2026-10-01 2026, Volume 11 Issue 5

  • Select all
  • research-article
    Chen Zhang, Anh Hoang Phuc Nguyen, Yongping Zheng, Sai Ho Ling
    2026, 11(5): 1321-1332. https://doi.org/10.1049/cit2.70157

    Scoliosis assessment has been increasingly supported by AI-assisted tools, enabling more efficient and standardised clinical workflows. However, existing ultrasound-based pipelines often suffer from limited generalisation and unstable performance. These limitations primarily arise from two challenges: (1) suboptimal segmentation under low-contrast and noisy imaging conditions, and (2) unreliable anatomical landmark identification due to missing or ambiguous features. To address these issues, we propose Auto-ISC, a novel two-stage automated ultrasound scoliosis assessment system. The first stage employs the ISC-Swin segmentation model, enhanced by an Inter-Sample Contrastive Bank (ISCB) to improve cross-sample feature consistency and class-level discrimination. The second stage introduces a dynamic pairing logic algorithm for accurate matching of thoracic vertebrae, enabling reliable downstream angle estimation. Experimental results demonstrate strong segmentation performance (Dice: 86.10%, IoU: 76.08%) and high agreement with manual measurements. The proposed method achieves low average absolute differences (ABS) (thoracic: 2.83° ± 2.16°; lumbar: 2.89° ± 2.55°) and strong correlations (R2= 0.91 for thoracic and 0.81 for lumbar). Overall, Auto-ISC provides a reliable and radiation-free solution for automated scoliosis assessment, with improved clinical applicability.

  • research-article
    Jiexing Li, Shuai Li
    2026, 11(5): 1333-1344. https://doi.org/10.1049/cit2.70175

    The k-winners-take-all (kWTA) problem, as a competitive selection strategy, is crucial in multi-agent systems. However, existing noise-robust kWTA networks mainly address noise while overlooking communication disturbances. The coexistence of both factors significantly degrades solution accuracy and hinders distributed consensus estimation. To address these issues, this paper first proposes an accelerated-convergence and noise-tolerantkWTA (ACNT-kWTA) network by introducing a nonlinear activation function and an error-integral term. Second, a robust consensus estimator is developed and integrated with the proposed ACNT-kWTA network to construct an enhanced-robustness competitive coordination (ERCC) model, which simultaneously suppresses noise and communication disturbances. Theoretical analyses demonstrate the global convergence of the proposed ACNT-kWTA network and the robust consensus estimator. Ultimately, the effectiveness and superiority of the proposed model are demonstrated through comparative numerical simulations and physical experiments.

  • research-article
    Zhengya Sun, Bojie Fan, Bo Yin, Jie Nie, Anan Liu
    2026, 11(5): 1345-1359. https://doi.org/10.1049/cit2.70160

    Typhoons pose a significant threat to both human safety and economic stability. As a crucial metric for assessing their destructive potential, typhoon intensity (TI) prediction has become an important research focus, with numerous methods developed. However, effectively combining two-dimensional typhoon structure domain-expert knowledge (2D-TSDK) with three-dimensional typhoon structure data-driven knowledge (3D-TSDK) for accurate TI prediction remains a challenge. In this paper, we propose a novel cascaded deep learning framework that combines domain expertise with data-driven refinement for typhoon intensity prediction. The two-stage architecture first generates initial intensity estimates by fusing 2D-TSDK features with established physical wind-pressure relationships via a specialised loss function. These estimates are subsequently refined by a vision transformer that extract fine-grained spatial patterns from 3D wind field reanalysis data. Both stages incorporate temporal attention mechanisms with dual (absolute and relative) position encoding, providing comprehensive modelling of typhoon evolution dynamics. Extensive experiments on CMA–BST and ERA-interim datasets demonstrate consistent performance improvements over baseline methods, confirming the framework's effectiveness for TI prediction.

  • research-article
    Weiqing Yan, Tianhao Xie, Liang Liao, Wujie Zhou, Chang Tang
    2026, 11(5): 1360-1374. https://doi.org/10.1049/cit2.70183

    Recent advances in SLAM systems based on 3D Gaussian Splatting have enabled dense and photorealistic scene reconstruction. However, 3DGS-based methods still suffer from limited geometric awareness due to the lack of normal rendering support, view inconsistency and weak depth constraints, ultimately hindering high-fidelity geometric reconstruction and robust pose tracking. In this paper, we present Geo2DGS-SLAM, a novel dense RGB-D SLAM framework that uses a geometry-aware 2D Gaussian representation for scene modelling, featuring a hybrid tracking framework, 2DGS-based submap construction and a mesh-binding submap merging strategy. Unlike prior approaches that rely solely on 3DGS appearance modelling, our 2DGS representation enables more accurate rendering of both depth and surface normals, facilitating high-quality and geometrically consistent reconstruction. To further enhance robustness and global consistency, the hybrid tracking framework integrates local bundle adjustment, loop closure and global bundle adjustment, which jointly mitigate accumulated drift and improve trajectory stability, particularly in long or challenging sequences. Extensive experiments on public indoor RGB-D datasets demonstrate that Geo2DGS-SLAM achieves competitive performance in geometry-aware mapping, rendering fidelity and pose estimation accuracy among recent Gaussian-based SLAM methods. Codes and additional experimental results will be available at https://github.com/Able1231/Geo2DGS-SLAM.

  • research-article
    Murad Ali Khan, Naeem Iqbal, Syed Shehryar Ali Naqvi, Muhammad Faseeh, Do-Hyeun Kim
    2026, 11(5): 1375-1395. https://doi.org/10.1049/cit2.70155

    Accurate load forecasting and reliable anomaly detection are critical for the stable operation of modern smart grids (SGs), which increasingly rely on cyber-connected infrastructures. However, the integration of smart metres and two-way communication exposes SGs to data integrity attacks that can manipulate consumption measurements, degrade forecast accuracy and compromise grid stability. This study proposes RES-AD, a robust AutoML-driven ensemble framework designed for real-time detection of abnormal consumption patterns and accurate short-term energy forecasting. RES-AD consists of an Optimal Ensemble Classifier (OEC) for anomaly detection and an Optimal Ensemble Regressor (OER) for forecasting, both selected and optimised automatically using H2O AutoML. Five types of synthetic data integrity attacks are injected to simulate realistic adversarial conditions, whereas SMOTE and correlation-based feature selection ensure balanced and informative training data. The revised framework further incorporates stealthy adversarial attack simulation evaluated using FGSM and demonstrates strong robustness with only a 1.7% degradation under perturbation, further mitigated through adversarial augmentation. Additionally, real-time readiness is validated through latency and memory profiling, where the OEC ensemble requires only 23.36 ms per inference with a footprint below 0.05 MB, making it suitable for edge or SCADA-integrated deployment. Experimental results using public residential smart metre data show that OEC achieves 99.7% (hourly) and 99.9% (daily) detection accuracy, whereas OER improves energy forecasting with a MAPE reduction of 11.11% and 11.72%, respectively. These outcomes demonstrate that RES-AD provides a scalable, attack-aware and deployment-ready solution that enhances both the reliability and resilience of modern smart grid operations.

  • research-article
    Yaxin Liu, Yan Liu, Xinyu Song, Ruiqing Han, Haiyong Liu, Bo Huang, Yufeng Yao, Ming Zhong
    2026, 11(5): 1396-1412. https://doi.org/10.1049/cit2.70136

    Sequential manipulation is the process by which robots perform multiple interdependent steps to accomplish composite tasks, demanding tight integration of perception, planning and execution. Existing methods incorporate explicit features such as category, semantics, 6D pose or affordance to enhance consistency, yet single-feature representations limit generalisation and adaptability across diverse task stages. To overcome these limitations, we propose a shape- and function-grounded keypoint (SFK) representation that provides interpretable, task-relevant object abstractions. The SFK representation preserves topological consistency, enabling the inference of functional semantics, 6D pose, spatial constraints and task-oriented grasp configurations. Building upon this, a keypoint-driven hierarchical framework is developed to unify perception, planning and execution, thereby enabling task-level action reasoning conditioned on keypoint representations and coherent multistep action execution. For evaluation, a keypoint dataset containing over 1600 RGB images of 29 household objects is constructed, and extensive experiments demonstrate task-relevant keypoint generation with reduced computational time, as well as higher task execution success rates and broader task coverage.

  • research-article
    Sultanul Arifeen Hamim, Mubasshar U. I. Tamim, S. M. Hasan Mahmud, M. F. Mridha, Zeyar Aung
    2026, 11(5): 1413-1429. https://doi.org/10.1049/cit2.70170

    Brain tumour classification is a critical task in medical imaging that requires accurate and interpretable solutions to assist in clinical decision-making. In this paper, we present GraphConvNet-X, a novel hybrid model that integrates convolutional neural networks (CNNs) for spatial feature extraction with graph neural networks (GNNs) that employ a novel hinge attention network (HAN) for robust relational aggregation. Leveraging Explainable Artificial Intelligence (XAI) techniques, such as gradient-weighted class activation mapping (Grad-CAM) and SHAP, the model provides transparent insights into its predictions, aligning them with clinically significant regions in magnetic resonance imaging (MRI) scans. The integration of HAN enhances the GNN's ability to selectively focus on the most salient tumour-related graph features, improving overall discriminative power. The proposed architecture was evaluated on three benchmark datasets, achieving impressive accuracies of 98%, 97% and 96%, surpassing state-of-the-art methods. Extensive experiments have demonstrated its robustness across various configurations and highlighted its adaptability to diverse scenarios. By combining precision, adaptability and interpretability, GraphConvNet-X sets a new standard for AI-driven diagnostic tools, showing its potential to revolutionise brain tumour classification and contribute significantly to modern healthcare.

  • research-article
    Jiale Ren, Mingxin Tan, Mengyuan Liu, Yingxuan You, Hong Liu
    2026, 11(5): 1430-1443. https://doi.org/10.1049/cit2.70176

    Relative pose estimation under the query–reference paradigm has emerged as a practical alternative for estimating the 6D pose of unseen objects without relying on CAD models or extensive annotations. However, many existing approaches remain difficult to deploy in practice, as their reliance on geometric matching makes robustness highly sensitive to matching quality. In this paper, we present MRPose, a framework for relative 6D object pose estimation under a multi-robot setting, where query and reference views are independently captured by different robots. Without requiring CAD models, shared calibration across robots or pose annotations, MRPose directly estimates the full relative transformation between unseen objects from RGB images. Built upon the power of vision foundation models, MRPose introduces a multiplex feature learning strategy that jointly exploits patch-level appearance correspondence and sparse geometric cues through attention-based cross-view feature aggregation. Instead of relying on traditional geometric solvers such as PnP or essential matrix estimation, the proposed framework directly regresses the full relative 6D pose in an end-to-end manner. Extensive experiments on the LINEMOD and YCB-Video benchmarks demonstrate that MRPose compares favourably with existing RGB-based relative pose estimation methods. Real-world experiments further validate its applicability in practical scenarios.

  • research-article
    Tawfeeq Shawly, Ahmed A. Alsheikhy
    2026, 11(5): 1444-1462. https://doi.org/10.1049/cit2.70171

    The Riyadh Metro represents a significant project in Saudi Arabia, designed to transform urban transportation and reduce traffic congestion within the city. With six metro lines and 85 stations, the network is expected to serve millions of passengers daily, necessitating innovative digitally driven maintenance approaches to ensure reliable efficient operation. We hypothesise that integrating Internet of things (IoT) sensors, digital twin technology, and a temporal graph neural network (TempGNN) enables real-time health monitoring. This paper introduces a real-time digital health monitoring framework that incorporates IoT sensors, digital twin technology, a proposed TempGNN and a new module single encounter-emphasise attention neural network (SEEA- NN). TempGNN exploits both temporal patterns and spatial relationships in these sensor data to feed SEEA-NN with its required data. SEEA-NN is designed to detect and assess the importance of rare single-occurrence failure precursors, often overlooked by conventional models, alongside recurring patterns. The proposed framework contains an input gating module to filter out noisy sensor data, a multihead encounter attention module that employs a specific attention head to isolate and enhance the signal from anomalous single-encounter events, and a temporal synthesis layer that combines these emphasised features with sequential data for thorough analysis. This enables the framework to predict potential failures with high accuracy of approximately 98%, allowing for proactive maintenance scheduling and decreased unplanned downtime by more than 30% and maintenance expenses by around 25%, whilst also enhancing energy efficiency and passenger satisfaction in simulations. This approach demonstrates a scalable model for the digital transformation of large-scale transport infrastructure and for implementing intelligent predictive maintenance in metro systems worldwide and in logistics fleets, container handling systems and freight rail networks.

  • research-article
    Youheng Bai, Zitao Liu, Mingliang Hou, Teng Guo, Weiqi Luo
    2026, 11(5): 1463-1479. https://doi.org/10.1049/cit2.70158

    Cognitive diagnosis aims to infer learners' knowledge states from their exercise responses, enabling personalised education at scale. Existing methods represent exercises solely by coarse-grained knowledge component annotations, overlooking semantic content and step-level cognitive processes. This causes diagnostic confounding, in which exercises requiring fundamentally different reasoning pathways become indistinguishable in learned representations. Such confounding degrades diagnostic accuracy by conflating distinct learning deficiencies. To address this, we propose H2CD, a semantic-enhanced heterogeneous hypergraph framework that unifies LLM-based cognitive decomposition with structural learning. First, we leverage a large language model to decompose exercises into step-by-step solution paths and extract fine-grained step-level KCs, capturing procedural knowledge beyond predefined annotations. Second, we construct a heterogeneous hypergraph with two types of hyperedges: knowledge hyperedges that connect exercises to their associated multi-granular KCs, preserving compositional knowledge structures, and interaction hyperedges that capture personalised student-exercise relationships. Third, a two-stage hypergraph neural network performs knowledge structure propagation on static exercise-KC relationships and interaction propagation on dynamic student-exercise patterns. Finally, a KC-aware prediction layer projects representations to a learnable diagnostic axis, enabling interpretable mastery assessment at individual KC granularity. Extensive experiments on three real-world educational datasets demonstrate that H2CD outperforms state-of-the-art methods in both AUC and accuracy.

  • research-article
    Falah Mohammed Mohammed, Osama Nayel Al Sayaydeh, Taha H. Rassem, Chee Peng Lim
    2026, 11(5): 1480-1495. https://doi.org/10.1049/cit2.70177

    This study investigates early detection of heart diseases using a new Enhanced Fuzzy Min-Max (EFMM) neural network. Two key modifications are introduced: a refined EFMM contraction procedure during the learning phase to reduce data distortion and information loss, and a new weighted validation method to optimise hyperbox selection and improve prediction accuracy. The proposed model is evaluated with heart disease data sets subject to varying expansion constraints, training data sizes, and cross-validation settings. The results indicate that the proposed EFMM model consistently outperforms other models, including the original Fuzzy Min-Max (FMM), EFMM, and non-FMM classifiers such as J48, MLP, IBK, AdaBoost, and others. The test accuracy rates achieve 82.11% with 5-fold cross-validation and 81.56% with 10-fold cross-validation. Compared with other FMM variants, the proposed model achieves 80.26% accuracy, surpassing GFMM, IOL-GFMM, Kn-EFMM, and RFMM models. Using real-world heart disease data, the proposed model outperforms FMM and EFMM in terms of accuracy, sensitivity, and specificity. Statistical test results indicate the effectiveness of the proposed model in performance statistically, as compared with those FMM, EFMM, and other classifiers. The improved EFMM model is generic and applicable not only to heart disease diagnosis but also to other general data classification tasks.

  • research-article
    Yirui Liu, Xueqiang Gu, Caiming Zheng, Xin Xu, Zhangling Wang, Wanpeng Zhang, Yixing Lan
    2026, 11(5): 1496-1510. https://doi.org/10.1049/cit2.70127

    Policy heterogeneity is crucial for achieving sophisticated coordination in complex collaborative tasks, which has emerged as one of the key challenges in multi-agent reinforcement learning (MARL) in recent years. Notably, the grouping paradigm has made remarkable progress in addressing policy heterogeneity. However, most existing grouping methods require predefining the number of groups or the composition and quantity of members within each group, which need to be individually configured for each scenario and are difficult to set without sufficient expert knowledge. By contrast, we propose a novel MARL grouping algorithm named Credit-driven adaptive Grouping (CreateG) which divides the entire training process into multiple phases and reallocates poorly adapted (low-credit) individuals at each training stage. With the help of the mechanism we designed, an environment-adaptive grouping is ultimately formed. Furthermore, we design a hierarchical hypernetwork architecture to accommodate this adaptive grouping mechanism. Experiments conducted on StarCraft II micromanagement hard and super hard tasks, Google Research Football and TAG scenarios show CreateG achives state-of-the-art MARL performance. Moreover, extensive ablation studies elucidate the operational mechanism of the grouping strategy and other components demonstrate how they enhance overall performance.

  • research-article
    Tao Chen, Jie Gao, Yong Xu, Weihong Qiu, Kunhong Liu, Yijie Wu, Weimin Ye
    2026, 11(5): 1511-1528. https://doi.org/10.1049/cit2.70174

    Fatty liver disease (FLD) is a prevalent chronic condition that can progress to clinically significant liver injury if left untreated. Given its high prevalence, FLD poses a substantial public health burden. Traditional Chinese medicine suggests that the appearance of tongue reflects hepatic and metabolic status, enabling artificial intelligence-based tongue diagnosis to provide rapid low-cost prescreening for FLD severity. In practice, the development of high-quality models is hindered by costly disease labelling, scarcity of positive cases, and severe class imbalance. In this work, we propose DualSS, a dual-diffusion synthetic sampling framework to mitigate these limitations. DualSS first trains a disease-agnostic dual-diffusion generator (DDG) on tongue-image data, leveraging abundant negative samples to learn unified tongue representations. Building on DDG, we design a latent space synthetic sampling pipeline that synthesises realistic positive-class images to construct a more balanced augmented training dataset. We then couple this pipeline with a noise-robust training strategy on the augmented dataset to further enhance the learning of downstream diagnostic models. Extensive experiments on FLD severity classification using tongue images and physiological indicators show that DualSS consistently outperforms existing rebalancing baselines, offering an effective and generalisable solution to alleviate positive sample scarcity in tongue image-based medical diagnosis.

  • research-article
    Yanlin Li, Kaibo Shi, Yue Yu, Xiangkun Wang, Ming Zhu, Xiao Cai
    2026, 11(5): 1529-1545. https://doi.org/10.1049/cit2.70139

    This paper proposes a boundary control method for nonlinear distributed parameter systems (DPSs) with limited boundary measurements (BMs), as typically encountered in networked cyber-physical processes with spatially distributed dynamics such as thermal and biomedical diffusion systems. The method aims to address cyber-physical threats and hybrid attacks whilst improving state estimation accuracy and system performance. First, a partial differential equation (PDE)-based state observer is established, with an embedded adaptive event-triggered mechanism (AMET) that optimises communication through dynamic sampling adjustment. Second, a Takagi–Sugeno (T–S) fuzzy noncollocated boundary sampled-data controller (NBSDC) is designed. It adopts a spatially separated architecture to mitigate observation-actuation mismatch under partial observability. Furthermore, the controller is designed to handle hybrid attacks with inconsistent modes, including denial-of-service (DoS) interruptions and deceptive signal alterations. Their impacts are bounded by energy constraints to ensure practical feasibility. In addition, a Markov jump process is employed to characterise stochastic controller failures, capturing abrupt mode transitions and enhancing fault tolerance under adverse conditions. Third, stability analysis based on Lyapunov–Krasovskii functionals (LKFs) provides theoretical guarantees for system stability and estimation error convergence. Finally, the effectiveness and applicability of the proposed method are validated through simulations on a tumour cell diffusion model.

  • research-article
    Chang Liu, Yapeng Li, Zengmao Wang, Dewen Soh, Bo Du
    2026, 11(5): 1546-1566. https://doi.org/10.1049/cit2.70120

    Human-centric sensing tasks such as visual sentiment analysis and smartphone-based human activity recognition (HAR) exhibit strong client heterogeneity from feature, label, and concept shifts. Among these, concept shift-how individual users interpret and generate human-centric signals-is particularly challenging yet underexplored in federated learning. We propose MAPFL (Modular Adaptive Personalised Federated Learning), a personalised federated framework that performs encoder-level personalisation while preserving standard communication. Each client model is decomposed into a shared encoder and a local classifier head, and equipped with both global and personalised encoders. A gradient-aligned personalisation controller learns client-specific mixing coefficients that adaptively balance the two encoders based on the alignment of their gradients, allowing each client to decide how much to share versus personalise at the representation level. This design supports both image- and sensor-based tasks. On an affective-computing benchmark built from three in-the-wild visual sentiment datasets (ARTphoto, ABSTRACT, PARA) and on a smartphone-based HAR benchmark derived from UCI HAR, MAPFL improves AUC and accuracy by up to 3.3% over strong federated and personalised baselines and converges faster and more stably. These results show that gradient-aligned encoder personalisation is an effective and practical mechanism for handling multi-dimensional heterogeneity in human-centric sensing.

  • research-article
    Xinlin Xiao, Xiangtao Zheng, Jianhua Tao, Xiaoqiang Lu
    2026, 11(5): 1567-1584. https://doi.org/10.1049/cit2.70159

    Test-Time Batch Adaptation (TTBA) aims to adapt a pre-trained source model to small, unlabelled target batches at test time. The TTBA methods focus on adapting the pre-trained model or the target data in a one-to-one alignment paradigm. However, these one-to-one alignment paradigms assume that the source domain may share the same knowledge with the target domain. It becomes ineffective when the target domain is different from the source domain. In this paper, a multi-source batch normalisation method is introduced. Specifically, spectral similarity clustering selects a small set of representative source domains, spectral entropy-guided weighting combines source statistics based on their relevance to the target domain, and batch-wise re-initialisation with unsupervised refinement stabilises updates. Extensive experiments on multiple benchmark datasets verify the effectiveness of the proposed method, achieving superior performance over state-of-the-art TTBA methods, especially in scenarios with small target batches.

  • research-article
    Zhengyu Zou, Jingfeng Li, Hao Li, Xiaolei Hou, Jinwen Hu, Jingkun Chen, Lechao Cheng, Dingwen Zhang
    2026, 11(5): 1585-1599. https://doi.org/10.1049/cit2.70086

    Neural radiance fields (NeRFs) implicitly model continuous three-dimensional scenes using a set of images with known camera poses, enabling the rendering of photorealistic novel views. However, existing NeRF-based methods encounter challenges in applications such as autonomous driving and robotic perception, primarily due to the difficulty of capturing accurate camera poses and limitations in handling large-scale dynamic environments. To address these issues, we propose vision-only dynamic NeRF (VDNeRF), a method that accurately recovers camera trajectories and learns spatiotemporal representations for dynamic urban scenes without requiring additional camera pose information or expensive sensor data. VDNeRF employs two separate NeRF models to jointly reconstruct the scene. The static NeRF model optimises camera poses and static background, whereas the dynamic NeRF model incorporates the 3D scene flow to ensure accurate and consistent reconstruction of dynamic objects. To address the ambiguity between camera motion and independent object motion, we design an effective and powerful training framework to achieve robust camera pose estimation and self-supervised decomposition of static and dynamic elements in a scene. Extensive evaluations on mainstream urban driving datasets demonstrate that VDNeRF surpasses state-of-the-art NeRF- based pose-free methods in both camera pose estimation and dynamic novel view synthesis.

  • research-article
    Shengkai Sun, Yutong Li, Jianfeng Lv
    2026, 11(5): 1600-1612. https://doi.org/10.1049/cit2.70164

    In this article, a new distributed k-winner-take-all (Dk-WTA) network is designed and used to construct a dynamic allocation scheme for competitive coordination tasks of multiple manipulators as embodied agents. The scheme is constructed based on the following two aspects. On one hand, with reference to a perspective that the Laplacian matrix is introduced into a k-WTA network, a quadratic programming (QP) problem used for the competitive coordination of multiple manipulators is designed. Then, inspired by a noise-suppressing neural dynamics (NSND) method, a Dk-WTA network is established by solving this QP problem. On the other hand, combined with the forward kinematics, a Dk-WTA dynamic allocation scheme is constructed. In addition, theoretical analyses on the constructed scheme are conducted regarding convergence and robustness. Finally, for verification of the feasibility of the proposed D k-WTA network, simulations and experiments based on multiple manipulators for trajectory-tracking tasks are conducted. Notably, the proposed scheme has the potential to be extended to other embodied multi-agent systems.

  • research-article
    Chao Liu, Xizhao Wang, Farhad Pourpanah, Sam Kwong
    2026, 11(5): 1613-1625. https://doi.org/10.1049/cit2.70166

    Drug–Drug Interaction (DDI) prediction is critical for ensuring patient safety, particularly under long-tailed distributions, where frequent (head) interactions dominate whereas rare (tail) interactions remain underrepresented. Conventional loss functions such as cross-entropy often tend to overfit head classes while they underperform on rare but clinically important classes. Recent studies have shown that sharpness-aware minimisation (SAM) improves generalisation by encouraging solutions that lie in flatter regions of the loss landscape. Given that the loss function plays a fundamental role in model training and that uncertainty estimation is crucial for handling ambiguous or difficult samples, we propose the uncertainty and SAM (USAM) loss for improving DDI prediction under long-tail scenarios. The USAM loss integrates predictive uncertainty with SAM via two key components: (i) a dynamic reweighting strategy based on uncertainty, which leverages prediction entropy to emphasise difficult samples and (ii) a class-balanced SAM regularisation term, which encourages flatter minima and enhances generalisation across imbalanced classes. Extensive experiments on four long-tailed DDI benchmarks demonstrate that the USAM loss consistently outperforms existing methods in terms of both F1 score and recall. These results highlight the effectiveness of incorporating uncertainty modelling with sharpness-aware optimisation in addressing long-tailed problems.

  • research-article
    Guodao Zhang, Yupeng Li, Yanjie Lu, Genfu Yang, Xiaotian Pan, Jing Wang
    2026, 11(5): 1626-1643. https://doi.org/10.1049/cit2.70172

    This paper presents a comprehensive closed-loop brain-eye movement framework, spanning parametric modelling, dynamic analysis and real-time FPGA implementation. A simplified first-order oculomotor model was derived from the classical Rob inson formulation to preserve the essential neural-mechanical feedback structure while enabling efficient hardware realization. Through systematic parametric variations of neural and mechanical properties and multiple classes of brain-generated inputs (step, ramp, sinusoidal and noisy), the system exhibited stable closed-loop behaviour with fast convergence and low steady-state errors. Time-domain and phase-plane analyses confirmed that the eye accurately tracked diverse target trajectories across wide parameter ranges, with all equilibrium points classified as asymptotically stable. The discrete-time model was implemented on a Xilinx Virtex-4 FPGA using only adders, subtractors and shifters, completely avoiding multipliers. The design achieved a maximum clock frequency of 118 MHz, operated at 100 MHz with a closed-loop latency of nine cycles (90 ns), and consumed approximately 280 mW total power while utilising less than 4% of the device resources. To validate the system in a practical setting, a real-time hardware platform with a 3 × 3 target grid (nine fixation points) was developed. Experimental results demonstrated that the closed-loop brain-eye architecture reliably steered the eye towards each target with minimal tracking error and smooth dynamic response. Overall, this work bridges physiological oculomotor models and neuromorphic hardware, offering a scalable resource-efficient platform for real-time brain-eye closed-loop emulation and control.