The rise of digital healthcare systems has generated large, fragmented medical data for enhanced illness prediction. However, modern centralized machine learning algorithms distribute sensitive patient data, raising privacy, security, and legal concerns. Federated learning allows collaborative model training while protecting raw healthcare records, but heterogeneous medical data distributions, limited feature representation, and insufficient secure and reliable model-sharing mechanisms among healthcare institutions often present challenges. This study proposes a Blockchain-Enabled Federated Learning (BCFL) framework for privacy-preserving multi-disease prediction by integrating several complementary optimization and deep learning techniques. The framework employs the Humboldt Squid Optimization Algorithm (HSOA) for optimal feature selection, the Geometric Algebra Transformer (GATr) for enhanced feature representation, the Federated Graph Multilayer Edge Attention Network (FGMEAN) for graph-aware federated disease prediction, and the Great Wall Construction Algorithm (GWCA) for adaptive hyperparameter optimization. Blockchain technology is incorporated to strengthen security and data integrity within the federated learning environment, thereby supporting secure and privacy-preserving multi-disease prediction. The proposed framework is evaluated using the Heart Disease, Diabetes, and Chronic Kidney Disease datasets. The experimental results demonstrate prediction accuracies of 99.3%, 99.5%, and 99.3% for heart disease, diabetes, and chronic kidney disease, respectively. The framework also lowered latency, communication overhead, execution efficiency, and security compared to current techniques. Component-wise study shows that feature optimization, geometric feature embedding, graph-aware federated learning, and adaptive hyperparameter tuning improve predictive performance. Experimental results show that the BCFL-FGMEAN-GWCA framework provides a secure, privacy-preserving, and precise multi-disease prediction solution in distributed healthcare settings, enabling next-generation intelligent healthcare systems.
Neural Machine Translation (NMT) has achieved state-of-the-art performance using autoregressive Transformer. However, the trade-off between translation quality and decoding speed remains a central challenge. This problem is further intensified for low-resource languages such as Urdu, where limited parallel data and complex linguistic structure result in poor translation quality and reduced model effectiveness. To overcome this problem Non-Autoregressive Transformer (NAT) and Non-Autoregressive Transformer with Iterative Refinement (iNAT) are proposed. This paper investigates NAT and iNAT for low resource language Urdu-English machine translation. We analyze translation accuracy and their effectiveness for low-resource language pairs. Experimental results show that both NAT and iNAT outperform the baseline Autoregressive Transformer (ART) in terms of translation accuracy. However, the overall translation quality remains limited due to the complex linguistic structure and resource scarcity of the Urdu-English pair. To address this limitation, we further refine these architectures by introducing an additional graph-based learning layer that captures latent relationships among tokens through attention-driven graph representations. The proposed graph-enhanced model learns contextual dependencies by dynamically constructing graphs over attention representations. Experimental evaluation demonstrates that the proposed architecture improves the translation performance of both NAT and iNAT models. On the Urdu-English translation task, the proposed model achieves a 3.0% higher BLEU score than the baseline models. The findings highlight the significance of incorporating graph-based layer to enhance translation quality, making NMT systems more effective for low-resource and linguistically complex languages such as Urdu.
Falls remain a major healthcare concern, particularly among older adults, since they frequently result in serious injuries, prolonged rehabilitation, and reduced quality of life. Although numerous inertial sensor-based fall detection methods have been reported, many are validated on a single dataset, limiting their ability to generalise across diverse sensing environments in real-world applications. This study proposes a cross-dataset Adaptive Hybrid Ensemble Framework (AHEF) for robust wearable fall detection. The framework combines heterogeneous Machine Learning (ML) classifiers Random Forest (RF), XGBoost (XGB), Extra Trees (ET), and Elastic Net Logistic Regression (ENet-LR) within a probability-calibrated stacking architecture, followed by adaptive decision refinement using activity-aware thresholds and temporal confirmation. Performance was evaluated using three publicly available benchmark datasets (SisFall, MobiFall & KFall) under both subject-independent cross-validation and Leave-One-Dataset-Out (LODO) protocols. Experimental results indicate that the proposed framework achieved 98.7% accuracy, 99.1% sensitivity, and 98.3% specificity during within-dataset evaluation while maintaining strong cross-dataset performance, with average sensitivity exceeding 98% across unseen datasets. The framework also reduced the False Positive Rate (FPR) to 1.7%, compared with 4.0–6.4% obtained by the individual base classifiers, without compromising detection sensitivity. Overall, these findings demonstrate that combining cross-dataset evaluation with adaptive ensemble learning improves both prediction reliability and generalisation, supporting deployment of wearable fall detection systems in heterogeneous real-world environments.
Security is one of the most severe problems in the Internet of Things (IoT) environment. Early detection of attacks is crucial, as it can significantly enhance network security. Convolutional neural networks (CNNs) represent a significant advancement in image classification; however, the application of CNNs to the classification of tabular data remains less explored for Intrusion Detection Systems (IDSs) compared to their success in image-based tasks. In this paper, we develop and evaluate a novel method for the classification of data using a modified VGG16 model by transforming tabular data into images. This transformation handles each block of samples as an image. A modified VGG16 model is then trained to classify the transformed images. We applied this method to the CICIoT2023 cybersecurity dataset, and the model achieved an accuracy of 98.05% with precision, recall, and F1-score of 98.15%, 98.05%, and 98.04%, respectively. The experimental results show that the proposed method achieves superior detection performance and greater classification robustness compared to other state-of-the-art methods. The significance of this study lies in detecting intrusions in IoT networks effectively and efficiently by utilizing tabular to image transformation and VGG16 model.
Clickbait headlines have become a persistent issue in online media, where attention often takes priority over informational accuracy. While numerous computational approaches have been proposed to address this issue, the field remains fragmented across datasets, modelling strategies, and linguistic representations. This study presents a systematic literature review of clickbait detection research published between 2016 and 2025, following the Protocol, Search, Appraisal, Synthesis, Analysis, and Report framework to ensure a structured and reproducible review process. A total of 646 records were initially identified from Scopus, IEEE Xplore, Web of Science, and SpringerLink, from which 43 peer-reviewed journal articles were selected for in-depth analysis. Across these studies, a clear shift is evident from traditional feature-based machine learning toward deep learning and transformer-based models. Despite this progression, linguistic cues remain central, either through explicit feature engineering or implicitly learned representations. However, the review reveals a persistent gap: linguistic feature modelling and embedding-based approaches are often developed in isolation. To address this limitation, this study proposes a conceptual framework that integrates language-aware preprocessing, explicit linguistic features, and hybrid embedding strategies within a unified framework. By synthesising these components, the framework offers a more comprehensive perspective on clickbait detection, positioning it as an integrated linguistic and contextual modelling task rather than a purely classification problem. The findings and proposed framework provide a foundation for developing more robust, interpretable, and adaptable clickbait detection systems, particularly in multilingual and low-resource settings.
