Macrofungi signify an extensive reservoir of bioactive metabolites. Still, their chemical description is stalled by fragmented analytical pipelines and low reproducibility. This study introduces SysFungiNet, a unified, FAIR-compliant systems-bioinformatics framework designed to accelerate bioactives discovery in non-model fungi. Unlike existing pipelines, SysFungiNet integrates LC-MS/MS metabolomics, transcriptomics, and genomic data with explainable artificial intelligence and molecular docking validation. By reconstructing biosynthetic pathways for Ganoderma lucidum and Craterellus cornucopioides, the framework achieved a Pathway Completeness Index of 0.86. An ensemble machine-learning model predicted bioactivity with an F1-score of 0.91, identifying 312 annotated metabolites. Crucially, decision-making was transparent, with SHAP analysis identifying specific chemical substructures driving predicted immunomodulation. The framework prioritized five high-confidence candidates, including a putative novel terpenoid, Cornucopiolide, which showed high binding affinity (−9.4 kcal/mol) to human immune receptors in silico. Benchmarking against MetaFungi and PhytoOmics demonstrated superior annotation accuracy and reproducibility. SysFungiNet offers a scalable, open-source ecosystem for transforming fungal bioprospecting into an evidence-driven process.

Fatigue-related impairment is a major contributing factor to road accidents; however, detecting early visual indicators of driver tiredness remains challenging under realistic driving conditions. This study introduces an artificial intelligence–based system for distinguishing between alert and fatigued drivers using facial images captured in natural driving environments. A total of 41,793 annotated facial images from the Driver Drowsiness Dataset (DDD) were used in the experiments. Although the dataset reflects realistic driving scenarios captured by dashboard-mounted cameras, the proposed system was evaluated offline and not tested in live traffic environments. Deep visual features were extracted using the SqueezeNet architecture and subsequently classified using three supervised learning models: Artificial Neural Networks (ANN), Random Forests (RF), and Support Vector Machines (SVM). Among the evaluated classifiers, ANN achieved the highest performance with an accuracy of 99.97%, followed by RF with 99.78% and SVM with 96.33%. The results indicate that combining lightweight deep feature extraction with classical machine learning classifiers can yield highly accurate fatigue detection while maintaining computational efficiency. The proposed framework provides valuable insights into the development of efficient, real-time driver fatigue monitoring systems with potential applications in accident prevention and road safety enhancement.

Generative AI models can now produce artwork that is virtually indistinguishable from human-created art, posing critical verification challenges for art galleries, educational institutions, and digital content platforms. Existing detection approaches face notable limitations: handcrafted feature–based methods offer interpretability but achieve limited accuracy, while deep learning approaches typically require substantial computational resources and provide minimal explanation for their predictions. We propose an attention-guided fusion framework that integrates Discrete Cosine Transform (DCT)–based frequency-domain features with deep learned representations through a learned attention mechanism, enabling both improved detection performance and interpretable decision-making grounded in signal processing theory. The attention module dynamically weights each feature modality based on input-specific reliability patterns, allowing adaptive fusion across diverse artistic styles and generation methods. We evaluate the proposed framework using two convolutional backbones: the lightweight MobileNetV2 for efficiency-critical deployment scenarios and the higher-capacity ResNet50 for settings where computational resources permit stronger feature extraction. Experiments are conducted on 18,288 artwork images spanning traditional paintings, digital illustrations, and outputs from multiple generative models, using stratified train–validation–test splits and five random seeds for statistical robustness. Under frozen-backbone settings, MobileNetV2-based attention fusion achieves an F1-score of 90.1%, while ResNet50-based attention fusion reaches 90.5%. When backbones are fine-tuned end-to-end, incorporating handcrafted features via attention fusion yields consistent performance gains: MobileNetV2 improves from 93.9% to 94.5% F1-score (+0.6%), and ResNet50 improves from 94.2% to 95.1% F1-score (+0.9%). Feature importance analysis further reveals that low-frequency DCT energy is the most discriminative handcrafted feature, confirming that frequency-domain characteristics effectively distinguish AI-generated from human-created artwork across diverse artistic styles. These results demonstrate that attention-guided fusion of signal processing features and deep learned representations provides consistent accuracy–efficiency benefits across both lightweight and heavyweight architectures, offering a practical and interpretable solution to the emerging challenge of AI-generated artwork detection.

Retrieval-Augmented Generation (RAG) enhances the capabilities of Large Language Models (LLMs) by integrating external knowledge retrieval into the generation pipeline, enabling responses that are grounded, adaptive, and up to date. While RAG can improve factual accuracy compared to models relying solely on pre-trained data, its effectiveness in practice depends strongly on the quality, relevance, and interpretability of retrieved context, and does not eliminate hallucinations entirely. Recent architectures such as Fusion-in-Decoder, Atlas, and ColBERT-RAG demonstrate measurable gains in retrieval precision, scalability, and cross-domain generalization. However, persistent challenges remain, including retrieval noise that can override model reasoning, hallucinations that persist even with high-quality evidence, latency constraints that hinder real-time deployment, and fragile domain adaptation. Moreover, although partial metrics and task-specific benchmarks exist, the absence of a unified evaluation framework for retrieval–generation grounding and robustness complicates fair comparison and reproducible progress. Rather than offering an exhaustive survey, this review provides a focused analytical perspective on retrieval design and architectural evolution in RAG systems. It consolidates representative architectures while critically examining structural limitations related to retrieval–generation coupling as a design choice, context over-reliance, and privacy-preserving computation. Building on these insights, the paper outlines future research directions, including structured knowledge integration via GraphRAG, modular agent-based orchestration in Agentic RAG, improved retrieval filtering, and unified evaluation methodologies. As RAG architectures continue to evolve rapidly in a pre-standardization phase, a more analytically grounded understanding of their design trade-offs is essential for advancing trustworthy and domain-adaptive language systems

This review examines the rapidly expanding landscape of Generative Artificial Intelligence (GenAI) in healthcare, focusing on how models such as GANs, VAEs, diffusion models, and large language models are being explored across medical imaging, clinical documentation, synthetic data generation, drug discovery, and decision-support workflows. Despite GenAI's growing influence, persistent challenges, including limited annotated datasets, concerns over model generalizability, privacy risks, and the opacity of generative architectures, underscore the need for careful evaluation and governance. Accordingly, this study aims to map current applications, assess methodological and ethical constraints, and identify future research opportunities. Using a structured search across ScienceDirect, Scopus, and other sources, the study follows a structured narrative review complemented by quantitative descriptive analysis of the literature. The review also adopts PRISMA-guided screening and standardized data extraction, the review synthesizes evidence from 110 studies published up to October 2025. The findings indicate that the literature frequently reports improvements in imaging quality, data augmentation, molecular modeling workflows, and clinical documentation through generative approaches, particularly in technically constrained settings; however, evidence of clinically validated impact remains uneven across domains. While issues of bias, hallucination, and limited interpretability persist as significant obstacles, imaging-focused applications appear comparatively more mature than decision-support and patient-level modeling tasks. Across domains, diffusion models are commonly associated with higher visual fidelity in biomedical image generation, whereas LLMs demonstrate promise in narrative-oriented tasks but require stronger factual grounding and external verification mechanisms. Overall, the evidence suggests that GenAI’s potential in healthcare is highly context-dependent and contingent upon robust validation frameworks, transparent governance, and human-in-the-loop oversight. The review concludes that responsible integration of GenAI—guided by ethical, legal, and clinical safeguards, will be essential for ensuring safe, equitable, and sustainable adoption in healthcare research, delivery, and policy.

Face detection in low-light conditions remains challenging due to underexposure, noise, and unstable contrast, which significantly degrade the performance of convolutional-based models. Conventional enhancement techniques, such as Histogram Equalization, Contrast-Limited Adaptive Histogram Equalization (CLAHE), and Low-Light Image Enhancement (LIME), often improve brightness but introduce visual artifacts and lack robustness for face detection. This study proposes Low-light Unified Multi-stage Enhancement Network (LUMEN), a lightweight multi-stage image enhancement pipeline designed as a preprocessing module to improve RetinaFace-based face detection under low-light conditions. LUMEN integrates Multiscale Retinex with Color Restoration, adaptive gamma correction, CLAHE-based local contrast enhancement, controlled image fusion, and Non-Local Means denoising to jointly stabilize illumination, preserve texture, and maintain visual naturalness. The method is evaluated on a low-light subset of the Human Faces Object Detection Dataset using RetinaFace. Detection performance is assessed using detection rate and confidence score, while visual quality is evaluated using no-reference metrics such as the Natural Image Quality Evaluator (NIQE) and the Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE). Experimental results show that LUMEN achieves a face detection rate of 91% with a high confidence score (0.9545), rep-resenting an improvement of 33 percentage points over raw low-light images (58%), while maintaining detection performance comparable to CLAHE, which ranks as the second-best method, LUMEN delivers superior perceptual quality, evidenced by the lowest BRISQUE score, as well as more stable visual appearance with reduced noise amplification and fewer contrast artifacts in low-light facial regions. LUMEN also achieves the lowest BRISQUE score, indicating superior texture preservation and perceptual stability. Ablation studies confirm that Retinex and CLAHE are the most critical components for detection robustness, while gamma correction, fusion, and denoising mainly contribute to visual naturalness. These results demonstrate that LUMEN provides an effective and practical preprocessing solution for low-light face detection.

This study investigates how Artificial Intelligence (AI)-based Tutoring Systems can be used to provide secondary school students with personalized learning feedback and to enhance their academic performance. The two main empirical objectives were to assess individualized formative assessment in classroom settings and to examine teachers’ and students’ perceptions of AI adoption. A mixed-methods quasi-experimental design was employed, integrating quantitative pre-/post-tests, interviews, and focus groups across six schools (n = 600). This approach enabled triangulation between measurable learning outcomes and contextual perception data for robust validation. Quantitative data were analyzed using Python (t-tests, ANCOVA), while thematic coding in NVivo was applied to qualitative data. Expert review, pilot testing, and Cronbach’s α (>0.80) were used to validate the instruments and ensure reliability, including pre-/post-tests and engagement scales. Findings revealed that students who received AI-based interventions achieved significantly higher academic performance (Cohen’s d = 1.05) and engagement (d = 0.72) compared with control groups. Teachers with AI exposure reported greater preparedness (mean = 3.4) and fewer perceived barriers. The study provides empirical evidence on the pedagogical viability of AI tutoring in under-resourced contexts, contributing to self-regulated and socio-technical learning theories. It also recommends enhanced systemic teacher education, ethical leadership, and structural support to foster equitable adoption of AI in education. The findings carry strong implications for policy development and educational innovation in promoting data-driven, inclusive learning within Nigeria’s secondary education system

Sentiment analysis is a critical task in natural language processing (NLP) that classifies text into sentiment categories, such as positive, negative, or neutral. This task is particularly challenging for languages like Persian due to the complexity of their linguistic structure and the scarcity of high-quality labeled datasets. Previous studies on Persian sentiment analysis have largely relied on TF-IDF representations or deep learning models, often overlooking handcrafted statistical and stylistic features that capture subtle textual patterns. This limitation reduces their effectiveness, especially when dealing with informal or noisy text data. Experiments in this study were conducted on a dataset of Persian product reviews from Digikala.com, labeled according to user ratings to indicate positive, negative, or neutral sentiment. In this paper, we propose a novel approach to Persian text sentiment analysis by combining statistical and stylistic (surface-level) features with traditional text-based features such as Term Frequency–Inverse Document Frequency (TF-IDF). Unlike prior works that rely solely on TF-IDF or deep learning representations, our method integrates stylistic and statistical cues to capture expressive nuances in informal Persian text. Additionally, the Support Vector Machine (SVM) classifier is optimized using RandomizedSearchCV to enhance performance. The proposed system utilizes both statistical and textual features to improve classification accuracy. We compare its performance with four baseline models, i.e., Naïve Bayes, Logistic Regression, Random Forest, and Decision Tree, that rely solely on TF-IDF features. The experimental results demonstrate that the proposed approach outperforms the baseline models in terms of accuracy, F1-score, recall, and precision. Specifically, the proposed system achieved the highest accuracy (0.8354), significantly improving negative sentiment detection while maintaining strong performance in positive sentiment classification.

Bone fractures are among the most common global health problems and require rapid and accurate diagnosis to prevent complications and reduce healthcare costs. Traditional diagnostic methods rely on the manual examination of X-ray images by experienced radiologists, a process that is time-consuming and prone to human error, especially in emergency departments with high patient volumes. This study proposes a lightweight hybrid approach that integrates deep learning–based feature extraction with classical machine learning algorithms for automatic bone fracture detection. The dataset used comprises 4,840 fractured and 4,623 non-fractured bone X-ray images obtained from the Kaggle platform, in which data augmentation was previously applied by the dataset developer. A compact convolutional neural network architecture, MobileNetV2, was employed exclusively for deep feature extraction. The extracted feature vectors were then used to train and evaluate five ensemble learning algorithms: Random Forest, Extra Trees, Gradient Boosting, AdaBoost, and Bagging Classifier. Model performance was assessed using 10-fold cross-validation based on accuracy, precision, recall (sensitivity), and F-score metrics. Experimental results show that the Extra Trees algorithm achieved the highest performance with 99.14% accuracy, 99.22% precision, 99.03% recall, and 99.12% F-score, while also yielding the shortest training time of 199 seconds. Random Forest followed with an accuracy of 98.38%. The superior performance and computational efficiency of the Extra Trees algorithm demonstrate its suitability for real-time clinical applications. These findings confirm that combining deep feature extraction with ensemble learning can effectively support automated diagnostic systems for bone fracture detection, enabling healthcare professionals to make faster and more reliable decisions.

We introduce DiFACE2 causal DAG, a framework for generating Diverse, Feasible, and Actionable Counterfactual Explanations under causal Directed Acyclic Graph (DAG) constraints. Leveraging background knowledge from an educational intervention program in post-conflict Nigeria, we construct an application-specific causal DAG to identify causal relationships and immutable features during counterfactual generation for black-box ML predictions. Using this DAG, we develop a causally constrained sparse ML model that integrates prior frameworks to generate diverse, feasible, and actionable counterfactuals for the Strengthening Education in Northeast – Early Grade Reading Assessment (SENSE-EGRA) dataset. DiFACE2 produces up to four counterfactuals per prediction, demonstrating high feasibility and actionability for categorical variables. While numeric features occasionally yield infeasible recommendations, this is mitigated by additional non-causal constraints. Although the sparse ML model yields lower accuracy and F1 scores than benchmark frameworks, this stems from the higher complexity and noisiness of the real-world dataset rather than methodological weakness. This trade-off underscores that meaningful, interpretable counterfactuals are preferable to near-perfect accuracy on simplified benchmarks. Therefore, when fairly evaluated on the same complex dataset, DiFACE2 outperforms prior methods across all counterfactual quality metrics.