Robust and reliable medical diagnosis using artificial intelligence is crucial, yet real-world clinical environments present significant challenges due to dynamic covariate shifts affecting multi-modal data (images, text, tabular). Existing methods, including the single-modal robust classifier LaDiNE, often fail under these complex, multi-modal shifts, lacking mechanisms for cross-modal invariance, dynamic modality fusion, and fine-grained uncertainty attribution. To address this gap, we propose DyMoLaDiNE (Dynamic Multi-Modal Latent Diffusion Nested-Ensembles), a framework designed for reliable medical diagnosis under dynamic multi-modal covariate shifts. DyMoLaDiNE introduces four key innovations: (1) a Cross-Modal Invariant Feature Extractor leveraging multi-modal Vision Transformers and contrastive learning to derive robust latent representations, (2) a Dynamic Modality Weighting Mechanism that adaptively adjusts modality contributions based on instance-specific reliability scores, (3) a Robust Multi-Modal Diffusion Ensemble utilizing conditional diffusion models conditioned on multi-modal inputs and reliability scores for flexible, calibrated density estimation, and (4) Modality-Attributed Uncertainty Quantification to decompose predictive uncertainty by input source. Extensive evaluations on diverse datasets (MedMD&RadMD, MultiCaRe, PadChest, TCIA RE-MIND, BRaTS, Camelyon16, PANDA) demonstrate that DyMoLaDiNE significantly outperforms (p 0.005) state-of-the-art methods (LDM, CMCL, CGMCL, CIIM, DTTL, FFL, ALDM, LaDiNE) in terms of classification accuracy, robustness under dynamic perturbations, confidence calibration (ECE), and precise uncertainty quantification (CPIW, CNPV), while providing superior modality attribution fidelity. Ablation studies confirm the necessity of each component. DyMoLaDiNE represents a significant advancement in trustworthy, robust multi-modal medical AI. Code supporting this study DyMoLaDiNE .

Breast ultrasound tumor segmentation is critical for early diagnosis and treatment planning. However, due to the high similarity between tumors and background tissue in ultrasound images, achieving accurate segmentation poses significant challenges. To address this, we propose a Masked Separable Attention Network (MSA-Net), a segmentation model based on an encoder-decoder architecture. The model employs PVTv2 as the feature extraction backbone encoder and introduces our designed Masked Separable Attention (MSA) module. The core innovation of the MSA module lies in separating the multi-head self-attention mechanism into three function-specific subgroups: the foreground attention group focuses on the tumor region, the background attention group focuses on the surrounding tissue, and the global attention group captures the overall image information. This structured attention mechanism aims to more effectively model the contextual relationships between the tumor region, background region, and the entire image, thereby significantly enhancing the model's ability to distinguish between tumors and background tissue. Extensive experiments demonstrate that our MSA-Net achieves competitive performance compared to state-of-the-art breast tumor segmentation methods. Ablation studies further confirm the effectiveness and complementary of each component in our MSA-Net. The code is available at https://github.com/chenwang1701/MSA-Net.

Existing intelligent diagnostic methods in Traditional Chinese Medicine (TCM) heavily rely on clinical experience and rule-based models, which often lack generalizability and overlook medical text data. To address this, we propose a multi-instance multi-label graph neural network model for TCM diagnosis. We frame TCM diagnosis as a multi-instance multilabel problem: each medical record is treated as a bag, where instances are symptom-cluster subgraphs generated via random walks, labeled with syndrome elements (zhengsu). Features are extracted from these sub-instances, aggregated into a bag-level representation, and subsequently mapped to zhengdu labels through a classifier by a graph neural network. Our model achieves a Hamming Loss of 2.58 %, Ranking Loss of 2.56 %, One-Error of 21.55 %, Coverage of 4.565, and Average Precision of 81.55 %, outperforming classic MIML methods including DeepMIML, AttentionMIML, Fast-MIML, and Lnn-MIML across these metrics. These results demonstrate that combining multiinstance multi-label learning with graph neural networks effectively captures symptom-zhengsu relationships, offering a promising direction for intelligent and objective TCM diagnostic research.

Electrocardiogram (ECG) arrhythmia identification is critical for early cardiovascular disease diagnosis and monitoring in Internet of Things (IoT) industry. Still, it is difficult due to complicated waveforms, individual variability, and the requirement for real-time analysis on resource-limited equipment. Traditional approaches sometimes fail to detect complicated spatial-temporal correlations in ECG data, limiting their efficiency in identifying arrhythmias. Furthermore, deploying these models in tinyML contexts, such as edge and IoT devices limited by large computational and memory needs, emphasizes the importance of lightweight, accurate models for real-time applications. Our suggested solution consists of three main components: SigNet, DualGCNN, and FusionGCNN. SigNet uses Separable Convolution layers to effectively extract local spatial features, making it ideal for IoT-based healthcare deployment. DualGCNN combines dual Graph Convolutional layers with spatial attention, allowing the model to capture local and global dependencies for better classification of arrhythmia. FusionGCNN combines the capabilities of GCN and SigNet with an effective feature fusion technique to improve feature representation while remaining computationally economical. Ablation tests show that FusionGCNN improves performance considerably, with greater accuracy (0.9641), lower training error (0.0004), and a higher F1 Score (0.9645) across a variety of ECG patterns. FusionGCNN, with its low training error, high stability, and computational economy, is well-suited to tinyML requirements, allowing implementation on edge and IoT devices for scalable, real-time ECG monitoring in healthcare.

Multiview data classification remains a challenging problem in machine learning, particularly in effectively integrating and representing data from different views. This article introduces contrastive multiview low-rank latent subspace self-representation and classification network (CMvLSCN), a novel end-to-end multiview discriminant learning framework that addresses classification from view, sample, and subspace levels. CMvLSCN employs contrastive learning to enhance interview consistency within categories while differentiating between categories. It imposes a low-rank latent self-representation structure on the unified subspace, capturing intrinsic data relationships. Additionally, sample-level contrastive constraints in the latent space further boost the representation’s discriminative power. Extensive experiments demonstrate CMvLSCN’s superior performance across various multiview classification tasks, notably maintaining robustness even with limited training data. Our code and datasets are publicly available on https://github.com/DeyuTsang/CMvLSCN

This article reports the design and validation of a novel predefined-time barrier function adaptive sliding-mode control (PTBFASMC) strategy for robust control of disturbed systems. The PTBFASMC strategy is established by integrating the time base generator along with the barrier function. Unlike existing similar works, the proposed method enables global predefined-time convergence, i.e., the system trajectory returns to the ultimate bound even if an escape occurs at a certain time instant. Besides, the convergence time can be predefined by the user, which is independent of the initial conditions and disturbance. Moreover, the reaching phase is eliminated and the magnitude of initial control output is zero. Another attractive feature of the proposed method lies in that the ultimate bound can be predefined, i.e., the ultimate bound is independent of the upper bound of disturbance. To avoid large control magnitude, a modified control strategy is provided, which extends the proposed scheme to different scenarios. The stability of the control system is demonstrated, and its superiority is verified through numerical simulations and experimental investigations on a piezoelectric actuator.

When applying the space-time coding technique into the ultrasound imaging system, the signal-to-noise-ratio (SNR) can be improved. However when the space-time coding combines with the conventional synthetic aperture techniques can result some ghost images because the conventional synthetic aperture technique cannot suppress the axial direction of the side-lobe which emerges after pulse compression and thereby reducing the image quality. To solve the problem of ghost imaging, this paper introduces a new method, which combines the techniques of space-time coding and ultrasonic back-propagation synthetic aperture. The main idea of the method is that different locations using different coding pulse emission. It can be achieved by using the Truncated long code[1]. The synthetic aperture technique based on the ultrasonic back-propagation is then applied for processing the echo signals to further improve the SNR as well as suppress the axial side-lobe. The crosstalk among echo signals can also be reduced. Simulation shows positive results of the new method. Both detection sensitivity and robustness against noise are improved and overall performance of the image system is enhanced.

Recently nonnegative Matrix Factorization (NMF) has been proven a powerful method in clustering analysis of gene expression data. There exist two popular loss functions for minimization in decomposition: one is Euclidean distance and the other generalized Kullback-Leibler divergence. Both loss functions can be derived from a linear model with additive noise, and the Euclidean distance loss corresponds to Gaussian noise while the generalized Kullback-Leibler divergence corresponds to Poisson noise. However real data is not only Gaussian or Poisson, or not both. In order to take into account complex type of noise, we combine both loss functions for NMF according to regularization method. We compared NMF based on Euclidean distance, the generalized Kullback-Leibler divergence, and our regularized version, with application in clustering gene expression data. The experimental results demonstrate the effectiveness of the proposed method.

Nonnegative matrix factorization is used extensively for feature extraction and clustering analysis. Recently many sparsity/sparseness constraints, such as L1penalty, are introduced for sparse nonnegative matrix factorization. Inspired by sparsity measures from linear regression model, this paper proposes to integrate nonnegative matrix factorization with another sparsity constraint, the elastic net. The experimental results of clustering analysis on three gene expression datasets demonstrate the effectiveness of the proposed method.

Automatically selecting a subset of genes with strong discriminative power is a very important step in classification problems based on gene expression data. Lasso and Dantzig selector are known to have automatic variable selection ability in linear regression analysis. This paper employs Lasso and Dantzig selector to select most informative genes for representing the class label as a linear function of gene expression data. The selected genes are further used to fit linear classifiers for cancer classification. On 3 publicly available cancer datasets, the experimental results show that in general, Lasso is more capable than Dantzig selector in selecting informative genes for classification.

Due to the enormous amounts of data contained in hyperspectral imagery, the main challenge for hyperspectral image classification is to improve the accuracy with less computation complexity. Hence, dimensionality reduction (DR) is often adopted, which includes two different kinds of methods, feature extraction and feature selection. In this paper, discrete wavelet transform (DWT) and affinity propagation (AP), which belong to feature extraction and feature selection respectively, are combined together to accomplish the DR task. Firstly, DWT-based features are extracted from the original hyperspectral data; secondly, AP is applied to select representative features from the obtained ones. Experimental results demonstrate that, compared with some other DR methods which only make use of feature extraction or feature selection, the features acquired by the hybrid technique make the classification results more accurate.

Classification of gene expression samples is a core task in microarray data analysis. How to reduce thousands of genes and to select a suitable classifier are two key issues for gene expression data classification. This paper introduces a framework on combining both feature extraction and classifier simultaneously. Considering the non-negativity, high dimensionality and small sample size, we apply a discriminative mixture model which is designed for non-negative gene express data classification via non-negative matrix factorization (NMF) for dimension reduction. In order to enhance the sparseness of training data for fast learning of the mixture model, a generalized NMF is also adopted. Experimental results on several real gene expression datasets show that the classification accuracy, stability and decision quality can be significantly improved by using the generalized method, and the proposed method can give better performance than some previous reported results on the same datasets.

Recently Saul and Lee proposed a mixture model for discriminative classification of non-negative data via non-negative matrix factorization for feature extraction. In order to improve the generalization, this paper considers a sparse version of the model. The basic idea is to minimize the sum of the weights of un-normalized mixture models for posterior distributions according to regularization method. Experiments on CBCL face database and USPS digit data set assess the validity of the proposed approach.

This paper addresses the well-known problem of natural image matting. It proposes a whole new framework that could effectively deals with the confused boundaries such as hair, furs and other complicated situations. We take the natural image matting problem as a pattern recognition problem and use the recently developed non-negative matrix factorization technique to solve it. Experimental results show that our approach could properly handle the confused boundaries. Compared with other algorithms visually, the results of our algorithm are comparable to the algorithms that are the best of nowadays.

This paper combines linear spun coding and nonnegative matrix factorization into sparse non-negative matrix factorization. In contrast to non-negative matrix factorization, the new model can learn much sparser representation via imposing sparseness constraints explicitly; in contrast to a close model -non-negative sparse coding, the new model can learn parts-based representation via fully multiplicative updates because of adapting a generalized Kullback-Leibler divergence instead of the conventional mean error for approximation error. Experiments on MIT-CBCL training facts data demonstrate the effectiveness of the proposed method.

In this paper, the linear support vector machine (LPSVM) algorithm is used to construct an over complete set of weak classifiers, and AdaBoost algorithm are adopted to select part of them to form a strong classifier. During the course of feature extraction and selection, the new method can minimize the classification error directly, whereas most previous works cannot do this. An important difference between this method and other methods is that the sparse features are learnt from the training set instead of being arbitrarily defined. Experiments demonstrate that the new algorithm performs well.