Peripheral neuropathy is a disease that significantly impacts the quality of life, and it is a common side effect of chemotherapy treatments for cancer patients. Therefore, continuous monitoring of drugs, their outcomes, and associated symptoms is crucial for detecting any signs or symptoms of neuropath, which is typically linked to chemotherapy agents. This research paper aims to investigate the relationship between chemotherapy exposure and the development of neuropathy through the emergence of associated side effects. It utilized the fundamentals of AI for intelligent analysis; for instance data preprocessing, feature extraction, pattern matching, signal detection, and classification. This target was achieved by using data that belong to USA Food and Drug Administration’s Adverse Event Reporting System (FAERS). Intelligent analytics was conducted on a dataset of pharmacovigilance reports collected from 2015 to 2026. This database underwent to data mining steps which were the preprocessing to remove missing and duplicate values, and standardize text by normalization. Exposure to chemotherapy was identified by feature extraction and matching a set of keywords of the patient’s drugs recorded in FAERS database. As a classification, and to compare neuropathy-related interactions with chemotherapy exposure, a 2 X 2 correlation table was constructed. Non proportional analysis was performed using Reporting Odds Ratio (ROR), with a 95% confidence level. Chi-square was used for evaluation. Over 527,000 reports were analyzed from FAERS database to investigate the relationship between chemotherapy exposure and neuropathy. Among all the data, 4.519 reports showed a correlation between the two, with 259 reports demonstration a correlation. ROR ratio was 6.28, with a 95% confidence interval between 5.5 and 7.1, indicating a significant disproportion between chemotherapy exposure and peripheral neuropathy. The Chi-square value (p < 0.001) showed a strong correlation between the analyzed variables. The paper results demonstrated a significant association between chemo exposure and the occurrence of neuropathic reactions in FAERS database. The paper highlighted the importance of ongoing pharmacovigilance data and continuous assessment of neurotoxicity resulting from chemotherapy exposure.
In this work we propose a new advanced scheme of the encryption and decryption of the satellite images using a dynamically generated substitution box (S -box) matrix integrated with the RC4 stream cipher. The satellite images whose data is usually huge and contains strong spatial correlations among adjacent pixels are challenging to conventional encryption techniques that do not, in many cases, provide a high level of security. To reduce such problems, the developed method will use a dynamic S-box based on a secret key, which will increase the nonlinearity and the strength of the system in confusion. To achieve adequate diffusion of the image data, the RC4 algorithm is used to produce a pseudo-random keystream. In the encryption process, pixel values are initially operated on by an XOR operation with the keystream, and then operated on by a nonlinear substitution via the dynamic S -box. These steps are reversed in the decryption process, which uses the inverse S -box, and reusing the key stream, which would guarantee the exact reproduction of the original image, without the loss of information. The security performance of the framework was tested using comprehensive experimental tests in Python in the Google Collaboratory environment. We looked at such metrics as uniformity of histograms, entropy of information, adjacent-pixel correlation coefficients (horizontal, vertical, and diagonal), and indicators of differential attacks (NPCR and UACI). The findings reveal that the encrypted images have a very uniform distribution of histograms, the entropy measures are also near to the theoretical maximum of 8 and the correlation coefficients are virtually near to zero. Moreover, the scheme is highly resistant to the differential cryptanalysis with NPCR values of over 99% and UACI values of nearly 33%. Therefore, the developed method is effective in terms of the ability to unite high security and sensible calculation, making it an secure storage and transmission of satellite imagery during remote sensing processes.
Despite the emergence of Large Language Models (LLMs), graph-based extractive summarization and specifically the TextRank algorithm have retained architectural significance in high-stakes domains including legal, medical, scientific, and journalistic text processing. This survey presents a systematic review of 121 references spanning 2002 to 2026, tracing the evolution of TextRank-based approaches into hybrid LLM pipelines and advancing three qualified arguments.First, TextRank provides architectural advantages that persist even as LLM context windows grow: source-level provenance that grounded generation methods only partially replicate; approximately 75% token consumption reduction at any fixed per-token rate; and deterministic, auditable extraction consistent with emerging regulatory transparency requirements in legal and clinical domains.Second, LLMs exhibit hallucination rates of 15–30% on domain-specific text and produce no source audit trail; TextRank pre-filtering substantially mitigates both limitations, while grounded generation alternatives address them at the generation layer. Third, hybrid TextRank→LLM pipelines achieve ROUGE-1 = 0.41 on CNN/DailyMail and 0.54–0.76 on domain-specific benchmarks; cross-benchmark ROUGE comparisons are not valid across different benchmark types and should not be interpreted as a unified performance ranking. The performance comparisons are derived from the reviewed literature; the layered pipeline architecture, limitations taxonomy, and research agenda are original contributions of this survey."
The short-term residential electricity demand forecasting is very important for sustainable electricity use management in smart cities. The purpose of this study is to compare the artificial intelligence models for short-term electricity load forecasting in a realistic residential condition in a rigorous comparison. The study uses half-hourly smart-meter data from 10 households for two years to assess the forecasting performance. The leakage-free chronological data splitting approach was used to maintain temporal causality and reliable out-of-sample evaluation. Three machine learning models, standalone XGBoost, Long Short-Term Memory (LSTM), and a hybrid model combining LSTM and XGBoost, were systematically compared to two naive models: persistence and seasonal benchmark models. The outcomes illustrate that XGBoost can always deliver the highest prediction accuracy in most homes and the hybrid model can boost the performance of LSTM, but does not always outperform XGBoost. The errors in predictions were larger when the electric demand was higher, thus practical evaluation under realistic operating conditions is needed. The results emphasize the need for selecting the most appropriate model based on data to accurately predict energy consumption in smart homes. Moreover, predicting energy demand for individual households is difficult because of its nonlinear, stochastic and human behavior characteristics. There is, however, a limitation in the number of households and the results obtained may not be directly transferable to a larger-scale smart city setting.
This study aims to build explainable Machine Learning (ML) for classifying migraine type where classification interpretation is along with clinical protocol International Classification of Headache Disorders, 3rd edition (ICHD-3). Migraine diagnosis it a challenge due to the complex overlapping between its types. For thus, we aim to train models based on valid 400 cases collected and labeled by expert medical staff at Hospital Materno Infantil de Soledad. Dataset contains (7) classes and (24) features that describe patients’ symptoms such as aura, speech complication and pain severity. Models that used in this study vary from simple models such as Logistic Regression (LR) and more complex such as Random Forest (RF) and Multi-Layer Perceptron (MLP). Top achieved models are LR, MLP and Support Vector Machine (SVM) with F1 score around (87%). A more in-depth study was conducted on these top-performing models to interpret their outcomes using SHAP (SHapley Additive exPlanations) values and LR coefficients. Results show that LR, besides its high-performance achievement, it also aligned with ICHD-3 clinical guidance as both SHAP values and coefficients shown highest match in their ranked features.
Convolutional neural networks (CNNs) have achieved promising results in the iris recognition by learning discriminative features directly from images. This study proposes a feature-level combination approach that combines representations extracted from two pretrained CNN architectures, Inception-v1 and ResNet-101. Iris images are first segmented and normalized, after which features are extracted independently from the two networks. The resulting feature vectors are concatenated to form a combined representation, which is classified using a support vector machine (SVM). The method is evaluated on three public datasets: CASIA-Iris-Interval (3.0), CASIA-Iris-Thousand (4.0), and IITD Delhi Iris. The proposed combination method achieves recognition accuracies of 96.50%, 97.63%, and 98.42% on these datasets, respectively, outperforming the corresponding single-network baselines in the reported experiments. These results indicate that combining complementary deep representations from Inception-v1 and ResNet-101 can improve iris-recognition performance. However, the evaluation protocol, multiclass SVM configuration, and train/test splitting procedure should be reported in sufficient detail to support reproducibility.
Embryo quality assessment in IVF remains challenging because conventional morphology-based grading is partly subjective and may vary across reviewers, laboratories, and imaging conditions. This study proposes XMV-EmbryoNet, an explainable deep learning framework for binary embryo-quality classification from microscopy images. The model uses three complementary image representations: a global view, an adaptive center-focused view, and a patch-bag view. The center-focused representation is generated by a learnable adaptive cropping module rather than manual embryo centering. These views are processed through a shared compact backbone and integrated using multi-scale decision heads, adaptive scale-gating, view-gating, and patch-attention pooling. The model was trained and evaluated on the public Hung Vuong Hospital embryo microscopy dataset, where embryos were classified into good-quality and poor-quality groups. XMV-EmbryoNet achieved 0.88 accuracy, 0.76 precision, 0.76 recall, 0.76 F1-score, 0.88 ROC-AUC, and 0.80 PR-AUC. Lowering the decision threshold increased recall to 0.85 and F1-score to 0.78. Within this single-dataset evaluation, the proposed model showed improved discrimination and more stable training behavior than EfficientNet-B0, DenseNet121, and ResNet50. The main contribution is the integration of adaptive local cropping, multi-view reasoning, multi-scale fusion, patch-attention pooling, and intrinsic interpretability signals within one embryo-grading framework. Grad-CAM and internal view-, scale-, and patch-level signals provided qualitative insight into model decisions.
Recent developments in MRI processing include reconstruction, demising, enhancement, segmentation, analysis from images, and under sampled k-space data using deep learning techniques. This article review new findings in the MRI processing with the use of deep learning features and the application of brain MRI segmentation and evaluation. As per the paper, the field-wise classification will take place at each phase of the pipeline like Preprocessing, Augmentation, Model building, Post-processing and Evaluation, etc. The paper examines various critical architectural families, including CNN Vision Transformers, Hybrid CNNs, Transformer Models, and Generative Approaches. A short overview of BraTS 2019 and BraTS 2020, including a literature matrix of 35 studies. Additionally, we present quantitative comparison of 18 representative segmentation methods based on their Dice scores and Dice IoU values., When something is not directly presented. Examination of the evidence indicates that performance depends on a generation of things, including the model architecture, the way of preprocessing and loss design, data quality, and way of validating. Hybrid CNN-transformer models have reported strong performance in the brain tumor segmentation task. Nonetheless, it is important to take caution when studying one study to another one because of the different datasets and evaluation protocols. Subsequent clinical validation should focus on generalization across sites, uncertainty-aware inference and federated learning, foundation models and reproducibility.
Phishing attacks are considered one of the highest risks to cybersecurity, as they exploit users by pretending to be a valid site and gaining access to sensitive data. Traditional methods of detection, including blacklist-based systems, cannot uncover newly created phishing sites (zero-day attacks), need regular updates to their databases, and can generally not offer real-time protection. The paper suggests implementing a hybrid phishing site detection system involving machine learning with rule-based analysis to improve the accuracy and reliability of detection. The model proposed employs feature extraction using URLs, such as lexical and structural features like URL length, frequency of special characters, use of IP addresses, and occurrence of suspicious keywords. Three supervised machine learning algorithms, Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression, where RF was indicated to achieve the highest performance. Moreover, a rule-based system of brand impersonation detection is also implemented to make the system resistant to targeted phishing attacks. Experimental findings prove that the hybrid methodology enhances classification accuracy and lowers false negatives compared to conventional ones. The proposed hybrid model represents an accuracy of 96%, with a recall of 94% and F1 of 95% ; it can be deployed in real-time and gives interpretable results by using confidence scoring. The solution suggested is a feasible and effective method of phishing detection and can be expanded to more sophisticated intelligent cybersecurity systems.
The traditional networking topologies of Internet of Things (IoT) systems are based on explicit topology building, neighbor discovery, and routing state maintenance. Although they are effective in less dense and stable networks, these methods are inefficient in dense, dynamic, and unstructured IoT networks because they increase the control overhead, energy usage, and scalability. This study proposes a topology-free networking paradigm that does not require topology awareness, routing tables, node identifiers, or neighbor discovery. Rather, data propagation is realized via implicit coordination of nodes as decided by locally visible cues, such as signal strength, reception timing, and channel contention. An opportunistic forwarding based on a suppression strategy is presented, in which each node calculates a forwarding score based on these normalized parameters and applies opportunistic forwarding to the node. The proposed method was evaluated through simulations under varying node densities and traffic demands. The performance is evaluated by means of several metrics, such as packet reachability, end-to-end delay, transmission redundancy, scalability, control overhead, and energy consumption. The findings indicate that the proposed topology-free approach offers reliable data delivery, with a substantially reduced overhead and high efficiency in dense IoT contexts. These results suggest the promise of topology-free networking as a scalable and energy saving alternative to large-scale and resource-constrained IoT systems in the future.
The title should concisely encapsulate the overall content; however, producing a strong academic title that reflects the core contribution remains challenging. This paper proposes and evaluates a pipeline-based model for academic paper title generation: abstracts collected from the NIPS dataset are preprocessed, condensed via extractive summarization, and then Standard YAKE is applied to extract weighted keywords. Titles are generated from a fixed Top-6 keyword budget while preserving the YAKE ranking order; in this work, the proposed model refers to the end-to-end pipeline configuration rather than a newly trained neural architecture. We compare four summarization algorithms (Luhn, LSA, Edmundson, and KL) based on their YAKE-weighted keywords, and adopt Luhn as default because it produces more topic-relevant YAKE-weighted keywords than the other summarizers. We illustrate the pipeline by using three qualitative examples and compare three generator-Ateeqq/keywords-title-generator, KeyToText (k2t), and GPT-2 (gpt2)-under identical Top-6 constraints. Experiments on 300 papers using both lexical and semantic similarity metrics (ROUGE-L, TF-IDF cosine similarity, BERTScore-F1, and SciBERTScore-F1) indicate that the adopted Luhn→YAKE→Top-6→LLM pipeline produces the most semantically aligned titles under identical input constraints.
Agriculture worldwide, particularly in arid and semi-arid regions, is severely affected by water scarcity, which has become a major global challenge, this paper describes development of the proposed artificial intelligence-based system, Smart Drop that optimizes irrigation water to address this issue. The system was tested and deployed through a 4.25-kilometer (internal diameter: 800 millimeters) pipe network irrigating 147 hectares of farmland at the design flow rate of 0.25 cubic meters per second as described in Case Study. The combination of machine learning models with genetic algorithms and particle swarm optimization was applied for prediction of irrigation needs and optimized irrigation water supply. Three models were established through the use of machine learning: random forest (R²=0.9965; RMSE= 0.0491 millimeters per day), Gradient Boosting (R² = 0.9976, and RMSE = 0.0487 millimeter/day) as well as neural networks models (R2 = 0.9943, and MSE = 0.0514). Compared with the traditional irrigation methods, this system reduced water use by 23.42% in irrigation (797,731) cubic meters over a period of 1,000 days), maintaining optimal crop conditions, the savings to total cost was estimated at US$121,036.62 (23.4%), indicating that the combination of reduced water use along with reduced energy consumption produced an economically viable alternative. The hydraulic performance results show acceptable pressure distribution through pipeline and overall losses within system are 1.0 meter drops across each section 9.78 kilopascals. This research is a data-driven precision irrigation solution for sustainable application in different agricultural environments.
Over the past decade, the rapid digital transformation of infrastructures to digital forms, such as cloud computing, Internet of Things (IoT), and large-scale interconnected network systems, has made the threat of cybercrime much more pronounced. Of those, Zero-Day attacks are regarded as the most serious since they are previously unseen and so the traditional signature-based intrusion detection systems are useless. This paper presents an in-depth overview of machine learning (ML) and deep learning (DL) methods of detecting Zero-Day attacks. The methodology is based on reviewing, analyzing, and synthesizing recent literature, which is applied to ML, DL, and hybrid methods, threat intelligence integration, and real-time intrusion detection systems. The findings suggest that both ML and DL methods have high detection accuracy but have a number of weaknesses including high computational complexity, data imbalance, scarce availability of labeled data, and susceptibility to adversarial attacks. Moreover, this review shows some of the main gaps in research, especially in coping with the unknown attack patterns, the development of lightweight and real-time detection models, and the enhancement of the generalization abilities. Finally, the research paper shows that it is crucial to establish adaptive, scalable, and hybrid intelligent systems to improve the detection of Zero-Day attacks. To enhance actual cybersecurity applications in the future, future studies should focus on efficient learning mechanisms, strong adversarial defenses, and data-efficient models.
Industrial surface defect classification is an important part of automated quality inspection systems. For these systems to work, they need to be able to detect surface defects accurately and efficiently to improve product reliability and reduce manufacturing costs. Traditional manual inspection methods are often time-consuming, subjective, and not suitable for fast-paced industrial environments. This study proposes a lightweight convolutional neural network (CNN)-based system for classifying industrial surface defects. The model was made to work well for classification and still be fast enough for real-world use. Keras Tuner was used in the proposed method to find the best hyperparameters. The proposed model was evaluated on the Northeastern University (NEU) Surface Defect Database using 5-fold stratified cross-validation. The results obtained from the experiments are promising since the system yields stable performance on all folds with accuracy scores of 99.72%, 99.17%, 100.00%, 99.17%, and 98.61% respectively. The mean accuracy score is calculated as 99.33%. Also, Grad-CAM visualization revealed that the network focuses on defective regions when processing an input image which supports the reliability and interpretability of the classification process.
Early diagnosis and clinical decision-making depend on accurate brain tumor classification using magnetic resonance imaging (MRI). However, traditional deep learning methods usually rely on centralized medical data, which raises privacy concerns and limits the use of distributed clinical data. This research proposes a privacy-preserving federated learning framework for MRI image-based binary brain tumor classification using a decentralized ResNet-18 architecture that enables collaborative training without sharing raw patient data. To reflect realistic clinical conditions, the framework integrates heterogeneous multi-source datasets in different image formats (PNG and JPG) and evaluates performance under both IID and non-IID settings. Experiments were conducted using the Kaggle Brain Tumor MRI dataset and Mendeley Data distributed across five simulated institutions. Within the evaluated experimental setup, the proposed framework achieved approximately 92% accuracy under IID conditions and 91.5% under non-IID settings, with an F1-score of approximately 0.90. Client-level evaluation demonstrated the model’s ability to handle data heterogeneity, while convergence analysis indicated stable training behavior across communication rounds. In addition, Grad-CAM visualization was employed to provide visual interpretability, showing that the model focuses on clinically relevant anatomical regions during prediction. Overall, the results demonstrate that combining federated learning with heterogeneous multi-source MRI data can preserve privacy, maintain robustness and interpretability, and achieve competitive classification performance, highlighting the potential of federated deep learning as a practical and scalable solution for privacy-aware medical image analysis in realistic clinical environments.
Malware detection is one of the most important cybersecurity issues because the traditional signature-based methods cannot resist polymorphic threats and obfuscated ones. This paper explores the dynamic API call sequences as behavioral characteristics and contrasts the two representation methods, integer-based feature encoding into the traditional machine learning models and graph-based models using Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT). Unlike prior studies, this paper conducts a systematic head-to-head comparison of these approachs and introduces a newly collected balanced dataset of 2,000 malware and 2,000 benign sampels, for this paper Two datasets were employed, one of which was a large public dataset with 42,797 malware and 1,079 benign samples, and the other was a novel developed dataset consisting of 2,000 malware and 2,000 benign samples that were collected according to this research by means of sandboxed execution. To facilitate 10-fold cross-validation, API calls were pre-encoded into fixed length sequences of integers and call graphs directed to allow fair evaluation. The findings indicate that ensemble and tree-based models achieved competitive results (≈92% on the public dataset and ≈90% on the novel dataset), but the graph-based ones were more accurate with GCN coming to 98.76% and GAT at 98.33%. Because graph neural networks can capture relational dependence and contextual patterns in API call behavior, they generate a richer representation and stronger categorization than integer encodings also the best feature of graph-based models is that they learn not only features but also the connectivity of API calls, which gives much richer and more accurate representation than integer-only encodings. Unlike prior studies, this work conducts a systematic head-to-head comparison of these approaches and introduces a newly collected balanced dataset of 2,000 malware and 2,000 benign samples.
We introduce a smart multi-objective optimization framework towards the performance enhancement of the 4G LTE network using an integrated Artificial Intelligence (AI) and Whale Optimization Algorithm (WOA) setup. This approach aims to integrate tested regression-based KPI prediction along with decision-variable-based meta-heuristic optimization in a reproducible experimental environment within MATLAB R2023b. A weighted multi-objective formulation for optimizing the performance of 4 key indicators, i.e., throughput, latency, packet loss ratio, and energy efficiency, is jointly executed. By comparison against existing baseline LTE configuration, we observe a 25.3% throughput improvement, 28.9% reduction in latency, 38.7% reduction in packet loss ratio, and 23.4% improvement in energy efficiency. We validate robustness with weight sensitivity analysis (±10% and ±20%), population scalability testing (N = 20, 30, 50), decay-strategy comparison (linear vs. exponential), and Pareto-front approximation (2,960 non-dominated solutions from 3,000 samples). The above results corroborate that, after approximately 70 iterations, convergence is stable and performance is consistent over the span of 30 independent runs. Thus, the proposed AI–WOA construction offers an organized and replicable process to fine-tune LTE performance optimization, a basis for downstream implementations towards a more distributed and next-generation wireless environment.
Cardiovascular diseases (CVDs) are the leading cause of mortality in the world. This paper constructs models of heart disease classification using a single-hidden-layer Multilayer Perceptron (1HL-MLP) that is trained using a bi-level model that utilizes Bayesian Hyperparameter Optimization (HPO) and three variations of evolutionary strategy-based Learner Performance-Based Behavior (LPB): LPB-MLP, aLPB-MLP and mLPB-MLP. Each of the four heart disease datasets was processed in a common pipeline based on schema alignment, median imputation, one-hot encoding, per-fold Z-score normalization, and Outof-Fold (OOF) threshold tuning and their performance was checked by stratified K-fold and external testing on an independent dataset. Results have shown that the Modified LPB model (mLPB-MLP) performed better, and it has the highest discrimination and calibration (AUC = 0.9782, AUPRC = 0.9732) and the overall accuracy (93.66%), F1-score (93.53%), recall (94%), specificity (93%), and the lowest BCE loss (0.193). These findings indicate consistent optimization processes, reasonable probability tuning and sensitivity-specificity compromise. In general, the Bayesian HPO in combination with LPB-family evolutionary training using data will lead to a clinically robust, well-calibrated, and reproducible heart disease risk prediction model.
Wireless Sensor Networks WSNs are widely adopted and cost-effective way of implementing intelligent solutions in many low-resource settings. Both known Clustering/Routing Protocols such as Low-Energy Adaptive Clustering Hierarchy LEACH and Threshold-Sensitive Energy-Efficient Network Protocol TEEN as well as other related methods such as WOA-LEACH, are severely limited by their inability to support a long term network cycle for use in real world applications. This is because they all experience rapid energy exhaustion because of their inability to adapt to varying levels of energy and traffic demand. In addition, failure to distribute energy usage evenly throughout the network. In order to resolve these issues, we will develop a new Hybrid Approach that combines the Whale Optimization Algorithm WOA for determining energy aware cluster heads, and Deep Reinforcement Learning DRL for providing an adaptive multi-hop routing protocol. The proposed hybrid DRL-WOA solution will make joint optimizations of cluster heads and routing nodes to determine optimized routes to minimize energy use while maximizing energy efficiency through the optimization of hop distances, thereby creating longer lasting and more reliable communication processes. Results from simulations run on a 100 node WSN environment demonstrate the hybrid DRL-WOA solution achieves better performance than LEACH, TEEN, WOA-LEACH and a DQN-based only routing solution, including 22% less total energy consumption, 60% extended First Node Death FND, and PDR improvements of 5-15% in comparison to each of the above mentioned base line protocols. All in all, the experimental results clearly demonstrate that the proposed Hybrid DRL-WOA approach leads to a considerable improvement of the energy efficiency, network lifetime and the reliability of data delivery of static WSNs.
In an article, the title should cover the whole content with a few important words. Several automated title-generating tools are available in the Internet. This paper is aiming to design and implementation a new model to generate title using a list of keywords. The model using a new dataset that is generated from NIPS dataset with Configure training arguments. The proposed system preprocesses keyword data, trains on a curated dataset, and produces coherent, contextually relevant titles through controlled text generation. The proposed model shows a strong generative capability by accurately producing research titles from list of important keywords. Its efficient fine-tuning strategy enables high performance with minimal training resources. The experimental results show that the proposed fine-tuned T5 title generation model can produce titles that are very close to the original scientific titles. For the paper “Learning to Play the Game of Chess”, the generated title matched the original exactly. As a result, all evaluation metrics reached 1.0, including cosine similarity, ROUGE-L, BERT_F1, and SciBERT_F1, indicating complete lexical and semantic agreement between the generated and reference titles.
