Prediction of unerupted canine and premolars width using artificial intelligence compared with Tanaka-Johnston and Moyers methods: A retrospective observational analytical study on Iranian population.

Voice analysis as a digital biomarker: A machine learning approach for automated multiple sclerosis classification.

Signal-based feature analysis of behavioral trajectories for predicting calving time and classifying assistance needs

Phishing is a common cyber security threat in which attackers attempt to deceive users into disclosing personal information such as passwords, credit card numbers, and other sensitive data. With the rapid advancement of technology, phishing techniques have become increasingly sophisticated and harder to detect using traditional methods. Therefore, it is essential to develop detection techniques capable of identifying phishing websites with high accuracy. This study aims to optimize phishing detection performance by integrating variable correlation analysis for feature selection and applying imbalanced learning techniques to address data imbalance. The research stages include Data Collection, Data Preprocessing, and Data Exploration, which involve correlation analysis, removal of low-correlation features, and data visualization. In the Model Building and Training phase, the dataset is split into features and labels, followed by training and the application of data balancing techniques, ending with Model Evaluation. The evaluated algorithms include Logistic Regression, Naive Bayes, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Multi-Layer Perceptron, Decision Tree, Random Forest, Gradient Boosting, and CatBoost. The results show that the KNN algorithm delivers the best performance, achieving an accuracy of 91.25% and optimal scores in Precision (0.906943), Recall (0.927858), and F1-Score (0.922141), along with the lowest Hamming Loss at 0.0875. In contrast, the SVM algorithm recorded the lowest performance among the tested models. The implementation of this method is expected to contribute to the development of more reliable and accurate phishing detection systems in the future.

Venous congestion is a major cause of postoperative free flap compromise, and early detection is crucial for improving flap salvage rates and patient outcomes. This study aimed to develop and validate a deep learning (DL)-integrated remote monitoring system with a smartphone application for real-time, quantitative assessment of free flaps, with a specific focus on the early detection of venous congestion. This diagnostic study was conducted at our institution. Patients aged 18-60 years who underwent free flap reconstruction between January 2019 and June 2025 were included. The study was divided into three phases: a 5-month model training phase for DL model development and internal validation, a 5-month external validation and clinical application phase, and a 4-month clinical comparison phase. The DL model was developed using TensorFlow Lite and a Flap Segmentation Network (FS-Net). Performance was evaluated through accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and clinical outcomes including time to detection and flap survival. A total of 1649 photographs from 615 patients were analyzed in the development and validation of the DL classification model between January 2019 and February 2025. During model development, the model achieved an accuracy of 86.8%, sensitivity of 92.4%, specificity of 79.7%, and AUC of 0.88. Internal validation improved these metrics to 87.6%, 95.0%, 81.1%, and 0.92, respectively. External validation demonstrated the model's generalizability, with an accuracy of 89.3%, sensitivity of 96.2%, specificity of 84.9%, and AUC of 0.93. The clinical application phase showed that the system had an overall accuracy of 92.16%, sensitivity of 95.18%, false-positive rate of 2.62%, and false-negative rate of 4.82%. A total of 113 patients were included in clinical comparison between March 2025 and June 2025. The remote monitoring group exhibited a trend towards a shorter mean time to congestion detection, higher flap survival rate, and shorter mean time to re-exploration, although these differences were not statistically significant. The DL-integrated remote monitoring system demonstrated high accuracy and reliability in detecting venous congestion. It provided an objective and real-time tool that may help reduce clinical burden and support timely intervention in free flap management. However, its impact on definitive clinical outcomes required further validation in larger studies.

Inference tasks based on multimodal data from the Internet of Things (IoT) play an important role in intelligent management. Due to the limited resources of IoT devices, existing edge frameworks struggle to achieve low-energy, high-efficiency accurate inference. This paper introduces a Federated Learning (FL) and semantic communication empowered multimodal terminal collaborative inferencing framework for IoT businesses (FLSC-CI). Firstly, we propose an FL-based Customized model Training Algorithm (FL-CTA) for semantic encoder-decoder models and business inference models. In the semantic extraction phase, high-quality terminals perform local model training, model aggregation, and semantic extraction, while low-quality terminals perform only local model training or semantic extraction. In the business inference model training phase, the edge server synchronously performs multimodal model training by utilizing semantics transmitted from terminals. Furthermore, this paper proposes a Heterogeneous Resource Dynamic Allocation Strategy (HRDAS) for FLSC-CI based on multi-agent deep deterministic policy gradient to manage FL training process. Intelligent agents at cluster heads make customized allocation decisions of system bandwidth and power according to terminals' service capabilities and model features within the cluster. Simulation results demonstrate that FLSC-CI significantly improves resource utilization and communication efficiency while maintaining high inference accuracy, making it suitable for large-scale heterogeneous IoT deployments.

The Arabic language occupies fifth place in the ranking of spoken languages, meaning that approximately 420 million people speak it. People have been biometrically identified using fingerprints, faces, and other similar features. In this paper, a biometric identification model for Arabic handwriting is proposed, as many Arabic letters have very similar shapes and can only be distinguished by the location of one or more dots, either above or below the letter. A novel and efficient offline Arabic handwriting identification model is presented. Its basis is the combination of several methods, such as the Harris corner detector, Shi-Thomasi, and convolutional neural networks (CNNs). Data augmentation is used during the model training phase to improve the data quality, without the need to segment words/characters. Leveraging a large collection of handwritten Arabic documents, such as KHATT and AHAWP, accuracy rates of 99% and 98% were reached, respectively.

Fast yet versatile machine unlearning for deep neural networks.

Graph neural networks have shown impressive performance in a variety of biomedical application tasks due to their powerful graph representation capabilities. Although GNN has achieved great success, the data noise and data scarcity problems commonly faced in real psychiatric disease prediction scenarios may affect the training and prediction of graph learning models. At present, there is no relevant work to obtain a reasonable solution. Data augmentation, which allows limited data to produce value equivalent to more data without substantially increasing the data, is considered a practical approach to addressing the problem of noisy data and data scarcity. In this work, we propose a method based on graph data augmentation for solving the problem of noisy data and data scarcity in mental illness prediction. To mitigate the negative effects of label noise, we use edge predictors to optimize the graph topology, enhance links to nodes with high similarity, remove erroneous noisy edges, and enhance the model robustness by adding adversarial perturbations in the feature space. In addition, a confident self-checking mechanism allows accurate pseudolabeling to be obtained, providing more supervision for the model training phase and further reducing the effect of label noise. Extensive experiments on two multimodal real mental illness datasets show that the proposed approach has better performance. Sufficient ablation experimental studies were conducted to assess the effectiveness of each component. The experimental results validate the effectiveness and scalability of our framework for population-based disease prediction, even under challenging conditions of data noise and sparsity. The implementation code is publicly available at: https://github.com/jiachengpan98/GDA-GCN.

Electromyography (EMG) signals, as important bioelectric signals reflecting human muscle activity, have broad application prospects in fields such as human-computer interaction, prosthetic control, and motion recognition. However, electromyographic signals themselves are susceptible to noise interference such as motion artifacts and power interference, and traditional filtering methods are difficult to effectively distinguish signals from complex background noise. Meanwhile, traditional feature engineering relies on manual experience, which limits the model's generalization ability and real-time processing performance. In response to the above issues, this article proposes an end-to-end electromyographic signal processing method based on Convolutional Neural Network (CNN), which is used to simultaneously achieve signal denoising and motion pattern recognition. Firstly, the original electromyographic signal is segmented using a sliding window method and standardized; Subsequently, the processed data is input into a CNN model consisting of multiple layers of convolution and pooling structures, which automatically extracts time-domain features and completes classification. During the model training phase, the introduction of cross entropy loss function and Adam optimization algorithm improves the convergence speed and classification accuracy of the model. We used an open-source surface Electromyographic (sEMG) dataset on the PhysioNet platform for validation and compared its performance with traditional support vector machine (SVM) methods. The experimental results show that the proposed method is significantly better than traditional methods in key indicators such as accuracy and F1 score, and has better robustness and real-time performance. This study demonstrates the potential of deep learning in electromyographic signal processing, providing technical support for future intelligent prosthetic control and high-precision human-machine interaction systems.

The precise prediction of football match outcomes holds significant value in the sports domain. However, traditional prediction methods are limited by data complexity and model capabilities, struggling to meet the demands for high accuracy. Quantum neural networks (QNNs) leverage the unique quantum properties of quantum bits (qubits) such as superposition and entanglement. They have enhanced information processing capabilities and potential pattern mining abilities when dealing with vast, high-dimensional, and complex football match data. This makes QNNs a superior choice compared to traditional neural networks and other advanced models for football match prediction. This study focuses on a deep learning (DL)-based QNN model, aiming to construct and optimize this model to analyze historical football match data for high-precision predictions of future match outcomes. Specifically, detailed match records from 2008 to 2022 of major European football leagues were obtained from the “European Football Database” public dataset on Kaggle. The data includes various factors such as match outcomes, team information, player stats, and match venues. The data are cleaned, standardized, and feature-engineered to meet the input requirements of neural network models. A multilayer perceptron model consisting of an input layer, multiple hidden layers, and an output layer is designed and implemented. During the model training phase, gradient descent is used to optimize weight parameters, and quantum algorithms are integrated to continuously adjust network weights to minimize prediction errors. The model is trained, parameter tuning is completed, and performance is evaluated using the training, validation, and independent test sets. The model’s effectiveness is measured using indicators such as F1 score, accuracy, and recall. The study results indicate that the optimized QNN model significantly outperforms other advanced models in prediction accuracy. The optimized QNN model has an improvement of more than 20.5% in precision, an enhancement of over 23.2% in recall, and an increase of over 22.3% and 21.8% in accuracy and F1 score. Additionally, the model predicts the championship probabilities for Spain, France, England, and the Netherlands in the European Championship as 31.72%, 27.61%, 22.58%, and 18.09%, respectively. This study innovatively applies the optimized QNN model to outcome prediction in football matches, validating its effectiveness in the sports prediction field. It provides new ideas and methods for football match outcome prediction while offering valuable references for developing prediction models for other sports events. By integrating public data with DL technology, this study lays the foundation for the practical application of sports data analysis and prediction models, holding significant theoretical and practical value. Furthermore, future research can further explore the integration of QNN models with mathematical analysis systems, expanding their application scenarios in the real world. For example, sports betting agencies are provided with more accurate risk assessments, assisting teams in formulating more scientific tactical strategies, and optimizing event organization arrangements, to fully leverage their potential value.

Abstract Valve leakage can cause reciprocating air compressors to lose operating performance or even shut down, and its leakage size cannot be directly measured. Therefore, in order to be able to achieve an accurate assessment of the valve leakage size, this paper proposes a multi-scale fusion convolutional neural network (MSFCNN) based air compressor valve leakage size inversion model for quantifying the size of the valve leakage. The model is designed with local feature and global feature extraction channels. The global-feature channel takes the acoustic emission (AE) data features with strong correlation coefficients with the valve leakage size as inputs, while the local-feature channel uses wavelet decomposition to obtain the different frequency features of the AE signals, which can extract more detailed feature information. Then, through the network structure of MSFCNN, the features of these two channels are fused to achieve the comprehensive extraction and analysis of multi-scale features of AE signals, and finally the inversion of valve leakage size. In the model training phase, an automatic windowing technique is used to ensure the integrity and accuracy of the input signals. The experimental results show that the proposed model exhibits excellent performance in AE-based leakage size inversion compared to other models, which provides an important basis for valve leakage condition assessment and predictive maintenance of reciprocating air compressors to ensure their normal and efficient operation.

The emergence of Edge Computing has led to an increasingly intricate and widespread issue of network infiltration in Edge Computing devices. Therefore, it is crucial to explore intelligent, automated, and resilient methods for network intrusion detection. Graph neural network-based techniques for network intrusion detection have been proposed by many scholars recently. However, the graph construction methods of these approaches cannot fully adapt to real network intrusion datasets, leading to problems such as overfitting and insufficient graph information mining. Furthermore, the methods employed in the model training phase are relatively limited, failing to consider the characteristic presence of grouping within networks. These shortcomings result in a lack of high accuracy in intrusion detection systems, especially in multi-class classification scenarios. This research suggests a graph neural network technique based on behavior similarity employing a graph attention network (BS-GAT) to handle the aforementioned problem. To address overfitting and inadequate graph information mining, a behavioral similarity-based graph creation method is first presented through an analysis of real datasets. Subsequently, the inclusion of edge behavioral relationship weights into the GAT leverages the relationship between data flow and graph structural information, enhancing the performance of the trained model. In the final stage, experiments were executed using the most recent datasets to assess the effectiveness of the proposed behavior similarity-based graph attention network in network intrusion detection. The findings indicate that the proposed approach is highly effective and significantly outperforms existing solutions. In binary classification, recall, precision, f1-score, and accuracy all exceed 99%. For multi-class performance, the recognition accuracy exceeds 93%, the weighted recall surpasses 91%, and the weighted f1-score exceeds 92%, showing a substantial improvement in multi-class recognition effectiveness.

Unsupervised Person Re-identification (Re-ID) aims to identify the same person shot from non-overlapping cameras without any annotated data. In this task, attributes such as contrast, saturation, and resolution of the camera cause the deviation in target features. Since the camera label is readily available, they are employed to achieve the constraints across cameras and smooth the deviations during the model training phase. However, features from the same camera are prone to generating false positives due to the identical camera properties, which induce camera deviations on pseudo-label assignment. To address this problem, this paper proposes a novel camera-unbiased method named Camera Deviation Elimination Learning (CDE-Learning). In CDE-Learning, the Camera Deviation Compensation (CDC) module is designed to align data distributions from disparate cameras to decouple camera information from identity information during the pseudo-label allocation. Our Camera Deviation Balancing (CDB) module integrates different camera constraints in a united loss and adjusts camera constraints by constructing contrastive pairs between intra-camera and inter-camera. After explicit constraints, the Camera Attribution Auxiliary (CAA) task predicts whether a pair of images originates from the same camera to implicitly enhance the capacity to distinguish the camera deviation. We demonstrated the superior performance of the proposed CDE-Learning on benchmark datasets.

This paper presents a novel method for automatically recognizing people's apparent personality traits as perceived by others. In previous studies, apparent personality trait recognition from multimodal human behavior is often modeled to directly estimate personality trait scores, i.e., the ``Big Five'' scores. In the model training phase, ground-truth personality trait scores were often determined from personality test results scored by many other people using fine-grained questionnaires, however, rich information in the personality test results have not been leveraged for anything other than determining the ground-truth Big Five scores. The scores assigned to each questionnaire item are thought to include more meta-level differences in personality characteristics. Therefore, we propose joint modeling methods that can estimate not only the Big Five scores but also questionnaire item-level scores. This enables us to improve awareness of multimodal human behavior. In addition, we present a newly created self-introduction video dataset with 50-item Big Five questionnaire results since previous apparent personality trait recognition datasets do not provide such personality test results. Experiments using the created dataset demonstrate that our proposed joint modeling methods with a multimodal transformer backbone can improve to estimate Big Five scores and effectively estimate questionnaire item-level scores. We also verify that the estimation performance reached human evaluation performance.

Backdoor attacks pose a significant threat during the model's training phase. Attackers craft pre-defined triggers to break deep neural networks, ensuring the model accurately classifies clean samples during inference yet erroneously classifies samples added with these triggers. Recent studies have shown that speaker recognition systems trained on large-scale data are susceptible to backdoor attacks. Existing attackers employ unnoticed ambient sounds as triggers. However, these sounds are not inherently part of the training samples themselves. In essence, triggers can be designed to maintain an intrinsic connection with the original speech to enhance stealthiness. Our paper presents a novel attack methodology named Speed Master, which undermines deep neural networks by manipulating the speed of speech samples. Specifically, we execute poison-only backdoor attacks using speed or tempo adjustment. Changes in speech rate have become a common occurrence, as seen on platforms that allow users to adjust playback speed. In real-world scenarios, people naturally adjust their speaking rate depending on the context. As a result, changes in a speaker’s speech rate are typically perceived as normal and are unlikely to raise suspicion. Furthermore, detecting such subtle adjustments becomes challenging for users without reference speech. Our comprehensive experiments demonstrate that Speed Master can achieve an ASR over 99% in the digital domain, with only a 0.6% poisoning rate. Additionally, we validate the feasibility of Speed Master in the real world and its resistance to typical defensive measures.

This paper proposes a cross-domain intelligent traffic signal control method based on federated Proximal-Policy Optimization (PPO) for distributed joint training of agents across domains for typical intersections, aiming at solving the problems of slow learning speed and poor model generalization when deep reinforcement learning (RL) is applied to cross-domain multi-intersection traffic signal optimization control. The proposed method improves the model generalization ability of different local models during global cross-region distributed joint training under the premise of ensuring information security and data privacy, solves the problem of non-independent and homogeneous distribution of environmental data faced by different agents in real intersection scenarios, and significantly accelerates the convergence speed of the model training phase. By reasonably designing the state, action and reward functions and determining the optimal values of several key parameters in the federated collaboration mechanism, the RL model could ensure high learning efficiency and fast convergence in the face of the gradual increase of road network size and the exponential increase of state and action space with the number of intersections. In addition, the new state interaction method and the reward function allow the agents to collaborate with each other, which greatly improves the information interaction efficiency between the federated learning local agents and the central coordinator, and improves the access efficiency of the road network and reduces the amount of communication data transmitted. Finally, through experimental comparisons, the proposed method can significantly reduce the average vehicle waiting time by up to 27.34% compared with the existing fixed timing method, and under the same convergence height, the convergence speed is up to 47.69% faster compared with the individual PPO trained in a single local environment, and up to 45.35% faster than the aggregated PPO trained jointly using all local data. The proposed method effectively optimizes intersection access efficiency with excellent robustness under various traffic flow settings.

ABSTRACT Accurate extraction of large-area urban built-up areas is of significant importance for urban development and map updating. Traditional methods relying on remote sensing imagery and deep learning exhibit low data utilization during the model training phase and significant discrepancies between extraction accuracy in application scenarios and during training, posing challenges in practical applications. In response to these issues, this paper proposes an urban built-up area extraction method based on deep learning semantic segmentation technology and large-size remote sensing imagery. This method, grounded in the sliding window theory, introduces the Overlapping Cutting and Joining (OCJ) approach. Within the OCJ method, this paper develops the Overlapping Cutting (OC) module to increase the quantity and richness of training samples. In application scenarios, the OC and Overlapping Joining (OJ) modules are utilized to circumvent large-size cross-sampling, thereby preserving original image information and enhancing segmentation accuracy. Additionally, to eliminate stitching seams and optimize detailed segmentation effects, this paper proposes the HCBM-DenseCRF method, which employs DenseCRF to eliminate stitching seams while utilizing the Hard Constraint Buffer Mechanism (HCBM) to protect complex built-up area foreground extraction results from destruction, achieving high-precision urban built-up area extraction. This paper takes the Xception-PSPNet deep learning semantic segmentation model as the base model instance and compares it with four widely used models and three ablation models, including the initial Xception-PSPNet model. OCJ-HCBM-DenseCRF achieved superior results compared to other methods across smaller, medium, and larger remote sensing image instances, particularly leading other models by 1.1%-47.3% on larger-size imagery, with a precision decrement of only 1%-2% across the three sizes. This indicates the high application value of the proposed method in the extraction of large-size imagery.

The co-pyrolysis process is an essential method for energy extraction from waste biomass and coal although the co-pyrolysis technology of biomass and coal presents a complex engineering challenge. To address these challenges, modern data-driven ensemble and tree-based machine learning approaches offer a promising solution. This study provides a comprehensive analysis of various machine learning techniques, including linear regression (LR), decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), and adaptive boosting (AdaBoost) to predict the outcome models of pyrolysis oil yield, syngas yield, char yield, and syngas lower heating value from co-pyrolysis of biomass and coal. The models are evaluated using different statistical metrics. The DT-based pyrolysis oil yield model outperformed the other four models (LR, RF, XGBoost, and AdaBoost) in predicting pyrolysis oil with robust accuracy, achieving an R 2 of 0.999 and a mean squared error (MSE) close to zero during the model training phase. Similarly, the DT-based syngas yield model showed a high R 2 of 0.999 and near-zero MSE while the based char yield model excelled the others with a high R 2 of 0.999 and negligible MSE during the model training phase. In the subsequent phase, explainable artificial intelligence-based Shapley additive explanation (SHAP) values were estimated for feature importance analysis. The SHAP analysis identified key features for pyrolysis oil and syngas yield, with biomass blending ratio and reaction time being the most crucial, while reaction time and temperature were the most important for the syngas LHV model.

Mobile device user authentication technologies have been studied for decades in the context of personal information security. To strike a balance between security, privacy, and usability, authentication methods based on motion sensors have gained widespread attention in recent years. However, these methods still face several challenges, such as the limited training samples, the finite scene coverage, and the high-cost models. Therefore, there is an urgent need to develop more efficient and reliable solutions to enhance the user experience. To address these challenges, we introduce, which offers the following features: 1) It uses a novel data augmentation method (AUTHGANS) to expand the dataset. 2) It employs a differential attention mechanism to reduce noise interference, improve model scene coverage, and simultaneously reduce the model size during the model training phase, and improve the model’s accuracy. 3) It uses a model distillation strategy (AuthFusion), ensuring high accuracy while reducing the model’s computational requirements on devices. Experiments on a dataset with 1,513 users and noise show that achieves high accuracy while requiring less computational power than other state-of-the-art authentication methods.