Introduction
Breast cancer (BC) is among the most common malignancies worldwide, with metastasis being the main cause of cancer-related death. Accurate prediction of distant metastasis risk is crucial for improving patient outcomes. Based on the literature, we selected a panel of actin cytoskeleton-related genes as potential biomarkers. This study aimed to develop machine learning models predicting distant metastases in primary BC using their expression profiles.

Methods
We analyzed TCGA Breast Invasive Carcinoma data (Firehose Legacy) via cBioPortal. The following genes were studied: TAGLN, ANXA6, ANXA2, ECM1, PFN1, MYH9, CFL1, EZR, ACTN4, GSN, and FSCN1. Expression was normalized as log2(TPM+1). Metastasis status was defined by AJCC staging (M0: absent; M1: present). Due to class imbalance (902 M0 vs. 22 M1), SMOTE was applied. Data were split 80/20 into training/test sets with per-gene standardization. We tested k-NN, logistic regression, decision trees, random forest, and XGBoost. Hyperparameters were tuned with Grid Search and Optuna (10-fold CV). Models were evaluated by Accuracy, F1 Score, Precision, Sensitivity, and Specificity. Gene importance was assessed.

Results
All models except logistic regression (Accuracy = 0.65) performed well (Accuracy ≥ 0.89). Random forest achieved the best results (Accuracy = 0.983, Kappa = 0.967), followed by XGBoost (Accuracy = 0.961, Kappa = 0.922). CFL1, ANXA2, and MYH9 were the top predictors. TAGLN, FSCN1, and ECM1 showed higher importance in XGBoost, aligning with prior cell line studies.

Conclusions
Expression of cytoskeleton-related genes demonstrates strong potential for predicting distant metastasis in BC and warrants further validation in other cancer types.

The growing interconnectivity of digital systems has led to the massive collection and centralization of sensitive data, raising serious concerns about confidentiality and compliance with privacy regulations. Biometric authentication systems, such as offline signature verification, are particularly vulnerable. Federated Learning (FL) provides a promising framework by enabling model training without exposing raw client data. However, data scarcity remains a significant barrier to building robust Deep Learning (DL) models in such settings. This work investigates privacy-preserving Writer-Dependent (WD) Offline Signature Verification (OSV) within an FL framework. To address limited biometric datasets, we explore complementary techniques such as data augmentation, transfer learning, knowledge distillation, and meta-learning. These methods are integrated into federated training pipelines to enhance model generalization while preserving data confidentiality. Preliminary experiments suggest that combining FL with data scarcity mitigation strategies improves the robustness of signature verification systems. Augmentation and transfer learning, in particular, reduce overfitting and enhance classification performance, while knowledge distillation enables the use of lighter yet accurate models suitable for distributed environments. FL offers a viable pathway for secure and effective biometric authentication by keeping sensitive data local. Incorporating advanced data efficiency strategies further strengthens the reliability of offline signature verification systems. The proposed WD-OSV system was trained and evaluated on the popular CEDAR signature dataset, from which an average Area Under Curve (AUC) of 88.93%, along with an average binary accuracy (ACC) of 80.12% were reported as preliminary results. Beyond biometrics, the findings are extendable to healthcare applications, where privacy and data scarcity pose parallel challenges.

Nervousness is a key emotional factor affecting student performance in high-stakes evaluations. Excessive anxiety before or during exams can impair concentration, reduce problem-solving efficiency, and compromise outcomes. Measuring students’ nervousness is therefore essential for fair and effective assessment. This work proposes a neural network-based framework to estimate and monitor nervousness in pre-exam scenarios. A convolutional neural network (CNN), trained for facial emotion recognition, analyzed real-time video streams, classifying seven emotions: anger, disgust, fear, happiness, neutrality, sadness, and surprise. Each emotion was translated into quantitative nervousness scores using a weighted scoring model, allowing continuous tracking of emotional tension during question answering.

The CNN achieved around 70% accuracy in training, ensuring reliable emotion detection and nervousness estimation. Nervousness scores ranged from 0 to 100, derived from weighted associations of emotions such as fear, anger, and sadness, based on psychology and education studies highlighting their impact on learning and test performance. The system produces per-question and overall assessments, culminating in a readiness report indicating if a student is in an adequate emotional state to proceed. By combining emotion recognition with an interpretable scoring model, the framework provides educators with a practical tool to monitor emotional readiness and identify students at risk of underperforming due to anxiety. Preliminary findings show the approach effectively captures variations in nervousness and offers insights into learners’ emotional states. This research contributes to affective computing in education, demonstrating the potential of neural networks to enhance fairness, well-being, and adaptability in assessment environments.

Effective communication in emergency medical services is critical in high-stakes scenarios, where information must be conveyed with speed, precision, and clarity. However, background noise, stress-induced speech patterns, and the use of specialized medical terminology frequently hinder comprehension. Improving the reliability of emergency communication is therefore a pressing challenge for both clinical outcomes and operational efficiency. This paper introduces a robust multi-model information extraction pipeline designed to enhance the accuracy and efficiency of emergency medical communication. The pipeline integrates advanced Speech-to-Text (STT) systems with Large Language Models (LLMs) to both improve transcription fidelity and extract mission-critical medical data. It comprises four modules: (1) audio capture of simulated German emergency communications under varied acoustic conditions, (2) STT transcription using Whisper, Azure, and IBM Watson, (3) LLM-driven refinement of transcriptions with GPT-4 to correct grammatical and terminological errors, and (4) structured information extraction with GPT-4, LLaMA 3.2, and Mixtral-8, guided by Chain-of-Thought and role-based prompting. The whole pipeline is evaluated using Word Error Rate (WER), BLEU, ROUGE-L, and semantic similarity, alongside accuracy, completeness, and relevance of extracted data. Azure STT with GPT-4 proved optimal, achieving the lowest post-refinement WER (0.1812, a 32.7% improvement), and high semantic similarity (0.9736), ROUGE-L (0.8802), and BLEU (0.7457). GPT-4 reached near-perfect extraction accuracy (0.995), surpassing LLaMA 3.2 and Mixtral-8, though Mixtral-8 remained highly competitive (0.980 accuracy with Whisper). Overall, the proposed pipeline demonstrates how combining STT and LLMs can transform noisy emergency dialogues into precise, structured clinical data, advancing responsive and reliable emergency management systems.

This study explores the identification of student profiles with parental responsibilities at the National University of Altiplano, Puno, using clustering algorithms. A total of 206 records of students with parental responsibilities were analyzed, employing a variety of sociodemographic, academic, and family-related variables. Dimensionality reduction was performed using Principal Component Analysis (PCA), retaining 10 key components to enhance computational efficiency and interpretability. Following this, three clustering algorithms—K-Means, DBSCAN, and Agglomerative Clustering—were implemented to segment students and evaluate the effectiveness of these methods in identifying distinct behavioral profiles based on their responsibilities and academic challenges.

The K-Means algorithm proved to be the most effective, generating two distinct clusters with a Silhouette Score of 0.1611, a Davies-Bouldin Index of 2.1475, and a Calinski-Harabasz Index of 33.7629. Cluster 0 included students with greater academic stability and fewer interruptions, while Cluster 1 represented students facing greater challenges, such as frequent study pauses and lower academic performance. DBSCAN identified a noise cluster, while Agglomerative Clustering produced intermediate results with less defined clusters.

These findings underscore the usefulness of clustering techniques in understanding the academic dynamics of students with parental responsibilities, offering valuable insights for developing personalized interventions. This approach helps fill a gap in the existing literature and provides opportunities for future research with larger datasets and additional variables, ultimately improving support strategies for this unique student population

Diagnosing heart disease is a complex and critical task that requires extracting meaningful patterns from large electrocardiogram (ECG) datasets. As the demand for faster and more reliable diagnostic tools increases, deep learning has emerged as a transformative solution, enabling automated ECG interpretation and reducing the risk of human error. However, traditional models often face challenges such as high computational complexity and limited adaptability to the dynamic nature of ECG signals.

In this study, we investigate the potential of transformer-based architectures to overcome these limitations and enhance classification performance. We explore two distinct strategies: the first employs a standalone transformer encoder to classify ECG signals into five categories—Normal beats (N), Unknown beats (Q), Ventricular ectopic beats (V), Supraventricular ectopic beats (S), and Fusion beats (F)—achieving an accuracy of 91%. The second approach integrates a Convolutional Neural Network (CNN) with the transformer encoder, where the CNN extracts relevant features that are subsequently refined and classified by the transformer, resulting in a significantly higher accuracy of 98%.

These findings demonstrate the effectiveness of transformer models, particularly when combined with CNNs, in improving the precision and robustness of ECG signal classification. This research contributes to the growing field of AI-assisted healthcare and highlights the promise of hybrid deep learning frameworks in supporting more efficient and accurate cardiac diagnostics.

Fish and meat products are known to transmit foodborne pathogens. Vernonia amygdalina, commonly called bitter leaf in English, “ewuro” in Yoruba, and “onugbu” in Igbo have been traditionally used as vegetable, food preservative and insect repellant. This study aimed to assess the susceptibility pattern of methanol extract of bitter leaf against bacteria isolated from abattoir meat and delineate its bioactive compounds. Methanol extract was tested against Bacillus sp, Alcaligenes faecalis1, Enterococcus faecium, Paenibacillus sp, Alcaligenes faecalis2, Lysinibacillus sp, and Lysinibacillus sphearicus, using agar well diffusion and broth dilution methods. Marked activity of 18mm and 13mm zone of inhibition were observed in Bacillus sp, and Lysinibacillus sphearicus respectively at 50mg/mL, as well as 18mm and 14mm in Alcaligenes faecalis1 and Alcaligenes faecalis2 respectively at the same concentration when compared to other isolates. Meanwhile, a range of 25-36mm zone of inhibition was noted for ciprofloxacin (5µg positive control). The minimum inhibitory concentration (MIC) of 25mg/mL was recorded against Lysinibacillus sphearicus, and 50mg/mL against Bacillus sp, Alcaligenes faecalis1and Alcaligenes faecalis2, while the minimum bactericidal concentration (MBC) was observed at 50mg/mL against Lysinibacillus sp. Gas chromatography-mass spectrometry (GC-MS) analysis revealed that Vernonia amygdalina possesses pest controlled Heneicosane with quality match (QM) of 86%. Other compounds such as pentadecane, nonanoic acid methyl ester, octacosane, pentacosane, benzenepropanoic acid, and squalene identified in the biter-leaf extract are known to have anti-inflammatory, anticancer, antioxidant, and antimicrobial properties. The study revealed that the extract can inhibit the growth of pathogenic bacteria.

Eaton Canyon in California, serves as the focal point for a comprehensive post-wildfire ecological impact assessment. This study employs an approach integrating satellite imagery from the European Space Agency's Sentinel constellation to study a burn area severity. The data include radar and multispectral data, providing a multidimensional view of the affected landscape. The analysis leverages the power of the Random Forest algorithm. Firstly, three widely-used indices the Difference Normalized Burn Ratio (dNBR), Relative Burn Ratio (RBR), and Relative Difference Normalized Burn Ratio (RdNBR) – were calculated and compared based on their accuracy and Kappa Index. Secondly, we developed a machine learning algorithm to create a fire severity map using wildfire indices independently based on their accuracy and Kappa Index and then developed a fusion approach to create a precise fire severity map by classifying the affected area into distinct severity classes. Finally, we compared our results obtained with the analysis of NASA predictions. The results showed a perfect 100% accuracy and Kappa index for all predictions. An area that did not burn due to the topography. Areas classified as low severity showed minimal damage with minimal tree mortality. Low severity to moderate showed regions with partial crown burns and tree mortality. Moderate to high severity areas represented significant tree mortality. High severity illustrated a complete tree mortality and significant loss of vegetation cover. Which may lead to a future work, classification and distribution of vegetation types before and after wildfires using deep earning.

Introduction. Tumor hypoxia is a major driver of cancer progression and therapy resistance. However, its cellular consequences can vary with metastatic potential (MP) of tumor cells. We compared the effects of chemically induced hypoxia on cell death and morphology of breast cancer lines MCF-7 and MDA-MB-231, having low and high MP, respectively.

Methods. Cells were exposed to CoCl2 (100μM, 20h) and analyzed immediately (T0) and 24h later (T1). Necroptosis was quantified by propidium iodide imaging (≥10 fields, three wells per condition). Morphometric parameters (area, circularity, aspect ratio, solidity, eccentricity, roundness) were extracted from at least 100 cells/condition. HIF-1α expression was assessed by immunofluorescence following fixation, permeabilization, and Hoechst counterstaining.

Results. In both untreated groups, cell death increased from 1±1% (T0) to 16±4% at T1. Under hypoxia, low-MP cells exhibited 33±5%, whereas high-MP cells showed only 4±1% necroptotic cells at T1. In non-treated high-MP cells, HIF-1α expression were higher than in low-MP cells and further increased after CoCl2 treatment. Morphology of low-MP cells was unaffected, while high-MP cells displayed increased circularity and roundness with reduced eccentricity (p<0.05) after CoCl2 treatment. At T1, CoCl2 treatment further decreased area, perimeter, and aspect ratio of high-MP cells, indicating a shift toward less elongated and more spindle-like forms compared with normoxia.

Conclusions. Chemically induced hypoxia elicits distinct subtype-specific responses: low-MP cells undergo necroptosis, whereas high-MP cells are less prone to this type of death exhibiting mesenchymal-like morphological traits. The combined use of HIF-1α detection and quantitative morphological analysis enables a thorough examination of the diverse responses to hypoxia.

Introduction: Column internals design is an integral part of chemical engineering applications. Traditionally reliant on computational/time-intensive optimization, this process is often limited in terms of efficiency and adaptability. With the recent emergence of A.I. and reinforcement learning, integrating these tools into the design process offers opportunities to further efficiency and optimization. The main challenge faced in this integration process is the navigation of complex, hybrid action spaces that contain, or combine, both continuous and discrete variables. Methods: By using custom reinforcement learning environments integrated with Aspen Plus, a digital twin framework was developed, allowing a machine learning agent to interact with process simulations. Two reinforcement learning algorithms were implemented, a hybrid Soft Actor–Critic and Deep Q-Network approach, which allocates continuous actions to the SAC algorithm and the discrete actions to the DQN algorithm, and a more unified Parametrized Deep Q-Network approach, which integrates discrete–continuous actions into one architecture. In both cases, the reward function is based on the percent approach to flooding at each section of the column, providing insight into the hydraulic stability of the column. While this is the metric chosen for current studies, the framework can be extended to include others. Results: Our results indicate that both reinforcement learning strategies navigated the hybrid action space, generated hydraulically feasible designs, and adapted to different column configurations. Conclusions: This research indicates that reinforcement learning is a plausible option for optimizing distillation internals design. The reinforcement learning strategies develop a pathway for scalable, multi-objective optimization in process design.