PeerJ Computer Science最新文献

The rapid development of the Internet of Things technology has led to a boom in the adoption of intelligent healthcare management systems in the healthcare industry. However, it has also highlighted key issues such as security, privacy, and efficient query of medical data. Traditional methods for querying medical data suffer from severe data leakage risks, low query performance, and excessive storage space. This article proposes a comprehensive Secure ENcrypted Search for Health Scheme (SENSH) solution based on consortium blockchain and searchable encryption to address these challenges. SENSH enables efficient authorization management through Bloom filters, ensuring fast querying of large datasets by authorized users while saving storage space. It uses off-chain Advanced Encryption Standard (AES) and on-chain storage management for data protection, significantly reducing the likelihood of data exposure. The system is also enhanced with event triggering and logging mechanisms to support real-time monitoring and data tracing to meet audit compliance requirements. It provides version control and timestamping to accommodate dynamic data updates, employs an obfuscationfactor to prevent tag-based original data content leakage, and supports dynamic updating of tags to accommodate different access requirements. Experimental results show that SENSH excels in authorization management, privacy protection, defense against tampering, and anti-replay and Distributed Denial of Service (DDoS). Compared with existing schemes, SENSH has significant advantages in terms of gas consumption, computation cost, and execution time. It is particularly suited for the protection and efficient query of medical and health data.

With the rapid expansion of educational data, institutions face increasing pressure to adopt advanced predictive models that can enhance academic planning, resource allocation, and student support. This study presents a novel educational data mining approach designed to forecast student performance levels categorized as low, medium, and high by analyzing historical and behavioral trends. This work proposes XSEJNet, an innovative hybrid model that integrates ResNeXt architecture with squeeze-and-excitation (SE) attention mechanisms, and employs the Jaya optimization algorithm to refine hyperparameters and boost predictive accuracy and computational efficiency. The model works with structured and unstructured academic data, effectively capturing complex, high-dimensional features to support accurate classification. Through extensive simulations and comparative evaluations, XSEJNet consistently outperforms conventional machine learning models and recent existing techniques such as reinforcement learning co-evolutionary hybrid intelligence (RLCHI), Enhanced AEO-XGBoost, convolution-based deep learning (Conv-DL), and dual graph neural network (DualGNN). The model achieves a high prediction accuracy of 97.98% while also demonstrating faster convergence and reduced computational overhead, making it a scalable and practical solution for real-world educational settings. The findings underscore XSEJNet's ability to support early intervention, strengthen e-learning platforms, and inform institutional decision-making. By advancing predictive capabilities in education, this work makes a meaningful contribution to developing inclusive, data-driven, and sustainable academic systems.

Edge intelligence is fast becoming indispensable as billions of sensors demand real-time inference without saturating backbone links or exposing sensitive data in remote data centres and emerging artificial intelligence (AI)-edge boards such as NVIDIA CPUs, 16 GB RAM, and microcontrollers with chip neural processing unit (NPU) (<1 W). This article introduces the Energy-Smart Component Placement (ESCP) algorithm of fog devices like fog cluster manager nodes (FCMNs) and fog nodes (FNs), allocates modules to fog devices, and saves energy by deactivating inactive devices framework transparently distributes compressed neural workloads across serverless. To optimize the deployment of AI workloads on fog edge devices as a service (FEdaaS), this project aims to provide a reliable and dynamic architecture that guarantees quality of service (QoS) and quality of experience (QoE). The cloud, fog, and extreme edge layers while upholding application-level QoS and QoE. Two machine learning (ML) methods that fuse eXtreme Gradient Boosting (XGB)-based instantaneous QoS scoring and long short term memory (LSTM) forecasting of node congestion, and a meta-heuristic scheduler that uses XGB for instantaneous QoS scoring and LSTM for short-horizon load forecasting. Compared with a cloud-only baseline, ESCP improved bandwidth utilization by 5.2%, scalability (requests per second) by 3.2%, energy consumption by 3.8% and response time by 2.1% while maintaining prediction accuracy within +0.4%. The results confirm that low-resource AI-edge devices, when orchestrated through our adaptive framework, can meet QoE targets such as 250 ms latency and 24 h of battery life. Future work will explore federated on-device learning to enhance data privacy, extend the scheduler to neuromorphic processors, and validate the architecture in real-time intensive care and smart city deployments.

Periodontitis is a common dental disease that results in tooth loss, if not diagnosed and treated in time. However, diagnosing bone loss due to periodontitis from panoramic radiographs is a time-consuming and error-prone process, requiring extensive training and expertise. This work addresses the research gap in automated periodontitis bone loss diagnosis using deep learning techniques. We have proposed a modified version of You Only Look Once (YOLO)v2, called YOLOv7-M, that includes a focus module and a feature fusion module for rapid inference and improved feature extraction ability. The proposed YOLOv7-M model was evaluated on a tooth detection dataset and demonstrated superior performance, achieving an F1-score, precision, recall, and mean average precision (mAP) of 92.5, 91.7, 87.1, and 91.0, respectively. Experimental results indicate that YOLOv7-M outperformed other state-of-the-art object detectors, including YOLOv5 and YOLOv7, in terms of both accuracy and speed. In addition, our comprehensive performance tests show that YOLOv7-M outperforms robust object detectors in terms of various statistical evaluation measures. The proposed method has potential applications in automated periodontitis diagnosis and can assist in the detection and treatment of the disease, eventually enhancing patient outcomes.

Software maturity models can be utilized by organizations to evaluate and enhance their development processes. Established and recognized models such as the Capability Maturity Model Integrated (CMMI) and ISO/IEC 15504 (Software Process Improvement and Capability Determination (SPICE)) have proven their value. However, many new software maturity models exist, and their quality and potential value remain questionable until they are properly assessed before adoption. Without such an assessment, organizations can implement poor or ineffective models, resulting in wasted resources and failed improvement initiatives. Our research aims to address this challenge by developing a measurement framework based on ISO/IEC 15504-3 standards to assess the quality of developed software maturity models. We derived our quality assessment criteria through literature analysis, analyzing four main categories: basic model information, structural design, assessment methods, and implementation support. After developing this framework, we validated it with expert reviews to assess its design and usability and through a series of case studies. Feedback from academics and industry practitioners confirmed the framework's utility, especially recognizing its clear structure and comprehensiveness of evaluation criteria. Case studies also revealed the framework's effectiveness in identifying strengths and areas of improvement, finding that evaluated models had quality scores ranging from 83.3% to 93.2%. Our study enhances software maturity models' practical utility and adoption across different software contexts, providing professionals and academics with a structured way to evaluate and enhance maturity models.

Detecting cloned and impersonated profiles on online social networks (OSNs) has become an increasingly critical challenge, particularly with the proliferation of AI-generated content that closely emulates human communication patterns. Traditional identity deception detection methods are proving inadequate against adversaries who exploit large language models (LLMs) to craft syntactically accurate and semantically plausible fake profiles. This article focuses on the detection of profile cloning on LinkedIn by introducing a multi-stage, content-based detection framework that classifies profiles into four distinct categories: legitimate profiles, human-cloned profiles, LLM-generated legitimate profiles, and LLM-generated cloned profiles. The proposed framework integrates multiple analytical layers, including semantic representation learning through attention-based section embedding aggregation, linguistic style modeling using stylometric-perplexity features, anomaly scoring via cluster-based outlier detection, and ensemble classification through out-of-fold stacking. Experiments conducted on a publicly available dataset comprising 3,600 profiles demonstrate that the proposed meta-ensemble model consistently outperforms competitive baselines, achieving macro-averaged accuracy, precision, recall, and F1-scores above 96%. These results highlight the effectiveness of leveraging a combination of semantic, stylistic, and probabilistic signals to detect both human-crafted and artificial intelligence (AI)-generated impersonation attempts. Overall, this work presents a robust and scalable content-driven methodology for identity deception detection in contemporary OSNs.

Acute lymphoblastic leukemia (ALL), one of the common diseases of our day, is one of the most common hematological malignant diseases in childhood. Early diagnosis of ALL, which plays a critical role in medical diagnosis processes, is of great importance especially for the effective management of the treatment process of cancer patients. Therefore, ALL cells must be detected and classified correctly. Traditional methods used today prolong the detection and classification processes of cells, cause hematologists to interpret them according to their expertise, and delay medical decision-making processes. In this study, the performance of the hybrid model developed with different deep learning models for ALL diagnosis was comparatively analyzed. In the proposed ALL detection architecture, blood cell images were processed using the center-based cropping strategy and irrelevant areas in the images were automatically removed. The dataset was divided into training, validation, and test sets, and then features were extracted with deep hyperparameters for convolution, pooling, and activation layers using a model based on Xception architecture. The obtained features were optimized to the advanced Extreme Gradient Boosting (XGBoost) classifier and model classification results were obtained. The results showed that the proposed model achieved 98.88% accuracy. This high accuracy rate was compared with different hybrid models and it was seen that the model was more successful in detecting ALL disease compared to existing studies.

Background/objectives: Lung cancer is the leading cause of cancer-related deaths worldwide. While computed tomography (CT) scans provide more comprehensive medical information than chest X-rays (CXR), the high cost and limited availability of CT technology in rural areas pose significant challenges. CXR images, however, could serve as a potential preliminary diagnostic tool in diagnosing lung cancer, especially when combined with a computer-aided diagnosis (CAD) system. This study aims to enhance the accuracy and accessibility of lung cancer detection using a custom-designed convolutional neural network (CNN) trained on CXR images.

Methods: A custom-designed CNN was trained on an openly accessible CXR dataset from the Japanese Society for Radiological Technology (JSRT). Prior to training, the dataset underwent preprocessing, where each image was divided into overlapping patches. A t-test was applied to these patches to distinguish relevant from irrelevant ones. The relevant patches were retained for training the CNN model, while the irrelevant patches were excluded to enhance the model's performance.

Results: The proposed model yielded a mean accuracy of 83.2 ± 2.91%, demonstrating its potential as a cost-effective and accessible preliminary diagnostic tool for lung cancer.

Conclusions: This approach could significantly improve the accuracy and accessibility of lung cancer detection, making it a viable option in resource-limited settings.

Automatic pain assessment involves accurately recognizing and quantifying pain, dependent on the data modality that may originate from various sources such as video and physiological signals. Traditional pain assessment methods rely on subjective self-reporting, which limits their objectivity, consistency, and overall effectiveness in clinical settings. While machine learning offers a promising alternative, many existing approaches rely on a single data modality, which may not adequately capture the multifaceted nature of pain-related responses. In contrast, multimodal approaches can provide a more comprehensive understanding by integrating diverse sources of information. To address this, we propose a dual-stream framework for classifying physiological and behavioral correlates of pain that leverages multimodal data to enhance robustness and adaptability across diverse clinical scenarios. Our framework begins with masked autoencoder pre-training for each modality: facial video and multivariate bio-psychological signals, to compress the raw temporal input into meaningful representations, enhancing their ability to capture complex patterns in high-dimensional data. In the second stage, the complete classifier consists of a dual hybrid positional encoding embedding and cross-attention fusion. The pain assessment evaluations reveal our model's superior performance on the AI4Pain and BioVid datasets for electrode-based and heat-induced settings.

With the rapid advancement of artificial intelligence, the demand for personalized and adaptive learning has driven the development of intelligent educational systems. This article proposes a novel adaptive learning-driven architecture that combines multimodal behavioral modeling and personalized educational resource recommendation. Specifically, we introduce a multimodal fusion (MMF) algorithm to extract and integrate heterogeneous learning behavior data-including text, images, and interaction logs-via stacked denoising autoencoders and Restricted Boltzmann Machines. We further design an adaptive learning (AL) module that constructs a student-resource interaction graph and dynamically recommends learning materials using a graph-enhanced contrastive learning strategy and a dual-MLP-based enhancement mechanism. Extensive experiments on the Students' Academic Performance Dataset demonstrate that our method significantly reduces prediction error (mean absolute error (MAE) = 0.01, mean squared error (MSE) = 0.0053) and achieves high precision (95.3%) and recall (96.7%). Ablation studies and benchmark comparisons validate the effectiveness and generalization ability of both MMF and AL. The system exhibits strong scalability, real-time responsiveness, and high user satisfaction, offering a robust technical foundation for next-generation AI-powered educational platforms.