Sleep apnea is a sleep disorder that can have serious health consequences. Its detection and differentiation between the types obstructive (OSA), central (CSA), and mixed (MSA) is crucial for selecting appropriate therapy. The aim of this study was to compare three feature selection methods: Particle Swarm Optimization (PSO), Neighbourhood Components Analysis (NCA), and Principal Component Analysis (PCA) in the context of detecting sleep apnea and its types using single-channel EEG signals. In the study, the EEG signals were pre-processed, divided into 30-second segments, and analyzed using a two-stage feature extraction approach. Feature selection methods (PSO, NCA, and PCA) were then applied to reduce data dimensionality and identify the most informative parameters. Parameter optimization was also conducted for each method. Classification was performed using the k-NN algorithm. The results showed that the PSO method achieved the highest average classification accuracy of 98.03%, reducing the number of features from 379 to 134, while NCA achieved an accuracy of 97.96%, reducing the number of features from 424 to 127. Although PCA was effective in dimensionality reduction, it achieved a lower accuracy of 85.56%. The applied methods enabled clear differentiation between normal breathing and sleep apnea episodes, with classification errors occurring only in distinguishing between apnea types.

We introduce Automata Pattern (AutPat), a feature extractor for EEG, and embed it in an explainable feature engineering (XFE) pipeline.We evaluated AutPat on three tasks: EEG artifact classification, stress detection, and mental performance detection. The pipeline computes AutPat features from raw EEG, selects informative variables with cumulative weighted iterative neighborhood component analysis (CWINCA), and performs classification using a t-algorithm-based k-nearest neighbors (tkNN) classifier. For interpretability, we map the selected features to Directed Lobish (DLob) symbols and derive DLob strings and cortical connectome diagrams.The AutPat-based XFE achieved > 88% classification accuracy on all datasets. CWINCA reduced the feature space while maintaining accuracy, and the DLob layer yielded dataset-specific symbolic outputs and 8 × 8 connectome matrices.AutPat, combined with CWINCA and tkNN, provides a compact and accurate EEG pipeline with inherent symbolic explanations. The results indicate that AutPat-based XFE is a practical option for EEG analysis when both performance and interpretability are required.

Accurate and timely identification of faults in photovoltaic (PV) panels is critical for maintaining system efficiency and ensuring safe operation. In this study, a hybrid classification framework is proposed that integrates deep feature fusion with an advanced feature selection method to detect PV panel faults using thermal infrared imagery. Feature representations were extracted using four pre-trained lightweight convolutional neural networks: MobileNet, MobileNetV2, MobileNetV3Small, and MobileNetV3Large resulting in a 3840-dimensional concatenated feature vector. To reduce redundancy and improve discriminative power, the Cumulative Weight-based Iterative Neighborhood Component Analysis (CWINCA) was employed, selecting 142 informative features. These were subsequently classified using a linear Support Vector Machine (SVM). Experiments were conducted on the publicly available PVF-10 dataset, comprising 5,579 thermal images across ten fault categories. The proposed method achieved an overall classification accuracy of 86.49%, outperforming several individual CNN based architectures. The results demonstrate that combining feature-level integration with targeted selection significantly enhances classification performance while maintaining low computational complexity. This framework offers a promising and scalable solution for UAV-based PV inspection systems.

Content-based Image Retrieval (CBIR) systems have been used frequently in recent years, along with developing technology. Especially in large datasets, retrieval-based systems produce more successful results. This study created a dataset consisting of 27 different Euphorbia seed types belonging to the same genus. It is difficult for Convolutional Neural Network (CNN) architectures to produce successful results in the created dataset. In addition, the high computational and memory requirements of CNN architectures have further increased the need for CBIR systems in large datasets. Therefore, a hybrid retrieval system was developed to make inferences from 27 different seed images. In the developed system, feature extraction was performed using Darknet53, Xception, and Densenet201 architectures. These extracted features were concatenated to bring together different features of the same image. Then, unnecessary features were eliminated from the combined features with the Neighborhood Component Analysis (NCA) method. The cosine similarity measurement metric was used to measure the similarity between the query image and other images. Precision-recall curves and Average Precision (AP) metrics were used to measure the performance of the proposed retrieval-based system. In the study, an average AP value of 0.96809 was obtained. The morphology of the seeds is a critical characteristic of Euphorbia, and this work has validated the artificial intelligence methodology.

Background/Objectives: Depression is a mental disorder that can lead to self-harm or suicidal thoughts if left untreated. Psychiatrists often face challenges in identifying the most effective courses of treatment for patients with depression. Two widely recommended depression-related therapies are selective serotonin reuptake inhibitors (SSRIs) and repetitive transcranial magnetic stimulation (rTMS). However, their response rates are approximately 50%, which is relatively low. This study introduces a computer-aided decision (CAD) system designed to determine the effectiveness of depression therapies and recommends the most appropriate treatments for patients. Methods: Each channel of the EEG is plotted in two-dimensional (2D) space via a novel technique called the amplitude polar map (APM). In each channel, the 2D plot of APM is utilized to extract distinctive features via the binary pattern of five successive lines method. The extracted features from each channel are fused to generalize the pattern of EEG signals. The most relevant features are selected via the neighborhood component analysis algorithm. The chosen features are input into a simple feed-forward neural network architecture to classify the EEG signal of a depressed patient into either a respondent to depression therapies or not. The 10-fold cross-validation strategy is employed to ensure unbiased results. Results: The results of our proposed CAD system show accuracy rates of 98.06% and 97.19% for predicting the outcomes of SSRI and rTMS therapies, respectively. In SSRI predictions, prefrontal and parietal channels such as F7, Fz, Fp2, P4, and Pz were the most informative, reflecting brain regions involved in emotional regulation and executive function. In contrast, rTMS prediction relied more on frontal, temporal, and occipital channels such as F4, O2, T5, T3, Cz, and T6, indicating broader network modulation via neuromodulation. Conclusions: The proposed CAD framework holds considerable promise as a clinical decision-support tool, assisting mental health professionals in identifying the most suitable therapeutic interventions for individuals with depression.

The rapid growth in database size due to technological advances has led to difficulties in locating and accessing specific data components. While deep learning and other machine learning architectures are promising in retrieving data components, their effectiveness is more pronounced when addressing groups of diseases. On the contrary, this effectiveness decreases when large data sets are accessed. Content-based Image retrieval (CBIR) methods are used in large data sets. In this study, knee osteoarthritis detection was performed using a developed hybrid CBIR-based system. Knee Osteoarthritis is the wear and tear of the cartilage in the knee joint. Knee osteoarthritis is a disease whose incidence increases, especially after a certain age. In this study, CBIR techniques were preferred to detect knee osteoarthritis. In the proposed method, feature extraction was performed using DarkNet53, Histogram of Oriented Gradients (HOG), and Local Binary Patterns (LBP). These features are combined to leverage the benefits of different aspects of the same image. To enhance the proposed model's speed and effectiveness, a hybrid model was developed utilizing the Neighborhood Component Analysis (NCA) method. Seven different distance measurement metrics were used in the developed CBIR model. Current deep learning architectures published in the literature struggle to achieve comparable success rates in distinguishing between closely related but distinct disease groups. The study highlights the challenges that increasing class diversity poses for the performance of deep learning architectures. In addition, the developed system aims to overcome the limitations of existing deep learning models in distinguishing similar disease groups.

Extracorporeal shock wave lithotripsy (ESWL) is one of the most often employed therapy methods for managing kidney stones. In our work, we sought to assess the efficacy of the artificial intelligence model developed using non-contrast computed tomography (CT) images in predicting stone-free rates for ESWL. The main difference between this study and other studies is that it proposes an artificial intelligence-based model that predicts the success of ESWL treatment using artificial intelligence methods. Data from 910 patients who underwent ESWL between January 2016 and June 2021 were analyzed retrospectively. Since the local binary pattern (LBP) and histogram of oriented gradients (HOG) feature extraction methods gave more successful results than other methods, a new feature map was obtained using the neighborhood component analysis (NCA) dimension reduction method after combining the features obtained using these methods. Then, the reduced feature map was classified into classifiers. In conclusion, we analyzed the effect of ESWL treatment using different artificial intelligence methods and found that the prediction accuracy was 94% on average. Results were obtained from seven different convolutional neural networks (CNNs) and two textural-based models in the study. Since textural-based models achieved the highest success among these models, these models were used as the base in the proposed model. The proposed model achieved better results than nine different models used in the study. When the results obtained from the proposed hybrid model for ESWL prediction are examined, this model will guide experts in the treatment of the disease.

Abstract With the widespread application of rotating machinery in intelligent manufacturing, aerospace, and other industrial fields, accurate and reliable fault diagnosis and maintenance have become increasingly critical for ensuring system safety and operational efficiency. However, existing domain-adaptation-based cross-domain intelligent fault diagnosis methods primarily focus on achieving feature transfer at the global domain level, often overlooking the complexity, imbalance, and significant class-level variability arising from the simultaneous distribution of samples across the source and target domains. This oversight can lead to inaccurate recognition of fine-grained class-level features, thereby limiting diagnostic accuracy. To address these challenges, this paper presents a class-level domain alignment method (CDD_DANN) that combines Classifier Deterministic Difference (CDD) loss with a dual-classifier structured Domain-Adversarial Neural Network (DANN), effectively improving class-level feature alignment and transfer in cross-domain fault diagnosis. Additionally, to effectively address the challenge of sparse marginal samples at deeper levels, we propose the PMCDAN method, which replaces CDD with a proxy-based metric learning approach, Proxy Neighborhood Component Analysis (ProxyNCA), to capture deeply shared features between the source and target domains more robustly. This enables global domain alignment and class alignment under challenging conditions. Furthermore, to tackle the data imbalance, this paper incorporates a Diffusion-GAN-based fault sample augmentation method, which facilitates both domain and class-level alignment when data is scarce, thus enabling more accurate fault diagnosis. The effectiveness and superiority of the proposed approach are validated through experimental evaluations against existing methods using the Paderborn University bearing dataset and a self-collected gear fault dataset. The proposed method provides valuable insights and practical guidance for fault diagnosis in complex real-world industrial scenarios.

Effective speech emotion recognition (SER) poses a significant challenge due to the intricate and subjective nature of human emotions. Recognizing emotional states accurately from speech signals has a broad spectrum of practical applications, such as healthcare, human-computer interaction, and social robotics. This study introduces an innovative approach that merges deep learning with metaheuristic algorithms to boost the efficiency of SER systems. Specifically, a stacked autoencoder (SAE) serves as the primary model, and its performance is fine-tuned using a nature-inspired hybrid algorithm that combines particle swarm optimization (PSO) with Grass Fibrous Root Optimization (GFRO). The proposed model adeptly extracts spectral and pitch features from speech signals, encompassing spectral crest, spectral entropy, spectral flux, and harmonic ratio, to capture emotional cues effectively. The model’s performance is evaluated on a standard emotion recognition dataset, comparing with some state-of-the-art models, including Convolutional Neural Network (CNN), Support Vector Machine (SVM), Deep Learning (DL), CNN and Iterative Neighborhood Component Analysis (CNN/INCA), VGG-16 achieving high accuracy in identifying various emotional states.

There are many products obtained from grapes. The early detection of diseases in an economically important fruit is important, and the spread of disease significantly increases financial losses. In recent years, it is known that artificial intelligence techniques have achieved very successful results in image classification. Therefore, the early detection and classification of grape diseases with the latest artificial intelligence techniques and feature reduction techniques was carried out within the scope of this study. The most well-known convolutional neural network (CNN) architectures, texture-based Local Binary Pattern (LBP) and Histogram of Oriented Gradients (HOG) methods, Neighborhood Component Analysis (NCA), feature reduction methods, and machine learning (ML) techniques are the methods used in this article. The proposed hybrid model was compared with two texture-based and four CNN models. The features from the most successful CNN model and texture-based architectures were combined. The NCA method was used to select the best features from the obtained feature map, and the model was classified using the best-known ML classifiers. Our proposed model achieved an accuracy value of 99.1%. This value shows that our model can be used in the detection of grape diseases.

Abstract An injury, chronic illness, obesity, infection, and more can negatively affect the hip joint. Surgery and implant placement are the standard treatments for moderate to severe hip issues. These treatments, however, can alter a patient’s gait patterns. Gait patterns must be assessed clinically by a qualified physician and a specialized examination is required to detect and monitor these changes. By contrast, Machine Learning (ML) techniques assist in diagnosing a wide variety of anomalies and illnesses. In addition to being extremely accurate, it reduces subjectivity in clinical expert evaluations. Gait anomalies can also be quickly identified and monitored inexpensively and quickly using ML. Three open-source datasets (GaitRec, Gutenberg, and Orthoload) were utilized in this study for the gait cycle conditions for healthy control, hip surgery, and hip implant patients. This study classifies individuals into two classes: Healthy control/Gait Abnormality and three classes: Healthy control/ Hip surgery/ Hip implant by gait cycle conditions using only vertical ground reaction forces (vGRFs) from these datasets, which consist of 3D GRFs. The essential steps in data preparation include filtering, denoising, normalizing, resampling, and augmenting. The purpose of these efforts was to improve the model performance in classification and reduce biases. We used several feature extraction techniques, focusing on excluding highly correlated features. The final analysis utilized five widely recognized feature selection algorithms (Minimum Redundancy Maximum Relevance (mRMR), Neighborhood Component Analysis (NCA), Multi-Cluster Feature Selection (MCFS), Chi-square, and Relief) to arrange the features systematically. Based on a comprehensive examination of five machine learning classifiers (k-nearest neighbor (KNN), artificial neural network (ANN), decision tree (DT), support vector machine (SVM), and Naïve Bayes (NB)), The KNN classifier exhibited the highest level of accuracy. The two and three-class classification’s overall accuracy, precision, sensitivity, and F1 score are 95.48%, 96.13%, 95.48%, 95.63% and 89.18%, 89.30%, 89.18, and 87.15%, respectively. With the proposed solution, clinicians can more easily identify gait abnormalities based on vertical ground reaction forces.

Proteins are the material foundation of life, and they are organic macromolecules that make up the basic organic matter of cells. Therefore, proteins can be considered as the main bearers of life activities. Proteins are important components that make up all cells and tissues in an organism. All critical elements of an organism require the participation of proteins, and the most important thing is that they are related to life phenomena. The transportation and localization of proteins within organelles is a complex and delicate process that involves multiple steps and mechanisms. Organelle proteins are an essential element in several biological processions. In this work, we proposed TAPE_selection methods to reduce the useless information of the Tasks Assessing Protein Embed-dings (TAPE) feature in some organelle proteins, which mainly include plant vacuole proteins (PVPs) and peroxidase ones. In order to reduce the useless information, we employed some feature selection Strategies, including the Chi-Squared Test, Minimum Redundancy Maximum Relevance(mRMR), and Neighborhood Components Analysis (NCA). With the selected feature, the Proper Orthogonal Decomposition (POD) and t-distributed Stochastic Neighbor Embedding (t-SNE) were employed to reduce the reconstructed feature scale.

This study focuses on machine learning classification of the sounds of equipment operating in the construction site environment to support improved construction management and site supervision processes. The research utilizes a large, openly available, open access audio dataset of seven different types of equipment, collected in the field under real conditions in urban regeneration projects initiated after a major earthquake in Elazığ. The dataset consists of 15,588 sounds recorded from vehicles such as bulldozers, excavators, dump trucks, graders, loaders, mixer trucks and rollers used on the construction site. In the developed classification system, discriminative features were first extracted from equipment sounds using Local Binary Patterns (LBP) and statistical moments. In the feature selection stage, the Neighborhood Component Analysis (NCA) and Chi-Square (Chi2) method is applied to identify the most significant features and dimensionality reduction is achieved. In the final stage, Support Vector Machines (SVM) and k-Nearest Neighbor (kNN) classifiers are used to discriminate the equipment types with high accuracy. The findings show that the proposed method makes a significant contribution to construction management objectives such as effective monitoring of vehicles and equipment on the construction site, resource management and process tracking. In addition, the transparency and reproducibility provided by the open dataset provides a strong basis for further studies in the related field.

Background: Intestinal metaplasia (IM) is a precancerous gastric condition that requires accurate histopathological diagnosis to enable early intervention and cancer prevention. Traditional evaluation of H&E-stained tissue slides can be labor-intensive and prone to interobserver variability. Recent advances in deep learning, particularly transformer-based models, offer promising tools for improving diagnostic accuracy. Methods: We propose ViSwNeXtNet, a novel patch-wise ensemble framework that integrates three transformer-based architectures—ConvNeXt-Tiny, Swin-Tiny, and ViT-Base—for deep feature extraction. Features from each model (12,288 per model) were concatenated into a 36,864-dimensional vector and refined using iterative neighborhood component analysis (INCA) to select the most discriminative 565 features. A quadratic SVM classifier was trained using these selected features. The model was evaluated on two datasets: (1) a custom-collected dataset consisting of 516 intestinal metaplasia cases and 521 control cases, and (2) the public GasHisSDB dataset, which includes 20,160 normal and 13,124 abnormal H&E-stained image patches of size 160 × 160 pixels. Results: On the collected dataset, the proposed method achieved 94.41% accuracy, 94.63% sensitivity, and 94.40% F1 score. On the GasHisSDB dataset, it reached 99.20% accuracy, 99.39% sensitivity, and 99.16% F1 score, outperforming individual backbone models and demonstrating strong generalizability across datasets. Conclusions: ViSwNeXtNet successfully combines local, regional, and global representations of tissue structure through an ensemble of transformer-based models. The addition of INCA-based feature selection significantly enhances classification performance while reducing dimensionality. These findings suggest the method’s potential for integration into clinical pathology workflows. Future work will focus on multiclass classification, multicenter validation, and integration of explainable AI techniques.

ABSTRACT This study developed a label-tracking feature ranking method (LTFRM) to monitor spindle preload in a computer numerical control machine tool by using vibration analysis. Unlike traditional feature ranking methods, which primarily focused on discriminative power, our LTFRM evaluates both a feature’s ability to differentiate between preload conditions and the consistency between feature variation trends and predefined preload variation trends. The proposed LTFRM was validated through extensive milling experiments conducted on a three-axis machining center under seven preload conditions ranging from 120 to 2500 N. Our experimental results indicated that the proposed LTFRM outperformed conventional feature ranking methods such as random Forest, neighborhood component analysis, and Fisher’s score, achieving a root mean square error of 9.59 N and having high computational efficiency. Combining the proposed LTFRM with genetic algorithm optimization enhanced model performance, leading to higher prediction accuracy (mean square error = 12.8) than that achieved with traditional genetic algorithm – based feature selection and optimization. In summary, the proposed LTFRM offers a practical solution for real-time spindle preload monitoring, contributing to the advancement of smart manufacturing systems and predictive maintenance strategies.

Background and ObjectiveStroke ranks among the leading causes of disability and death worldwide. Timely detection can reduce its impact. Machine learning delivers powerful tools for image‑based diagnosis. This study introduces StrokeNeXt, a lightweight convolutional neural network (CNN) for computed tomography (CT) and magnetic resonance (MR) scans, and couples it with deep feature engineering (DFE) to improve accuracy and facilitate clinical deployment.Materials and MethodsWe assembled a multimodal dataset of CT and MR images, each labeled as stroke or control. StrokeNeXt employs a ConvNeXt‑inspired block and a squeeze‑and‑excitation (SE) unit across four stages: stem, StrokeNeXt block, downsampling, and output. In the DFE pipeline, StrokeNeXt extracts features from fixed‑size patches, iterative neighborhood component analysis (INCA) selects the top features, and a t algorithm-based k-nearest neighbors (tkNN) classifier has been utilized for classification.ResultsStrokeNeXt achieved 93.67% test accuracy on the assembled dataset. Integrating DFE raised accuracy to 97.06%. This combined approach outperformed StrokeNeXt alone and reduced classification time.ConclusionStrokeNeXt paired with DFE offers an effective solution for stroke detection on CT and MR images. Its high accuracy and fewer learnable parameters make it lightweight and it is suitable for integration into clinical workflows. This research lays a foundation for real‑time decision support in emergency and radiology settings.

Automatic violence detection is one of the most important research areas at the intersection of machine learning and information security. Moreover, we aimed to investigate violence detection in the context of neuroscience. Therefore, we have collected a new electroencephalography (EEG) violence detection dataset and presented a self-organized explainable feature engineering (SOXFE) approach. In the first phase of this research, we collected a new EEG violence dataset. This dataset contains two classes: (i) resting, (ii) violence. To detect violence automatically, we proposed a new SOXFE approach, which contains five main phases: (1) feature extraction with the proposed distance matrix pattern (DMPat), which generates three feature vectors, (2) feature selection with iterative neighborhood component analysis (INCA), and three selected feature vectors were created, (3) explainable results generation using Directed Lobish (DLob) and statistical analysis of the generated DLob string, (4) classification deploying t algorithm-based k-nearest neighbors (tkNN), and (5) information fusion employing mode operator and selecting the best outcome via greedy algorithm. By deploying the proposed model, classification and explainable results were generated. To obtain the classification results, tenfold cross-validation (CV), leave-one-record-out (LORO) CV were utilized, and the presented model attained 100% classification accuracy with tenfold CV and reached 98.49% classification accuracy with LORO CV. Moreover, we demonstrated the cortical connectome map related to violence. These results and findings clearly indicated that the proposed model is a good violence detection model. Moreover, this model contributes to feature engineering, neuroscience and social security.

Comparison between five pattern-based approaches for automated diagnostic classification of mature/peripheral B-cell neoplasms based on standardized EuroFlow flow cytometry immunophenotypic data.

Detection of β-Thalassemia trait from a heterogeneous population with red cell indices and parameters.

Customer churn poses a significant challenge to e-commerce businesses, impacting revenue and hindering long-term growth. Predicting customer churn is paramount for the long-term viability of e-commerce businesses, enabling proactive retention strategies and minimizing costly customer attrition. This research introduces a novel meta-model approach for e-commerce customer churn prediction. Our approach integrates multiple deep learning architectures, including TabNet, and leverages Neighborhood Component Analysis (NCA) to select the most relevant features for churn prediction. We evaluate our approach on two distinct e-commerce datasets: the Olist Online dataset and the REES46 dataset, which has not been previously used for churn prediction. By effectively combining the strengths of individual base models, our proposed approach surpasses the limitations of traditional models and accurately captures both feature-specific and hierarchical relationships within the data. Our proposed model obtained an area under the curve (AUC) of 0.98, an F1-score of 98.78%, an accuracy of 99.62%, a precision of 99.32%, Matthew’s Correlation Coefficient (MCC) of 0.987540 and a recall of 98.66% in the Olist dataset exceeding the results obtained from traditional models. An AUC of 0.95, an accuracy of 88.63%, a precision of 81.54%, a recall of 83.47%, an MCC of 0.731223, and an F1-score of 82.49% were obtained on the REES46 dataset demonstrating clear superiority over baseline models. Experimental results demonstrate superior performance on both datasets, with significant improvements over baseline and state-of-the-art methods. By enabling businesses to proactively identify at-risk customers, our model empowers them to implement targeted retention campaigns, optimize marketing spend, and ultimately enhance customer satisfaction and long-term profitability.