Deceptive reviews on hospitality platforms can undermine consumer trust and affect the reliability of online reputation systems. This study presents a comparative empirical evaluation of two deep learning pipelines for deceptive review detection: a DistilBERT-based pipeline fine-tuned end-to-end on review text and complemented with rating and sentiment features, and a Perceiver-based pipeline operating on fixed sentence embeddings generated by a frozen encoder. The experiments were conducted on a combined English-language corpus of 3,322 hotel reviews assembled from the Myleott benchmark corpus and a Kaggle dataset of authentic Ritz-Carlton New York reviews, comprising approximately 2,520 genuine and 800 deceptive reviews. The corpus was evaluated using stratified 80/20 train–test splits, with 2,657 reviews used for training and 665 for testing in each split, and the deep learning models were trained for 10 epochs. To assess the robustness of the findings, the evaluation was repeated across three independent random seeds and compared with majority-class and TF-IDF plus logistic regression baselines. DistilBERT achieved the most stable performance, reaching an average accuracy of 0.9348 ± 0.0014, an F1-score of 0.9569 ± 0.0013, and a Matthews correlation coefficient of 0.8240 ± 0.0012. In contrast, the original Perceiver pipeline showed a strong tendency to collapse toward the majority class, which limited its ability to identify deceptive reviews despite apparently high recall. Although class-weighted training improved Perceiver's behavior, its results remained less reliable and more variable than those obtained with DistilBERT. The balanced TF-IDF plus logistic regression baseline performed close to DistilBERT, indicating that well-configured traditional natural language processing methods remain competitive on datasets of this size. The ablation analysis further indicates that the rating and sentiment features contribute primarily to the stability of the model across different train–test splits, rather than providing a substantial independent gain in discriminative performance. The results indicate that the fine-tuned DistilBERT pipeline provides the most robust option in the evaluated setting, while also emphasizing the importance of reporting corpus composition, test-set size, class imbalance, and imbalance-aware metrics when developing deceptive review detection systems for hospitality review data.
IntroductionIncomplete judicial records constrain reliable crime analytics because missing attributes, heterogeneous case descriptions, mixed variable types, and imbalanced charge distributions reduce their analytical value. This study developed a reconstruction-aware framework for drug-crime charge classification from incomplete judicial records.MethodsDrug-related judicial documents from Chengdu, China, covering 2014-2021 were screened, cleaned, and converted into structured person-level observations. The final dataset comprised 12,620 valid documents and 15,184 observations with 16 structured features. A Heterogeneous Graph Convolution Temporal Autoencoder (HetGConv-TAE) was developed to jointly model heterogeneous case-attribute relations and temporal changes in case composition. Performance was evaluated under 10%, 20%, and 30% controlled missingness, followed by XGBoost charge classification and SHAP interpretation.ResultsHetGConv-TAE achieved the best mean reconstruction performance across conventional, static graph, temporal, and relational graph baselines. XGBoost trained on HetGConv-TAE reconstructed data achieved weighted F1 scores close to 80% and ROC-AUC values above 90%. SHAP analysis identified drug weight, place category, and administrative district as the most influential predictors across charge categories.DiscussionCombining label-excluded heterogeneous graph reconstruction, temporal encoding, downstream classification, and interpretable analysis improves the analytical usefulness of incomplete judicial archives. The framework is intended to support data-quality assessment and aggregate public-safety research rather than automated legal decision-making.
Financial correlation networks under infrastructure shocks undergo curvature-driven interference that standard estimators cannot detect. This study introduces Heat-Kernel Deep Instrumental Variables (HK-DeepIV), a geometric diagnostic and early-warning framework modeling shock propagation as heat diffusion on the Riemannian manifold of Johannesburg Stock Exchange (JSE) asset-return correlations. Three formal results underpin the architecture: a non-parametric identification result for the projection of the structural function onto the leading heat-kernel eigenfunction (k = 1; a scalar instrument can identify at most one linear combination of the eigenfunctions, and no claim is made beyond that projection); double robustness via Neyman orthogonality; and Corollary 3.2, showing that unit-level distinguishability collapses exponentially in diffusion time whenever Ollivier-Ricci curvature is positive; a purely geometric statement providing the basis for the fragility diagnostics developed here. Applied to 60 JSE tickers (2,832 trading days, 2015–2025) with Eskom load-shedding as the treatment (2022–2025), the trained model produces an ATE of +383 bp, reported as an overfitting artifact. Five independent baselines converge on smaller, mostly negative estimates. The most credibly conditioned conditional-association estimate is double machine learning [−20.75 bp, heteroskedasticity-and-autocorrelation-consistent (HAC)-corrected 95% CI [−51.62, 10.12] bp, p = 0.188], which is not significant at conventional levels; consistent with the instrument exogeneity caveat (5-day lagged return balance test, p = 0.007) and the descriptive framing of all estimates. The instrument [48-h-ahead Eskom stage forecast, Corr(Zt, Dt)≈0.71, first-stage F = 47.3] is distinct from the treatment (realized Stage ≥2 binary), but exogeneity is not confirmed; all estimates are conditional associations. The Fiedler eigenvalue (mean 0.3827, minimum 0.1819) and mean Ollivier-Ricci curvature (0.5517, persistently positive) provide computable real-time fragility indicators. Across the four most severe load-shedding quarters, Fiedler and Ricci diagnostics offer comparable early-warning signals (mean lead-time difference −0.5 trading days); neither is systematically superior, but their combination is more informative than either alone. The framework contributes toward real-time spectral monitoring of infrastructure-driven systemic risk, supporting Uited Nations Sustainable Development Goal (UN SDG 9) in emerging markets.
The unprecedented increase in the number of Internet of Things (IoT) devices has widened the attack surface of modern-day networks, making them vulnerable to various cyberattacks. Conventional intrusion detection mechanisms face challenges in identifying the subtle correlations between network traffic characteristics and providing consistent detection rates in varying heterogeneous environments. With this context, this research aims at introducing HyperIDS, an innovative Intrusion Detection System (IDS) which combines Hypergraph Learning, Quantum-Inspired Feature Selection and Optimization, and Transformer-based Ensemble Classifier for intelligent IoT cyberattack detection. First, Dual-Fitness Enhanced Gaussian Quantum Particle Swarm Optimization (DFE-GQPSO) approach is utilized to select the most relevant traffic characteristics while reducing feature space and computational cost. These selected traffic features are then converted to a hypergraph form, allowing higher order relations between different entities of the network to be captured. Following this step, Hypergraph Neural Networks (HGNN) is deployed to generate structural and relational representations from the hypergraph structure of the dataset. Long-term dependencies and attack patterns are subsequently extracted using a transformer encoder. The final classification process involves combining CatBoost and XGBoost using stacking ensemble method. In addition, a SHAP-based explainability module is incorporated to ensure transparency and trustworthiness of the developed system. In order to evaluate the proposed framework, HyperIDS is experimentally tested against two commonly used cyber security datasets, namely, TON_IoT and Bot-IoT. Experimental findings have shown superior effectiveness of HyperIDS in detecting cyber-attacks with 98.92, 98.81, 98.76, and 98.78% of accuracy, precision, recall, and F1 score, respectively on the TON_IoT dataset. Similarly, accuracy, precision, recall, and F1 scores of HyperIDS reach 99.14, 99.05, 99.01, and 99.03% on the Bot-IoT dataset. Comparison with conventional machine learning (ML), deep learning (DL), and hybrid intrusion detection techniques have proven HyperIDS's superiority in detecting cyberattacks on IoT infrastructure.
IntroductionDeep learning (DL) shows great potential for predicting biomarkers from routine histopathological slides of gastrointestinal (GI) cancers. Yet most existing models are validated on limited patient cohorts, while pathological image annotation and molecular marker standardization demand substantial professional expertise. To address these gaps, we constructed the Gastrointestinal Cancer Pathological Image Archive (GICPIdb, gicpidb.shubuzuo.top), a dedicated database and web platform covering seven major GI cancer types.MethodsHigh-quality hematoxylin and eosin (H&E)-stained whole-slide images were collected from multiple sources and uniformly processed. Image annotations were performed by board-certified pathologists following standardized protocols. GICPIdb offers five interactive web modules for data uploading, quality control, feature extraction, online annotation and AI-based prediction. Its intuitive interface supports data browsing, retrieval, visualization and downloading.ResultsThe database houses 2,863 pathologist-annotated, uniformly processed, high-quality H&E stained images collected from 2,655 patients. Of these, 1,699 patients were sourced from The Cancer Genome Atlas (TCGA), 182 from the Clinical Proteomic Tumor Analysis Consortium (CPTAC), and 424 from China-Japan Friendship Hospital and 350 from Chifeng Municipal Hospital in Inner Mongolia, China. It also integrates data on over 50 key molecular markers (e.g., MSI, TMB) and prognostic labels related to survival, recurrence and metastasis.DiscussionGICPIdb aims to promote the development of DL-driven AI tools for cancer research and clinical translation. The multi-institutional data collection and standardized annotation pipeline are expected to enhance the generalizability and reproducibility of AI-based prediction models across diverse patient populations.
IntroductionAccurate retail sales forecasting in modern data-intensive environments requires models that not only achieve high predictive accuracy but also scale efficiently as data volume and feature complexity increase. This study proposes a novel cognitively inspired hybrid framework, XGB–ANN–Attn, for interpretable and scalable retail analytics.MethodsThe model introduces feature-level fusion by integrating XGBoost-derived leaf embeddings with deep neural representations, enabling joint modeling of low-order statistical dependencies and high-order nonlinear feature interactions. A lightweight attention mechanism dynamically assigns instance-specific feature importance, enhancing both predictive performance and interpretability while maintaining linear computational complexity with respect to feature dimensionality. Unlike conventional ensemble approaches that operate at the decision level, the proposed framework integrates representations at the level of representations, reducing redundancy and improving computational efficiency. The model is designed for scalability, combining the log-linear complexity of gradient boosting with the linear scaling properties of neural networks and attention mechanisms.ResultsExperimental evaluation on the BigMart and Walmart datasets demonstrates superior performance compared to state-of-the-art models, achieving RMSE = 0.1584 and R2 = 0.9946 on BigMart, and RMSE = 0.8652 and R2 = 0.9568 on Walmart. Furthermore, the framework supports parallelization and distributed deployment, making it suitable for large-scale retail systems.DiscussionThe alignment between attention weights and SHAP explanations provides transparent and actionable insights. The results confirm that the proposed approach offers a scalable, interpretable, and high-performance solution for AI-driven decision support systems in big-data retail environments.
BackgroundSocioeconomic alignment between interviewers and respondents may play a critical role in shaping the accuracy of reported reproductive histories. This study investigates how interviewer characteristics contribute to the reporting of adoption of traditional contraceptive methods using reproductive event calendar data from the fifth round of the National Family Health Survey (2019–21) in India.Data and methodsUsing contraceptive calendar data, a cross-classified multilevel model was employed to assess the interviewer's influence on the reporting of traditional contraceptive method use. The analysis included women aged 15–49 years who reported contraceptive use in the calendar month(s), and the interviewers interviewed them.ResultsMultilevel cross-classified models show that interviewer characteristics significantly influence reporting of traditional contraceptive use following periods of non-use, childbirth, or pregnancy termination. Women interviewed by older interviewers (≥30 years; adjusted odds ratio (AOR) = 1.40, p < 0.01) and those who shared the same religion (AOR = 1.21, p ≤ 0.01) and marital status (AOR = 1.42, p < 0.01) as the interviewer were more likely to report traditional contraceptive method use. While women of the same age group (AOR = 0.89, p < 0.05), same-state residence (AOR = 0.82, p < 0.01), and same caste (AOR = 0.94, p < 0.01) as the interviewer were less likely to report traditional contraceptive method use. Older interviewers (AOR = 0.86, p < 0.01), those with secondary education (AOR = 0.54, p < 0.01), those belonging to the Muslim religion (AOR = 0.43, p < 0.01), and those with a lack of prior experience in conducting surveys (AOR = 0.90, p < 0.10) were associated with lower reporting. Moderate (AOR = 1.35, p < 0.01) and high (AOR = 1.22, p < 0.05) levels of prior experience with the same survey schedule in conducting the survey significantly improved reporting. Distribution of variance across different levels indicates substantial variation at the interviewer level (28.50%), followed by the district (18.50%) and PSU (13.46%) levels, suggesting that interviewer effects play a significant role in shaping the outcome.Discussion and conclusionInterviewer characteristics such as education, age, and prior survey experience influence the reporting of traditional contraceptive use. Similarities or differences between interviewers and respondents in characteristics such as age, marital status, religion, caste, and state of residence also affect the accuracy and completeness of reporting traditional contraceptive use in contraceptive calendar data.
This paper proposes and studies the Discrete Exponentiated Exponential (DEE) distribution, a new two-parameter discrete lifetime model obtained by applying the survival discretization technique to the exponentiated exponential distribution. The DEE distribution possesses a key structural advantage: the single shape parameter α simultaneously controls both the hazard rate shape (increasing for 0 < α < 1, decreasing for α > 1) and the dispersion regime (over-dispersion for small α, under-dispersion for large α), a dual flexibility that most competing discrete models cannot replicate without structural modification. These properties are rigorously established in closed form via the Glaser technique and numerical investigation of the dispersion index. Closed-form expressions are derived for the probability mass function, cumulative distribution function, survival and hazard rate functions, moment generating and characteristic functions, raw and central moments up to the fourth order, and the mean residual life function. Parameter estimation is carried out via maximum likelihood, and a comprehensive Monte Carlo simulation study assesses the finite-sample performance of the estimators and benchmarks it against four competing discrete models under identical settings. The practical advantage of the DEE distribution is demonstrated through three real-world count datasets from oncology, education, and nephrology; the DEE model exhibits competitive and frequently superior performance relative to seven benchmark discrete distributions. Finally, a neutrosophic generalization (DNEE) extends the framework to count data characterized by vagueness, incompleteness, or indeterminacy, with accompanying estimation and illustrative numerical analysis.
Hallucination has become a serious concern in large language models (LLMs), as these models can generate useful yet incorrect or misleading information, which has led to growing research interest in retrieval-augmented generation (RAG) as a mitigation approach. RAG provides LLMs with access to external information, such as databases or documents, which can help them to answer users' questions. Currently, several benchmarks released measure RAG performance on various LLMs; however, evaluations of the noise robustness ability in Arabic are absent. In this paper, we systematically investigated the capabilities of state-of-the-art multilingual LLMs with regard to two essential RAG abilities, noise robustness and negative rejection. To accomplish this, we generated an Arabic benchmark consisting of 300 questions along with 6,196 documents. Then, we assessed the performance of six LLMs in relation to the two aforementioned RAG abilities. The results reveal that all six LLMs were negatively affected when the noise ratio in the external documents was increased. Under the highest noise settings at 80%, the best LLM performance was for Claude-4 sonnet, in which their performance decreased by only 4.67 percentage points. Furthermore, when it comes to the negative rejection task, there has been a significant impact on all six models. The best two models, Claude-4 sonnet and Llama-4, scored 90.67% and 85.67%, respectively, while smaller models, like GPT-3.5, scored 69.33%. Furthermore, our manual analysis reveals that many errors made by LLMs are attributed to over-caution behavior. LLMs often decline to respond probably due to training mechanisms designed to reduce hallucinations. Additionally, other errors occur when there is a high lexical similarity between the question and the words of noisy documents, which causes the model to rely on irrelevant content instead of the correct information.
Real-time data analysis plays an important role in the operation of digital urban systems, where the behavior of residents and the load on services can change over short periods of time. At the same time, traditional analytical approaches based on batch data processing often do not allow timely detection of such changes, which leads to delayed and not always accurate management decisions. This paper introduces an artificial intelligence-driven smart city system (AISSC) for real-time data trend analysis. The proposed AISSC framework processes streaming data in a smart city digital environment. The proposed approach encompasses real-time feature generation and statistical techniques for identifying significant changes. The proposed AISSC solution is executed on the Python platform. The results demonstrate that the proposed AISSC solution achieves a directional accuracy of 98.6% for trend prediction, together with strong numerical forecasting performance with a MAPE of 6.3% and a WMAPE of 7.8%. The framework detects statistically significant deviations within 2.4 s at the sliding-window level while maintaining an end-to-end system update cycle of approximately 5 min. These results demonstrate the capability of the proposed framework to support reliable real-time trend analysis and short-term forecasting for smart city decision-making. This shows that the framework can reliably identify trends and accurately estimate demand for real-time smart city decision-making. The results demonstrate consistent performance improvements compared to representative baseline methods under identical streaming and computational constraints. The proposed AISSC facilitates real-time observation of urban dynamics and enhances short-term forecasting accuracy for informed decision-making in smart city management systems.
IntroductionInternet of Things (IoT) botnet detection faces significant challenges due to the growing intricacy and decreased transparency of Machine Learning (ML) models.MethodsIn this work, we provide an ensemble-based detection system that makes use of a voting classifier made up of a Boosted Decision Tree and a Bagged Random Forest. Using manual feature extraction, the model is trained and assessed using the N-BaIoT dataset. The popular Explainable AI (XAI) method, SHapley Additive exPlanations (SHAP), is used to analyze feature contributions across models while taking important factors like consistency, sensitivity, and monotonicity into account in order to improve the interpretability of the features. SHAP eases the black-box characteristic of sophisticated machine learning models by quantifying the influence of specific features, hence facilitating transparent model interpretation.ResultsAccording to experimental data, the ensemble model is more sensitive than individual classifiers. Additionally, dynamic changes in SHAP values are shown by the weight adjustments made within the voting classifier, highlighting the impact of weight tuning on feature importance.DiscussionThis study highlights the effectiveness of integrating SHAP-based XAI into ensembled models, enhancing the transparency, interpretability, and reliability of IoT botnet detection systems.
IntroductionFraud user identification in telecommunications is hindered by scarce, noisy, and imbalanced labels, while expert rules may provide ambiguous or contradictory evidence.MethodsWe propose DKFraudNet, a knowledge-guided framework that integrates domain knowledge regularization, an attention-adaptive conditional generative adversarial network, and virtual category learning. Expert rules are organized into a Deterministic-Ambiguous-Contradictory evidence taxonomy for controlled pseudo-labeling. Class-conditioned augmentation alleviates data imbalance, and kernel-based similarity refines ambiguous samples. The framework was evaluated using two real-world city-level telecommunications datasets containing 59,015 and 52,183 users, respectively.ResultsDKFraudNet consistently improved downstream classification performance under weak supervision. On the independent City 2 validation set, the CatBoost instantiation achieved an accuracy of 0.911 and an F1-score of 0.909. For operational review prioritization, DKFraudNet-XGBoost required reviewing 40.2% of unverified users to capture 80% of fraud users, corresponding to a 49.71% workload reduction relative to random review, with an AUPRC of 0.954. In operator-side blind verification, genuine-member identification and false-member screening achieved accuracies of 98.9% and 93.3%, respectively.DiscussionThe results show that combining structured domain evidence, adaptive generative augmentation, and uncertainty-aware refinement improves robustness, data efficiency, and operational usefulness for fraud detection under weak supervision.
AI research agents increasingly support ideation, literature search, coding, experimental execution, analysis, and manuscript drafting across the scientific workflow. This progress advances automated discovery, but workflow automation is not scientific cognition. Scientific reasoning is cumulative and path-dependent: it depends on what a researcher has written, read, learned from critique, absorbed through experience, and used as habitual standards for judging novelty, rigor, feasibility, and significance. We propose Mnemo as a framework for modeling scientific cognition in AI research agents. First, scientific cognition may require differentiated memory, organized into distinct spaces for authored work, external reference, critique, experience, and judgment. Second, provenance should be treated not as passive metadata but as memory routing, because source origin helps determine cognitive function in reasoning. Third, new ideas may be better modeled as controlled collisions across memory spaces, filtered by judgment, rejection, and epistemic calibration, than as generic recombination from model priors. Mnemo motivates a research agenda for AI in science centered on routing fidelity, critique use, judgment alignment, rejection quality, and epistemic calibration.
IntroductionNews coverage of politically sensitive events involves framing choices that accumulate across thousands of headlines, years, and media institutions. Likewise, no single reader or journalist can track these patterns as they happen. This study examines how English-language news media framed Rodrigo Duterte's anti-drug campaign in the Philippines from June 2016 to June 2025.MethodsThe study applies BERTopic topic modeling and VADER sentiment analysis to 1,133 headlines collected through Google News's RSS endpoint. Institutional Theory organizes the results around three pressures: normative, coercive, and mimetic.ResultsTen topic clusters emerged from the corpus, centered on human rights violations, International Criminal Court (ICC) legal proceedings, and post-presidency accountability narratives. Domestic Philippine outlets covered human rights themes in a larger share of headlines than international outlets during the campaign's early years, but that pattern reversed after Duterte left office in 2022. By 2024 and 2025, international outlets covered human rights themes in roughly 64 percent of headlines, compared with about half for domestic outlets. Sentiment analysis classified 807 headlines as negative, 203 as neutral, and 123 as positive. Negative coverage rose sharply in early 2025, around Duterte's arrest on an ICC warrant. Headline language also clustered around three institutional events: the 2018 ICC filing, the 2020 United Nations (UN) Human Rights Report, and the 2021 ICC investigation authorization, each marked by distinct word patterns across positive, negative, and neutral coverage.DiscussionNormative pressure explains why domestic outlets led on human rights framing before any international body had acted. Coercive pressure from the ICC and UN corresponds to the sentiment spikes around each event. Mimetic pressure explains how framing conventions moved between domestic and international outlets over time, first from domestic to international sources and later in the other direction. The study also considers how Google News's own curation choices shape which outlets and frames appear in the corpus. These findings suggest computational analysis is both a research method and a way to hold a media record open to scrutiny after the news cycle moves on.
Against the backdrop of electrification and supply chain resilience, the global automotive industry is in an era of great transformation. Using component supply information provided by MarkLines, this study investigated the impact on inter-firm dependencies among major global automakers from 2018 to 2024. Specifically, it analyzed changes in the community structure of the interdependency network between manufacturers, based on common suppliers for each model year and component category. The results revealed the impact of electrification: a reduction of roughly two-thirds (about 64%) in transactions for internal combustion engine (ICE) powertrains and an approximately eight-fold increase in e-powertrain transactions. Furthermore, the results of community detection also revealed a structural reorganization: while geographical clustering among manufacturers intensified for ICE components, new cross-border interdependencies formed for e-powertrain components, leading to the fragmentation of the traditionally integrated Japan-US-Europe bloc and the emergence of a China-centric ecosystem. This study provides new empirical evidence on the structural realignment of the global automotive value chain, offering important implications for management strategy and industrial policy in an era of great technological and geopolitical change.
This study explored the role of artificial intelligence in higher education in Saudi Arabia. It aims to identify trends, key contributors, influential papers, collaborations, and important research areas from January 2022 to February 2026 to guide future research. The researchers used bibliometric and content analyses. This combined quantitative descriptive methods and network analysis with qualitative content analysis of the most-cited articles. They extracted data from Scopus and the Web of Science, resulting in 66 documents after removing duplicates, editorials, and notes. The analytical techniques included the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA), co-word analysis, citation analysis, co-authorship analysis, and bibliographic coupling. VOSviewer supported the visualization. The key findings show that King Abdulaziz University, Qassim University, and King Saud University are the top contributors. A total of 52 papers have been published in journals indexed by WOS and Scopus, while 14 papers are indexed only by Scopus. Additionally, five major groups revealed significant correlations among various word pairs: High-education, Artificial intelligence, AI Chatbot, learning systems, and ChatGPT. The leading journals include Sustainability, Acta Psychological, and Systems, with notable authors such as Al-Harbi. Content analysis highlights the potential of AI to improve learning, boost administrative efficiency, and drive innovation in Saudi Arabia. This study offers practical recommendations for students, universities, and policy makers. This study has several limitations that should be addressed in future studies.
IntroductionARP spoofing poses a major security threat to Internet of Medical Things (IoMT) networks by enabling man-in-the-middle attacks that compromise the integrity of life-critical communications. Existing intrusion detection methods fail to simultaneously address temporal attack dynamics, unequal medical safety requirements, and explicit control of false negative rates.MethodsThis study proposes the Self-Healing IoT-Optimized Random Forest (SH-IORF) framework, which integrates temporal behavioral feature engineering, validation-guided cost-sensitive learning, and medical safety-constrained threshold optimization. To ensure methodological rigor and prevent information leakage, a stratified three-way partitioning strategy consisting of training, validation, and completely held-out testing datasets was employed. Class penalty weights and operating thresholds were determined exclusively from the validation dataset.ResultsExperimental evaluation on the CICIoMT2024 benchmark demonstrated that the proposed SH-IORF framework achieved 99.90% accuracy, 99.83% recall, 99.95% precision, a 0.9989 F1-score, and an AUC-ROC of 0.9996. The framework limited the false negative rate to 0.17%, satisfying the predefined medical safety constraint (FNR ≤ 0.5%), corresponding to 40 missed detections among 23,390 attack samples and 12 false alarms across 28,768 benign traffic instances.DiscussionThe results demonstrate that the proposed framework provides stable and safety-oriented intrusion detection capability under heterogeneous IoMT deployment conditions while maintaining strict testing independence and robust performance under rigorous evaluation settings.
BackgroundObstructive Sleep Apnea-Hypopnea Syndrome (OSAHS) has a high global prevalence and is prone to causing various serious complications. Our objective is to develop severity stratification of OSAHS by integrating multiple commonly available clinical features based on machine learning (ML).Materials and methodsThis study collected data from 432 cases at Qujing Central Hospital in Yunnan Province, integrating 25 clinical feature variables. The cases were randomly split into training (70%) and validation (30%) sets. The importance of the 25 features was analyzed.ResultsIt showed that the HCY, TBIL, BMI, GGT, and Age made significant contributions to OSAHS severity. We established five machine learning models-Multilayer Perceptron (MLP), Random Forest, XGBoost, LightGBM, and Support Vector Machine (SVM)-by integrating 25 clinical features. Through cross-validation and continuous adjustment of model parameters, the optimal predictive model was determined. By calculating model accuracy and F1-score, XGBoost was identified as the best-performing model, achieving an area under the curve (AUC) of 0.63, an accuracy of 75% and an F1-score of 65.60.ConclusionIn this study, we established a predictive model for the severity stratification of OSAHS based on machine learning algorithms. The XGBoost model demonstrated superior predictive performance.
The rapid growth of IoT and IIoT expands the cyber-attack surface of interconnected and safety-critical systems, and, as such, IDSs have become a fundamental security mechanism. Although very impressive results have been reported for machine learning and deep learning-based IDS in benchmark datasets, these gains often do not generalize to real-world deployments owing to dataset design limitations, realism deficits, and evaluation biases, rather than inherent flaws in detection algorithms, which can lead to significant vulnerabilities in actual operational environments. This study presents a dataset-centric review of widely used intrusion detection datasets from the IIoT, IoT, and traditional network domains. A unified taxonomy differentiates datasets based on the domain context, traffic representation, protocol semantics, and attack modeling assumptions. Based on a common analytical framework, each dataset was reviewed regarding its realism, coverage of the threats, class imbalance, temporal continuity, and modern ML/DL-based evaluation of the IDS. The cross-dataset analysis conducted in this study shows that, in addition to the fact that model architecture and feature engineering play a major role, several studies indicate that the simplicity of the datasets, the class imbalance, and the repetitive attack patterns as well as the evaluation methods can affect accuracy of the IDS. This work further underlines the remaining gaps, such as zero-day and adaptive attacks, limited encrypted traffic, weak temporal evolution, poor support for federated learning, and sparse annotations for explainable IDSs. Finally, this study presents future directions for dataset design aligned with the requirements of next-generation IDSs by highlighting digital twin-based IIoT environments, edge-cloud collaborative data generation, sequential traffic modeling, and explainability-oriented annotations that can ensure robust, trustworthy, and deployment-ready IDS solutions.
Attention-related Pilot Performance Decrements (APPD) contribute substantially to aviation incidents, yet existing electroencephalography (EEG)-based monitoring methods often lack generalization, robustness to noise, and effective temporal-spectral integration. We propose a temporal-spectral fusion transformer (TF-T) combining multi-scale preprocessing, dual-stream temporal and spectral feature extraction, and transformer-based fusion with enhanced temporal-spectral integration and multi-resolution feature processing for multiclass cognitive state recognition. Three variants (TF-T1-TF-T3) are evaluated on controlled and ecologically realistic EEG datasets under clean and noise-augmented (Gaussian, Uniform, COMBO) conditions, using chronological partitioning to avoid temporal leakage. TF-T2 achieves the highest clean-data accuracy (99.2%), while TF-T3 offers superior robustness, improving Macro-F1 by ~4.5-4.7 points across all noise types and outperforming state-of-the-art baselines by up to +8 Macro-F1 under COMBO noise, supporting its deployment in perturbation-prone aviation environments.
