Dementia, particularly Alzheimer’s disease (AD), poses a major global health challenge, currently affecting over 55 million people and projected to reach 152 million by 2050. While neuroimaging-based approaches have demonstrated high diagnostic performance, there is growing interest in inexpensive, non-invasive alternatives such as handwritingderived digital biomarkers. Handwriting analysis, as provided by the DARWIN dataset, can capture subtle motor and cognitive decline using widely available digital devices. Recent machine learning (ML) and DL studies on handwriting-based AD detection show promise but often rely on complex architectures, large datasets, and provide limited transparency regarding feature importance. In this work, we propose an interpretable deep learning framework for AD diagnosis using the DARWIN handwriting dataset. Our approach integrates Recursive Feature Elimination (RFE) with a Feedforward Neural Network (FNN) to identify the most discriminative handwriting features, reduce dimensionality, and enhance generalization. Remarkably, the model achieves strong performance with only 20 selected features, demonstrating robustness in small-sample learning scenarios, as evidenced by a 10-fold cross-validation accuracy of 0.9833 ± 0.0356, ROC-AUC of 0.9987 ± 0.0039, F1-score of 0.9841 ± 0.0331, recall of 0.9818 ± 0.0545, and precision of 0.9889 ± 0.0333. Unlike data-intensive CNN, RNN, or Transformer models, the FNN attains perfect accuracy, sensitivity, and precision on DARWIN while remaining lightweight and scalable. To strengthen clinical trust, we incorporate SHAP-based explainability to highlight critical handwriting biomarkers and optimize assessment protocols by identifying less informative tasks. Comparative evaluations against established ML and DL baselines confirm consistent performance gains and enhanced interpretability, underscoring the potential of handwriting-based approaches for practical and transparent early AD detection.
Alzheimer's disease (AD) is a progressive neurodegenerative disease that requires early diagnosis to effectively treat and slow down the progression of the disease. To deal with all these issues, we come up with the Self-Aware Quantization Adaptive Federated Transfer Learning framework with FedProx-Optimized Inception-Transformer Networks (SAQ-AFTL-ITN) that is capable of finding accurate AD diagnosis, without hindering the privacy of patients' data. The SAQ-AFTL-ITN framework achieves phenomenal computational efficiency by slashing model size by 45% and communication overhead by 35%, breaking the bottleneck for decentralized healthcare environments. By using FedProx, the stabilization of the training on non-independent and identically distributed (non-IID) datasets is achieved, while TL helps to improve the generalization power of the model, particularly for clients having very limited data. Without noticing, this research shows that this framework, by combining Inception modules with Transformer encoders and using advanced aggregation methods, has already attained state-of-the-art results on the OASIS dataset, attaining an accuracy of 98.0%, precision of 97.6%, recall of 98.15%, and an F1-score of 98.0%. This work unveils the game-changing capabilities of FL integrated with SAQ, thereby forming new benchmarks for systems for medical AI that are privacy-preserving, scalable, and efficient.
Extractive Question Answering (EQA) in the biomedical domain presents significant challenges due to complex medical terminology and intricate relationships between entities. Existing systems often perform suboptimally compared to their counterparts in general domains. Current biomedical QA models struggle with accurately identifying named entities and extracting meaningful relationships within medical contexts, thereby limiting their practical utility. This study introduces a novel transformer-based approach for biomedical EQA that incorporates Named Entity Recognition (NER) to identify key medical terms, Relation Extraction (RE) to understand their interconnections, and a BiLSTM layer to enhance contextual comprehension. These components are integrated within an ensemble framework that leverages pretrained biomedical language models. Our methodology systematically evaluated six transformer encoder models and identified Bio+ClinicalBERT as the optimal foundation for developing three progressively advanced biomedical EQA models. The final model — Bio+ClinicalBERT_NER_RE_BiLSTM — incorporating ensemble learning techniques, demonstrated substantial performance improvements: an F1 score of 91. 69%, exact match of 88. 35%, and lenient accuracy of 0.84, marking a 5. 45% performance gain over baseline methods. These results highlight the model’s superior ability to navigate complex biomedical language and extract accurate information from medical texts. The developed system directly benefits biomedical researchers, clinical practitioners, medical students, and healthcare professionals by delivering accurate and rapid access to precise information within extensive medical literature. By bridging the gap between intricate medical texts and effective information retrieval, this work significantly advances automated biomedical knowledge extraction and supports evidence-based medical decision-making.
This paper proposes a novel approach for improving chest X-ray image classification by integrating self-supervised representation learning with a new margin-based ensemble classifier. A major limitation of supervised learning in medical image analysis is its strong dependence on large labeled datasets, which are expensive and time-consuming to obtain. To address this issue, we leverage self-supervised learning to exploit abundant unlabeled X-ray images and learn informative visual representations before fine-tuning with limited labeled data. In addition, we introduce a new classification algorithm, termed Random Forest of Biggest Margin Trees (RF-BMT), designed to enhance multi-class classification by combining large-margin decision boundaries with ensemble learning. RF-BMT is trained on features extracted from a linear fine-tuned Momentum Contrast (MoCo) model with a ResNet50 backbone, and serves as an alternative to the conventional softmax classifier in deep networks. This design improves classification robustness and provides stronger generalization, particularly in limited-data settings. We evaluate the proposed approach on chest X-ray datasets containing multiple conditions, including normal, COVID-19, edema, mass-nodule, and pneumothorax cases, under varying proportions of labeled training data. Experimental results demonstrate that the proposed method consistently outperforms standard classifiers and significantly improves the performance of the MoCo-based representation. In particular, RF-BMT achieves 86.2% accuracy using only 5% of the labeled training data (3,467 images), corresponding to a 7.4% improvement over the baseline, highlighting its effectiveness in limited-data scenarios. When trained with the full dataset, the model reaches 88.4% accuracy, demonstrating its robustness and scalability. These results show that combining self-supervised learning with the proposed RF-BMT classifier provides an effective and practical solution for medical image classification with limited labeled data.
This paper addresses the formal verification of Reconfigurable Discrete-Event Control Systems (RDECSs), modeled by Reconfigurable Timed Net Condition/Event Systems (R-TNCESs). Formal verification ensures system correctness, but faces challenges due to the complexity and redundant calculations introduced by reconfigurability. This complexity, exacerbated by internal similarities within system configurations, and external similarities between different systems, often leads to state explosion during model checking, complicating analysis. The novelty of the proposed ontology alignment approach lies in its ability to reduce verification complexity by identifying internal and external similarities, thus minimizing redundancy. A key contribution is the enhancement of ontology alignment approaches by introducing a new similarity function tailored for RDECSs, coupled with a Multi-Criteria Decision Analysis (MCDA) method for objectively weighting this function. Importantly, all similarity scores are computed deterministically in closed form from ontology content; the MCDA weighting step is applied once, non-iteratively, to the resulting decision matrix — this approach is therefore fundamentally different from iterative machine-learning-based methods. This approach significantly increases scalability and accuracy, overcoming limitations of existing solutions. The added value of this work is reflected in the substantial gains it offers in formal verification efficiency. Numerical validation on two benchmark production systems, FESTO MPS and Threading Hole System, shows significant improvements. While precision remains at 1.0, recall improves by 13.33%, enabling the detection of more relevant similarities and reducing the state-space size. This reduction leads to notable decreases in execution time, demonstrating the effectiveness of the proposed approach.
Accurate differentiation of mononuclear cells is essential for diagnosing hematological conditions. Nonetheless, their similar morphology can lead to misclassification. The application of artificial intelligence enables the extraction of discriminative features from cell images through advanced mathematical modeling. While a compact model is necessary to enable dynamic-monitoring in clinical practice, achieving high accuracy typically requires large, well-balanced datasets, presenting significant challenges in balancing model complexity and classification performance. This study compares traditional features like Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG), and Gray-Level Co-occurrence Matrix (GLCM) with transfer learning using fine-tuned Vision transformer (ViT) for capturing features of images. Using a dataset of 4,315 images focusing on mononuclear cells, we integrated class balancing techniques and tested models with ten classifiers. Results showed HOG with XGBoost achieved 0.77 accuracy, while fine-tuned ViT with augmentation and XGBoost reached 0.82. The computational intensity of ViT is offset by its real-time monitoring potential.
The increasing sophistication and frequency of cyberattacks make Intrusion Detection Systems (IDSs) essential, yet many machine learning-based IDSs studies still rely on legacy benchmarks such as KDD99, raising concerns about real-world generalization. This study optimizes and comparatively evaluates supervised IDS models on KDD99 and BKIDS2025, a contemporary dataset collected from a production university network (1,463,847 labeled flows over 30 days) covering modern protocols and usage patterns, including TLS 1.3, DNS-over-HTTPS, and IoT traffic. Five supervised models - Logistic Regression (LR), Random Forest (RF), XGBoost (XGB), Voting, and Stacking - are assessed under a unified pipeline using precision, recall, F1-score, and false alarm rate. Ensemble methods consistently outperform individual classifiers across both datasets. The Stacking Classifier (SC) achieves the best results, reaching 99.8% recall on KDD99 and 96.0% recall with a 97.2% F1-score on BKIDS2025. A Wilcoxon signed-rank test supports the statistical significance of the ensemble advantage, and runtime profiling indicates sub-millisecond inference latency, enabling real-time deployment. The performance drop on BKIDS2025 highlights the limitations of benchmark-only evaluation and motivates routine validation on up-to-date, real-world datasets for robust IDS design.
Image segmentation is an essential process in fungi microscopic image analysis that visually represents fungi morphology. Recent research indicates that deep learning excels in image segmentation for microscopy images. U-Net is a standard method for deep image segmentation, but challenges arise due to limited training data on fungi images. Researchers are modifying the U-Net architecture to address these challenges and enhance segmentation results. The modifications focus on the encoder component to improve feature extraction, skip connections to enhance spatial features, and data augmentation to expand data diversity. This study proposes an enhanced VGG-UNet (VU-Net) approach called Attention VGG-UNet (AVU-Net), which combines hard and soft attention using attention gates (AGs) and augmentation for fungi microscopy image segmentation. This study utilized several data augmentation methods to find the best result. By utilizing data augmentation, Focal loss, and the AdamW optimizer, AVU-Net shows improved segmentation results with better details and reaches a mIoU score of 0.833. The results show that VGG is the most significant enhancement for the model, while AGs significantly improve VU-Net. However, AVU-Net, which implements soft attention, becomes less effective when hard attention is added through data augmentation.
Recognizing natural enemies is crucial in ecological management and agricultural pest control, but existing methods face challenges in feature extraction and classification accuracy. This paper proposes a novel approach to enhance feature extraction based on Transformer models, leveraging their powerful global attention mechanism, representation learning capabilities. The proposed method integrates the strengths of convolutional neural networks (CNNs) and Transformers to improve the robustness of recognition in diverse environments. A significant contribution of this work is the development of a standardized benchmark dataset, carefully curated to represent a wide range of natural enemy species and environmental conditions. The dataset includes high-quality annotated images, ensuring its applicability in training and evaluating deep learning models. Experimental results demonstrate that the Transformer-based feature extraction model outperforms state-of-the-art methods not only in terms of computational efficiency but also in terms of accuracy and robustness. This research provides a comprehensive approach to the recognition of natural enemies, with potential applications in sustainable agriculture and monitoring biodiversity.
The study examines in detail how deep learning models (MobileNet, ResNet50, EfficientNetB4, InceptionV3, Xception, and DenseNet201) support the automated detection of vitiligo in medical images. Further, this research also seeks to design a solid transfer learning algorithm to get optimum results by joining DenseNet and Firefly Optimization, which is known as Firefly-Optimized DenseNet201 and it efficiently deals with classifying a range of images from the dataset. The dataset was obtained from Kaggle, which is based on two classes, i.e. vitiligo and healthy. Training the model on 3,628 skin images was accomplished, where the accuracy and loss were reviewed on the training and validation sets as each epoch finished. MobileNet reached high training accuracy quickly (93.86%), but it started overfitting after epoch 4. ResNet50, on the other hand, displayed steady validation accuracy of 87.4%. Using bio-inspired methods to tune hyperparameters, the Firefly-Optimized DenseNet201 achieved both high training accuracy (97.77%) and validation accuracy (89.58%). This model may be useful as a computer-aided tool for dermatologists, particularly in areas where there are shortages and prompt and precise vitiligo detection is crucial. The findings suggest useful ways to use deep learning in dermatology by focusing on both efficient processing and correct results.
Deploying deep neural networks (DNNs) on resource-constrained devices requires hybrid device-cloud inference in which end-to-end latency and data transmission must be optimized jointly. Existing Dynamic Split Computing (DSC) methods primarily target latency and overlook the computation-communication trade-off; as a result, varying bandwidth and batch size can lead to suboptimal operating points (e.g. excessive transmission for marginal latency gains). We propose Genetic Algorithm Multi-Objective (GA-MO), a genetic-algorithm framework that constructs Pareto fronts across diverse bandwidth scenarios and supports adaptive runtime selection. GA-MO jointly reduces inference latency and transmission load to satisfy application-specific performance requirements, including common real-time constraints. Beyond single-objective baselines, GA-MO is further shown to outperform a weighted-sum multi-objective approach, highlighting the effectiveness of evolutionary search in exploring the split-computing design space. Experiments on EfficientNet-B0 and VGG16 show that GA-MO outperforms DSC, achieving 55-60% lower latency and 52-62% bandwidth savings, with a 100% win rate across all tested conditions. These results indicate that GA-MO is an effective and scalable approach to multi-objective optimization for hybrid edge-cloud inference.
Breast cancer remains a leading cause of cancer-related deaths among women, highlighting the need for accurate computer-aided diagnosis systems (CADs). Convolutional neural networks (CNNs) have demonstrated substantial progress in medical image analysis, significantly improving diagnostic accuracy. This paper introduces Tri-ResNet, a triple-input model composed of parallel fine-tuned ResNet-based branches using transfer learning (TL) for efficient breast cancer classification. The model simultaneously processes full mammogram images (FMs), regions of interest images (ROIs), and contrast-enhanced ROI images (CLAHE-enhanced ROIs) using Contrast-Limited Adaptive Histogram Equalization (CLAHE). Extensive experiments were conducted using multiple pre-trained models across single-input and multi-input architectures. Tri-ResNet achieved outstanding results on the Mini-DDSM, MIAS, and INbreast datasets, with peak performance on MIAS reaching 99.62% accuracy for normal-abnormal classification and 99.14% for benign-malignant classification, while maintaining competitive results on Mini-DDSM and INbreast. The model consistently outperformed single-input models and state-of-the-art approaches, demonstrating the effectiveness of multi-input CNNs for enhancing automated breast cancer diagnosis.
The graph coloring problem (GCP) is a classical NP-hard problem that aims to assign different colors to adjacent nodes while minimizing the total number of colors used. While previous studies have used graph neural networks (GNNs) to solve GCP, they rely only on randomly initialized node features or a trainable embedding layer, leaving other alternative node feature extraction methods unexplored. Therefore, this study explores eight node feature extraction methods, including positional and structural node features. We assess their impact on GNN performance for GCP and provide insights into why certain methods outperform others. Across 12 COLOR graphs and three large citation graphs, experimental results show that both the trainable embedding layer and node2vec achieve the strongest performance. Under different hyperparameter settings, embedding layer demonstrates consistent effectiveness in minimizing conflicts across GNN architectures, while node2vec demonstrates greater average performance and stability on large graphs. Compared to the embedding layer baseline, node2vec reduces mean conflicts by 43.63% and standard deviation by 27.23% on large citation graphs. However, their performance gains involve a computational trade-off: embedding layer requires backpropagation, and node2vec requires pre-computation for its biased random walks. Furthermore, positional node features give better prediction performance than structural ones, having approximately 9.3 & times; lower mean conflicts in the best case.
The Internet of Health Things (IoHT) is a network of healthcare equipment, software, and systems that enable remote monitoring and healthcare services. Real-time health data are gathered via sensors. Even IoHT offers many benefits for modern smart healthcare, security concerns are increasing since IoHT devices lack appropriate processing power, storage capacity, and self-defense capabilities. In the healthcare sector, the use of Machine Learning (ML) for malware detection is vital for saving patients sensitive data. Therefore, it is essential to improve the accuracy and effectiveness of detection methods. ML models have been utilized to enhance the efficiency of malware detection. The main objectives of the attackers are to obtain personal information and take advantage of device flaws. Scientists are also devising diverse methods for identifying and analyzing malware to address these challenges. Given the continuous introduction of new malware by developers, it is highly tough to construct comprehensive algorithms for detecting such malware. Researchers have developed several ML and Deep Learning (DL) algorithms. The precision of these models will mainly contingent upon the amount of the training dataset. In addition, our work is divided into three primary stages: feature selection, prediction, and pre-processing. This work introduces feature selection technique that integrates two approaches, the first one Pearson correlation, to assess the correlation between features and identify significant features and Embedded method. These selected features are subsequently utilized in a classification model. Our method utilizes a soft voting classifier that combines multiple machine learning models (decision tree, logistic regression, gradient boost, random forest, and support vector machine) to detect malware. This approach creates a single model that incorporates the strengths of the combined models, resulting in the highest prediction accuracy. The proposed methodology surpasses previous research by reaching a 99.6% accuracy rate, an F1 score of 0.9972% a recall rate of 0.9998, a precision rate of 0.9947.
Protein-protein interactions (PPIs) are essential to understand biological processes and cellular functions. Experimental methods for identifying PPIs are often time-consuming and costly, motivating the development of accurate computational prediction approaches. In this paper, we combine embeddings from the state-of-the-art protein language model ESM-2 with handcrafted descriptors including short-linear motifs representation based on the ELM database. The integration of both deep learning-based and biologically meaningful handcrafted features aims to capture global and local characteristics of protein sequences. These extracted features are subsequently fused and used to train machine learning classifiers to determine interaction likelihoods between protein pairs. Experimental results on benchmark PPI datasets demonstrate that our hybrid feature strategy significantly improves prediction performance, highlighting the complementary strengths of learned embeddings and handcrafted descriptors. This work provides a robust and extensible framework for sequence-based PPI prediction, paving the way for scalable and accurate interactome mapping.
Deep learning has emerged as a transformative approach in medicinal plant identification, addressing the critical need for accurate and scalable solutions to support biodiversity conservation, traditional medicine, and sustainable healthcare practices. This systematic literature review examines 30 papers on deep learning for medicinal plant identification, revealing diverse approaches across global contexts. Convolutional neural networks emerge as the primary technique, achieving high accuracy, particularly with leaf-based identification. Data collection methods vary, with manual fieldwork predominating. The review highlights challenges in scaling to larger species sets and using crowdsourced data, though strategies like data augmentation show promise. Plant state and maturity impact model performance, warranting further investigation. The geographical distribution of studies emphasizes the global relevance of this research, with India and China contributing the most. Mobile applications offer potential for deployment and data collection but lack robust user feedback mechanisms for model refinement. The review identifies gaps in continuous model updating and suggests exploring incremental and zero-shot learning. Overall, the field shows promise but requires more balanced datasets and context-aware approaches to maximize real-world impact in medicinal plant identification.
Nowadays, interpreting human emotions through speech has attracted great attention in human–computer interaction and artificial intelligence. Speech emotion recognition (SER) systems have become a significant field of research. SER is one of the interesting directions in speech processing, is to predict the expressed emotional state. SER systems encounter numerous challenges, such as the availability of appropriate emotional databases, the identification of suitable speech features, and the choice of the appropriate classification method. SER systems are mostly implemented in English, French, German, Indian, and Chinese languages. However, SER for the Arabic language is still in the growing phase. In this work, a literature review on the SER in Arabic has been presented in terms of emotional databases, speech features, and classification algorithms. This review contributes to filling the gap in the works on emotion recognition available in the Arabic language and constitutes a valuable resource for researchers in this field.
This study investigates the predictive power of learning management system (LMS) log data for student performance in higher education. Analyzing interactions from 114 students in a sports pedagogy course, we compared linear regression (LR), random forest regression (RFR), and support vector regression (SVR), each paired with mutual information (MI) and backward elimination (BE) feature selection. Results show LMS log data alone can effectively predict final grades, with SVR[Formula: see text]BE performing best ([Formula: see text], MAE [Formula: see text] 4.54). Feature selection, particularly BE, consistently improved model performance across all algorithms. Key findings include: LMS interactions strongly predict academic performance; SVR outperforms other algorithms in capturing complex educational data relationships; and BE’s superiority highlights the importance of feature interactions. This research advances educational data mining (EDM) by identifying optimal modeling approaches for LMS data, contributing to the development of early warning systems in online and blended learning environments.
Phishing is a common malicious cybercrime in which attackers trick people into revealing their sensitive information. Phishing attacks are serious, as they permit the attacker to perceive and steal the victim's personal and sensitive information, such as credit card numbers and passwords, in a scam way while the victim is browsing the phishing website. Security experts are responsible for developing and continually improving their algorithms to save the community and guard its information. Traditional security techniques fail to do this effectively; experts are looking for new and robust methods. In this research, we use machine learning (ML) to improve phishing detection tasks. This technology efficiently influences the recognition of hidden patterns in large data inputs. The CMLM algorithm, used for detecting phishing website attacks, suffers from a high rate of indiscernibility relations, which reflect the lack of understanding of the outputs. Additionally, it does not account for imbalanced data. This paper proposes two new versions of the CMLM algorithm that effectively address these issues by integrating methods such as K-means clustering, Stability-correlation and correlation (ScC), rough set (RS) theory, principal component analysis (PCA), decision tree (DT), and deep learning (DL) in a controlled manner. The results show that the proposed methods demonstrate higher accuracy in detecting phishing websites than CMLM, achieving accuracies of 100%, 99.96%, and 99.81% across three key datasets. Compared to CMLM, the improvement margins are 0.47%, 3.08%, and 3.45% for DS1, DS2, and DS3, respectively.
Emotion detection on social media, especially related to mental health, has emerged as a significant topic in natural language processing (NLP). This study fine-tunes Transformer-based models - IndoBERT, RoBERTa, and DistilBERT - to identify emotions linked to depression in Indonesian tweets. A dataset of 5271 tweets was compiled using domain-specific keywords derived from mental health literature and expert inputs. Preprocessing follows standard NLP practices to ensure linguistic consistency and minimize noise. Several feature extraction techniques, including TF-IDF, Bag of Words (BoW), and Word2Vec, were evaluated using traditional machine learning (ML) classifiers under a 10-fold cross-validation scheme. The best performance was achieved by DistilBERT combined with TF-IDF and support vector machine (SVM), yielding an F1-score of 0.699 and accuracy of 0.705. Although the results remain moderate compared to high-resource languages, they mark a meaningful step toward modeling emotional indicators of depression in Indonesian social media contexts. The study underscores the importance of domain-specific and culturally aware approaches in advancing NLP applications for mental health.
