<?xml version="1.0" encoding="utf-8"?>
    <rss version="2.0">
      <channel xmlns:content="http://purl.org/rss/1.0/modules/content/">
        <title>Frontiers in Artificial Intelligence | New and Recent Articles</title>
        <link>https://www.frontiersin.org/journals/artificial-intelligence</link>
        <description>RSS Feed for Frontiers in Artificial Intelligence | New and Recent Articles</description>
        <language>en-us</language>
        <generator>Frontiers Feed Generator,version:1</generator>
        <pubDate>2026-04-05T12:22:51.989+00:00</pubDate>
        <ttl>60</ttl>
        <item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1796099</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1796099</link>
        <title><![CDATA[Dynamic-focus transformer for point cloud segmentation]]></title>
        <pubdate>2026-04-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ziwen Wang</author><author>Xiaoting Fan</author><author>Mei Yu</author><author>Jianlu Liu</author><author>Shuai Wang</author><author>Yonghua Wang</author><author>Chuanfu Wu</author>
        <description><![CDATA[Transformer-based methods have significantly advanced 3D point cloud segmentation by effectively capturing long-range dependencies. However, the global or fixed-window self-attention mechanisms they often employ suffer from computational redundancy and overfitting due to processing excessive, potentially irrelevant key-value pairs for each query. To address this, we propose the Dynamic-Focus Transformer, a novel architecture that introduces a data-dependent adaptive attention mechanism. Through learned soft point masks, we selectively sparsify keys and values to focus on semantically critical regions. Our method enables flexible, input-adaptive receptive fields without the heavy memory overhead associated with per-point offset learning in deformable designs. Furthermore, when integrated into a U-Net-style encoder-decoder, our method attains a highly efficient balance between modeling capability and computational cost. Extensive experiments on S3DIS and ScanNetv2 benchmarks demonstrate that our method achieves state-of-the-art performance with notably improved efficiency, validating its effectiveness for large-scale point cloud understanding.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1755151</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1755151</link>
        <title><![CDATA[The silent accumulation: AI as mental contaminant]]></title>
        <pubdate>2026-04-02T00:00:00Z</pubdate>
        <category>Hypothesis and Theory</category>
        <author>Vladimír Šucha</author>
        <description><![CDATA[This paper introduces a theoretical framework for understanding how seemingly innocuous AI systems create cumulative effects on human cognition, emotion, and agency that current governance approaches fail to address. Drawing on environmental health science, we propose that “low-risk” AI applications, those falling below regulatory thresholds in frameworks such as the EU AI Act, function as cognitive environmental contaminants whose collective and sustained presence may reshape human psychological capacities. We operationalise cumulative AI exposure along five dimensions (frequency, duration, intensity, diversity of systems, and developmental timing) and identify five pathways through which cumulative effects may manifest: attention erosion, emotional dependency, social connection alteration, decision-making dependency, and identity fragmentation. For each pathway, we distinguish empirical regularities documented in existing research, plausible mechanisms through which cumulative effects may operate, and speculative population-level hypotheses that require empirical testing. Situating our framework against adjacent literatures including technostress, cognitive offloading, hypernudging, and automation bias, we argue that the distinctive contribution lies in the cumulative, cross-system, population-level analytical paradigm and its governance translation. We propose three governance mechanisms—cumulative impact assessment extending existing algorithmic auditing frameworks, cognitive-social environmental monitoring using validated psychometric instruments, and economic valuation of cognitive-social ecosystem services, accompanied by a phased validation strategy. The framework is offered as a complement to existing risk-based governance, addressing a specific gap: the systematic invisibility of effects that emerge from the interaction of multiple AI systems over extended time periods.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1771088</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1771088</link>
        <title><![CDATA[Few-shot deployment of pretrained MRI transformers in brain imaging tasks]]></title>
        <pubdate>2026-04-01T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Mengyu Li</author><author>Guoyao Shen</author><author>Chad W. Farris</author><author>Xin Zhang</author>
        <description><![CDATA[IntroductionTransformer-based deep learning has shown great potential in medical imaging, but its real-world applicability remains limited due to the scarcity of annotated data. This study aims to develop a practical framework for the few-shot deployment of pretrained MRI transformers across diverse brain imaging tasks.MethodsWe employ a Masked Autoencoder (MAE) pretraining strategy on a large-scale, multi-cohort brain MRI dataset comprising over 31 million 2D slices to learn transferable representations. For classification tasks, a frozen MAE encoder with a lightweight linear head (MAE-classify) is used. For segmentation, we propose MAE-FUnet, a hybrid architecture that fuses pretrained MAE embeddings with multi-scale CNN features. Extensive evaluations are conducted on multiple datasets, including NACC, ADNI, OASIS, NFBS, SynthStrip, and MRBrainS18, under controlled few-shot settings.ResultsThe proposed framework achieves state-of-the-art performance in MRI sequence classification, reaching an accuracy of 99.24% with only 6,152 trainable parameters. For segmentation tasks, MAE-FUnet consistently outperforms strong baselines, achieving superior Dice and IoU scores across skull stripping and multi-class anatomical segmentation benchmarks. The model also demonstrates enhanced robustness and stability under data-limited conditions, with lower performance variance compared to competing methods.DiscussionThese results highlight the effectiveness of pretrained MAE representations for few-shot medical imaging tasks. The proposed framework enables efficient, scalable, and adaptable deployment of transformer-based models in data-constrained clinical environments. The fusion of global transformer embeddings with local CNN features provides a generalizable design paradigm for a wide range of medical imaging applications.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1776546</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1776546</link>
        <title><![CDATA[AI-augmented reliability in CI/CD: a framework for predictive, adaptive, and self-correcting pipelines]]></title>
        <pubdate>2026-04-01T00:00:00Z</pubdate>
        <category>Hypothesis and Theory</category>
        <author>Rohit Dhawan</author><author>Mohit Dhawan</author>
        <description><![CDATA[Modern CI/CD pipelines face a critical challenge. While AI tools accelerate code generation, static pipelines have become the primary bottleneck to delivery velocity. Flaky tests and pipeline noise create a persistent challenge, with reported failure rates ranging from 11 to 27 percent for test flakiness and 5–16 percent for noise-induced build failures. This forces teams to spend more time investigating false failures than building features. As systems scale across regions and dependencies, these problems compound and threaten the fundamental promise of continuous delivery. We introduce a framework that transforms CI/CD pipelines from deterministic scripts into intelligent, adaptive systems. At its core is the Sense-Analyze-Predict-Act-Learn loop, which we call SAPAL. This loop extends classical adaptive models with CI/CD specific capabilities including flakiness characterization, dependency risk scoring, multi-region awareness, and developer feedback. We operationalize this loop through a five-layer architecture spanning data collection, reliability intelligence, predictive modeling, adaptive execution, and human-AI collaboration. Three novel metrics quantify pipeline intelligence. Pipeline Health Index measures overall reliability. Test Stability Score identifies flaky patterns. Failure Prediction Confidence validates model accuracy. Three scenarios demonstrate application to real CI/CD challenges. Intelligent retry strategies, grounded in empirical studies of flaky test detection and resolution, project 60 percent reduction in flaky-induced build failures. ML-based test selection techniques from recent literature suggest 50 to 80 percent reduction in feedback time. Stability-aware deployment orchestration adapts rollout strategies to regional reliability patterns. These projections synthesize findings from published component studies rather than measurements from unified framework deployment. By enabling pipelines to learn from executions, predict with calibrated confidence, and adapt to behavior patterns, this framework provides a practical path toward reliable delivery at scale where intelligence is essential, not optional.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1770922</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1770922</link>
        <title><![CDATA[Benchmark datasets for predictive maintenance challenges in steel manufacturing]]></title>
        <pubdate>2026-04-01T00:00:00Z</pubdate>
        <category>Data Report</category>
        <author>Jakub Jakubowski</author><author>Szymon Bobek</author><author>Grzegorz J. Nalepa</author>
        <description></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1749527</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1749527</link>
        <title><![CDATA[Explainable AI in healthcare: a systematic review of XAI use cases in imaging, diagnostics, and rehabilitation]]></title>
        <pubdate>2026-04-01T00:00:00Z</pubdate>
        <category>Systematic Review</category>
        <author>Apoorva Aravindkumar</author><author>Marimuthu Ramadoss</author><author>Saqhibuddeen Ahmed Fakhruddin Ahmed</author><author>Vidhya Sampath</author><author>Kishor Lakshminarayanan</author>
        <description><![CDATA[BackgroundExplainable artificial intelligence (XAI) is used in healthcare to make machine-learning outputs more transparent and clinically usable. This is important because many machine learning models work like a “black box” which can hide bias, reduce trust in the model. XAI addresses this problem by showing which features or image regions influenced a result, either for one patient or across a dataset.ObjectivesOur objective is to provide a clear, systematic review of how XAI is being used in healthcare. We summarize the main XAI methods, the data and models they are paired with, and how these explanations support clinical understanding across imaging, diagnosis, and rehabilitation.MethodsWe performed a systematic review with narrative synthesis (2020–2025) of 36 empirical studies across three verticals–Imaging (n = 10), Diagnosis (n = 16), and Rehabilitation (n = 10) that are identified via PubMed/MEDLINE, IEEE Xplore, and Google Scholar, following PRISMA 2020 guidelines. We included research studies that employed XAI in the three mentioned verticals. We excluded review articles and viewpoint studies. Screening numbers were - records identified 1,481; duplicates removed 647; other removals 187; screened 647; excluded 532; reports sought 115; not retrieved 31; assessed 84; full-text excluded 48; included 36. From each study we extracted ML models, XAI methods, study design, methodologies, and dataset/source. Meta-analysis was not undertaken due to heterogeneity.ResultsAcross 36 studies, SHAP was used in 21 studies, Grad-CAM in ~12/36, and LIME in ~11/36. A clear method-modality fit emerged with Imaging predominantly using saliency/heat-map methods, especially Grad-CAM, for spatial evidence. Diagnosis and Rehabilitation were dominated by feature-attribution tools like SHAP and LIME for global and case-level explanations. Many papers combined ≥ 2 explainers to cross-check interpretations namely SHAP+LIME, and Grad-CAM + LIME.ConclusionRecent healthcare XAI demonstrates consistent method-modality fit and frequently combine two or more methods, helping translate opaque predictions into clinician-oriented reasoning. To enable trustworthy deployment, future work should pair these practices with standardized XAI reporting, faithfulness/stability assessments, and external, cross-site validation.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1766899</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1766899</link>
        <title><![CDATA[A multi-layer annotated corpus for information extraction in Russian clinical NLP]]></title>
        <pubdate>2026-03-31T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Anar Sultangaziyeva</author><author>Madina Sambetbayeva</author><author>Nurzhan Mukazhanov</author><author>Bayangali Abdygalym</author><author>Sandugash Serikbayeva</author>
        <description><![CDATA[IntroductionClinical exome sequencing reports contain valuable genetic and phenotypic information but are typically stored in unstructured text form, making automated biomedical information extraction challenging. For the Russian language, publicly available annotated corpora for genetic report analysis remain extremely limited.MethodsWe present GENEXOM, the first multi-level annotated corpus of Russian-language clinical exome sequencing reports designed for biomedical information extraction. The corpus includes 5,318 reports (318 authentic and 5,000 synthetic) and comprises 16 entity types and 7 relation types aligned with HGVS, OMIM, ClinVar, and ACMG/AMP standards. Annotation was performed in the Label Studio platform by expert geneticists. Baseline transformer models (RuBERT, RuBioBERT, ModernBERT) were fine-tuned for Named Entity Recognition (NER) and Relation Extraction (RE).ResultsThe annotation achieved span-level F1-IAA = 0.83 and macro κ = 0.79 ± 0.04, indicating substantial inter-annotator agreement. Among the evaluated models, ModernBERT achieved the best performance with F1 = 0.88 ± 0.03 for NER and F1 = 0.836 ± 0.04 for RE on the held-out test set.DiscussionThe GENEXOM corpus provides a linguistically and clinically adapted resource for Russian medical NLP and supports downstream tasks such as variant interpretation, phenotype–disease mapping, and biomedical knowledge graph construction. The corpus and accompanying code are publicly available for research purposes.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1749517</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1749517</link>
        <title><![CDATA[From simulated empathy to structural attunement: Realtime Editable Memory Topology and the evolution of emotionally grounded AI]]></title>
        <pubdate>2026-03-31T00:00:00Z</pubdate>
        <category>Perspective</category>
        <author>John Albanese</author>
        <description><![CDATA[Large language models (LLMs) and retrieval-augmented generation (RAG) systems have achieved remarkable linguistic fluency, and many now implement persistent cross-session memory at the application layer. However, these mechanisms typically rely on external storage and reinjection of stored content rather than structural reorganization of memory relationships. As a result, they remain limited in their ability to integrate affective salience into a dynamically evolving internal memory topology capable of supporting coherent long-term behavior. To address this gap, we introduce Realtime Editable Memory Topology (REMT), an architectural framework for imbuing conversational agents with persistent autobiographical memory organized as an evolving graph of emotionally valenced nodes. REMT formalizes synthetic neuroplasticity through explicit update rules governing edge reinforcement, decay, and pruning, and introduces a bounded Mood Index that modulates retrieval bias and response generation as a function of accumulated affective experience. In this Perspective, we argue that memory-grounded architectures integrating insights from cognitive science, affective computing, and memory-augmented neural systems are necessary for building adaptive conversational agents with stable long-term interactional tendencies. We conclude by outlining a roadmap for empirical validation using an internally developed evaluation framework, with results to be reported in a future Original Research article.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1702756</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1702756</link>
        <title><![CDATA[A deep learning-based approach for detecting anomalous behavior in safety-critical spaces]]></title>
        <pubdate>2026-03-31T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Aqib Anees</author><author>Syed Asim Jalal</author><author>Hassan Jalil Hadi</author><author>Naveed Ahmad</author><author>Mohamad Ladan</author>
        <description><![CDATA[Wrong-turn violations in safety-critical spaces such as road roundabouts are a type of traffic violation that can lead to traffic congestion and increase the risk of road crashes. Although many researchers have focused on detecting various traffic violations, wrong-turn violations have not received enough attention. This may be due to a lack of relevant datasets. This study aims to address this gap. We developed a deep learning–based approach to detect wrong-turn traffic violations at roundabouts. The proposed system captures video from strategically placed cameras at roundabouts, which is then fed into an artificial intelligence (AI) model capable of detecting vehicles committing wrong-turn violations in real time. For this purpose, we utilized the popular You Only Look Once (YOLO) algorithm. Due to the absence of an existing dataset for this specific type of violation, we created our own. Images were collected and annotated from local roundabouts in Peshawar, Pakistan. The YOLO model was trained on this dataset and evaluated using standard performance metrics, including accuracy and recall. The results suggest that the proposed approach has strong potential for refinement and real-world implementation.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1819135</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1819135</link>
        <title><![CDATA[Correction: LSML-SF: a lightweight stacked ML approach for spreading factor allocation in mobile IoT LoRaWAN networks]]></title>
        <pubdate>2026-03-30T00:00:00Z</pubdate>
        <category>Correction</category>
        <author>Arshad Farhad</author><author>Muhammad Ali Lodhi</author><author>Farhan Nisar</author><author>Hassan Jalil Hadi</author><author>Naveed Ahmad</author><author>Mohamad Ladan</author>
        <description></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1794925</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1794925</link>
        <title><![CDATA[Hybrid expert system for lifestyle recommendations in hypertensive patients]]></title>
        <pubdate>2026-03-30T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Miguel Angel Valles-Coral</author><author>Lloy Pinedo</author><author>Richard Injante</author><author>Jorge Raul Navarro-Cabrera</author><author>Karen Luz Quintanilla-Morales</author><author>Sarita Saavedra</author><author>Jorge Valverde-Iparraguirre</author><author>Flor Enith Leveau-Barrera</author><author>Nerida Gonzalez-Gonzalez</author>
        <description><![CDATA[IntroductionHypertension management requires personalized lifestyle interventions, yet clinical decision-making often relies on manual assessment and limited decision-support tools. This study presents a hybrid clinical decision-support system that integrates unsupervised machine learning with rule-based expert reasoning to generate personalized lifestyle recommendations for hypertensive patients.MethodsA real-world dataset of 615 clinical records obtained from routine healthcare services was analyzed. A preprocessing pipeline including data imputation, normalization, and dimensionality reduction was applied prior to patient stratification. Principal Component Analysis (PCA) preserved the dominant latent structure of the dataset, followed by K-Means clustering to identify patient profiles. The resulting clusters were integrated into a rule-based inference engine structured across six lifestyle intervention domains: physical activity, stress management, nutrition, sleep patterns, therapeutic adherence, and general health behaviors. Recommendations were generated using a dual-weighting strategy that prioritizes individual patient attributes while incorporating cluster-level contextual information. System performance was evaluated through blind expert validation involving cardiologists and clinical nutritionists.ResultsK-Means clustering identified three clinically interpretable patient profiles with a Silhouette coefficient of 0.5608. Agreement between automated recommendations and expert clinical consensus reached 78.3%, with a Cohen’s Kappa coefficient of 0.742, indicating substantial concordance. No statistically significant differences were observed between system outputs and expert judgments (χ2 = 8.347, p = 0.908).DiscussionThe findings demonstrate that combining unsupervised patient stratification with explicit clinical reasoning enables interpretable and scalable decision support for non-pharmacological hypertension management. This approach may be particularly valuable in healthcare environments with limited labeled data and constrained clinical resources.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1598741</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1598741</link>
        <title><![CDATA[Artificial intelligence anxiety, digital well-being, and future career concerns among engineering and information technology students in Jordan]]></title>
        <pubdate>2026-03-30T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Mais AL-Nasa’h</author><author>Luae Al-Tarawneh</author><author>Ola Alhwayan</author>
        <description><![CDATA[IntroductionThe rapid advancement of artificial intelligence (AI) is fundamentally transforming educational and employment landscapes, generating increasing psychological concerns among students in technology-intensive fields. This study examines AI-related anxiety, digital well-being, and career uncertainty among engineering and information technology (IT) students, with a focus on their prevalence, interrelationships, and demographic variations.MethodsA cross-sectional quantitative design was employed using a structured survey administered to 820 undergraduate students from four Jordanian universities. Standardized measures were used to assess AI anxiety, digital well-being, and career-related concerns. Statistical analyses included descriptive statistics, correlation analysis, and group comparisons based on gender and academic discipline.ResultsThe findings indicated elevated levels of AI anxiety (M = 5.26, SD = 0.32), low levels of digital well-being (M = 1.75, SD = 0.20), and moderate levels of career concerns (M = 4.07, SD = 0.34). AI anxiety was strongly negatively correlated with digital well-being (r = −0.849, p < 0.01) and positively correlated with career concerns (r = 0.680, p < 0.01). Female students reported significantly higher AI anxiety and career concerns than male students (p < 0.001). Additionally, IT students exhibited higher levels of AI anxiety and career uncertainty compared to engineering students (p < 0.001).DiscussionThese findings highlight the psychological impact of AI integration on students, emphasizing the need for targeted AI literacy programs, digital well-being interventions, and career guidance strategies. Addressing gender disparities and discipline-specific differences is essential to enhance students’ resilience, adaptability, and readiness for an AI-driven labor market.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1754000</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1754000</link>
        <title><![CDATA[A federated multimodal deep learning framework for brain tumor classification using MRI]]></title>
        <pubdate>2026-03-30T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>K. Lakshmi Vasanthi</author><author>J. Sree Darshne</author><author>Pattabiraman Venkatasubbu</author><author>Parvathi Ramasubramanian</author>
        <description><![CDATA[IntroductionBrain tumor classification using MRI plays a critical role in early diagnosis and treatment planning. However, traditional centralized approaches require sharing sensitive medical data, which raises serious privacy concerns. Additionally, the distribution of data across multiple hospitals limits effective model training and utilization. Therefore, there is a strong need for privacy-preserving and distributed learning methods that ensure both security and accuracy in classification.MethodsIn this work, a federated learning framework is proposed to enable collaborative model training without sharing raw data. To improve efficiency, a layer skipping mechanism is applied, which reduces communication cost during training. The FedPropSAG aggregation method is used to enhance convergence and overall model performance. Furthermore, Differential Privacy (DP) and Secure Aggregation (SA) techniques are incorporated to ensure data privacy and secure communication.ResultsThe proposed model achieves high classification accuracy across distributed datasets, demonstrating its effectiveness. The communication cost is significantly reduced due to the implementation of the layer skipping mechanism. The model performs well even under non-IID data distributions, which are common in real-world scenarios. Importantly, the integration of privacy-preserving techniques does not degrade the overall model performance.DiscussionThe proposed approach provides an efficient and scalable solution for distributed medical data analysis. It ensures patient data privacy while still enabling collaborative learning across multiple institutions. The reduction in communication overhead makes the framework suitable for practical deployment in healthcare systems. Overall, the model successfully balances accuracy, efficiency, and privacy, making it a strong candidate for real-world applications.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1744410</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1744410</link>
        <title><![CDATA[Advances in artificial intelligence and thermal analysis for brain tumor detection: a review of models, methods, and modalities]]></title>
        <pubdate>2026-03-30T00:00:00Z</pubdate>
        <category>Review</category>
        <author>Abedalmuhdi Almomany</author><author>Uzair Soomro</author><author>Anwar Al Assaf</author><author>Alaa Abd-Alrazaq</author><author>Rafat Damseh</author><author>Muhammed Sutcu</author><author>B. S. Ksm Kader Ibrahim</author><author>Barış Yıldız</author>
        <description><![CDATA[Brain tumors pose a major challenge in neuro-oncology due to their high mortality rates and complex diagnosis. This review summarizes recent advances in using artificial intelligence (AI), particularly deep learning, in conjunction with thermal imaging and simulated thermal mapping for brain tumor detection. AI methods such as convolutional neural networks (CNNs), hybrid architectures, and bioheat transfer models, including the Pennes equation, are evaluated to determine how temperature variations, tumor biology, and image preprocessing influence malignancy classification. Traditional imaging techniques, such as Magnetic Resonance Imaging (MRI) and Computed Tomography (CT), provide detailed structural information but are often costly, invasive, and limited in their ability to capture physiological data. Recent studies indicate that integrating AI with thermal imaging, either through direct infrared thermography or simulated thermal maps derived from MRI, enables non-invasive, physiology-aware diagnosis. The review examines current approaches to thermal data preprocessing, simulation, deep learning-based tumor segmentation, and malignancy prediction, as well as key evaluation metrics, model interpretability tools, and recent performance outcomes. Despite ongoing progress, challenges remain, including limited availability of multimodal datasets, variability in thermal signatures, and the need for clinical validation. Future research directions include large-scale data collection, advanced thermal modeling, multimodal fusion frameworks, and the development of explainable AI tools that meet clinical standards. In resource-limited settings, AI-powered thermal imaging may serve as a valuable supplement to traditional diagnostics, offering safer, more precise, and more accessible brain tumor detection. This technology has the potential to improve patient outcomes and transform neuro-oncology practices by integrating anatomical and functional insights. This review critically evaluates current evidence and identifies the challenges that must be addressed to facilitate the translation of promising research into clinical practice.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1694192</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1694192</link>
        <title><![CDATA[Metacognition of ChatGPT in confidence judgements]]></title>
        <pubdate>2026-03-27T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Shun Yoshizawa</author><author>Ayako Onzo</author><author>Shinichi Nozawa</author><author>Tsugumi Takano</author><author>Tetsuo Ishikawa</author><author>Ken Mogi</author>
        <description><![CDATA[Recent advances in Large Language Models (LLMs) have raised critical concerns regarding AI alignment and safety, particularly with respect to the reliability of their outputs. In humans, metacognition plays a key role in making cognition robust and adaptive. LLMs frequently express high confidence in their responses, raising the question of whether such confidence reflects human-like metacognitive capability. In this study, we systematically compared humans and GPT-4 across multiple task formats to examine how confidence relates to performance. GPT-4 consistently outperformed humans in task accuracy. This advantage was not accompanied by human-like confidence behavior: Human confidence closely tracked variations in accuracy, while GPT-4 was not. Humans adjusted their confidence more sensitively to changes in accuracy, whereas GPT-4 showed a shallow confidence–accuracy mapping. Humans exhibited higher and more stable metacognitive sensitivity and efficiency, while GPT-4 showed condition-specific variability. These findings reveal a dissociation between task-level performance and metacognitive behavior in GPT-4, suggesting that its confidence reflects structural properties of its outputs rather than genuine internal uncertainty monitoring. Taken together, these findings suggest that GPT-4 lacks robust metacognitive abilities compared to humans, or at least that its metacognitive processes differ significantly from those of humans.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1770564</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1770564</link>
        <title><![CDATA[Evaluation of large language models in generating and optimizing educational materials for neonatal home oxygen therapy]]></title>
        <pubdate>2026-03-27T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Zhendong Liu</author><author>Xiaoping Yang</author><author>Yu Zhang</author><author>Yujing Xu</author><author>Yue Xiang</author><author>Hongyan Wang</author>
        <description><![CDATA[BackgroundNeonatal Home Oxygen Therapy (NHOT) is a critical treatment for premature infants with Bronchopulmonary Dysplasia (BPD). However, existing health education materials are generally difficult to read, particularly for grandparent caregivers with lower educational backgrounds. This study aimed to systematically evaluate the capacity of six major Large Language Models (LLMs) to generate and optimize NHOT health education materials.MethodsSix LLMs were included: ChatGPT-5.1, Claude 4.5 Sonnet, Gemini 2.5 Pro, Grok-4.1, Qwen-3-Max, and DeepSeek-V3.2. Each model generated 20 texts under three prompting strategies—baseline (Prompt A), simplification (Prompt B), and rewriting (Prompt C)—yielding 360 texts in total. Twenty WeChat public health articles served as the human-authored baseline. Subjective evaluation employed C-DISCERN, C-PEMAT (understandability and actionability), and a medical accuracy Likert scale, supplemented by objective linguistic analysis using the Alpha Readability Chinese (ARC) tool.ResultsAll models demonstrated superior medical accuracy compared to the human baseline (Likert median 1.0, against 2.0 for the original articles). Under baseline conditions, Qwen achieved the highest content quality (C-DISCERN median 57.0), while Claude attained perfect actionability scores. The simplification prompt (Prompt B) significantly reduced C-DISCERN scores across all models (all p < 0.001) without meaningfully improving understandability or actionability. In the rewriting task (Prompt C), all models significantly enhanced the understandability of original texts (p < 0.01), with Grok and Qwen additionally improving content quality and actionability. Linguistic analysis revealed that prompt optimization improved semantic accuracy and reduced semantic noise, but at the cost of decreased lexical richness.ConclusionLLMs demonstrate significant potential for optimizing existing health education materials, performing more reliably in rewriting mode than in de novo generation. Simplistic “plain language” instructions risk compromising content quality, highlighting the need for carefully designed prompts that balance accuracy, clarity, and completeness. All AI-generated materials require rigorous review by qualified clinical professionals prior to distribution.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1799522</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1799522</link>
        <title><![CDATA[AI algorithms and IoT platforms for anomaly and failure prediction in industrial machinery—systematic review]]></title>
        <pubdate>2026-03-26T00:00:00Z</pubdate>
        <category>Systematic Review</category>
        <author>Mario Esteban Marín Vásquez</author><author>Juan Carlos Blandón Andrade</author><author>Alonso Toro Lazo</author><author>Jesús Alfonso López Sotelo</author>
        <description><![CDATA[Predictive Maintenance (PdM) focuses on anticipating potential failures in industrial machines by the monitoring key parameters. Artificial Intelligence (AI) provides algorithms that can be used for this purpose. Specialized literature mentions that some companies need to adopt more proactive and predictive strategies in managing of industrial maintenance. This study aims to conduct a Systematic Literature Review (SLR) on Artificial Intelligence algorithms and software and IoT platforms used for anomaly and failure prediction. The method includes of six main phases: (i) defining the research questions; (ii) conducting a search process; (iii) establishing exclusion and inclusion criteria; (iv) performing a quality assessment of studies; (v) collecting data; and (vi) analyzing the data. The findings show that the main AI techniques for PdM are classified as: (i) machine Learning-based methods; (ii) neural networks-based methods; and (iii) knowledge transfer-based methods. Nine software and IoT technologies were identified to support maintenance operations. Additionally, it is discussed how Machine Learning and Deep Learning algorithms perform well in fault classification, prediction, Remaining Useful Life (RUL) estimation, and diagnostic tasks. They can also be applied in earlier stages, such as data preprocessing and feature extraction. Finally, it is shown that knowledge transfer can improve AI algorithms when sudden changes occur in data and their relationships. In conclusion, the AI technologies identified can significantly contribute to predicting failures in industrial machinery.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1761336</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1761336</link>
        <title><![CDATA[GAN-based bone suppression using a combined loss function]]></title>
        <pubdate>2026-03-26T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Lukáš Jochymek</author><author>Markéta Vašinková</author><author>Vít Doleží</author><author>Petr Gajdoš</author>
        <description><![CDATA[IntroductionAccurate analysis of chest radiographs (X-rays) is essential for diagnosing diseases such as pneumonia and lung cancer, yet bone structures often obscure critical soft tissues and lesions. From an artificial intelligence perspective, bone suppression can be formulated using different modeling paradigms that reflect distinct assumptions about the task.MethodsIn this study, the problem is addressed as a comparative methodological investigation, and three conceptually different approaches are systematically evaluated within a unified experimental framework: denoising-based regression using autoencoders, structured image-to-image transformation using U-Net architectures, and distribution-based generative modeling using adversarial learning. In addition, the impact of different loss configurations and training regimes on reconstruction quality is examined. An enhanced generative adversarial network (GAN) with improved generator and discriminator components and a combined loss function (Wasserstein, L1, perceptual, and Sobel losses) is proposed to improve structural consistency and preserve soft-tissue appearance.ResultsModel performance was assessed using the peak signal-to-noise ratio (PSNR) and the multi-scale structural similarity index measure (MS-SSIM). Among the evaluated approaches, the GAN achieved the best performance, reaching a PSNR of 44.09 dB and an MS-SSIM of 0.9968, and outperformed recently published methods evaluated on the same dataset.DiscussionThese results highlight the importance of both modeling paradigm selection and loss formulation for achieving structurally consistent bone suppression in chest radiographs.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1782405</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1782405</link>
        <title><![CDATA[Technical evaluation of language models adapted for the automation of legal contracts: clause extraction, classification, and summarization]]></title>
        <pubdate>2026-03-26T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Jaime Govea</author><author>Iván Ortiz-Gárces</author><author>Pablo Palacios</author><author>Alexandra Maldonado Navarro</author><author>Santiago Acurio Del Pino</author><author>William Villegas-Ch</author>
        <description><![CDATA[The growing demand for automation in legal contract management exposes a persistent limitation of current language models: insufficient adaptation to the semantic, structural, and regulatory constraints of legal language. While large language models perform well on general NLP tasks, their direct application to legal document classification, clause extraction, and contract summarization often yields unstable, legally unreliable outputs. This work presents a structured methodological pipeline for evaluating and adapting language models for legal contract automation, combining domain-specific fine-tuning of open-source models with a controlled comparative assessment against large general-purpose LLMs used exclusively in inference mode. The methodology integrates legal corpus curation, clause-level annotation, and efficient adaptation techniques, and is evaluated across three core tasks: contract document classification, normative clause extraction, and regulatory summarization. The evaluation protocol is explicitly designed to disentangle the effects of supervision from deployment constraints arising in regulated legal settings. Experimental results show consistent and statistically significant performance gains for legally adapted models over general-purpose baselines, achieving Macro-F1 of 0.921 in classification, span-level F1 of 0.903 in clause extraction, and ROUGE-L of 0.886 in summarization (p < 0.01). Robustness analysis and cross-validation confirm stability across heterogeneous private-sector contract types. The findings should be interpreted under the evaluated comparison regime and highlight that, in legally constrained multi-stage workflows, task-aligned supervision provides measurable structural benefits that are not reducible to model scale alone when general-purpose LLMs are restricted to inference-only deployment.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1741144</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1741144</link>
        <title><![CDATA[Portable electrochemical impedance biosensing with DRT-enabled machine learning for detecting E. coli O157:H7 in poultry meat]]></title>
        <pubdate>2026-03-26T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Yang Tian</author><author>Ziyu Liu</author><author>Chaitanya Pallerla</author><author>Siavash Mahmoudi</author><author>Ramesh Bahadur Bist</author><author>Yiting Xiao</author><author>Terry Howell</author><author>Jeyamkondan Subbiah</author><author>Dongyi Wang</author>
        <description><![CDATA[Ensuring food safety requires rapid and accurate detection of pathogens such as Escherichia coli O157:H7. Here, we report a portable electrochemical immunosensor coupled with machine learning (ML) that enables quantitative prediction even when the impedance response is not strictly linear with concentration. The sensor employs protein A-mediated oriented antibody immobilization on a gold electrode and measures target binding using electrochemical impedance spectroscopy (EIS) and cyclic voltammetry. To move beyond single-parameter equivalent-circuit fitting, we apply distribution of relaxation times (DRT) deconvolution to resolve the impedance spectrum into mechanistic contributions associated with charge-transfer kinetics, double-layer charging, and transport-limited (diffusion/Warburg-type) processes, and use these DRT-derived features for concentration inference. Multiple ML models (partial least squares (PLS), Random Forest, histogram-based gradient boosting, support vector regression, ridge regression, and Gaussian process regression) were evaluated using leave-one-concentration-out cross-validation and independent hold-out testing, demonstrating accurate prediction on unseen concentration levels. Validation in poultry meat samples artificially inoculated with E. coli O157:H7 confirmed applicability in a food-relevant matrix, along with high selectivity against non-target bacteria and stable performance during storage. This work is novel in combining DRT-based mechanistic feature extraction with ML-based inference to deliver a field-deployable immunosensing platform for robust pathogen quantification in complex food samples.]]></description>
      </item>
      </channel>
    </rss>