Breast cancer is a common malignant tumor that poses a serious threat to women's health. The incidence and mortality rates of breast cancer have shown an increasing trend worldwide in recent years; therefore, an accurate assessment of breast cancer prognosis is crucial for the development of individualized treatment plans and for the improvement of survival quality of patients. The traditional prognosis assessment of breast cancer mainly depended on doctors' clinical experience and multidisciplinary comprehensive judgment, which lacks unified objective evaluation criteria. This study proposes an innovative cross-modal contrastive learning model PreGAT based on graph neural networks and attention mechanism. The proposed model can efficiently integrate features from multiple sources of patient data, including clinical features and constructed graph structure features, and significantly improve the performance of the model through a novel contrastive learning loss function. The PreGAT model achieves excellent performance on the public METABRIC dataset with an average accuracy of 92.9% and an AUC value of 0.969. This research provides a promising technique for breast cancer prognosis prediction in clinical practice, which can provide more accurate and reliable decision support for the development of precise treatment programs.
This study proposes a rapid and non-destructive method for ginseng seed classification at the individual seed level using hyperspectral imaging combined with transfer learning. A total of 900 seeds from three ginseng varieties were used, with 300 samples per variety. Spectral features were extracted from circular regions of interest, and classification models including PLS-DA, LSSVM, and random forest were built using preprocessed spectral data, achieving accuracies of 92%, 92%, and 93%, respectively. Principal component images (PC1–PC3) were fused into RGB format and fed into pre-trained deep learning models (AlexNet, GoogLeNet, Inception-V3, ResNet-50, SqueezeNet). Among them, the improved AlexNet model achieved the highest accuracy of 97%. The results demonstrate that transfer learning models integrating spectral and image features outperform traditional spectral models, offering an effective solution for precise classification and quality control of ginseng seeds.
Industrial data have always played a pivotal role in the development, management, and optimization of processes and products. Over time, the intrinsic characteristics of data (modality, acquisition frequency, granularity, uncertainty, etc.) have evolved significantly and so have the computational methods and technology to process and integrate them, transforming these inputs into valuable insights. By integrating multiple heterogeneous data sources, data fusion provides an updated and improved status of the operations, enhancing information quality and leading to more comprehensive, accurate, and insightful analysis of the systems under examination and, therefore, better decision-making. In this work, we present a review of the data fusion methodologies from the perspective of practitioners, including engineers and data analysts, and how it can be systematically organized into classes of methods dedicated to achieving specific goals. The terminology adopted in the field, not always used consistently and without ambiguity, is also discussed and clarified. Additionally, several case studies are provided to showcase some of the applications and potential advantages of adopting data fusion frameworks.
Multi-excitation fluorescence spectroscopy allows characterizing the distribution of fluorophores in samples of various natures and is a powerful tool for chemical, biological, and medical diagnostics. However, interpreting large sets of excitation-dependent fluorescence spectra remains a significant challenge, especially when the fluorophores are initially unknown. In optically thin samples, spectral decomposition can be performed using several statistical methods. In practice, however, various light propagation effects cause deviations from the pure component spectra and prevent the direct application of these methods. A sample-specific inner filtering correction is required prior to their use. Unfortunately, performing such a correction is not always feasible due to complex light propagation geometry or the difficulty of obtaining additional measurements. Here, we suggest an indirect procedure for optically thick samples with negligible fluorescence reabsorption that does not rely on any assumptions about the light illumination and collection geometry, by iteratively determining appropriate correction coefficients for excitation variations. Due to the correlated nature of fluorescence spectra in excitation-emission matrices, this approach enables decomposition when other methods fail. Model experiments with simulated spectral data and real fluorophores showed good agreement between the calculated spectra and the actual fluorescence spectra of the corresponding substances.
An overview of various prospective PLS-DA strategies is presented. PLS-DA is considered not only as a discrimination method per se but also as a method for extracting low-dimensional features suitable for discriminating high-dimensional data, which makes it similar to PCA. Unlike PCA, PLS-DA extracts two sets of features: the matrix of predicted dummy responses (Y) and the matrix of PLS scores (T). Both sets are then used as input to hard (deterministic) and soft (probabilistic) discriminators, providing a wide selection of combinations. Moreover, the features can be enriched by using additional information provided by the PLS model. A total of 18 different strategies are considered, including 8 based on Y-features and 10 based on T-features. Two data sets are used to illustrate the theory, yielding the following results. Both the T- and the Y-based approaches are comparable; additional information is ineffective. The cons and pros of each strategy are discussed.
This paper honors Paul Geladi with a new chemometric data visualization approach based on immersive analytics principles using virtual reality (VR). With the many technological advancements improving the accessibility of extended reality (XR), including VR, the frontier of immersive analytics is wide open for utilizing VR as a powerful chemometric data analysis tool. Immersion of human senses into a virtually generated three-dimensional (3D) world compels us to act more instinctively in complex data decision-making scenarios by using our inherent cognitive pattern recognition capacity learned over years of experience. Proposed is an immersive analytics application of VR for a hybrid human/computer-aided outlier detection process. In this application, each training set sample is virtually realized as a glyph to visually assess the inter- and intra-sample relationships present in datasets while mining for nonrepresentative samples. Sample glyph shape and size are developed using hundreds of sample similarity measurements for each sample based both on spectral and prediction property information. These sample glyphs are visually and spatially compared with one another in VR by the user. Outlier checking in VR safeguards against masking and swamping problems that are difficult to recognize with automatic algorithms. Results from dataset situations based on near infrared (NIR) and ultraviolet (UV) spectra and analyte reference values show the viability of using VR for data analysis and outlier detection. This VR application demonstrates the looming evolution of immersive analytics with an XR interface involving human reasoning in difficult chemometric data analysis settings.
Immersive analytics (IA) leverages virtual reality, augmented reality, and mixed reality to transform how users interact with complex datasets across domains such as science, industry, and education. These immersive technologies offer spatial and multimodal environments that foster intuitive exploration, but they also introduce challenges related to cognitive load, interface design, and system performance. This article presents a comprehensive review of visualization techniques, interaction models, and multimodal inputs utilized in IA. Drawing on case studies in scientific visualization, industrial training, and educational communication, we examine both the potential and limitations of current systems. Finally, we propose future research directions, focusing on real-time collaboration, adaptive user interfaces, and scalable data exploration strategies to advance the field.
In this work, bench-top and portable near-infrared (NIR) spectroscopy equipment are compared in the context of generating predictive models for the quantification of phosphorus (P), potassium (K), and nitrogen (N) components from corn silage samples. For this, 200 spectral samples resulting from bench-top and portable NIR are preprocessed by the following sequence of approaches: mean centering application for removing the spectral bias/offset, Savitzky–Golay filter for highlighting signal energy absorption in relation to spectral noise, interval partial least square (iPLS) for selection of spectral region, and Monte Carlo method for outlier detection and removal. Then, from the preprocessed spectra, predictive models were built using the partial least squares (PLS) and support vector regression (SVR) methods for each chemical component and NIR equipment. In this sense, six models are generated, three for each NIR spectroscopy (or two for each element). As a result, considering all components and machine learning (ML) methods, bench-top models achieved R2 values between 0.66 (quantification of P using PLS or SVR) and 0.81 (prediction of K and N using SVR regressors) during the validation, whereas portable ones achieved values between 0.50 (prediction of K using SVR) and 0.67 (quantification of N using PLS). Our results can be considered competitive, as robust and accurate predictors are also generated.
Chemometrics encompasses numerous facets such as experimental design, data collection and analysis, and many others. This paper, in honor of Paul Geladi, provides our perspective on growing the scientific intuition of multivariate analysis in conventional chemometric directions not generally practicing multivariate principles. The motivation for this perspective is to express our opinion on the need for chemometrics to expand the role of the Rashomon effect beyond “many models predict well” by integrating a more comprehensive consideration of the multivariate nature of matrix effects. Described are multiple chemometric techniques that have already been enhanced by broadening the application the Rashomon effect including model selection and explanation, figures of merit (FOM), sample similarity assessment for model reliability, outlier detection, and classification—all recent research topics from the authors. This expository discussion revolves around spectroscopic data such as near infrared and fluorescence, but the concepts are relevant to other chemometric data structures.
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：