In this work, we aimed to develop an automated asphalt fault detection model to classify pavement cracks caused by road surface deterioration. For this purpose, we used the open-access Pavement Image Dataset and constructed a balanced dataset containing 4500 labelled images from nine pavement distress classes: reflective crack, transverse crack, block crack, longitudinal crack, alligator crack, sealed reflective crack, lane longitudinal crack, sealed longitudinal crack and pothole. Before feature extraction, all images were resized according to the input requirements of the selected pre-trained deep networks. InceptionResNetV2 and ResNet50 were then used as transfer learning models to extract deep image features. Two feature sets were obtained from each network and concatenated to form a single feature vector. To select the most informative features, four iterative feature selection methods, namely IChi2, ImRMR, INCA and IRF, were applied. The selected features were classified using different classifiers. The best result was obtained using the IChi2 selector and LDA classifier. This combination achieved 98.80% accuracy, precision, recall and F-measure. These results show that the proposed deep feature-based model is highly effective for automated pavement crack classification.

IntroductionHuntington's disease (HD), a dominantly inherited neurodegenerative disorder caused by CAG repeat expansions in the HTT gene, manifests with progressive motor dysfunction, cognitive decline, and psychiatric disturbances. While current transgenic mouse models recapitulate key pathological features, they exhibit rapid disease progression, and early behavioural phenotypes are not analyzed comprehensively to understand their progression.MethodsWe employed a high-resolution 3-dimensional motion capture and unsupervised machine learning to dissect behavioral dynamics in the R6/1 HD mouse model at 8 weeks of age, a stage analogous to human pre-diagnostic HD.ResultsThrough unsupervised learning-based clustering analysis, we identified 40 major movement categories in mice. Using a subsequent supervised learning approach, we recognized 13 fundamental spontaneous behavioral movements and identified disrupted behavioral modules in R6/1 mice, including reduced locomotor fraction, increased pausing frequency, and altered exploratory patterns. Our key findings revealed that HD mice exhibited reduced velocity and increased stride length during running and trotting behaviors, mirroring bradykinesia and gait abnormalities observed in HD patients. These mice also showed preferential exploration of the peripheral zone and decreased sniffing frequency, which might suggest that they have displayed behaviors analogous to anxiety or depression.Furthermore, an escalating frequency of pausing was observed over 30-minute sessions, suggesting early-onset motor fatigue. Additionally, lower behavioral entropy and fewer transitions from exploratory or maintenance states to locomotion were detected, pointing to executive dysfunction. A LDA classifier integrating these core behavioral metrics achieved an AUC of 0.917, surpassing the performance of traditional coarse motor assessments.ConclusionThese results establish precision behavioral analytics as a sensitive platform for detecting premanifest HD pathology, providing a framework for evaluating presymptomatic therapeutics and scientific base for developing early diagnostic and treatment strategies for HD.

Effective weed detection for precise management remains a pertinent issue in modern agriculture. In this study, hyperspectral imaging (HSI) was combined with machine learning (ML) to differentiate between peanut plants and four common weeds found in peanut fields. Several spectral preprocessing methods—Moving Window Averaging (MWA), Median Filtering (MF), Gaussian Filtering (GF), and Savitzky–Golay smoothing (SGS)—were applied. Feature selection algorithms, including Correlation-based feature selection (CFS), Principal Components Analysis (PCA), and Wrapper Feature Selection (WFS), were then used to extract the most informative wavelengths. Among the various classifiers evaluated, the combination of MF preprocessing, WFS algorithm, and LDA classifier (MF-WFS-LDA) achieved the best performance, while the WFS method selected 12 optimal wavelengths from a total of 465. The accuracy, precision, recall, and RMSE values of this model in the training stage were 99.71%, 0.997, 0.997, and 0.054, respectively. These statistics were 96.67%, 0.967, 0.968, and 0.088, respectively, in the test stage. Furthermore, it successfully differentiated peanuts from each weed species using a minimal number of optimal wavelengths. These findings highlight the potential of integrating HSI with ML for precise weed detection in peanut cultivation. However, further validation under diverse environmental and field conditions is recommended.

To construct and validate an optimal machine learning (ML)-based clinical-radiomics model integrating clinical and radiomics features for predicting recurrence risk within 2 years after radical surgery in patients with sinonasal squamous cell carcinoma (SNSCC). This retrospective study included a total of 207 patients with pathologically confirmed SNSCC who underwent preoperative MRI. Patients were divided into a training cohort (n = 151) from the headquarter hospital and an independent testing cohort (n = 56) from the branch hospital. Radiomics features were acquired from diffusion-weighted imaging (DWI) sequences. Clinical, radiomics, and combined clinical-radiomics models were constructed and validated using various ML algorithms, with the areas under the receiver operating characteristic curves (AUCs) computed to evaluate their classification performance. The optimal clinical-radiomics model that integrates multimodal features was subsequently identified. During the two-year period after radical surgery, 93 patients (44.9%) experienced recurrence at follow-up. The clinical-radiomics model, developed with seven predictive features (Range, surgical margin status, Kurtosis, T4 stage, Idn, Robust Mean Absolute Deviation, and Ki-67 levels) using ANOVA feature selection and LDA classifier, achieved optimal discriminative performance. In the training cohort, the model achieved an AUC of 0.85, in contrast to 0.76 for the clinical model and 0.79 for the radiomics model. This superiority was maintained in the testing cohort, where it achieved an AUC of 0.83, compared to 0.71 for the clinical model and 0.77 for the radiomics model. The combined model consistently outperformed the standalone models across both cohorts. The present study developed an ML-based clinical-radiomics model that exhibited promising performance for predicting recurrence risk after radical surgery in SNSCC. Given its preliminary nature, this model has the potential to serve as an intelligent decision support tool, advancing precision medicine in SNSCC management.

Colorectal cancer (CRC) remains one of the leading causes of cancer-related mortality worldwide, emphasizing the urgent need for early, non-invasive, and accessible diagnostic tools. This study aimed to evaluate the effectiveness of a microelectromechanical systems (MEMS)-based electronic nose (E-nose) in combination with gas chromatography-mass spectrometry (GC-MS) for CRC detection through sweat volatile organic compounds (VOCs). A total of 136 sweat samples were collected from 68 volunteer participants. Samples were processed using solid-phase microextraction (SPME) and analyzed by GC-MS, while a custom-designed E-nose system comprising 14 gas sensors captured real-time VOC profiles. Data were analyzed using multivariate statistical techniques, including PCA and PLS-DA, and classified with machine learning algorithms (LDA, LR, SVM, k-NN). GC-MS analysis revealed statistically significant differences between CRC patients and healthy controls (COs). Cross-validation showed that the highest classification accuracy for GC-MS data was 81% with the k-NN classifier, whereas E-nose data achieved up to 97% accuracy using the LDA classifier. Sweat volatilome analysis, supported by advanced data processing and complementary use of E-nose technology and GC-MS, demonstrates strong potential as a reliable, non-invasive approach for early CRC detection.

Uncovering the neuronal mechanisms un-derlying optimal behavioral performance is essential to understand how the brain dynamically adapts to changing conditions. In Parkinson's disease (PD), these neuronal mechanisms are disrupted and lead to impairments in motor coordination and higher-order cognitive functions. This study investigates neuronal dynamics during a lower-limb pedaling task by analyzing the dynamical entropy of EEG signals in healthy controls (HC), PD patients, and PD patients with freezing of gait (PDFOG). We examined both average entropy changes and entropy variability across trials to characterize task-specific neural adaptations across disease progression. Results showed that PD and PDFOG patients exhibited decreased levels of permutation entropy in frontal and parietal regions, which may be associated with loss of cognitive adapta-tion due to altered information processing. Additionally, Vasicek's entropy variability in both PD groups was significantly diminished in occipital and left frontal regions, suggesting reduced cognitive capacity to dy-namically allocate neuronal resources during task engagement. We extended this analysis to the classification of groups using LDA and SVM classifiers, where entropy-derived features achieved a classification accuracy of up to 96.15% when distinguishing HC from PDFOG patients. This dynamical entropic framework provides a novel approach for capturing neural complexity changes during task performance, revealing subtle cognitive-motor impairments in PD. Understanding the maintenance of cognitive information processing and flexibility in response to motor and cognitive task demands could be a useful tool to track PD diagnosis and progression in addition to resting-state analyses.

Abstract Thermal comfort is a critical determinant of human health, productivity, and well-being in indoor environments. While numerous studies have utilised electroencephalography (EEG) to explore human physiological responses to varying thermal conditions, comprehensive analyses that synthesise the effectiveness of various machine learning (ML) approaches for interpreting EEG data remain limited. To address this gap, this study compares various EEG feature sets and ML algorithms using a single EEG dataset. The dataset consists of EEG signals collected from 40 participants exposed to two distinct thermal conditions: a baseline comfortable state and an overheating state induced by wearing heavy clothing. To this end, our objective is to investigate the most pertinent EEG signal features, such as mean power density, power spectral densities, and so on, and evaluate the performance of popular machine learning models for predicting thermal comfort. We examine classifiers including Support Vector Machines (SVM), Random Forests (RF), and various neural network configurations, comparing their efficacy in analysing EEG data. The results indicate that the LDA classifier demonstrates high accuracy when using mean power density features in each 1 Hz frequency range. The SVM classifier, utilizing power density ratios of EEG frequency bands, exhibits robustness in recall and F1 scores. Additionally, the CNN classifier effectively captures complex patterns in the EEG data, showcasing the potential of deep learning methods. These Gindings contribute to the optimization of indoor environmental controls and advance the Gield of environmental engineering by providing insights into the neurophysiological impacts of thermal conditions.

Scheduled maintenance and condition monitoring of power transformers in smart grids is mandatory to reduce their downtimes and maintain economic benefits. However, to minimize energy losses during inspection, non-invasive fault diagnosis techniques such as thermogram imaging can enable continuous monitoring of transformer health with minimal out-of-service time. Deep learning (DL) has proven to be a fast and efficient intelligent diagnostic tool. In this paper, a DL-based thermography method is proposed called Trans-Light for transformers’ interturn faults detection and short-circuit severity identification. Trans-light extracts deep features from two deep layers of a convolutional neural network (CNN) rather than depending on one layer, thus obtaining more intricate patterns. Moreover, a Dual-tree Complex Wavelet Transform method is adopted which offers two enhancements. First, it acquires time–frequency knowledge besides the already obtained spatial information and second, it reduces the huge deep features dimensionality. Trans-light combines extracted deep features, then a feature selection process is applied to further reduce features’ size, thus decreasing computation burden and reducing classification and training time. To validate the proposed scheme’s diagnosis performance and robustness, different combinations of two CNN models, two feature selection methods, and six classifiers were tested, applying the proposed Trans-light framework, under noise-free and noise-existing conditions. Experimental results indicated that the combination of the LDA classifier, applied with the ResNet-18 CNN model and trained with merged deep features undergoing the chi-square (χ2) selection approach, attained superior performance under noise-free conditions. Compared to its counterparts in previous work, this configuration outperforms their performance since it uses the fewest features’ number yet maintains 100% classification accuracy. Besides, it attained robust performance under two different noise natures again with minimal features’ dimension, thus minimizing computational load and implementation complexity.

Introduction: As an important vegetable crop, cauliflower (Brassica oleracea var. botrytis) is afflicted by many diseases that significantly reduce yield and quality. Although advancements have taken place in agricultural technologies, a disease detection system based on deep learning yet remains to be developed for the specific purpose of cauliflower. Objectives: The specific aim of this research is to develop an automated expert system that integrates the Internet of Things feature for cauliflower disease early detection with the use of deep learning. Such a system could facilitate farmers in identifying infections by employing mobile or portable-based devices to relay the results. Methods: The mobile and IoT-enabled devices have aided in the collection of 750 images of cauliflower plants. The Cat Swarm Optimization (CSO) technique was utilized to segment the affected areas visible in the given images. Feature extraction methods were explicitly engaged to acquire the statistical and co-occurrence features of the images. The study was limited to four diseases—in particular, bacterial soft rot, white rust, black rot, and downy mildew. The performance of the proposed model, termed GNN-PDP (Graph Neural Network-based Plant Disease Prediction), was observed against CNN, DNN, Random Forest, Decision Tree, LDA, and PCA classifiers. Results: The GNN-PDP model achieved almost 89% accurate output, outperforming other models in disease classification. Conclusion: An effective solution for real-time detection of cauliflower diseases, allowing early intervention and promoting sustainable agriculture, is offered by the proposed GNN-based system.

Abstract Electroencephalographic (EEG) signals have attracted much attention as a desirable new type of biometric traits due to their unique advantages of security, concealment and evasion. In this paper, we used the EMOTIV EPOC+ head-mounted EEG device to collect EEG signals from participants through 14 electrode channels, including AF3 and AF4. The processing methods of peak-to-peak amplitude detection and iterative averaging denoising are proposed for the presence of ocular electrical interference and industrial frequency interference in the signal, respectively. The processed signal data is fed into the authentication model of the deep convolutional recurrent neural network constructed in this paper to carry out security authentication test experiments. The average authentication accuracy of this authentication model can reach 92.60%, which improves the accuracy compared to the LDA classifier. In the “self-stranger” and “selfacquaintance” categorization tests, the accuracy of this paper’s method curve is always higher than that of random selection and outperforms PCA on some feature bands. The deep learning model of CNN fused with LSTM can make full use of EEG data features to defend against illegal users in real time.

Imbalanced data are often delegated issues in data sets as it has the power to affect the result and the performance of the classification algorithm. Such problems, if not handled well with good sampling techniques could lead to biased results, overfitting as well as a high rate of misclassification thereby favouring just one class among the two classes. Usually, when assigning sampling techniques, it is necessary to look at the nature of the dataset being studied. It is of a truth that the LDA classifier looking for an efficient performance when presented with imbalanced instances is not suitable to deal with imbalanced learning tasks, since it tends to classify all the data into the majority class, which is usually the less important class. This work explains the different approaches which have been employed by different researchers to resolve the issue of imbalanced data in LDA and the effect of the results obtained both positively and negatively. It should be noted that this single article cannot completely review all the works or research done on the topic, hence we hope that the references which was dually cited will be of help to the major theoretical issues.

Distinguishing between tumor recurrence and pseudoprogression (PsP) in high-grade glioma postoperatively is challenging. This study aims to enhance this differentiation using a combination of intratumoral and peritumoral radiomics. To assess the effectiveness of intratumoral and peritumoral radiomics in improving the differentiation between high-grade glioma recurrence and pseudoprogression after surgery. A total of 109 cases were randomly divided into training and validation sets, with 1316 features extracted from intratumoral and peritumoral volumes of interest (VOIs) on conventional magnetic resonance imaging (MRI) and apparent diffusion coefficient (ADC) maps. Feature selection was performed using the mRMR algorithm, resulting in intratumoral (100 features), peritumoral (100 features), and combined (200 features) subsets. Optimal features were then selected using PCC and RFE algorithms and modeled using LR, SVM, and LDA classifiers. Diagnostic performance was compared using area under the receiver operating characteristic curve (AUC), evaluated in the validation set. A nomogram was established using radscores from intratumoral, peritumoral, and combined models. The combined model, utilizing 14 optimal features (8 peritumoral, 6 intratumoral) and LR as the best classifier, outperformed the single intratumoral and peritumoral models. In the training set, the AUC values for the combined model, intratumoral model, and peritumoral model were 0.938, 0.921, and 0.847, respectively; in the validation set, the AUC values were 0.841, 0.755, and 0.705. The nomogram model demonstrated AUCs of 0.960 (training set) and 0.850 (validation set). The combination of intratumoral and peritumoral radiomics is effective in distinguishing high-grade glioma recurrence from pseudoprogression after surgery.

The objective of the article is to recognize users' emotions by classifying facial electromyographic (EMG) signals. A biomedical signal amplifier, equipped with eight active electrodes positioned in accordance with the Facial Action Coding System, was used to record the EMG signals. These signals were registered during a procedure where users acted out various emotions: joy, sadness, surprise, disgust, anger, fear, and neutral. Recordings were made for 16 users. The mean power of the EMG signals formed the feature set. We utilized these features to train and evaluate various classifiers. In the subject-dependent model, the average classification accuracies were 96.3% for KNN, 94.9% for SVM with a linear kernel, 94.6% for SVM with a cubic kernel, and 93.8% for LDA. In the subject-independent model, the classification results varied depending on the tested user, ranging from 91.4% to 48.6% for the KNN classifier, with an average accuracy of 67.5%. The SVM with a cubic kernel performed slightly worse, achieving an average accuracy of 59.1%, followed by the SVM with a linear kernel at 53.9%, and the LDA classifier at 41.2%. Additionally, the study identified the most effective electrodes for distinguishing between pairs of emotions.

This study investigates the feasibility of using UV–Vis spectroscopy coupled with machine learning methods to authenticate tea samples based on their geographical origins in a narrow longitudinal strip (200 km). Several preprocessing methods, such as standard normal variate (SNV), auto-scaling, multiplicative scatter correction (MSC), mean centring (MC), first derivative, and their combinations, were applied to eliminate the noninformative information. The partial least squares-linear discriminant analysis (PLS-LDA) model using first derivative spectra represented the following results, including 98.0% sensitivity, 99.5% specificity, and a mean accuracy of 98.0%. The support vector machine (PLS-SVM) classifier using first derivative spectra represented 94.0% sensitivity, 98.6% specificity, and a mean accuracy of 94.0%. The satisfactory results of the models depicted that the chemical components of tea, such as polyphenols, chlorogenic and fatty acids that absorb UV radiation are the chemical markers that can discriminate tea samples based on their geographical origin. Therefore, UV–Vis spectral fingerprinting combined with machine learning methods could be a practical, feasible, and simple method for classifying tea based on their geographical origins in a narrow longitudinal strip.

The song "Dangdut" is one of the most popular songs in Indonesia, having gained popularity from the 1960s until the present. It's even been acknowledged as authentic Indonesian music. There are both positive and negative effects on the pendengarnya of lagu dangdut. Positive dampening can lower stress levels, and negative dampening occurs when emotions are heightened. If this was brought up by a young child who was not yet fully grown, it would give them a hard time and negatively impact their journey. According to this framework, it is recommended that any eroticism in the lyrics of dance music be identified. It is therefore advised to look for signs of sexuality in the lyrics of dangdut songs. The intention is to restrict and filter the music that kids can listen to. Using LDA and QDA classifiers in conjunction with natural language processing is the suggested approach. According to research findings, LDA can identify more than QDA. The LDA examination yielded the following results: recall = 56.522%, accuracy = 56.522%, precision = 79.13%, and F1score = 65.942%. It has been demonstrated that discriminant analysis, particularly LDA, is useful for classification, as QDA has not shown itself to be the most effective method in this instance.

Vectorcardiography is an alternative form of ECG for measuring electrical activity of the heart. It achieves higher sensitivity and provides the cardiologist additional information that can contribute to early diagnosis. This study is focused on proposal of a methodology for the processing of directly measured and transformed VCG records by using Kors regression transformation. A total 16 VCG features were extracted, while 12 features showed relevant information based on the statistical analysis and the method of maximum relevance minimum redundancy. These features served as input to the LDA and decision trees classifiers, while LDA achieved the most accurate results with accuracy 91.5%, specificity 76.3% and sensitivity 94.8% for directly measured VCG and accuracy 90.9%, specificity 76.3% and sensitivity 94.0% for transformed VCG. We conclude that this proposed methodology and the results obtained from it can be beneficial for the early diagnosis of myocardial infarction within the framework of automated detection.

Fundamental to human movement is the ability to interact with objects in our environment. How one reaches an object depends on the object's shape and intended interaction afforded by the object, e.g., grasp and transport. Extensive research has revealed that the motor intention of reach-to-grasp can be decoded from cortical activities using EEG signals. The goal of the present study is to determine the extent to which information encoded in the EEG signals is shared between two limbs to enable cross-hand decoding. We performed an experiment in which human subjects (n = 10) were tasked to interact with a novel object with multiple affordances using either right or left hands. The object had two vertical handles attached to a horizontal base. A visual cue instructs what action (lift or touch) and whether the left or right handle should be used for each trial. EEG was recorded and processed from bilateral frontal-central-parietal regions (30 channels). We trained LDA classifiers using data from trials performed by one limb and tested the classification accuracy using data from trials performed by the contralateral limb. We found that the type of hand-object interaction can be decoded with approximately 59 and 69% peak accuracy in the planning and execution stages, respectively. Interestingly, the decoding accuracy of the reaching directions was dependent on how EEG channels in the testing dataset were spatially mirrored, and whether directions were labeled in the extrinsic (object-centered) or intrinsic (body-centered) coordinates.

Our study defines a novel electrode placement method called Functionally Adaptive Myosite Selection (FAMS), as a tool for rapid and effective electrode placement during prosthesis fitting. We demonstrate a method for determining electrode placement that is adaptable towards individual patient anatomy and desired functional outcomes, agnostic to the type of classification model used, and provides insight into expected classifier performance without training multiple models. FAMS relies on a separability metric to rapidly predict classifier performance during prosthesis fitting. The results show a predictable relationship between the FAMS metric and classifier accuracy (3.45%SE), allowing estimation of control performance with any given set of electrodes. Electrode configurations selected using the FAMS metric show improved control performance ( ) for target electrode counts compared to established methods when using an ANN classifier, and equivalent performance ( R2 ≥ .96) to previous top-performing methods on an LDA classifier, with faster convergence ( ). We used the FAMS method to determine electrode placement for two amputee subjects by using the heuristic to search through possible sets, and checking for saturation in performance vs electrode count. The resulting configurations that averaged 95.8% of the highest possible classification performance using a mean 25 number of electrodes (19.5% of the available sites). FAMS can be used to rapidly approximate the tradeoffs between increased electrode count and classifier performance, a useful tool during prosthesis fitting.

This article aims to determine the coefficients that will reduce the in-class distance and increase the distance between the classes, collecting the data around the cluster centers with meta-heuristic optimization algorithms, thus increasing the classification performance. The proposed mathematical model is based on simple mathematical calculations, and this model is the fitness function of optimization algorithms. Compared to the methods in the literature, optimizing algorithms to obtain fast results is more accessible. Determining the weights by optimization provides more sensitive results than the dataset structure. In the study, the proposed model was used as the fitness function of the metaheuristic optimization algorithms to determine the weighting coefficients. In this context, four different structures were used to test the independence of the results obtained from the algorithm: the particle swarm algorithm (PSO), the bat algorithm (BAT), the gravitational search algorithm (GSA), and the flower pollination algorithm (FPA). As a result of these processes, a control group from unweighted attributes and four experimental groups from weighted attributes were obtained for each dataset. The classification performance of all datasets to which the weights obtained by the proposed method were applied increased. 100% accuracy rates were obtained in the Iris and Liver Disorders datasets used in the study. From synthetic datasets, from 66.9% (SVM classifier) to 96.4% (GSA Weighting + SVM) in the Full Chain dataset, from 64.6% (LDA classifier) to 80.2% in the Two Spiral datasets (weighted by BA + LDA). As a result of the study, it was seen that the proposed method successfully fulfills the task of moving the attributes to a linear plane in the datasets, especially in classifiers such as SVM and LDA, which have difficulties in non-linear problems, an accuracy rate of 100% was achieved.

Abstract The electrocardiogram is a very valuable clinical tool which allows to retrieve information about the presence and location of arrhythmic foci as well as ischemic and scar tissue and disorder’s of the dedicated cardiac conduction system. In the presented study timing parameters computed by a delineating beat detector for identifying the P-Wave, QRS - complex and T-Wave are used to classify the individual beats. From a set of total 419 feature generated from these parameters 64 are used to train LDA classifier for discriminating 3 classes (Normal, Artifact, Arrhythmic) and 5 Classes (Normal, Artifact, Atrial and ventricular premature contractions and bundle branch blocks). Further it is investigated how the imbalance between normal beats and arrhythmic beats as well as the beats missed by the beat detector affect the classification results. In the case of 5 classes accuracies of 97.52 % in the imbalanced case and 96.38 r for the balanced data were obtained. For 3 classes accuracies of 97.76 % and 95.18 % were achieved. Considering in addition the beats missed by the detector the accuracies dropped to 96.68 %, and 95.54 % for 5 classes and 95.54 % and 96.92 % for 3 classes. These values are within the ranges for linear classifier reported in literature. This is quite promising for implementing a real-time classifier which exploits the parameters and values computed by the beat detector.