Chemotherapy-induced cardiotoxicity presents a major risk to cancer patients, often leading to severe cardiac complications such as heart failure, myocardial infarction, and arrhythmias. Early detection is crucial for preventing long-term damage and improving patient outcomes, yet existing diagnostic methods struggle to effectively capture the complexity of multimodal medical data and often lack interpretability. In this study, we propose an innovative approach that integrates multimodal deep learning with Explainable AI (XAI) techniques to enhance early cardiotoxicity detection. Our model combines clinical data (e.g., age and cardiovascular metrics) with Tissue Doppler Imaging (TDI), a functional imaging technique that captures myocardial velocity during the cardiac cycle. To overcome data limitations, we employed Conditional Generative Adversarial Networks (cGANs) and Conditional Tabular Generative Adversarial Networks (CTGANs) to augment the dataset, improving its diversity and balance for better model training. We developed three architectures that integrate Convolutional Neural Networks (CNNs) for feature extraction from TDI images with Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer models to capture temporal dependencies and enhance prediction accuracy. Additionally, we incorporated SHapley Additive Explanations (SHAP) to interpret the contribution of input features, increasing model transparency and clinical applicability. Our Transformer-based model achieved the highest accuracy of 96%, outperforming the GRU (94%) and LSTM (89%) models, significantly surpassing traditional approaches. These findings highlight the potential of transformer-based architectures in multimodal deep learning for precise cardiotoxicity prediction, supporting early intervention and personalized treatment strategies while improving interpretability through XAI techniques such as SHAP.
The main analyses in Social Network Analysis field are focusing on measuring the frequency of social interactions that are modeled only with numbers; 0 or 1 on the graph adjacency matrix. Therefore, ontologies have drawn attention to empower and enhance the Social Network Analysis with semantics. By exploring theses semantics, besides we better view and analyze the interactions between students and teachers inside the Learning Management System, the network graph structure and analyses change. In this paper, we propose a methodology to build Moodle’s OWL (Web Ontology Language) ontology for Social Network Analysis ends. The proposed ontology gives explicit meaning to the relationships inside Moodle instead of modeling them as nodes and edges with implicit semantics like the case in classical Social Network Analysis. The semantics of relationships influences the graph structure and statistics by making some nodes more central and influential and other nodes less central and influential. The built ontology is basically made by exploring the UML (Unified Modeling Language) class diagram that we created from Moodle Mount Orange School and is our main metadata source.
Structural Health Monitoring (SHM) is crucial for maintaining the sustainability and safety of civil infrastructure. The Z24 Bridge in Switzerland remains one of the benchmark datasets used to validate vibration-based damage detection methods. Traditional approaches based exclusively on modal parameters are frequently limited by data scarcity and environmental variability. Recent advances in artificial intelligence have enabled data-driven neural networks to learn discriminative features directly from raw measurements. Meanwhile, hybrid methods such as Physics-Informed Neural Networks (PINNs) incorporate governing physical laws into the learning process. This study presents a comparative analysis of three successive artificial Neural Network models (NN V1–V3) and One Physics-Informed Neural Network (PINN V1), all applied to the Z24 Bridge dataset. The NN models progressively improve in depth, optimization strategy, and regularization, achieving ≈97.7% validation accuracy and a macro AUC ≈1.00 with NN V3. However, they remain completely dependent on the quality and quantity of training data. In contrast, the PINN incorporates the differential equation of a damped oscillator into its loss function, balancing a data-driven term with a physics-based residual. This approach enables more stable learning with limited labeled data and ensures consistency with structural dynamics. Experimental results highlight the trade-off between accuracy and robustness: while NN V3 yields the highest predictive performance (≈97.7% validation accuracy, macro AUC ≈1.00),PINN V1 achieves slightly lower accuracy (≈92%) but offers improved stability and interpretability. This dual perspective demonstrates that hybrid physics-informed models provide a more reliable basis for decision-making in SHM. The findings underscore the potentialof combining machine learning with physical knowledge, paving the way for future developments such as hybrid PINNs (HPINNs), multi-sensor integration, and high-performance computing deployment.
3 March 2026 in International Journal of Engineering Pedagogy (iJEP)
Assessing competencies in engineering education increasingly requires digital assessment approaches that support learning regulation, instructional decision-making, and educational quality, rather than focusing solely on measurement efficiency. Computerized adaptive testing (CAT), grounded in Item Response Theory (IRT), provides a robust methodological foundation for personalized assessment. However, its pedagogical effectiveness in formative contexts depends critically on curriculum alignment, diagnostic capacity, and adaptive control strategies. This study proposes and evaluates a formative adaptive assessment framework for engineering education that integrates an IRT-based CAT engine with a Bayesian network– based diagnostic component. The framework is designed to support competency-oriented feedback, learning monitoring, and instructional interpretation within a curriculum-aligned assessment structure. Assessment relies on dichotomous multiple-choice items explicitly aligned with engineering learning outcomes, while item selection dynamically adapts to learners’ evolving proficiency estimates. In parallel, probabilistic diagnostic modelling prioritizes under-assessed competencies throughout the adaptive process. Item calibration was conducted using empirical data collected from 612 university students in computer science, and system performance was examined through a simulation-based evaluation involving 500 simulated learners. Results demonstrate high estimation accuracy (r = 0.912) and satisfactory reliability for formative use across most learner profiles. Reduced precision at the extremes of the proficiency continuum and imbalances in item exposure were also observed, highlighting structural limitations primarily related to item bank coverage and curriculum representation rather than to the adaptive algorithms themselves. Overall, the proposed framework positions adaptive assessment as a pedagogically grounded tool for formative learning support, instructional decision-making, and quality assurance in engineering education.
11 August 2026 in Engineering, Technology & Applied Science Research
Computerized Adaptive Testing (CAT) is a key component of large-scale digital assessment systems, where high measurement accuracy must be achieved under constraints of test length, computational cost, and sustainable item bank usage. Although classical 3PL-IRT Item Selection Rules (ISRs) are efficient and interpretable, their effectiveness varies across test stages and ability regions, especially under short-test stopping conditions. This paper presents ADAPT-CAT, an interpretable phase-aware hybrid control architecture that models CAT as a state-dependent adaptive control process rather than a static one-step optimization rule. The framework coordinates complementary psychometric ISRs across three phases: early stabilization, conditional adaptation, and final convergence. A machine-learning selector based on gradient-boosted trees (XGB-ISR) is implemented exclusively as an offline benchmarking policy. Large-scale Monte Carlo simulations using a real 50-item 3PL-calibrated bank show that Fisher Information (FI) yields the best overall accuracy (RMSE = 0.649) with an average length of 16 items, while XGB-ISR achieves competitive precision but increases test length (18 items) and concentrates exposure. ADAPT-CAT reduces mean test length to 13 items while maintaining monotonic convergence and more balanced exposure, at the cost of lower global accuracy (RMSE = 0.940). Tail precision loss is primarily driven by structural item bank information deficits rather than the selection mechanism. These results support interpretable hybrid control architectures as a practical alternative for scalable and resource-aware CAT deployment.