Depression significantly affects health, manifesting as alterations in typical emotional responses. Its diagnosis depends on subjective evaluations by clinicians, which are often time-intensive. Electroencephalogram (EEG) signals offer a viable solution for aiding diagnosis through computational means. However, current methods primarily focus on binary classification of depression, neglecting the quantification of depression risk. We propose the Multi-Cue EEG Network (MCEEGNet), which consists of parallel branches of EEGNet that extract features from various emotional stimuli to approximate scores on the Patient Health Questionnaire (PHQ-9). MCEEGNet aims to identify depression in patients using EEG signals and assess the severity of the condition. Our method achieved 91.13% accuracy in classification and reported Mean Squared Error (MSE) and Mean Absolute Error (MAE) of 20.45 and 3.28, respectively, in the Multi-Emotion Induced EEG Depression Database. The experimental outcomes suggest that MCEEGNet is highly effective in diagnosing subthreshold depression, offering a comprehensive system for evaluating clinical depression through EEG analysis influenced by multiple emotional cues, thereby meeting the need for quantitative depression evaluation.

Skin lesions are one of the most prevalent form of diseases existing among us. Early detection and classification of potentially malignant skin lesions can give us a lead in the fight against skin cancer. There are many lesion classification divisions on medical grounds; however, an automated system that detects and classifies a majority of these classes is not prevalent. In view of this scenario, our proposed study aims to classify the input skin lesion images into nine classes, namely, squamous cell carcinoma (SCC), Basal cell carcinoma (BCC), melanocytic nevi (NV), actinic keratoses and intraepithelial carcinoma (AKIEC), melanoma (MEL), seborrheic keratosis (SEK), dermatofibroma (DF), benign keratosis like lesions (BKL), and vascular lesions (VASC). The proposed methodology uses contrast stretching as an image enhancement technique to facilitate efficient Region of Interest (ROI) segmentation. The novelty of the proposed study lies in the first-hand implementation of Vision Transformer (ViT) for feature extraction in the domain of skin lesion detection. Finally, a light-weight multi-layer perceptron (MLP) composed of fully connected layers is used for multinomial classification. Combining the aforementioned techniques, the proposed method achieves training accuracy of 98% and testing accuracy of about 93.22%. The impressive performance across nine distinct categories represents a significant milestone. This success demonstrates the model’s scalability, suggesting it can be effectively extended to a broader array of diagnostic classes in future research.

Introduction Artificial intelligence (AI) is increasingly embedded in journalism, yet audience responses may depend on both AI provenance, meaning who or what is presented as having written the story, and transparency cues that disclose AI use. This systematic literature review synthesises empirical studies examining how AI provenance cues and AI disclosure cues in journalism affect perceived credibility and trust. Methods Following PRISMA 2020 and PRISMA-S, Scopus and Web of Science Core Collection were searched on 2 February 2026 for English-language, peer-reviewed journal articles and conference papers. Searches yielded 492 records. After deduplication and pre-screen exclusions, 290 records were screened at title/abstract level, and 47 studies with retrievable full texts were included. A structured narrative synthesis was conducted, guided by the Synthesis Without Meta-analysis (SWiM) guideline, to map study designs, cue operationalisations, outcome targets (message, source, outlet), and moderators. Results Across heterogeneous designs, AI provenance cues were not associated with a consistent “AI penalty”: most extractable results indicated no difference between AI-attributed and human-attributed news, and observed effects were typically conditional on topic, baseline trust, outlet/source cues, and whether human oversight was signalled. Evidence on disclosure cues was limited (10 studies) and was dominated by null or conditional findings. Scepticism appeared more likely when disclosures implied full automation without accompanying accountability or oversight information. Discussion A Cue–Inference–Target (CIT) framework is proposed to explain when AI cues shift epistemic-quality versus normative-legitimacy judgments. Future research should use factorial designs that separate provenance from disclosure and standardise reporting of cue wording, placement, and validated outcome measures.

This work presents a hybrid deep learning approach for identifying students who are likely to experience academic difficulties in virtual learning environments. The proposed framework is evaluated on the Open University Learning Analytics Dataset (OULAD) and combines two complementary types of information: temporal patterns of learner activity captured using Bidirectional Long Short-Term Memory (Bi-LSTM) networks and relatively stable student attributes modeled through a Multi-Layer Perceptron (MLP). To account for the fact that student engagement does not contribute equally across different phases of a course, a temporal attention mechanism is incorporated to emphasize the most informative periods over the 32-week learning timeline. The experimental results indicate that the hybrid model achieves stronger predictive performance than baseline approaches, with a ROC-AUC of 0.95 and a weighted F1-score of 0.90. Analysis of the attention weights suggests that both early engagement and activity toward the end of the course (Weeks 25–32) play an important role in predicting final outcomes. To address the imbalance between successful and at-risk learners, a cost-sensitive training strategy is adopted, resulting in a recall of 0.79 for the at-risk group. Overall, these findings suggest that integrating temporal behavioral signals with static student characteristics leads to more reliable risk prediction and provides a useful basis for data-informed academic support in online learning contexts.

[This corrects the article DOI: 10.3389/frai.2026.1719703.].

Background Detecting exploitative or unethical player behavior in online gaming platforms is challenging due to ambiguous gray-area actions that are neither clearly legitimate nor illegal. Methods This study presents an interpretable behavior analysis framework combining anomaly detection, synthetic data augmentation using Conditional Tabular GAN (CTGAN), and ensemble learning with post-hoc explainability. Datasets from a massively multiplayer online role-playing game (MMORPG) spanning 88 days (49,739 player sessions) were augmented to address class imbalance. Anomaly detection using an Encoder-Decoder GAN for Anomaly Detection (EGBAD) approach generated anomaly-aware features. A stacked ensemble model combining Random Forest, XGBoost, and Artificial Neural Networks was developed, with SHAP and LIME providing explanations for predictions. Results The proposed framework achieved 95.98% accuracy, 0.915 ROC-AUC, and 0.90 macro F1-score, outperforming baseline models. The integration of CTGAN improved minority class recall by 5–7 percentage points, while EGBAD-derived anomaly features enhanced gray-area detection. Human-in-the-loop triage for low-confidence predictions (6.8% of cases) achieved 75% human-AI agreement with reduced false positives (21% decrease) and false negatives (17% decrease). Discussion The framework successfully balances automated detection with human oversight, providing transparent, interpretable decisions for player behavior moderation while maintaining fairness and reducing wrongful enforcement actions.

There is a great need to categorize cancer types for early cancer detection and treatment. RNA-Seq data is essential for getting insight into the differentially expressed genes. Due to its high dimensionality and complexity, performing an analysis on RNA-Seq data is quite challenging. In the past, RNA-Seq data were analyzed for a single cancer type as a two-class problem (either positive or negative) and did not contain information from other classes of cancer types. To classify different cancer types and discover the most promising genes, RNA-Seq data for different types of cancer should be examined. Multiple repositories offer RNA-Seq-based cancer types data. The present study incorporates a dataset from the Mendeley repository for classification. RNA-Seq values are then converted to their respective 2D images using some transformations. The classification problem is handled by five Transfer Learning (TL) algorithms (VGG16, VGG19, Resnet50, Resnet101, and Resnet152). Four different splitting strategies are applied for each classifier presented in the results and discussion section. A comparative analysis is also carried out with and without data augmentation. Results show that classifiers perform best at a split of 70-30. VGG16 attained the best position on overall results by achieving an accuracy of 95%. Hence, VGG16 is the leading TL algorithm for classification among all the accessible models and is not difficult to execute and easy to comprehend.

Effective patient education is critical for informed decision-making and adherence in kidney transplantation. Generative artificial intelligence (AI), particularly large language models (LLMs), has the potential to enhance patient education in kidney transplantation; however, its factual accuracy and readability remain incompletely characterized. We evaluated the performance of the GPT-5.1 (2025) model in generating plain-language explanations for 100 clinically relevant kidney transplantation terms. Explanations were generated using a standardized prompt (first round) and a revised prompt explicitly requesting an eighth-grade reading level or lower (second round). Accuracy was assessed by expert reviewers using a 5-point Likert scale, while readability was evaluated using the Flesch Reading Ease (higher score indicated easier readability of the text) and Flesch-Kincaid Grade Level (higher score indicated higher education level required to understand the text) score. The study was conducted in November 2025. All AI-generated explanations demonstrated high accuracy, with no clinically significant errors. In the first round, the mean Flesch Reading Ease score was 23.6 ± 23.4, indicating very difficult readability, and 46% of explanations required a college-level reading ability (mean Flesch-Kincaid Grade Level 13.4 ± 4.8). Following prompt revision, readability improved substantially. The mean Flesch Reading Ease score increased to 62.4 ± 7.5, corresponding to standard readability, and all explanations were written at a middle school level or below (mean Flesch-Kincaid Grade Level 6.3 ± 1.1). GPT-5.1 generated highly accurate explanations of kidney transplantation terms across prompting strategies. Explicit readability-focused prompting substantially improved readability without compromising accuracy, underscoring the importance of prompt design when deploying LLMs for patient-centered education in transplantation.

Public acceptance of artificial intelligence (AI) in legal decision-making has been primarily explained through individual differences in personality traits and general attitudes toward technology. However, emerging evidence suggests that contextual features of legal disputes themselves may systematically influence preferences for AI versus human adjudicators. Across two studies with Japanese participants (N = 1,384 and N = 596), we examined whether psychological characteristics of dispute content—beyond demographics and individual traits—shape acceptability judgments for algorithmic adjudication. Study 1 employed exploratory factor analysis on acceptability ratings across 46 legal dispute vignettes, revealing a robust two-dimensional structure distinguishing interpersonal-relational disputes (where human adjudicators were strongly preferred) from institutional-procedural disputes (where AI acceptance was comparatively higher, though not surpassing human preference in most cases). Study 2 replicated this dimensional structure in an independent sample and demonstrated that experimentally manipulated contextual features—emotional involvement and prototypicality—systematically modulated acceptability judgments, with effects varying by dispositional trust, AI-specific attitudes, and gender. AI-specific expectations emerged as the strongest predictor of acceptance (η2 = 0.252), and a three-way interaction among emotional involvement, gender, and prototypicality indicated that contextual effects are moderated by individual characteristics. These findings suggest that the psychological features of dispute content constitute an overlooked dimension in AI acceptance research, extending beyond technology acceptance models to fundamental questions about how individuals construe social problems and allocate adjudicative authority. We discuss limitations related to measurement approaches, alternative psychological mechanisms, and directions for future research employing real-world case materials and direct assessment of cognitive processes.

BackgroundGenerative artificial intelligence (GenAI) is rapidly entering knowledge work, yet organizational psychology lacks a clear account of when and how GenAI functions as a job resource in the Job Demands–Resources (JD-R) model, especially in high-demand academic work.MethodsWe conducted a JD-R–guided directed qualitative content analysis of a de-identified human–AI interaction log generated during routine academic work. The log was segmented into interactional episodes and coded using function-first descriptors. Codes were mapped deductively to three JD-R–aligned resource domains (cognitive; structural/strategic; emotional/psychosocial) and a boundary-conditions stream (human oversight; data integrity/traceability).ResultsAcross 15 episodes, GenAI enacted cognitive functions that reduced informational complexity, structural/strategic functions that increased planning capacity and task structure, and emotional/psychosocial functions expressed through observable efficacy-reinforcing and action-orienting cues. Boundary-condition coding showed that benefits depended on human oversight and integrity routines; in episodes requiring substantial verification or traceability work, GenAI could shift rather than reduce demands.ConclusionGenAI can operate as a conditional job resource in demanding academic knowledge work, but sustainable benefit requires explicit human-in-the-loop oversight and data integrity practices that support reliable and responsible use.