paper has mentioned a framework that is scalable, interpretable and culturally significant. With the focus on Ayurveda, the methodology also mentions the potential for developing cross-cultural medical questions–answering systems, thereby bridging ancient wisdom with modern technological approaches.
