MAGED: Multimodal attentive graph learning with gene expression dynamics on knowledge graphs for TCM target prediction.

Missing value imputation in multivariate time series is a critical challenge in the field of data mining. Although Transformer-based methods excel in modeling long-range dependencies, their inherent point-wise attention mechanisms often lack explicit modeling of local inductive biases in time series, making it difficult to effectively capture local smoothness and evolutionary trends. Furthermore, existing feature embedding strategies struggle to fully decouple the internal temporal evolution of variables from complex cross-variable dependencies. To address these limitations, this paper proposes a novel dual-stage imputation framework named M-SAITS. This framework innovatively introduces a decoupled feature encoder based on large-kernel depthwise convolutions. By utilizing an extended effective receptive field, it explicitly enhances the model's perception of local trends. Additionally, it employs a grouped convolution structure to achieve decoupled modeling of intra-variable temporal patterns and inter-variable interaction features. On this basis, combined with a Diagonally-Masked Self-Attention mechanism, the framework physically blocks information leakage paths while achieving lossless global context aggregation. Relying on a "Preliminary Inference–Iterative Refinement" cascade strategy and a masked weighted joint optimization objective, the model achieves high-fidelity data reconstruction. Extensive experiments on multiple benchmark datasets, such as Electricity and Air Quality, demonstrate that this method significantly outperforms existing state-of-the-art models across multiple evaluation metrics. Notably, in high-dimensional electricity data imputation tasks, M-SAITS achieves substantial performance improvements over baseline models such as CSDI and Transformer, with the Mean Absolute Error significantly reduced (up to approximately 60% under low missing rates).

Self-assembled monolayers (SAMs) have emerged as a new generation of hole transport materials (HTMs) for perovskite solar cells (PSCs), particularly in inverted architectures. Compared to conventional HTMs, SAMs demonstrate superior power conversion efficiency (PCE) and enhanced operational stability. However, the current discovery of SAMs still relies heavily on empirical trial-and-error approaches, suffering from long development cycles, high costs, and low success rates. Here we present a novel machine learning (ML) platform for accelerated SAM discovery and design. We constructed a comprehensive feature space combining RDKit molecular descriptors and Morgan fingerprints, and then systematically evaluated various ML algorithms. Multiple evaluation metrics were used to assess model reliability. The results demonstrate that the RDKit-based XGBoost model achieved optimal performance with a root-mean-square error (RMSE) of 1.862, a coefficient of determination (R2) of 0.5058, a Pearson correlation coefficient (r) of 0.8161 and a mean absolute error (MAE) of 1.528. Then, SHapley Additive exPlanations (SHAP) analysis further elucidated the structure-property relationships between key molecular features and photovoltaic performance. The SHAP values revealed that the top five most important features were all RDKit descriptors, specifically EState_VSA5, fr_benzene, EState_VSA2, SlogP_VSA1, and Chi0v. The external validation using recently reported SAM molecules demonstrated remarkable prediction accuracy. The relative errors between predicted and experimental PCE values were mostly within 10%, with the minimum being only 0.55%. Meanwhile, three new SAM molecules were designed based on the model, with the highest predicted PCE approaching 27%. Therefore, this work provides an efficient digital solution for SAM development, offering valuable guidance for accelerating the discovery of next-generation photovoltaic materials.

UTMorph: A hybrid CNN-transformer network for weakly-supervised multimodal image registration in biopsy puncture.

To improve the identification of minor-effect molecular markers and genes associated with quantitative traits, addressing inefficiencies in traditional molecular marker mining and the limited impact of these markers in practical breeding, we analyzed over 15,000 soybean (Glycine max) genotypes across nine maturity groups using an AI-driven genomic approach, uncovering 2513 key genetic markers and 393 genes associated with yield. These included stable determinants unaffected by environmental variability and others with environment-dependent effects. The light gradient boosting machine (LightGBM) model was applied, and model performance was assessed using multiple evaluation metrics across the total datasets by fivefold cross-validation, yielding a mean squared error of 0.0862 and an R2 of 0.8369, which demonstrates a strong alignment between predicted and observed yield values. To further validate prediction accuracy, a t-test comparing predicted and actual yields produced a T-statistic of 0.5236 with a p-value of 0.60054, indicating no statistically significant difference between the two distributions. These results confirm the reliability and robustness of the LightGBM approach for yield prediction and underscore its potential utility in breeding programs focused on enhancing agricultural productivity. Our maturity group-based analysis revealed region-specific marker distributions, providing refined targets for precision breeding. Three key principles for marker-assisted selection (MAS) in quantitative trait breeding were introduced: marker complementarity, regional specificity, and quantification. Breeding programs should employ comprehensive sets of region-specific markers, while the effectiveness of MAS can be measured based on the number and SHapley Additive exPlanations value of markers utilized. Our approach to assessing marker sensitivity across environments offers valuable perspectives on genotype-by-environment interactions.

ESMO basic requirements for AI-based biomarkers in oncology (EBAI).

Abstract As a task at the intersection of computer vision and natural language processing, image captioning offers significant application value in domains such as intelligent human–computer interaction, accessibility support and multimedia content retrieval. The primary objective is to generate natural language descriptions by interpreting visual features, traditionally relying on heterogeneous single‐stream grid features and region features. However, existing approaches face limitations: grid features struggle to balance global semantic perception with local detail analysis, and region features exhibit weakened spatial modelling efficacy due to sparse semantic correlations. Furthermore, fusing heterogeneous visual features often lacks effective control over complementarity and redundancy, leading to descriptions prone to semantic bias or detail omission. To address these challenges, we propose a novel Multi‐Gated Dual‐Stream Visual Feature Fusion (MGDSF) for Image Captioning. Our approach enhances the semantic accuracy and completeness of generated captions through dual‐stream feature extraction and a multi‐gated fusion (MGF) mechanism. First, we employ a Mamba‐like linear attention mechanism to construct a grid feature network with hierarchical positional awareness. This network achieves global modelling while maintaining local sensitivity by dynamically modulating information flow. Second, based on the Detection Transformer (DETR) framework, we design a region feature extractor to provide complementary local object visual information. Finally, we introduce a MGF module that balances the complementarity of dual‐stream visual features and suppresses cross‐modal information redundancy via multiple context‐aware gates, thereby achieving fine‐grained visual‐semantic alignment. Experiments on MS COCO demonstrate that MGDSF surpasses existing methods on multiple evaluation metrics, achieving METEOR, ROUGE‐L and CIDEr scores of 30.0%, 59.8% and 140.1%, respectively. These results validate the effectiveness of our proposed method and indicate its broad application potential.

Intrusion Detection Systems (IDS) play a vital role in safeguarding modern networks against increasingly complex cyberattacks. However, the presence of high-dimensional, redundant, and noisy features in network traffic data often degrades detection accuracy and scalability. To address this challenge, this paper proposes GSFOIDS (Group Search Based Feature Optimization for Intrusion Detection System), a hybrid framework that integrates Group Search Optimization (GSO) with a Convolutional Neural Network (CNN) for efficient and accurate intrusion detection. Initially, the intrusion dataset is cleaned and preprocessed to remove irrelevant attributes and improve data quality. GSO is then employed to identify an optimal subset of discriminative features using cooperative producer, scrounger, and ranger search strategies. A CNN-based fitness function guides the optimization process by assessing classification accuracy. Experimental results demonstrate that GSFOIDS significantly outperforms existing models across multiple evaluation metrics

Traditional models for natural language-to-SQL translation in Electronic Medical Record (EMR) systems struggle with understanding medical terminology, handling complex queries, and bridging the syntax-semantics gap, leading to scalability and accuracy issues. Advanced solutions like Large Language Model (LLM) based approaches address these challenges by leveraging deep learning and domain-specific training to enhance performance and usability. Hence, this article introduces an advanced medical Text-to-Structured Query Language (SQL) paradigm that simplifies accessing EMRs by translating natural language queries into SQL commands. This model is built on the advanced Code-T5 (Text-to-Text Transfer Transformer) architecture, further enhanced with Low-Rank Adaptation (LoRA) and Quantized Low-Rank Adaptation (QLoRA) techniques; it effectively addresses the challenges posed by the complexity of traditional SQL queries enabling seamless access to critical healthcare data. The innovation of the proposed model lies in its exceptional performance across multiple evaluation metrics. It achieves a Bilingual Evaluation Understudy (BLEU) score of 81.68, significantly outperforming leading models like T5, Fine-Tuned Language Net (FLAN) T5, and Bidirectional and Auto-Regressive Transformers (BART) while excelling in Recall-Oriented Understudy for Gisting Evaluation (ROUGE) metrics, underscoring its proficiency in generating semantically accurate and coherent SQL queries. Furthermore, the proposed model attains a high token-level F1-score, ensuring a balanced precision and recall and a Jaccard similarity score of 0.83, surpassing T5, Flan T5, and BART. The proposed model excels in handling complex medical queries, bridging natural language and SQL to empower data-driven decisions and advance medical informatics.

Turbine blades serve as critical components in aircraft engines, yet casting defects inevitably arise during manufacturing. Therefore, accurate pre-service turbine blade defect detection is critical for aircraft engine safety. However, existing deep learning-based detection methods face several challenges: poor image quality, intraclass variance, interclass similarity, and irregular defect geometries. Moreover, most existing defect detection methods rely primarily on spatial-domain features, which are insufficient for capturing fine-grained texture information, limiting their ability to discriminate complex defect patterns. To address these challenges, we propose a novel Spatial-Frequency Complementary Fusion Network (SFCF-Net) that synergistically integrates spatial and frequency-domain features through complementary cross-modal fusion for accurate defect segmentation. First, a Selective Cross-modal Calibration (SCC) module is introduced that selectively calibrates spatial-frequency features through gated cross-modal interactions, effectively preserving fine-grained details under poor image conditions. Next, we propose a Cross-modal Refinement and Complementation (CRC) module that employs dual-stage attention mechanisms to model intra- and inter-modal feature dependencies, enabling robust discrimination between similar defect categories while maintaining consistency within the same defect class. Finally, we propose an Asymmetric Window Attention (AWA) module that employs bidirectional rectangular windows for accurate defect geometric characterization. Comprehensive experiments on the Aero-engine Turbine Blade Casting Defect Segmentation (ATBCD-Seg) dataset and a public benchmark demonstrate that SFCF-Net consistently outperforms state-of-the-art methods across multiple evaluation metrics, meeting practical requirements for automated quality control in blade manufacturing.

To address the limitations of existing annual load scenario generation methods, including insufficient ability to represent long-term trends, excessive randomness in generated scenarios, and inadequate consideration of special holiday conditions, in this paper, an annual load curve generation method is proposed that integrates Seasonal–Trend decomposition using Loess (STL) with an improved denoising diffusion probabilistic model (DDPM). In the proposed method, the STL algorithm is first applied to decompose the annual load curve into a trend component and a daily seasonal component. The trend component is used as a baseline to ensure that the generated load curves remain consistent with the actual long-term trend characteristics. On this basis, an improved diffusion-based denoising model is employed to achieve controllable generation of different types of daily load scenarios. Finally, the generated daily load scenarios are aggregated with the trend component on an hourly basis to construct annual load scenario curves that simultaneously preserve realistic trend behavior and stochastic fluctuations. A case study based on a city in China is used to evaluate the proposed method. The results demonstrate that both the generated daily load scenarios and annual load scenarios outperform existing benchmark methods across multiple quantitative evaluation metrics, thereby validating the effectiveness of the proposed load scenario generation approach.

To address the current challenges of deep learning music generation models in capturing long-range dependencies, ensuring generation diversity, and maintaining training stability, this study proposes an optimized music melody generation model—T-WGAN. The model deeply integrates Transformer and Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP). This study first preprocesses the large-scale Lakh MIDI dataset, extracts single-track main melodies, and converts them into symbolic sequences using a REMI-based event representation. On this basis, the model innovatively adopts a generator based on Transformer decoder to learn the long-range structure of melodies. It also uses a critic based on Transformer encoder for stable adversarial training under the WGAN-GP framework to enhance the diversity and authenticity of generated melodies. Experimental results show that T-WGAN performs excellently on multiple key evaluation metrics. T-WGAN achieves a Rhythmic Consistency Rate (RCR) of 85.17%, significantly higher than baseline models (e.g., Transformer’s 75.68%). Its score on Fréchet Distance for Music (FDM) drops to 31.02, proving that the generated melodies are closer to real music in feature distribution. The conclusion indicates that the proposed T-WGAN model successfully addresses the three core issues in melody generation—structural integrity, diversity, and training stability—synergistically. The findings provide an effective technical approach for generating high-quality music melodies with both structural logic and innovation.

In aerospace applications, critical metal additive manufacturing (AM) structures require rigorous damage tolerance certification, depending on reliable crack growth analysis. As the bridge from specimen-level experiments to structural assessments, the construction of crack growth models traditionally relied on extensive databases and engineering expertise. Yet the complexity of AM processes hinders the establishment of stable, comprehensive data sets, limiting conventional approaches. To address this challenge, this study proposes a Parametric Symbolic Regression framework for learning crack growth models of AM structures (PSR-AM), in which the symbolic regression is constructed upon parametric candidate models, enabling a unified representation of underlying physical mechanisms while capturing variations across experimental conditions through parameter shifts. Moreover, models’ parameter fitting incorporates both physical and global trend constraints, ensuring reliable and physically consistent modeling even under sparse data conditions. Guided by multiple evaluation metrics, PSR-AM effectively leverages diverse and even sparse AM experimental data sets to derive concise, robust, and generalizable crack growth models. Test on 178 AM Ti-6Al-4V data sets demonstrates that the PSR-AM learned crack growth model captures highly scattered experimental results under multifactor influences with concise mathematical forms and a small set of stable parameters, while outperforming the conventional empirical model in reflecting the intrinsic data characteristics.

Abstract Background: Care for HER is a campaign that serves Black breast cancer patients nationwide with free access to: 1) integrative care therapies and services, and 2) 24/7 culturally tailored patient navigation by Black nurses and social workers who are also breast cancer survivors. This study aims to evaluate program outcomes and satisfaction, as reported by program participants. Method: After participating in the Care for HER program, participants provided feedback online for multiple program evaluation metrics using a combination of scaled response questions: satisfaction and resource use (5-point scales), perceived benefits (4-point agreement scale and yes/no items), pre- and post-program distress (0-10 scale), and likelihood of recommending to others (0-10 scale). Descriptive and bivariate results are presented. Results: Between April and May 2025, 57 Black women completed the survey, with an average age of 52 (SD=9.1; range 33-71) and average time since diagnosis of 14 months (SD=11.6; range 6-71 months). Results show 98% report being satisfied or very satisfied with the wellness passport program of integrative care therapies and services, with the average likelihood of recommending to others a 9.4 out of 10; 93% of program participants report using resources provided (47% often or very often and 46% sometimes), and 7% rarely or never. As a result of the program, participants reported being better able to find support to reduce treatment side effects (75%) and reducing at least one aspect of financial strain (83%). Nearly all participants reported improvements in their health behaviors, with 95% saying they felt more motivated to make healthier food choices and 93% indicating a better understanding of the benefits of physical activity. Many also experienced gains in navigating the healthcare system and advocating for themselves: 89% reported a better understanding of the resources and services available to them, 88% felt better equipped to follow their treatment plan, and 84% felt more confident advocating for themselves in healthcare settings. Additionally, 69% of participants agreed or strongly agreed that they were better able to understand their diagnosis and treatment options. Average distress score dropped from 6.6 before the program to 2.7 after, reflecting a 3.9-point decrease. Overall, 85% of participants reported reduced distress, 11% saw no change, and 4% reported an increase. High distress (scores of 8-10) declined from 39% to 0%, while those reporting no distress (score of 0) rose from 5% to 21%. Frequency of resource use was significantly associated with reduction in distress (F=4.252, p<.02), such that those who reported using available resources often or very often had the largest reduction in distress after program participation (M=-3.8), compared to those who only use resources sometimes (M=-3.0). Additionally, more frequent use was significantly correlated with satisfaction (r=.43, p<.001). Conclusions: Participation in the Care for HER program was associated with improvements in health behaviors, managing side effects, healthcare navigation, and understanding of diagnosis and treatment, as well as a notable reduction in distress, suggesting the program effectively supports practical and emotional well-being for women with breast cancer. Further, program satisfaction and perceived benefits were significantly correlated with frequency of resource use, indicating that greater engagement enhances program outcomes. Programs like Care for HER provide essential supportive services and empower patients to take a more active role in their cancer care. Expanding access is crucial, as patients—especially those from historically underserved communities—cannot obtain these services through traditional healthcare systems. Citation Format: S. Weldon, V. Worthy, R. Fairley, E. Powers, G. Kelly, J. Meakim, E. Fortune. A new comprehensive integrative care and navigation model for enhancing outcomes for Black breast cancer patients: evidence from the Care for HER program [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS1-04-20.

Traditional non-negative matrix factorization (NMF) faces challenges when dealing with complex data, primarily characterized by sensitivity to noise, neglect of data geometric structure, and inability to effectively utilize supervised information. To address these limitations, this paper proposes a class-driven robust non-negative matrix factorization with dual-hypergraph regularization (CRNMFDH) framework. The core contributions of this framework include the following: Firstly, the design of a novel dual-hypergraph regularization term that symmetrically captures and preserves the higher-order geometric structures of both the sample space and feature space, establishing a mutually reinforcing topological relationship between them. Secondly, an introduction of a class-driven mechanism to effectively integrate label information into the decomposition process, significantly enhancing the discriminative capability of the low-dimensional representations. Finally, the employment of a loss function based on correntropy to replace the traditional Euclidean distance, thereby enhancing the model’s robustness against noise and outliers. Extensive experiments across nine datasets demonstrate that CRNMFDH significantly outperforms existing state-of-the-art algorithms in multiple clustering evaluation metrics and noise robustness, providing an effective new solution for complex data clustering tasks.

Semantic segmentation of metallographic micrographs is a key task for quantitative microstructural analysis in additive manufacturing, yet it remains challenging due to phase heterogeneity, complex morphologies, and the scarcity of annotated data. The MetalDAM dataset, composed of 42 labeled scanning electron microscopy images of steel microstructures, has been widely adopted as a benchmark, with U-Net commonly reported as the strongest supervised baseline. Nevertheless, the encoder–decoder structure of U-Net imposes architectural constraints that hinder the precise delineation of heterogeneous and irregular phase boundaries under severe data limitations. To address this limitation, this paper investigates a Fully Convolutional Network (FCN)-based architecture as an alternative approach for semantic segmentation on the MetalDAM dataset. The FCN is trained and evaluated under the same experimental protocol as the U-Net baseline, enabling a direct and fair comparison. Performance is assessed using multiple evaluation metrics, including Intersection over Union (IoU), precision, recall, and mean Average Precision at an IoU threshold of 0.5. The results show that the FCN achieves comparable overall IoU values (0.75) while delivering substantial improvements at the class level, particularly for minority and morphologically complex phases, with gains of up to 25–30% in class-specific IoU. Additional metrics confirm enhanced robustness, with consistently higher precision, recall, and mAP@0.5 values. These findings demonstrate that FCN-based architectures constitute a competitive and robust alternative to U-Net for metallographic segmentation in additive manufacturing scenarios characterized by limited annotated data.

Studying the three-dimensional (3D) structure of a genome, including chromatin loops and Topologically Associating Domains (TADs), is essential for understanding how the genome is organized, such as gene activation, cell development, protein-protein interaction, etc. Hi-C protocol enables us to study 3D genome structure and organization. Chromatin 3D structure changes dynamically over time, and modeling these continuous changes is crucial for downstream analysis in various domains such as disease diagnosis, vaccine development, etc. The high expense and impracticality of continuous genome sequencing, particularly what evolves between two timestamps, limit the most effective genomic analysis. It is crucial to develop a straightforward and cost-efficient method for constantly generating genomic data between two timestamps in order to address these constraints. In this study, we developed HiCInterpolate, a 4D spatiotemporal interpolation architecture that accepts two timestamp Hi-C contact matrices to interpolate intermediate Hi-C contact matrices at high resolution. HiCInterpolate predicts the intermediate Hi-C contact map using a deep learning-based flow predictor, and a feature encoder and decoder architecture similar to U-Net. In addition, HiCInterpolate supports downstream analysis of multiple 3D genomic features, including A/B compartments, chromatin loops, TADs, and 3D genome structure, through an integrated analysis pipeline. Across multiple evaluation metrics, including PSNR, SSIM, GenomeDISCO, HiCRep, and LPIPS, HiCInterpolate achieved consistently strong performance. Biological validation further demonstrated preservation of key chromatin organization features, such as chromatin loops, A/B compartments, and TADs. Together, these results indicate that HiCInterpolate provides a robust computer vision-based framework for high-resolution interpolation of intermediate Hi-C contact matrices and facilitates biologically meaningful downstream analyses. HiCInterpolate is publicly available at https://github.com/OluwadareLab/HiCInterpolate .

As spatially resolved transcriptomics (SRT) datasets increasingly span multiple adjacent or replicated slices, effective joint analysis across slices is needed to reconstruct tissue structures and identify consistent spatial gene expression patterns. This requires resolving spatial correspondences between slices while capturing shared transcriptomic features, two tasks that are typically addressed in isolation. Multi-slice analysis remains challenging due to physical distortions, technical variability, and batch effects. To address these challenges, we introduce Joint Alignment and Deep Embedding for multi-slice SRT (JADE), a unified computational framework that simultaneously learns spatial location-wise alignments and shared low-dimensional embeddings across tissue slices. Unlike existing methods, JADE adopts a roundtrip framework in which each iteration alternates between alignment and embedding refinement. To infer alignment, we employ attention mechanisms that dynamically assess and weight the importance of different embedding dimensions, allowing the model to focus on the most alignment-relevant features while suppressing noise. To the best of our knowledge, JADE is the first method that jointly optimizes alignment and representation learning in a shared latent space, enabling robust multi-slice integration. We demonstrate that JADE outperforms existing alignment and embedding methods across multiple evaluation metrics in the 10x Visium human dorsolateral prefrontal cortex (DLPFC) and Stereo-seq axolotl brain datasets. By bridging spatial alignment and feature integration, JADE provides a scalable and accurate solution for cross-slice analysis of SRT data.

In drug discovery and therapeutic research, the prediction of drug-disease associations (DDAs) holds significant scientific and clinical value. Drug molecules exert their effects by precisely identifying disease-related biological targets, systematically modulating the entire pharmacological process from absorption, distribution, and metabolism to final efficacy. Accurate prediction of drug-disease associations not only facilitates an in-depth understanding of molecular mechanisms of drug action but also provides critical theoretical foundations for drug repositioning and personalized medicine. While traditional prediction methods based on in vitro experiments and clinical statistics yield reliable results, they suffer from inherent drawbacks such as long development cycles, substantial resource consumption, and low throughput. In contrast, emerging machine learning techniques offer a promising solution to these bottlenecks, enabling the intelligent and efficient discovery of potential drug-disease association networks and significantly improving drug development efficiency. However, it is noteworthy that existing machine learning methods still face significant challenges in practical applications: the complexity of feature construction raises the threshold for data processing; data sparsity constrains the depth of information mining; and the pervasive issue of sample imbalance poses a severe challenge to the model's predictive accuracy and generalization performance. In this study, we developed an efficient and accurate framework for drug-disease association prediction named FKSUDDAPre. The model employs a multi-modal feature fusion strategy: on one hand, it leverages an ensemble of Mol2vec and K- BERT to deeply capture the semantic features of drug molecular fingerprints; on the other hand, it integrates Medical Subject Headings (MeSH) with DeepWalk to effectively reduce the dimensionality of disease features while preserving their relational structure. To address the class imbalance problem, FKSUDDAPre designed an optimization algorithm called AMDKSU, which combined clustering with an improved distance metric strategy, significantly enhancing the discriminative power of the sample set. For data processing, F-test was employed for feature importance ranking, effectively reducing data dimensionality and improving model generalization. For the predictive architecture, FKSUDDAPre proposed a novel ensemble framework composed of XGBoost, Decision Tree, Random Forest, and HyperFast. By employing a dynamic weight allocation strategy, this ensemble effectively harnesses the complementary strengths of these models to achieve significantly enhanced predictive performance. Rigorous validation demonstrated the system's outstanding performance across multiple evaluation metrics, with an average AUC of 0.9725, improving the AUC by approximately 3.88% compared to the best-performing baseline model. In the prediction of Alzheimer's disease and Parkinson's disease, 80% and 60% of the top 10 candidate drugs recommended by FKSUDDAPre, respectively, had been confirmed by literature, demonstrating the model's good practical application potential. Furthermore, we conducted a LIME-based feature importance analysis on the model's predictions, visualizing the correlations between features and the target variable to demonstrate the model's interpretability. A cross-platform, user-friendly visualization tool had also been developed using the PyQt5 framework.

With the rapid increase of automation and intelligent assistance in modern traffic operations, human–machine interaction (HMI) has become central to system effectiveness and operational safety. In such high-stakes environments, operators often need to manage complex information flows and dynamic decision-making processes, making situation awareness (SA) a critical determinant of performance. However, the growing reliance on automated systems raises the risk of SA degradation, potentially undermining safety and efficiency. Eye-tracking has emerged as a promising tool to monitor operator SA, but in real-world traffic control environments, data incompleteness and signal loss remain major obstacles. To address this challenge, we introduce a robust eye-tracking-enabled SA recognition framework, the Masked AutoEncoder for EYE-tracking data (MAEYE). MAEYE integrated CNN modules and Transformer layers to effectively capture both structural patterns and temporal dynamics of eye movements. Leveraging a self-supervised learning paradigm, it demonstrates strong resilience against incomplete data, outperforming state-of-the-art methods under varying levels of data loss. An SA-probe experiment with 26 participants validated its effectiveness, confirming reliable and accurate SA decoding from imperfect gaze input. By enabling dependable SA monitoring in traffic operations HMI, this work advances the development of safer, more resilient, and human-centred automation systems for future mobility and transportation management. Highlights A robust eye-tracking–based approach is proposed for SA recognition in traffic operations. Self-supervised learning enhances representation robustness under noisy or incomplete eye-tracking inputs. Adaptive masking strategies in autoencoder training are systematically evaluated for robust feature learning. The method consistently outperforms state-of-the-art baselines across multiple evaluation metrics. Reliable SA recognition supports adaptive HMI for safer traffic management.