<?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 | Machine Learning section | New and Recent Articles</title>
        <link>https://www.frontiersin.org/journals/artificial-intelligence/sections/machine-learning</link>
        <description>RSS Feed for Machine Learning section in the Frontiers in Artificial Intelligence journal | New and Recent Articles</description>
        <language>en-us</language>
        <generator>Frontiers Feed Generator,version:1</generator>
        <pubDate>2026-11-12T18:12:59.778+00:00</pubDate>
        <ttl>60</ttl>
        <item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1939804</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1939804</link>
        <title><![CDATA[Enhancing DeepIRT through counterfactual item-invariance regularization and joint learning]]></title>
        <pubdate>2026-11-12T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Christian Onyeke</author><author>Lijun Qian</author><author>Pamela Obiomon</author><author>Xishuang Dong</author>
        <description><![CDATA[Knowledge Tracing (KT) models are widely used in adaptive learning systems to estimate students’ evolving knowledge states. Despite recent advances in deep KT models, many approaches exhibit poor probabilistic calibration, producing overconfident predictions that misrepresent students’ knowledge. This paper presents Counterfactual Item-Invariance Regularized Knowledge Tracing (CIR-KT), a bias-aware approach for calibrated knowledge tracing. For each interaction the method constructs an item-level counterfactual, substituting the presented item while holding the observed response fixed, and enforces consistency between the factual and counterfactual latent states through a regularization objective. By reducing the model’s reliance on item-identity shortcuts, the framework improves probability calibration without altering the underlying KT architecture. On a severely miscalibrated private College of Engineering (COE) subset, counterfactual regularization substantially reduces Expected Calibration Error (ECE) and Brier score and improves ranking (AUC and PR-AUC) relative to the DeepIRT backbone, while on already well-calibrated public benchmarks its effect on discrimination and calibration is small. Skill-level analyses show that the calibration gains are broad-based across knowledge components rather than driven by a few items. These results indicate that item-level counterfactual regularization is a practical mechanism for improving calibration in knowledge tracing, supporting more dependable adaptive learning systems. At the validation-selected operating point, CIR-KT is the best-calibrated model among the knowledge-tracing backbones evaluated while remaining competitive in AUC, and it provides interpretable ability and difficulty estimates that attention baselines do not.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1953730</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1953730</link>
        <title><![CDATA[Sample-aware test-time adaptation for medical image-to-image translation]]></title>
        <pubdate>2026-11-12T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Irene Iele</author><author>Francesco Di Feola</author><author>Matteo Tortora</author><author>Rosa Sicilia</author><author>Valerio Guarrasi</author><author>Paolo Soda</author>
        <description><![CDATA[Image-to-image translation has emerged as a powerful technique in medical imaging applications, including image denoising and cross-modality synthesis. However, pretrained translation models may exhibit performance degradation when test samples differ from the source-domain data observed during training. We propose a sample-aware Test-Time Adaptation framework that dynamically determines whether and how each test sample should be adapted. The framework combines a Reconstruction Module, which measures compatibility with source-domain representations, with a Dynamic Adaptation Block that selectively modifies the internal feature levels of a frozen translation model. The output reconstruction error is used to trigger adaptation and to guide the sample-specific selection and optimization of lightweight feature-level adaptors. We evaluate the proposed approach on low-dose CT denoising and T1-to-T2 MRI translation under source-domain, anatomical-shift, and external-domain conditions. The proposed method improves translation quality under anatomical and external-domain shifts, while remaining comparable to the frozen task model in the predominantly in-distribution setting, where adaptation provides no measurable benefit. Our results show that combining selective activation with sample-specific adaptor selection provides a more reliable strategy than applying a fixed adaptation procedure uniformly across test samples. The code is available at: https://github.com/arco-group/Sample-Aware-TTA.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1887124</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1887124</link>
        <title><![CDATA[Hamiltonian-encoded quantum feature maps augment classical CNNs for pulmonary tuberculosis detection]]></title>
        <pubdate>2026-11-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Sonet Daniel Thomas</author><author>Aparna Rajan</author><author>Lijin Varghese</author><author>Naveen Joy</author><author>Aswathi Balakrishnan</author><author>Suhail Subair</author><author>Ivin Shajan</author><author>Muhammed Shahul</author><author>S. S. Vinodchandra</author><author>J. Abhithaj</author><author>K. Devdas Acharya</author><author>Manavalan Vijayakumar</author><author>Jalaluddin Akbar Kandel Codi</author><author>Rajesh Raju</author>
        <description><![CDATA[IntroductionEarly detection of pulmonary tuberculosis (TB) from chest radiographs remains challenging due to subtle and low-contrast pulmonary lesions. This study investigates whether quantum-enhanced feature representations can improve TB detection compared with classical deep learning models.MethodsWe developed a hybrid quantum-classical deep learning framework integrating 4-qubit parameterized quantum circuits with six classical CNN backbones: VGG16, ResNet50, DenseNet121, InceptionV3, EfficientNetB0, and EfficientNetV2-S. A controlled ablation study using classical Sobel-texture feature extraction was conducted. All models were evaluated using 5-fold stratified cross-validation on 662 chest X-ray images (326 normal and 336 TB-positive). Pairwise McNemar tests were used for statistical comparisons.ResultsQuantumHybrid_EfficientNetB0 achieved the highest AUC of 0.9686 ± 0.0074 (95% CI: [0.934, 0.968]), with a sensitivity of 0.9315 and specificity of 0.8624. Quantum-hybrid models consistently outperformed their classical and ablated counterparts across the evaluated backbone families, with statistically significant improvements in most comparisons (p < 0.05).DiscussionThese findings demonstrate the potential of Hamiltonian-encoded quantum feature maps to enhance classical CNN-based pulmonary TB detection and provide a benchmark for further investigation of quantum-augmented diagnostic systems.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1954939</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1954939</link>
        <title><![CDATA[Fancy GNG: localized color augmentation via Growing Neural Gas]]></title>
        <pubdate>2026-11-11T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Tigran Andrikyan</author><author>Hans-Oliver Hansen</author><author>Paul Schönbrunner</author><author>Amir Madany Mamlouk</author>
        <description><![CDATA[Deep artificial neural networks often benefit from large and diverse training datasets. As collecting sufficiently varied data can be costly, data augmentation has become a standard component of modern computer vision pipelines. While geometric augmentations are extensively studied, most existing color augmentation methods apply global photometric transformations that primarily simulate illumination changes or camera-related effects. However, in many real-world recognition tasks, chromatic variability can be heterogeneous within an image and may differ across object or material regions rather than follow a single globally uniform photometric shift. In this work, we introduce Fancy Growing Neural Gas (Fancy GNG), a novel cluster-specific color augmentation method that combines Growing Neural Gas clustering with PCA-based color perturbations. Instead of applying a single global transformation to the entire image, Fancy GNG partitions the image-specific RGB distribution into coherent color-space clusters and applies independently sampled, cluster-specific perturbations. This enables structured and image-adaptive chromatic variation without relying on semantic labels or explicit spatial segmentation, while preserving the overall image content. We evaluated the proposed method on the Flowers-102, CIFAR-10, and ImageNet-1k datasets using a ResNet-50 architecture. The results reveal a strong dataset-dependent behavior. Fancy GNG achieves its largest improvements on Flowers-102, a fine-grained and highly color-sensitive classification benchmark, while also providing consistent gains on ImageNet-1k. In contrast, only marginal improvements are observed on CIFAR-10, where chromatic information appears less critical for class discrimination. Furthermore, we show that the performance degradation caused by grayscale conversion serves as a practical indicator of how strongly a dataset is likely to benefit from color-aware augmentation. Overall, the results demonstrate that cluster-specific, color-distribution-aware augmentation can provide meaningful advantages over conventional global color perturbation methods, particularly for recognition tasks in which chromatic information is both discriminative and internally structured.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1929384</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1929384</link>
        <title><![CDATA[Single-level wavelet preprocessing is downsampling: measurement failures in cross-dataset polyp classification]]></title>
        <pubdate>2026-11-09T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Jenifer R.</author><author>Dhimant Tewari</author><author>Goutham Veerapu</author><author>Nisha J. S.</author><author>Gopinath Palanisamy</author><author>Saanji Kohli</author>
        <description><![CDATA[Models for colorectal polyp classification lose accuracy when applied to images from a different clinical center, and training on the low-frequency approximation subband of a discrete wavelet transform has been proposed as a remedy. We set out to test that hypothesis on Kvasir-SEG and PICCOLO, and report instead that three measurement practices common in this literature do not support the conclusions drawn from them. First, a single-level Haar LL subband is arithmetically identical to grayscale conversion followed by 2× decimation, verified to floating-point precision on both random and endoscopic data. At one decomposition level, wavelet decomposition and resolution reduction are not separable hypotheses, and the proposed mechanism is therefore not well-formed as posed. Second, image-level splitting—the standard remedy for patch-level leakage—is insufficient for video-derived datasets. PICCOLO's 3,433 frames derive from 70 physical lesions in 40 examinations; an image-level split leaves 58 of those lesions and 37 of those videos shared between training and test, and in-distribution AUC falls by 0.054 when measured against the partition the dataset authors supply. Third, the intersection-over-union metric used to assess attention localization inverts on high-coverage patches, where a polyp patch is almost entirely polyp: random attention scores 0.59 against 0.24 for trained heads. Measured without the area confound, localization is real but weak and confined to particular heads. We additionally quantify a reproducibility floor. In small models trained from scratch, runs receiving bit-identical inputs diverge by up to 0.177 AUC in the CNN/PICCOLO row through early-stopping checkpoint selection, and in one experimental arm this null control is itself statistically significant, so no claim from that arm is supportable at any p-value. A color-removal benefit observed in this regime (+0.067 AUC, 95% CI [+0.037, +0.097]) does not replicate on ImageNet-pretrained ResNet-18 and DeiT-Tiny backbones, where intervals exclude effects a fifth its size. We recommend reporting a null control alongside every comparison, and show that this identifies arms in which no claim can be supported.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1967344</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1967344</link>
        <title><![CDATA[A hybrid XGBoost-RF residual learning model for energy consumption forecasting]]></title>
        <pubdate>2026-11-05T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Monire Norouzi</author><author>Can Eyupoglu</author><author>Oktay Karakuş</author>
        <description><![CDATA[Recently, energy consumption forecasting plays an important role in power systems, smart grids, and power generation networks. Given the complex nature and time dependence of electricity consumption data, applying machine learning techniques improves energy forecasting accuracy and performance. In this research, a hybrid XGBoost–Random Forest (XGBoost–RF) model is proposed to make better the forecasting accuracy. To improve the performance of the proposed model, a residual learning-based forecasting framework is presented to learn the remaining prediction errors. In the forecasting procedure, important prediction metrics such as Mean Absolute Error (MAE), and coefficient of determination R2 are evaluated. The experimental results show that the proposed hybrid model achieves an MAE of 15.8683, Root Mean Square Error (RMSE) of 22.1588, R2 of 0.985807, and Mean Absolute Percentage Error (MAPE) of 1.3118%. Compared with the base XGBoost model, the proposed model is able to decrease MAE by 11.19% and RMSE by 11.30%. In addition, the SHapley Additive Explanations (SHAP) method is used to examine the effect of the input features on the model predictions and improve model interpretability. The SHAP results show that temporal features and lag-based variables have an important role in energy consumption forecasting. In particular, lag_1 and lag_168 are identified as the most important features. The experimental results confirmed that the proposed hybrid model achieved lower prediction errors and a higher R2 than the other evaluated models. These findings show that the proposed XGBoost-RF residual learning approach can improve energy consumption forecasting accuracy while providing useful information about the important forecasting features.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1892198</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1892198</link>
        <title><![CDATA[Composite Schatten-p norm factorization for efficient matrix completion]]></title>
        <pubdate>2026-11-04T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Jiaoyan Zhao</author><author>Shuangyan Yi</author>
        <description><![CDATA[Schatten-p norm optimization has emerged as an effective non-convex relaxation for low-rank matrix completion. However, the primary computational bottleneck arises from repeated singular value decomposition (SVD) operations required during optimization, which leads to high computational cost. In this work, we propose a novel factorized reformulation of the Schatten-p norm by decomposing it into combinations of nuclear and Frobenius norms. Specifically, we investigate three distinct decomposition schemes: the sum of two nuclear norms, the sum of two Frobenius norms, and the combination of a nuclear norm and a Frobenius norm. Based on these formulations, we develop an alternating gradient descent algorithm with rigorous convergence guarantees. By avoiding SVD, the proposed method significantly reduces computational complexity. Extensive experiments on synthetic datasets, image inpainting tasks, and recommender systems demonstrate that our approach consistently outperforms several is the state-of-the-art matrix completion methods in terms of both accuracy and computational efficiency.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1975513</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1975513</link>
        <title><![CDATA[Physics-guided multi-view hybrid depth-wise separable convolution and temporal LSTM network for fault diagnosis of rotor-bearing systems]]></title>
        <pubdate>2026-11-04T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Sandhya Mitta</author><author>Poornima Nedunchezhian</author>
        <description><![CDATA[IntroductionRolling-element bearings are critical components of rotating machinery, and their failures can affect equipment safety, availability, and service life. However, non-stationary vibration signals and background noise make reliable automatic fault diagnosis challenging. This study proposes a physics-guided Hybrid Depth-wise Separable Spatial Convolution and Temporal Long Short-Term Memory network (HDSSC-TLST) for rotor-bearing fault diagnosis.MethodsThe proposed approach derives the minimum sample length from the characteristic defect frequencies of the inner race, outer race, and rolling elements. A leakage-free chronological multi-view sliding-window scheme generates large windows to capture impact periodicity and centered small windows to resolve high-frequency transients. Both views are converted into continuous wavelet transform (CWT) scalograms using a complex Morlet wavelet. Depth-wise separable convolution blocks extract spatial time-frequency features, while a stacked LSTM models their temporal evolution to classify four conditions: normal, inner-race fault, outer-race fault, and ball fault.ResultsExperiments using the Case Western Reserve University (CWRU) 48 kHz drive-end dataset show that HDSSC-TLST achieves 94% accuracy at a signal-to-noise ratio (SNR) of 19 dB and outperforms the hybrid adaptive fusion network (HAFDL) and nonlinear auto-regressive model with exogenous inputs (NARX) at every tested SNR from 1 to 19 dB. Averaged over ten data sizes, the model achieves precision, recall, and F1-score values of 0.90, compared with 0.79, 0.78, and 0.79 for HAFDL and 0.75, 0.74, and 0.74 for NARX, respectively. It also reduces training time by 14.5% relative to HAFDL and 18.8% relative to NARX. On a leakage-free chronological split of noise-free 1 hp data, it achieves 96.0 ± 0.8% accuracy and a macro-F1 score of 0.960 ± 0.009. Its mean accuracy on unseen 0, 2, and 3 hp loads is 88.1% without retraining.DiscussionThe results indicate that combining physics-guided sample selection, multi-view time-frequency representation, depth-wise separable convolution, and temporal modeling can provide an effective and computationally efficient approach to bearing fault diagnosis under noisy conditions and changing operating loads.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1969037</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1969037</link>
        <title><![CDATA[Geography, geospatial analysis and artificial intelligence: between traditional methods and new spatial paradigms]]></title>
        <pubdate>2026-11-03T00:00:00Z</pubdate>
        <category>Opinion</category>
        <author>Cláudia M. Viana</author><author>Jorge Rocha</author>
        <description></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1933918</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1933918</link>
        <title><![CDATA[SACX: Suppression–Absence–Competition eXplanations for the why-not question in visual recognition]]></title>
        <pubdate>2026-11-03T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Adithya JK Annamalai</author><author>L. RamaParvathy</author>
        <description><![CDATA[IntroductionExisting attribution-based explainable artificial intelligence (XAI) methods for convolutional neural network (CNN) classifiers primarily address why a particular class was predicted, while providing limited insight into why a competing class was not selected. To address the question “Why class A and not class B?”, we introduce Suppression-Absence-Competition eXplanations (SACX), a framework for structured contrastive visual explanations.MethodsSACX characterizes the contrastive margin za−zb using three spatially aligned heatmaps: Hcomp, representing positive evidence for the predicted class; Hsup, representing negative attribution for the competing foil class; and Hmiss, identifying regions where expected foil-class features are absent through prototype-based similarity. We introduce three direction-aware faithfulness metrics: suppression-removal AUC (SR-AUC), feature-injection AUC (FI-AUC), and localization fidelity under injection (LFI@k). We additionally evaluate suppression using a fixed-budget placement test (SR@k). Experiments are conducted on CUB-200-2011 and Pattern Analysis, Statistical Modelling and Computational Learning (PASCAL) Visual Object Classes (VOC) 2007 using fine-tuned ResNet-50 models and a frozen DINOv2 ViT-B/14 backbone with a linear probe.ResultsOn fine-tuned ResNet-50, injecting features at top-ranked Hmiss locations outperforms random placement, improving the foil-class logit by Δzb = +0.41 (95% CI: 0.39–0.42) on CUB and +1.29 (95% CI: 1.23–1.36) on VOC. On the frozen DINOv2 backbone, all injection variants reduce the foil-class logit, but top-ranked Hmiss locations remain the least disruptive, achieving an LFI@20% top-minus-random gap of +0.68. The LFI gaps remain stable across three K-means and three probe seeds (standard deviation ≤ 0.04). The SR@k evaluation shows that Hsup outperforms random removal at tight budgets (+0.09 at k = 5% on CUB Top-2), although suppression effects are generally weak for false-negative cases.DiscussionSACX provides an interpretable structural decomposition of the existing contrastive attribution signal rather than introducing a new aggregate map, as Hcontra is mathematically equivalent to Contrastive Grad-CAM. Its prototype-based missing-evidence component preserves spatial localization rankings even under frozen self-supervised representations, although absolute injection gains require task-specific supervision. The suppression component is interpreted as negative attribution rather than a guarantee of causal suppression. These findings demonstrate the value of separating competition, suppression, and missing evidence when explaining why a competing class was not predicted.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1966982</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1966982</link>
        <title><![CDATA[Machine learning prediction model for respiratory failure in patients with asthma]]></title>
        <pubdate>2026-11-03T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Xiaojun Hao</author><author>Xiaohui Du</author><author>Si Tang</author><author>Hongxia Jiang</author>
        <description><![CDATA[BackgroundAsthma is one of the most common chronic respiratory diseases worldwide, and its global burden continues to increase. Although most patients can achieve good disease control with standardized treatment, some still experience severe acute exacerbations triggered by infection, environmental exposure, poor treatment adherence, or comorbid COPD, which may further progress to respiratory failure. Respiratory failure is one of the most serious adverse outcomes during hospitalization for asthma and is often associated with a greater need for mechanical ventilation, longer intensive care stays, and poorer clinical prognosis. Therefore, early identification of high-risk patients and the development of reliable prediction tools are of substantial clinical importance.ObjectiveTo develop and validate a machine learning-based risk prediction model using routinely available clinical data from hospitalized patients with asthma to identify those at high risk of developing respiratory failure during hospitalization, screen key predictive factors, and evaluate the model’s discrimination, calibration, clinical net benefit, and interpretability.MethodsThis was a single-center retrospective cohort study. A total of 2,098 patients hospitalized for asthma at Zhongnan Hospital of Wuhan University between November 2016 and December 2024 were retrospectively screened. After applying predefined inclusion and exclusion criteria, 1,405 patients were included in the final analysis, among whom 233 developed respiratory failure during hospitalization (respiratory failure group) and 1,172 did not (control group). Patients were randomly assigned by stratified sampling in an 8:2 ratio to a training set (n = 1,124) and an internal validation set (n = 281). In addition, 403 patients hospitalized for asthma at the same institution between January 2025 and June 2026 were screened using the same criteria, and 371 patients were ultimately included as an temporal validation cohort from an independent time period (72 with respiratory failure and 299 controls). A total of 62 candidate variables were collected, including demographic characteristics, smoking history, COPD history, and laboratory indicators such as complete blood count, coagulation function, liver and renal function, electrolytes, and blood lipids. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression combined with the Boruta algorithm. Eleven machine learning models were then developed based on the selected variables, including logistic regression, random forest, naive Bayes, multilayer perceptron, support vector machine, k-nearest neighbors, decision tree, gradient boosting decision tree (GBDT), LightGBM, CatBoost, and XGBoost. Model performance was comprehensively evaluated using receiver operating characteristic (ROC) curves, decision curve analysis (DCA), precision-recall (PR) curves, and calibration curves. Shapley additive explanations (SHAP) and a nomogram were further applied to interpret and visualize the optimal model.ResultsLASSO regression combined with the Boruta algorithm ultimately identified five core predictive variables from the 62 candidate features: D-dimer, bicarbonate concentration (HCO3−), neutrophil percentage, direct bilirubin, and lymphocyte percentage. Among the 11 candidate models, CatBoost showed the best overall performance. In the training set, the model achieved an AUC of 0.945 (95% CI: 0.928–0.962), accuracy of 0.923 (95% CI: 0.889–0.962), precision of 0.943 (95% CI: 0.896–0.989), specificity of 0.994 (95% CI: 0.989–0.996), recall of 0.556 (95% CI: 0.403–0.712), F1 score of 0.699 (95% CI: 0.678–0.714), and Matthews correlation coefficient (MCC) of 0.689. In the internal validation set, the model achieved an AUC of 0.796 (95% CI: 0.723–0.868), accuracy of 0.858 (95% CI: 0.812–0.899), precision of 0.842 (95% CI: 0.758–0.911), specificity of 0.987 (95% CI: 0.973–0.994), recall of 0.302 (95% CI: 0.217–0.381), and F1 score of 0.444 (95% CI: 0.369–0.513). In the temporal validation set, the model achieved an AUC of 0.878 (95% CI: 0.829–0.926), accuracy of 0.879 (95% CI: 0.832–0.924), precision of 0.909 (95% CI: 0.857–0.962), specificity of 0.989 (95% CI: 0.982–0.995), recall of 0.417 (95% CI: 0.333–0.485), and F1 score of 0.571 (95% CI: 0.549–0.602).ConclusionThe CatBoost machine learning model developed using routinely available clinical data can accurately predict the risk of respiratory failure in hospitalized patients with asthma. The model demonstrated good discrimination, acceptable temporal generalizability, and strong interpretability, and may provide useful support for early identification of severe progression and individualized management in patients with asthma.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1966841</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1966841</link>
        <title><![CDATA[StoneXAI-Net: an explainable weakly supervised framework for kidney stone classification, localization, and segmentation from CT images]]></title>
        <pubdate>2026-11-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>P. Karthick</author><author>Chiranji Lal Chowdhary</author>
        <description><![CDATA[IntroductionKidney stone disease (urolithiasis) is one of the most frequent urological diseases worldwide; it is highly recurrent and carries a significant clinical burden and economic cost. Non-contrast computed tomography (CT) is the primary diagnostic modality and offers the highest sensitivity and specificity. However, automated kidney stone analysis remains challenging because expert pixel-level annotations are scarce, hyperdense non-stone anatomy causes mislocalization in weakly supervised localization, and explainable artificial intelligence (XAI) methods are often used circularly, both for weakly supervised localization and for its subsequent interpretation.MethodsWe present StoneXAI-Net, an explainable, weakly supervised framework to classify, localize, and segment kidney stones from CT images. The framework consists of three modules: (i) a CT-physics augmentation strategy that diversifies training and mitigates overfitting; (ii) consensus pseudo-labeling that combines complementary explanation maps with uncertainty maps and anatomical priors to suppress false localizations; and (iii) a faithfulness audit module with deletion, insertion, and pointing-game metrics that objectively evaluates explanation quality. The resulting consensus pseudo-labels were used to train an attention-prior-conditioned Attention U-Net for segmentation. The framework was evaluated on 3,364 axial abdominal CT images (1,577 stone and 1,787 normal) using five-fold cross-validation.ResultsStoneXAI-Net achieved 97.4% accuracy, 98.4% precision, 96.0% recall, a 97.1% F1-score, and a 99.5% AUC. Against consensus pseudo-masks, the full segmentation configuration achieved a proxy Dice coefficient of 0.834, an IoU of 0.777, and an HD95 of 2.94 px; anatomical post-processing reduced HD95 by 6.26 px.DiscussionStoneXAI-Net is a feasible and interpretable framework for the automated analysis of kidney stones that substantially reduces the time-consuming and costly manual annotation of kidney stones by experts. Because segmentation was evaluated against pseudo-masks at the slice level, these results are proxy estimates that require confirmation with expert annotations and patient-level validation.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1860893</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1860893</link>
        <title><![CDATA[Revisiting dynamic batch scheduling for grokking: the role of batch size, sample budget, and regularization]]></title>
        <pubdate>2026-11-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Roland Fang</author>
        <description><![CDATA[Grokking—the sudden generalization of neural networks long after overfitting—has inspired training strategies aimed at accelerating its onset. This paper revisits dynamic batch size scheduling and asks whether larger batches provide intrinsically better optimization dynamics, or whether they appear better because they process more examples per update. We evaluate three dynamic schedules (linear, exponential, and step-wise increases from 32 to 512) against fixed batch sizes of 32 and 512 across four algorithmic tasks, each with 10 random seeds. When measured by update steps, the fixed-512 baseline is highly competitive: on modular division it reaches 99.6% final accuracy with 10/10 runs grokking at a mean of 3,245 steps. However, this advantage changes substantially after normalizing by total examples processed. Under a matched budget of 240,000 training examples (equivalently 960,000 input tokens), fixed-512 performs poorly and never groks on any task, while smaller-batch and dynamic strategies remain competitive on the modular tasks. A focused hyperparameter sensitivity study additionally shows that grokking depends strongly on weight decay, model size, and maximum batch size: without weight decay, none of the tested strategies grok, while intermediate or strong weight decay can make both fixed-large and dynamic strategies reliable. These results show that larger batches improve update-step efficiency but are not necessarily more sample-efficient, and that their effect is conditional on regularization and capacity. We therefore revise the original interpretation: batch size magnitude is a major confound in grokking experiments, and claims about acceleration should report both update-step and sample-normalized metrics.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1947204</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1947204</link>
        <title><![CDATA[Coarse-to-fine-grained insulator defect detection using two-stage deep learning framework]]></title>
        <pubdate>2026-11-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Sultan Daud Khan</author><author>Atif Naseer</author><author>Habib Ullah</author><author>Mohib Ullah</author>
        <description><![CDATA[IntroductionIn power transmission and power distribution systems, insulators are the fundamental components which provide mechanical support and electrical isolation between conductors and supporting towers. However, these insulators are continuously exposed to extreme weather conditions, mechanical stress, pollution and dust, which can lead to surface damage and abnormal operation.MethodConventionally, manual on-site inspection is performed to identify faulty insulators. However, this method is labor intensive, risky, and reactive. With the advancement in unmanned aerial vehicle (UAV) technology and deep learning method, automatic visual inspection has replaced the conventional methods. However, automatic visual inspection from UAV images is challenging task due to complex background and small size of the insulators. To address these challenges, we proposed a framework that consists of two stages. In the first stage, you only look once (YOLO)-based is used for extracting candidate region of insulator to get coarse class prediction. During the second stage, the candidate regions are then cropped from the image into image patches and fed to a lightweight patch-level classifier. In order to enhance inter-class separability, hard-pair contrastive classification loss is proposed by adding hard-pair contrastive term to the cross-entropy loss in the embedding space.DiscussionThe proposed technique is evaluated using an open-source insulator defect image dataset. The experiment results demonstrate the superior performance of the proposed framework compared with the YOLO variants and representative two-stage detectors.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1949474</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1949474</link>
        <title><![CDATA[MoFed-Guard: a mobility-aware federated and explainable lightweight Transformer framework for misbehavior detection in VANETs]]></title>
        <pubdate>2026-11-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Fida Muhammad Khan</author><author>Asim Zeb</author><author>Rex Bacarra</author><author>Jamil Abedalrahim Jamil Alsayaydeh</author><author>Rostam Affendi Bin Hamzah</author><author>Maqsood Mahmud</author>
        <description><![CDATA[Authenticated vehicles can disseminate false safety messages that pass cryptographic verification but remain behaviorally untruthful, making insider misbehavior a challenge in Vehicular Ad Hoc Networks (VANETs). This study proposes MoFed-Guard, a data-local federated framework combining OBU-TFT, a compact temporal Transformer, with reliability-aware aggregation and on-demand SHAP explanation. Fixed RSUs act as training clients and OBUs perform inference. MoFed-Guard weights client updates using vehicle-association stability, communication reliability, and leave-one-client-out update consistency; these indicators are evaluated empirically rather than treated as theoretically derived trust variables. On the trajectory-disjoint 20-class VeReMi Extension test partition, the framework achieved 95.20% accuracy, 94.75% macro-precision, 94.30% macro-recall, and 94.52% macro-F1. With architecture and training fixed, MoFed-Guard improved macro-F1 by 1.45 points over FedAvg and 0.61 points over FedAdam; the five-seed improvement was significant [t(4) = 6.30, p = 0.003, dz = 2.82]. Five-seed convergence reached 98.0 ± 0.2% validation macro-F1 at round 100, while held-out client-scale macro-F1 increased from 93.10% with three clients to 94.52% with ten. Explicit FL-layer delivery tests and a road-network-aware partition preserved the aggregation benefit, whereas coordinated-client corruption exposed the limitation of cosine-based consistency. The joint reliability score correlated positively with update utility (ρ = 0.71, p < 0.001). OBU-TFT contains 85,460 trainable parameters, occupies 0.34 MB, and achieved 5.10 ± 0.42 ms T4 inference latency. Parameter-matched comparison showed that OBU-TFT achieved higher macro-F1 than Linformer and Performer, although Linformer remained smaller and faster. These results support empirical reliability-conditioned federated weighting under simulation without implying universal convergence, Byzantine robustness, formal privacy, or field deployment readiness.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1942761</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1942761</link>
        <title><![CDATA[Sustainable Federated Continual Learning for environmental monitoring: a systematic review]]></title>
        <pubdate>2026-11-02T00:00:00Z</pubdate>
        <category>Systematic Review</category>
        <author>G. Suruthi</author><author>Rabindra Kumar Singh</author>
        <description><![CDATA[IntroductionThe use of distributed sensing infrastructures is increasingly becoming a key approach to environmental monitoring that is required to be mobile, flexible and adaptable to non-stationary conditions, preserve data privacy and consume as few resources as possible. The decentralized training paradigm provided by Federated Learning together with the ability of adapting sequentially over time without forgetting from Continual Learning has resulted in a concept of Federated Continual Learning (FCL) which offers a promising approach for privacy preserving environmental sensing systems that are capable of learning over time without forgetting. This systematic review provides an in-depth synthesis of the latest research, progress, environmental applications, sustainability, trustworthiness and future directions of Federated Continual Learning (FCL) for environmental monitoring.MethodsA total of 90 peer-reviewed publications were identified by conducting a systematic literature search using pre-defined inclusion/exclusion criteria, quality assessment, bibliometric analysis and qualitative evidence synthesis with the framework of PRISMA 2020.ResultsThe synthesized evidence suggests that Federated Learning has come of age in terms of optimization, communication complexity, robustness, and personalization, but Federated Continual Learning is still in its infancy with few applications in the real world. Typically, sustainability, trustworthiness, privacy, fairness, personalization and explainability are explored separately and not in parallel. Although most applications have been single-shot Federated Learning systems, without genuine continual adaptation, isolated examples include some applications from air quality to water quality, agriculture to disaster monitoring and remote sensing. Seven cross-cutting research gaps are identified: disjointed integration of research directions, few deployments in the real world, absence of standardized evaluation, unsolved multi-objective optimization, poor scaling in the long term, weak sustainability-trustworthiness coupling, and poor cross-domain generalization.DiscussionIt is not just about the evolution of single algorithmic breakthroughs anymore, but actual realization of Federated Continual Learning systems that can effectively assist in continual adaptation while ensuring communication efficiency, sustainability, trustworthiness, scalability and long-term deployment under real environmental conditions.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1979431</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1979431</link>
        <title><![CDATA[Measuring risk preferences in large language models using adaptive binary search]]></title>
        <pubdate>2026-11-02T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Idan-Chaim Cohen</author><author>Noya Littor</author><author>Aviad Elyashar</author><author>Odeya Cohen</author><author>Rami Puzis</author>
        <description><![CDATA[Large language models (LLMs) increasingly support decision-making in high-stakes domains, where a model's risk preference, its tendency to favor lower-risk or higher-risk options under uncertainty, can shape the decisions it supports. Prior studies measuring these preferences in LLMs have produced mixed results. Several used fixed designs, such as choice lists, which locate the point where a preference changes only as finely as their preset values allow. Adaptive procedures, which adjust each query to earlier responses, concentrate measurement near that point and are established for eliciting human risk preferences. We present such a procedure for LLMs: it poses a choice between a lower-risk and a higher-risk option, varies the expected value of the higher-risk option across steps, locates by binary search the model's tipping point, the point at which its outputs shift to favor the higher-risk option, and reports it as an index of risk preference under one probability structure. We applied the procedure to eight LLMs from four providers (OpenAI, San Francisco, CA, USA; Anthropic, San Francisco, CA, USA; Google, Mountain View, CA, USA; xAI, Palo Alto, CA, USA), using emergency and disaster management vignettes, and tested whether presenting each option's expected value explicitly, rather than leaving it to be inferred, changed the measured preference. Median tipping points for six of the eight models fell near the point of equal expected value under both presentations. Claude Sonnet 4.6's tipping point was higher when the expected value was left to be inferred, with a shift whose bootstrap interval excludes 0; Claude Haiku 4.5's was higher as a point estimate, but its shift interval includes 0. In individual queries, deviations from expected-value maximization concentrated in four models: those of the two Anthropic models all toward the lower-risk option and those of the two Google models almost all toward the higher-risk option. Over the range queried in both presentations, presenting the expected value explicitly reduced these deviations in all four models. Because the two presentations convey the same expected values, the reduction indicates that these models did not respond to them as equivalent. The procedure applies to any model accessible via a text interface and is a reusable tool for comparing and screening models for risk-related decision support.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1950048</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1950048</link>
        <title><![CDATA[Intelligent energy management of a novel non-isolated bidirectional dual-input single-output converter for electric vehicle applications]]></title>
        <pubdate>2026-10-29T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Faridha Banu M. R.</author><author>Jayapragash R.</author>
        <description><![CDATA[For use in electric vehicle (EV) applications, this paper suggests a non-isolated bidirectional Dual-Input Single-Output (DISO) DC–DC converter with an intelligent energy management method. The proposed converter integrates photovoltaic (PV) and battery energy sources in a single power conversion structure, which results in fewer components, higher efficiency and bidirectional power flow capability. The converter is designed based on a two-phase interleaved boost topology with a relay-assisted reconfiguration stage that facilitates flexible source selection, battery charging/discharging and regenerative braking operation. A Modified Deep Neural Network (MODNN) based intelligent energy management strategy is proposed to enhance energy utilization and adaptive power sharing. The proposed MODNN modifies a conventional single-task feedforward DNN by introducing a shared-trunk, multi-head architecture that jointly predicts the discrete operating mode and the continuous converter duty ratios through a composite classification-regression loss function, benchmarked against recent machine-learning baselines (k-NN, a bagged neural-network ensemble, and LSTM) under identical training/test conditions. In the proposed MODNN framework, the PV voltage, PV current, battery state-of-charge, battery voltage, and load demand information are used to decide the optimal operating mode and power flow condition under different operating scenarios. The converter provides flexible power distribution between available energy sources, productive use of regenerative braking energy, uninterrupted load supply and effective battery management. The performance of the proposed converter is verified by MATLAB/Simulink simulations, and the converter topology is experimentally validated with a laboratory prototype of 500 W. The experimental results show that the five operating modes operate successfully with a peak efficiency of 96.7%, which validates the effectiveness of the proposed converter for EV energy management applications.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1855069</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1855069</link>
        <title><![CDATA[Lightweight audio-visual sentiment analysis with dynamic mouth-state modeling and multimodal LLM benchmarking]]></title>
        <pubdate>2026-10-29T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Atitaya Yakaew</author><author>Teeradaj Racharak</author><author>Matthew N. Dailey</author><author>Watanee Jearanaiwongkul</author>
        <description><![CDATA[Sentiment analysis from video streams requires effective modeling of both temporal dynamics and multimodal cues. This paper presents a lightweight multimodal CNN-LSTM framework, designed for low-latency operation on live video streams, that classifies individuals' affective states as negative, neutral, or positive by jointly exploiting audio and visual information. The proposed system consists of four lightweight deep neural network models that fuse sequences of audio and visual features into a unified temporal stream, enabling the accumulation of sentiment evidence over time for robust prediction. To further enhance discriminative power, dynamic mouth state classification is incorporated, allowing the model to better capture subtle visual cues related to speech and sentiment. Experiments conducted on the Ryerson Audio-Visual Database of Emotional Speech (RAVDESS) demonstrate that the proposed approach achieves a test accuracy of 84.89% on the three-class sentiment recognition task, competitive with substantially larger audio-visual architectures, while using an order of magnitude fewer parameters and maintaining low-latency operation. In addition, we provide a comprehensive comparison with multiple multimodal large language models (LLMs), evaluated under diverse input configurations and prompting strategies. The results offer practical insights into the strengths and limitations of contemporary LLMs for audio-visual sentiment analysis and highlight the continued effectiveness of compact, task-specific neural architectures for streaming audio-visual input.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1952072</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1952072</link>
        <title><![CDATA[Dynamic receptive field and feature enhanced network for bridge defect detection]]></title>
        <pubdate>2026-10-28T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ziqi Tang</author><author>Yuang Ji</author><author>Bin Chen</author><author>Shan Huang</author><author>Yan Jin</author><author>Jichao Liu</author><author>Hui Li</author>
        <description><![CDATA[IntroductionAccurate defect detection in underwater and cross-domain marine infrastructure, such as bridge piers, submarine pipelines, and subsea equipment, is essential for structural safety assessment and intelligent ocean engineering. Existing methods, however, suffer from large scale variations, dense small targets, severe class imbalance, and texture loss in complex underwater environments.MethodsWe propose a multi–dimensional collaborative detection network based on YOLOv10n. In the backbone, switchable dilated convolutions (SAConv) adaptively adjust receptive fields, and a spatial-channel downsampling (SCDown) module preserves fine-grained details. At the backbone output, a Partial Spatial Attention and Multi-layer Latent Linear Attention (PSAMLLA) module captures global context with linear complexity via rotary and local position encoding. In the feature pyramid, a Conditional Identity Block Feature Fusion (C2fCIB) module enhances multi-scale fusion using deep large-kernel convolutions with structural re-parameterization. Finally, a Dynamic Detection Head (DyHead) integrates scale-aware, spatial-aware, and task-aware attention for adaptive inference.ResultsExtensive experiments on the public bridge-bkqxz/2 dataset demonstrate that our method achieves comparable or better results than several state-of-the-art detectors while maintaining low computational cost.DiscussionAlthough the current experiments are limited to bridge surface defect detection, the proposed techniques address challenges such as texture degradation and multi-scale object variability and may provide a useful foundation for future extension to underwater vision systems, which remain an important direction for subsequent investigation.]]></description>
      </item>
      </channel>
    </rss>