Proceedings XoveTIC 2025: Talento científico
Palabras chave:
Proceedings XoveTIC 2025Sinopse
O IX Congreso XoveTIC, organizado polo Centro de Investigación TIC da Universidade da Coruña (CITIC), terá lugar os días 15 e 16 de outubro de 2026. Está destinado a investigadores/as júnior (predoutorais e posdoutorais) que presentarán comunicacións orais ou pósters dos seus traballos científicos no ámbito das Tecnoloxías da Información e as Comunicacións (TIC). O obxectivo é construír un espazo de encontro para o debate científico e contribuír deste modo á súa formación.
O congreso, cuxa inscrición é gratuíta, suporá unha oportunidade ideal para que as persoas participantes compartan as súas investigacións nun ambiente colaborativo e distendido. As contribucións ao congreso poderán ser orixinais ou traballos adaptados xa presentados noutros foros. Todos aqueles traballos que se acepten, serán publicados nun libro de actas.
Capítulos
-
Methodological Advances in Robust Small Area Estimation
Small Area Estimation techniques address the growing demand for reliable disaggregated statistics by fitting models to unit-level or area-level data. While recent advances in unit-level mixed models are significant, the presence of outliers has driven the development of robust methods. M-quantile regression offers a promising alternative to mixed models for Small Area Estimation, though some theoretical aspects remain underexplored. We cover the optimal selection of the robustness parameters for bias correction, improving outlier detection. In addition, we propose bootstrap methods to approximate the distribution of area-specific M-quantile coefficients and optimal robustness parameters, enhancing inference and diagnostics.
-
Benchmarking a HAR(1) Fay-Herriot Model for Estimating Labour Indicators from the Spanish Quarterly Labour Cost Survey
national scale, we introduce a Divide-et-Impera strategy: models are fitted in parallel by Autonomous Community and optimally recombined using Fisher-information weights, delivering large speed-ups with negligible efficiency loss. For institutional coherence, we implement a fast triple-conformity benchmarking heuristic---with positivity constraints---aligning model-based totals with official aggregates across enterprise size, industry and territory; ratios are reconstructed from benchmarked numerators and denominators. Using 2024Q2 data and auxiliary covariates, the approach substantially stabilises estimates: median CVs are <=20% and seldom exceed 40%, meeting international dissemination standards. The pipeline---HAR(1) BFH + Divide-et-Impera + positivity-preserving benchmarking---offers a reproducible, policy-ready solution for releasing coherent labour- and wage-cost indicators at granular levels.
-
Presmoothed Latency Estimation of Cardiotoxicity in Breast Cancer Patients
Breast cancer remains the leading cancer diagnosis among women. While treatment advances have significantly improved patient outcomes, some of the adverse effects associated with these therapies have been linked to cardiotoxicity, potentially compromising heart function. Understanding the time when such toxicity emerges is crucial to optimizing patient care and monitoring strategies. This study aims to estimate the time until cardiotoxicity appears, known as latency. The idea is to apply an innovative adaptation of traditional cure models considering a presmoothing method.
-
Identifying Atmospheric Nucleation Events Using Machine Learning
models and deep neural networks are tested. Preliminary results show promising performance, highlighting the system's ability to identify positive events with high sensitivity, suggesting a possible future integration into atmospheric monitoring platforms.
-
Pricing with Rough Bergomi Model in Commodity Markets
propose a rough volatility model for pricing European options on commodity futures, based on the work of Nastasi et al. (2020). Furthermore, we present an efficient numerical scheme to simulate the model and calibrate it to real market data on WTI Crude Oil.
-
Extended Reality and Its Role in Medical Education: A Survey
As Extended Reality (XR) technologies become increasingly relevant in medical education,
this cross-sectional survey at ICBAS investigated undergraduate medical students’ perceptions
of XR’s usefulness, learning impact, and potential barriers. Results indicated generally
favorable attitudes, especially in anatomy, surgery, and emergency medicine, where spatial comprehension
and procedural skills are essential. Nonetheless, many respondents reported minimal
prior interaction with XR, frequently linking it to gaming rather than education. Students
expressed a preference for hands-on, purpose-driven applications. Perceptions varied by academic
year: junior students valued XR for consolidating theoretical knowledge, whereas senior
students highlighted its role in simulating clinical practice. A phased approach to curricular integration
is advised, starting with basic XR literacy and progressing toward targeted, disciplinespecific
implementations. -
Mixed Reality Simulator for Orthopedic Surgical Skills Training
Considering the growing need for effective and accessible methods in surgical training,
the potential of mixed reality (MR) was explored as an alternative to traditional methods.
In pursuit of this goal, an immersive simulator for laparoscopy and arthroscopy was developed, combining virtual environments with gamification techniques. The approach integrates a simulator incorporating clinical exercises and a web platform that allows a training scenario to be customised. The evaluation using the System Usability Scale (SUS) revealed a perception of high usability, accompanied by improvements in the participants’ autonomy and training experience. The results demonstrate that MR, together with gamification, is a viable and effective alternative to traditional surgical teaching methods. -
OptimizedWeb Scraping System Using Multithreading Techniques for Corporate Information Extraction
This study aims to automatically identify the various ICT innovations implemented by
companies on their websites. To achieve this, web scraping techniques are proposed as the most effective approach. However, the need to process large sets of companies presents challenges in terms of execution time and system performance. To address these issues, our system incorporates multithreading techniques, enabling the concurrent execution of analysis drivers that extract the targeted elements. As a complement to our extraction system, we developed a web application that allows users to visualize the obtained results, organized on a map of Spanish provinces and grouped by the analyzed companies. -
Procedure for the Deployment of Fog Computing Use Cases
Fog Computing (FC) emerges in response to the demands of the Internet of Things,
which requires lower latency, greater security, and the capacity to manage multiple devices. Its implementation poses challenges due to the diversity, mobility, scalability, and interoperability of the elements involved. This work aimed to define the key elements for designing and deploying fog computing systems. Astudy of the state of the artwas conducted, and a procedure composed of six stages was proposed, with technical aspects and recommendations in each. The procedure was validated through its application to a video surveillance use case, with satisfactory results -
Quantum Machine Learning for Financial Applications
The rise of quantum technologies has led to growing interest in exploring the use
of quantum circuits alongside ML models—also referred to as Quantum Machine Learning
(QML)—to enhance the development of new applications, such as through Parametrized Quantum Circuits (PQCs). Among the applications enabled by the use of PQCs in the financialworld, some of the most relevant include the computation of risk metrics, aswell as the option pricing of financial derivatives. This work aims to illustrate the study of the aforementioned applications, where the main idea consists in approximating, through the PQC, the underlying distribution of the assets and the payoff function, and from there, computing risk metrics and derivative prices. -
Green Algorithms: Environmental Impact, Legal and Ethical Framework in Artificial Intelligence
This article explores the concept of green algorithms and their connection to sustainability
in the field of artificial intelligence, analyzing the environmental impact of emerging technologies, with special attention to the energy consumption of AI models and the carbon footprint generated by data centers. The aim is to delve into the legal and ethical aspects related to the development of sustainable technologies, reviewing the international and European regulatory framework, as well as the existing legal gaps, while addressing the ethical principles that should guide responsible technological innovation. -
Quadruped Robot Control Using Immersive Interfaces in Emergency Contexts
This study presents an experimental system for the immersive teleoperation of a
quadruped robot in emergency scenarios. Based on the Unitree Go1 robot and integrates the
Meta Quest 3 headset for immersive visual feedback, along with its built-in joystick controllers for robot navigation. The user can control the robot’s movement intuitively via joystick inputs while receiving real-time video from the robot’s onboard cameras directly into the headset. Preliminary tests in controlled environments show that this approach improves user situational awareness and responsiveness during remote monitoring tasks. Designed as a modular and low-cost prototype, the system aims to explore new interaction paradigms in immersive human-robot collaboration. -
Accelerating Inference in Computer Vision Tasks with VCK190 and Vitis AI
The increasing use of Deep Neural Networks (DNNs) for real-time tasks has created
a need for specialized hardware accelerators.This paper presents a study of the Xilinx VCK190
development board, a next generation Adaptive Compute Acceleration Platform (ACAP) for artificial intelligence tasks, representing a first approach to this technology at the Universidade da Coru˜ña. The work details a complete workflow using Vitis AI, covering the quantization, compilation, and deployment of a neural network model. The results establish a solid technical baseline, intended to facilitate the integration of this advanced hardware into future academic and research projects. -
Application of Transformers for Sleep Stage Classification
Transformer architectures have revolutionized the field of artificial intelligence. This
work investigates the application of Transformers to the sleep stage classification problem.
De-spite promising results in recent studies, the clinical application of these methods remains
lim-ited. In this paper, we propose a novel model that replaces the LSTM component of a
state-of-the-art CNN+LSTM architecture with a Transformer encoder, comparing its performance against our baseline and several state-of-the-art models. The results showed faster convergence, reduced complexity, and enhanced performance. Leveraging the Transformer architecture, we propose and investigate an interpretability method based on attention mechanisms. Finally, we evaluate the inter-database generalization performance of our model. -
Developing an Open Source Tool for Man-in-the-Middle (MitM) Attacks on the MQTT Protocol
IoT has facilitated the connection of millions of devices, creating an interconnected
ecosystem of machines that collect, process, and transmit large amounts of data. This ecosystem is becoming a target of interest for the cybercriminals because many of the devices within it lack robust security protocols and measures. The resulting vulnerabilities, while already posing risks to users of these technologies, can have devastating impacts at the industrial level. The objective of this work is to develop a new tool that allows the alteration of MQTT traffic, a popular IoT protocol, in conjunction with other known technologies and techniques to carry out Man-in-the- Middle attacks, providing a newoption to understand this kind of attack and to test the resistance of the communications. -
Optimization of the Scheduling Problem in Cell-Free Massive MIMO Communication Systems
This work addresses the selection of subsets of access points (APs) to serve users cooperatively in Cell-Free Massive MIMO (Multiple-Input Multiple-Output) systems. Traditional
cellular architectures suffer from coverage and interference issues at cell edges. In contrast, Cell-Free networks distribute simpler APs uniformly across the area, enhancing spectral efficiency and connection quality. We simulate realistic communication scenarios to train a Deep Contextual Bandits (DCB) model that optimizes AP assignment based on user channel gains and interference.
Alternative data-driven approaches, such as loss-based models and fuzzy clustering,
are also developed and evaluated. Results show that DCB offers scalable, adaptive performance for next-generation wireless networks like B5G and 6G. -
CowApp: Mobile Application for Livestock Farm Management Using RFID Technology
In the complex world of livestock farming, producers face daily tasks ranging from animal
health and production management to record-keeping and documentation. Often, farmers
must handle administrative tasks in addition to ensuring the welfare and performance of their animals.
The application, developed in React Native, is designed to simplify and optimize daily farm operations while ensuring mobile compatibility. It features an intuitive interface for centralized management of animal and farm data. The backend, built with Spring Boot using JPA and Hibernate, provides a robust and efficient system for data storage and processing. MySQL is used to ensure data integrity and security.
The application covers everything from detailed monitoring of animal health and production
to replacing daily paper records, giving farm staff confidence that all essential information is
available at any time. Additionally, through RFID technology, farmers can quickly access relevant animal data, receive notifications for medicine administration, and export information as PDF files.
In summary, the application not only simplifies data collection and management but also provides tools for informed decision-making. By generating charts and visual representations of production, health, and other key metrics, it allows for deeper analysis of livestock performance, helping farmers identify trends and make decisions to improve animal health and productivity. -
Integration of Flamapy.js in a Software Product Line Engineering Platform to Enhance the Analysis of Feature Models
SPLALM is a Product Line Engineering factory development environment that introduces
two key innovations. It integrates with GitLab and git to enable consistent and coordinated
changes in complex product families. It also supports parsing, persisting, analyzing, and visualizing Feature Models (FMs) using the Universal Variability Language, a community-driven standard for variability representation. To support FM analysis directly in a browser, SPLALM uses Flamapy.js, a Pyodide-basedWASMwrapper of the Flamapy framework. This paper focuses on the integration of Flamapy.js and how it upgrades the analysis of FMs in SPLALM. Together, these features align SPLALM with ISO/IEC 26580 standards, offering a reliable solution for managing high-variability product lines. -
Quantum and Classical Kernels Applied to Classification of Unbalanced Datasets
This work investigates quantum kernel methods for classification on unbalanced datasets. The study considers the effect of techniques commonly used in classical machine learning, such as kernel centering and cost-sensitive learning, when applied to quantum models. Experiments are conducted on five datasets, including MNIST-1D, the real-world dataset Glass6, and synthetic datasets tailored for quantum classifiers. Results indicate that quantum kernel methods can address unbalanced classification tasks albeit none of the techniques studied showed a statistically significant improvement of the scoring metrics. Results also highlight the influence of hyperparameterization and in particular, quantum bandwidth is identified as a critical hyperparameter for the performance of quantum models.
-
Real-Time Line Detection via GPU-Based Hough Transform
This work investigates quantum kernel methods for classification on unbalanced datasets. The study considers the effect of techniques commonly used in classical machine learning, such as kernel centering and cost-sensitive learning, when applied to quantum models. Experiments are conducted on five datasets, including MNIST-1D, the real-world dataset Glass6, and synthetic datasets tailored for quantum classifiers. Results indicate that quantum kernel methods can address unbalanced classification tasks albeit none of the techniques studied showed a statistically significant improvement of the scoring metrics. Results also highlight the influence of hyperparameterization and in particular, quantum bandwidth is identified as a critical hyperparameter for the performance of quantum models.
-
Mixed Reality Game-Based Simulator for Arthroscopy Skills Training
The acquisition of arthroscopic surgical skills remains a significant educational challenge due to the technical complexity and limited accessibility of traditional training methods, which often entail high costs, ethical concerns, and low interactivity. This study presents a Mixed Reality (MR) game-based simulator, developed in Unity for the Meta Quest 3, aimed at enhancing arthroscopy training. The core task involves navigating a virtual maze using simulated surgical instruments, replicating the visuomotor coordination required in real procedures. A study with ten participants assessed usability, engagement, and perceived workload. Results indicated high satisfaction with the intuitive interaction, reinforcing the simulator’s potential as an effective and accessible training solution.
-
Multimodal Assessment of Cognitive and Motor Performance in Immersive Surgical Training
The integration of immersive technologies is reshaping surgical education. This study introduces a performance assessment system for immersive simulators in laparoscopic and arthroscopic training. By combining EEG and accelerometry, it monitors cognitive and motor functions in real time. A multidisciplinary framework enabled the definition of objective metrics for attention, stress, and motor control. A complete pipeline was developed for data acquisition, processing, and analysis, with results accessible via a web interface. Findings suggest that neurofeedback and motion tracking through VR controllers provide a reliable, objective alternative to traditional assessments, enhancing skill acquisition and enabling personalized, practice-based medical training.
-
HATHOR - Accessible Automation and Interaction in a Hydroponic Garden for People with Disabilities
This paper presents the design and implementation of an intelligent hydroponic garden, remotely controlled via mobile devices and automated through sensors and actuators integrated into Home Assistant. The system enables automated irrigation through programmable electromagnetic valves, as well as control of the hydroponic solution conditions, temperature via fans and light through motorized blinds and heat lamps. The developed app was specifically designed to facilitate interaction for individuals with intellectual disabilities, prioritizing accessibility, autonomy and positive reinforcement. This intervention was implemented in a Care Center for People with Disabilities aiming to enhance users’ motivation in meaningful activities related to plant care, irrigation and environmental control.
-
Study on the Implementation of AI Inference Services in a Business Environment
Diffusion-based generative models, such as Stable Diffusion, have revolutionized the generation of images from textual descriptions. However, their high computational cost represents an obstacle in production environments. This work analyzes strategies for their efficient and reproducible deployment in scalable clusters, with the technical support and resources provided by the company. The study is conducted in a real corporate context, where these models are already being used to accelerate design processes and optimize key stages of the value chain. To achieve this, different inference service strategies are compared in terms of latency, GPU utilization, and integration capabilities.
-
Integration of GUASOM into the SPACIOUS Computing Platform
The Science PlAtform Cloud Infrastructure for Outsize Usage Scenarios (SPACIOUS) project is developing a computational environment for astrophysical research using Big Data technologies, with the aim of enhancing the scientific exploitation of large volumes of data. To this end, data mining and Artificial Intelligence tools are being developed and integrated into the platform.
This paper presents the adaptation of GUASOM, a tool for the visualisation and analysis of Self-Organising Maps (SOM) aimed at analysing outliers from the Gaia mission. A Python application has been developed that allows these maps to be trained and loaded into GUASOM directly from SPACIOUS Jupyter Notebooks, facilitating their integration and use in the project ecosystem. -
Methods for the Redistribution of Tourist Expenditure in Spain: From EGATUR to Reallocation Matrices
To estimate the expenditures made by foreign visitors in Spain, the National Statistics Institute (INE) produces the "Tourist Expenditure Survey'' (EGATUR), which allocates all spending to the main destination region, even if it was not fully carried out there. Consequently, there arises a need to redistribute expenditure across the Autonomous Communities actually visited, for which we rely on supplementary information from transactions with foreign bank cards.
Our proposal is to estimate a reallocation matrix whose elements allow the redistribution of expenditure from the main destination regions to all those effectively visited. Several approaches are explored, including constrained least squares, overnight-stay proportion matrices, and Data Envelopment Analysis (DEA), among others. -
Automated Counting of Zebrafish: An Image Processing Approach
Accurate quantification of fish populations is a critical task in aquaculture and behavioral research. In this work, we present a lightweight image processing pipeline for the automatic counting of zebrafish (Danio rerio) in a multi-compartment aquatic system. The approach combines background subtraction, morphological refinement, and watershed segmentation to estimate fish counts directly from raw video without annotated training data. The method enables non-invasive and reproducible counting while remaining computationally efficient and interpretable. Although challenges remain under occlusion and limited visibility, the method reduces reliance on manual observation and demonstrates that classical image-processing techniques can provide efficient, interpretable, and accessible solutions for early-stage behavioral experiments.
-
Data Extraction and Transformation Methodology for Biometric Signals from Wearable Devices
The increasing use of consumer wearable devices, such as smart bands, provides new opportunities for health monitoring and clinical research. However, access to high-resolution data is often limited by proprietary formats and aggregated summaries that are unsuitable for detailed analysis. This work presents a reproducible methodology for data extraction and transformation from Xiaomi devices, applied in a clinical study with 179 participants suspected of obstructive sleep apnea. A two-stage pipeline was developed to convert exported files into structured, minute-level datasets, accessible through a graphical interface designed for non-technical researchers. The approach was evaluated in terms of data quality, robustness, and utility, successfully generating key metrics such as heart rate, oxygen saturation, respiration, steps, stress, and sleep stages. Results show that the methodology facilitates standardized access to physiological signals, supporting visualization and analysis in clinical and interdisciplinary research contexts.
-
Application of Artificial Intelligence Models in Predicting Clinical Outcomes after Nerve Block in Orthopedic Surgery
This project aims to analyze the effect of the anesthesiology technique of nerve block on the postoperative outcomes of patients undergoing hip and knee surgery. Using the collected data, which includes variables such as pain levels, medication intake, and satisfaction with the procedure, the goal is to develop an artificial intelligence model capable of determining which approach, using a nerve block or not, offers better clinical results. The main objective of this project is to demonstrate that the use of nerve blocks contributes to a less painful recovery and a lower consumption of analgesics.
-
Classification of Orthopoxvirus with Deep Learning in reduced data scenarios using resampling techniques
In recent years, monkeypox has become a growing global threat, where early diagnosis is essential for its control. This work explores the use of Deep Learning techniques applied to skin image analysis to improve the detection and classification of this disease compared to other similar ones. A highly unbalanced dataset of 770 images is used, so resampling techniques such as SMOTE and SMOTEENN are applied. The objective is not only to compare the performance of different Deep Learning models, but also to measure the impact on classification produced by the use of resampling strategies. It also seeks to identify the best combination to support automatic diagnosis in clinical and epidemiological contexts.
-
Prediction of Photovoltaic Energy Time Series: a Comparative Study between LSTM, Hybrid and XGBoost Models
In recent years significant progress has been made in the field of renewable energy, with photovoltaics standing out in particular. This is partly because the users usually try to use clean energy to protect the planet, as well as seeking energy independence and saving money. However, one of the main disadvantages of implementing photovoltaic systems is knowing how much energy will be generated and how to manage it. For this reason, multiple models have been created that are capable of making accurate predictions. This article uses historical data from a simulated installation with PVGIS located at the epef, based on which some of the most prominent methods for making these predictions are presented and analysed, specifically LSTM, XGBoost and hybrid models. These predict the power generated at two stages with acceptable accuracy, enabling the user to improve energy planning. Finally, a comparison will be made, concluding with a brief recommendation on which one to use depending on the context.
-
Open-source Secure NFC-based Physical Access Control System using EV2 Mutual Authentication
Currently, physical access control is dominated by private and proprietary solutions that are closed, inflexible and often insecure. To address these shortcomings, this work presents an open-source, modular NFC-based access control system. The system consists of two modules.
First, a hardware module, capable of reading NFC NTAG424 cards and unlocking a magnetic lock has been assembled. Second, a centralized management server, with an administration panel. The system supports Enhanced Version 2 mutual authentication protocol, time-based and role-based policies, allowing administrators to protect multiple zones within the same facility from a centralized management server.
This open solution provides a secure and cost-effective alternative that seeks greater transparency and control. -
Real-time Defect Detection in Conveyor Belts using Point Clouds Generated by a ToF Camera
A three-dimensional analysis system was developed using Time-of-Flight (ToF) infrared cameras to detect conveyor belt teeth and estimate their speed. From the captured point clouds, structural patterns are identified through geometric fitting and spatial segmentation. As a novel contribution, the system detects belt teeth, locating absences and deformations. Detection is performed by comparing the expected distribution with the observed one.
The speed of the conveyor belt is estimated by calculating the relative displacement between consecutive point clouds. The system was successfully tested, providing accurate and reliable analysis in dynamic industrial environments. -
System for the Acquisition and Analysis of Physico-Chemical Parameters of Surface Waters of Ferrol and A Coruña
The increasing pressure of urban and industrial activities on aquatic ecosystems, plus stricter water quality standards, highlights the need for continuous monitoring. This study aims to present the planned implementation of a data acquisition framework based on the Eureka Manta+35 multiparameter probe, designed to monitor surface waters in the municipalities of Ferrol and A Coruña. Real-time measurements of pH, temperature, conductivity, turbidity, dissolved oxygen, ammonium and nitrate will be sent to the cloud via IoT. Data processing is expected to include outlier filtering, drift correction, and cross-validation with laboratory analyses. Analytical techniques such as Fourier spectral analysis, regression, and correlation will be applied to identify spatiotemporal patterns and anthropogenic impact gradients. The anticipated outcomes will enable the quantification of pollution sources, support the assessment of ecological risks, and underpin adaptive management strategies in line with European water directives.
-
Protein Language Models for Predicting Mutational Effects in Plants
Predicting the effect of mutations is key for estimating genetic load—the accumulation of deleterious mutations that may reduce fitness of organisms. Traditional methods rely on predefined features, such as evolutionary conservation or the predicted impact on the protein sequence and structure. Protein language models (PLMs) offer a new data-driven alternative by learning functional constraints directly from protein sequences. In this study, we apply the PLM ESM-1v for predicting mutational effects, using 12,865 functionally annotated mutations from 433 plant species, a group with limited predictive tools. We also explore integrating ESM-1v embeddings as features in supervised machine learning models.
-
Application of Haptic Gloves in an Interactive Environment as Support for Hand Functional Intervention
This master's thesis explores the application of the Manus Quantum Metagloves as a tool for hand rehabilitation. Through data analysis, three key parameters were evaluated: joint range of motion, joint velocity, and abduction–adduction movements. The results show specific joint limitations, especially in the interphalangeal and metacarpophalangeal joints, as well as functional patterns consistent with anatomy and clinical practice. In addition, a virtual environment was designed for therapeutic purposes. This study validates the potential of using motion capture technologies as an objective complement in the functional assessment of the hand, providing value both in clinical and research settings.
-
Web Application Oriented to the Analysis of Multiple Trajectories in Football Matches Based on Georeferenced Data
Data analysis has become an essential tool for understanding and optimizing performance in football. Within this context, the study of player trajectories plays a key role in the tactical analysis of the game. This paper presents the development of a system designed to visualize, search and compare player trajectories during a match based on georeferenced data. Implemented in a web application, the system enables users to identify and explore movements at key moments using either static or dynamic visualizations. In addition, it integrates a system that allows the detection of trajectories with movement patterns similar to a given trajectory, which required the design and optimization of an efficient similarity search algorithm.
-
VR-ADAPT: A Virtual Reality Environment for Skill Development in New Wheelchair Users with Spinal Cord Injuries
Annually, hundreds of thousands of individuals worldwide experience mobility-limiting injuries that compromise not only their walking ability but frequently the functionality of their upper extremities as well. Such injuries constitute a profound life transformation, necessitating an intensive period of learning and adjustment to new physical realities. We introduce VR-ADAPT, a novel virtual reality platform conceived to support this critical transition toward greater independence. The system comprises three integrated components: an advanced electric wheelchair operation simulator, immersive digital recreations of home and professional environments, and a suite of therapeutic serious games embedded within these spaces. Through gamification strategies, users can safely acquire and refine the competencies needed to navigate their everyday environments with confidence. A comprehensive kinematic capture and analysis module continuously records detailed performance metrics throughout training sessions, furnishing rehabilitation professionals with objective, quantifiable data to track patient advancement. This evidence-based approach enables more precise therapy customization and improved clinical outcomes.
-
CARE-XR: Saving Lives in Combat Environments with a Virtual Reality Framework
The CARE-XR project (Combat Aid \& Response Environment – Extended Reality) presents a practitioner-driven framework for immersive training in Tactical Combat Casualty Care (TC3). Grounded in evidence from a prior systematic review and a Delphi study with Portuguese Special Operations Forces (SOF), CARE-XR combines XR technologies with bio-adaptive affective computing and participatory co-design to address operational needs for realism, adaptability, feedback, and teamwork. The manuscript outlines a phased development process, needs assessment, iterative co-design with SOF, and validation planning, and details the technical components: XR headsets, biofeedback wearables, haptic devices, authoring tools, affective algorithms and simulation management systems. Rather than presenting an efficacy trial, this work is framed as a design and analysis contribution, articulating design requirements, system architecture and ethical considerations. In practical terms, CARE-XR lays the groundwork for a modular and scalable foundation to evaluate transfer of TC3 skills and stress-regulation to live operational contexts.
-
Empowering Digital Health Education through SAP Analytics Cloud Gamification and Virtual Environments
This paper presents an innovative educational framework developed for the Bachelor's degree in Digital Health at the Polytechnic Institute of Porto. The model integrates SAP Analytics Cloud (SAC) for business intelligence and data visualization with immersive Virtual Reality (VR) environments and gamification strategies. Designed to bridge the gap between theoretical knowledge and practical digital competencies, the course "Health Data Analysis and Visualization" engages students in active learning through real-world health datasets, predictive modeling, and simulated public health scenarios. Preliminary results from pilot implementations indicate significant improvements in student engagement, data literacy, critical thinking, and interdisciplinary collaboration. This approach not only aligns with constructivist and experiential learning theories but also responds to international calls for digital upskilling in the healthcare workforce. The study concludes that the synergistic use of SAC, VR, and gamification offers a powerful, replicable model for transforming health data education and preparing a new generation of data-savvy healthcare professionals.
-
Ethical and Participatory Decision-Making in Military Contexts: A Delphi Method Approach
Decision-making in military contexts is inherently complex, characterized by high uncertainty, operational pressure, and profound ethical challenges. This article explores the intersection of ethical and participatory decision-making in military environments, emphasizing the potential of the Delphi Method as a structured, iterative approach to enhance moral reasoning and collective input. Drawing on adaptive leadership principles, the Delphi Method provides a framework for mitigating hierarchical and cognitive biases, promoting inclusive participation, and fostering ethical reflection among military experts. By integrating expert insights systematically, the method supports commanders in making morally grounded and operationally effective decisions. Implications for military professional ethics, organizational legitimacy, and operational performance are discussed, alongside considerations for future research and practical implementation. In addition, we clarify the specific contribution of Delphi to the validation and ethical governance of AI-based decision-support systems in defense.
-
Development of an Instant Messaging Application Based on MLS using QRNG-Generated Keys
Nowadays, instant messaging applications have become the main channel of communication between people. In response to demands of privacy by users, some applications have begun to incorporate end-to-end encryption to protect conversations. The Messaging Layer Security (MLS) is a recent protocol that offers end-to-end encryption designed for groups, providing both efficiency and security. In this work we develop a messaging application that uses MLS as the basis for secure communications between users. The application will use a peer-to-peer network to prevent messages from passing through servers that could store them and try to decrypt them. Furthermore, the cryptographic keys used will be generated by a Quantum Random Number Generator and accessed through an Entropy-as-a-Service platform.
-
MIRACS: An Unsupervised Multi-Omics Strategy for Identifying Patterns of Therapeutic Resistance in Cancer.
Drug resistance is a major challenge in cancer therapy, often driven by poorly characterized molecular states. We developed MIRACS (Multiomic Integration for Resistance Association via Correlation Strategy), an unsupervised multi-omics framework to identify resistance
programs across diverse cancer types and drug classes. Transcriptomic and proteomic profiles from 370 cell lines (22 tumor types, DepMap) were decomposed into 30 latent factors using Multi-Omics Factor Analysis (MOFA). Correlating these factors with drug response data from the PRISM dataset revealed over 700 significant factor–mechanism of action (MoA) associations across 10 cancer types.
Among these, a molecular program emerged as a cross-lineage resistance program linked to five drug classes—including topoisomerase, Bcr-Abl kinase, Aurora kinase, DNA synthesis, and HSP inhibitors—across eight cancer types. Functional analyses of gene and protein loadings highlighted stress response and inflammatory pathways (p53, NF-κB, TNF-α, JAK–STAT) and transcriptional regulators ATF6 and IRF1, consistent with a pre-existing stress-adapted state.
These results demonstrate that integrating multi-omics with drug perturbation data can uncover convergent, lineage-independent resistance programs, providing a systematic approach to identify targetable vulnerabilities across cancers. -
Mechanisms for Autonomous Sub-goal Discovery in Lifelong Robotic Learning
Lifelong learning in robotics seeks the continuous acquisition and reuse of knowledge to autonomously master complex tasks. A key challenge is discovering sub-goals without relying on explicit intermediate rewards, enabling robots to decompose tasks and transfer skills across domains. We propose a dual mechanism for sub-goal discovery. The first is a top-down strategy that builds hierarchical sub-goal chains from general goals via intrinsic motivations. The second is a bottom-up approach that uncovers latent links between previously learned goals and perceptual classes. Implemented in the e-MDB cognitive architecture, our method was tested in both simulation and a real-world robotic manipulation task. Results show efficient sub-goal generation, transfer, and generalization.
-
Motivational System for the Discovery of Missions in Autonomous Robots
This work introduces a motivational system for autonomous robots that allows the generation and learning of missions (and their associated drives) aligned with human purposes. The system is integrated into the e-MDB cognitive architecture, employing Large Language Models (LLMs) to interpret natural language instructions. The system employs three distinct LLM instances, each specialized for a specific function: semantic alignment, mission generation, and modeling of motivational drives. Experiments conducted within the Gazebo simulation environment demonstrated consistent alignment between robot behavior and human purposes, as well as high reliability in mission and drive validity.
-
Evaluating Factual Grounding Strategies in Large Language Models
Large Language Models (LLMs) often generate non-factual, or “hallucinated,” content, limiting their reliability in knowledge-intensive tasks. This challenge is particularly critical in multi-hop question answering (MHQA), where models must integrate and reason over multiple pieces of evidence. In this paper, we present an empirical study of prompting strategies aimed at improving the factual grounding of LLMs. Using the Llama-3.1 (8B) model on the HotpotQA benchmark, we evaluate five prompting techniques along three main design points: shot count (zero-shot vs. few-shot), context integration (with vs. without supporting documents), and output constraints (free-form vs. structured responses requiring evidence). We assess both answer accuracy and the precision of supporting fact identification, allowing us to analyze correctness from evidential grounding. Our results reveal different trade-offs: while few-shot prompting improves reasoning consistency, the gains diminish without high-quality supporting context. Similarly, structured outputs reduce variance and improve factual alignment, but their benefits depend critically on how evidence is presented and constrained. Compared to prior studies that focus primarily on accuracy, our analysis highlights the importance of balancing answer quality with verifiable evidence. These findings provide actionable guidance for the design of prompts in multi-hop QA and inform broader efforts to mitigate hallucinations in LLMs across retrieval-augmented and reasoning-intensive applications.
-
Immersive Virtual Reality Scenarios Applied to Static Bike Exercises to be Used in Rehabilitation and Therapeutic Settings
The advent of virtual reality has introduced a valuable tool for rehabilitation and therapy by providing safe and controlled environments that promote physical and cognitive activities in an intuitive and engaging manner. The aim of this project is to design a platform that enables dual-task training—combining physical and cognitive activities—as a novel intervention for individuals with cerebral palsy. The system integrates static bike exercises with cognitive challenges presented within an immersive and pleasant virtual environment. The project was developed using the Unity game engine and deployed on the Meta Quest 3 VR headset, with WEMOS ESP32 microcontrollers measuring the user’s pedaling cadence. This approach seeks to combine therapeutic effectiveness with user engagement through gamified feedback and interactive design.
-
Quantum Recurrent Neural Networks for Multivariate Time Series Prediction
Classical models for time series forecasting, despite their success, often face challenges related to training, generalization, energy consumption and interpretability. Unconventional computing paradigms, such as quantum computing, offer a promising avenue to address these limitations. Quantum Recurrent Neural Networks (QRNN) emerge as a powerful approach for multivariate time series prediction. Due to some features, the QRNN model we study is not supported in many quantum computers. We develop a specific-purpose emulator to address this barrier, and then we benchmark the model against datasets of varying complexities, involving realistic features such as noisy outputs. The results show significant performance compared to other well-known classical approaches for time series prediction.
-
Performance and Security Evaluation of Ethereum ERCs in an Industrial Environment
Ethereum use as a decentralized platform for executing smart contracts has driven the adoption of standards that optimize interoperability in industrial environments. Ethereum Requests for Comments (ERC) establish uniform patterns for smart contracts, facilitating their integration and operation in network nodes, which are essential for industrial applications such as supply chain management or process automation. However, this standardization can propagate vulnerabilities or inefficiencies in critical systems if the contracts are not optimized, affecting the reliability of industrial processes. Since operations in Ethereum consume computational resources (measured in gas) with associated economic costs, poor design can lead to significant losses in industrial settings. This paper evaluates the efficiency and security of ERC standards by examining their functional diversity and technical complexity. Thus, it analyzes existing implementations to identify common errors and proposes improvements to enhance the robustness and optimization of three of the most popular standards: ERC-20, ERC-1400 and ERC-3643. The ultimate goal is to support the effective adoption of ERCs in industrial applications. Considering the Ethereum network incentives on lower complexity logic and the obtained results, it is advisable to use simple standards, which also reduce error risks and ease maintainability.
-
Adaptation of Driving Videogames for People with Physical Disabilities
Video games can serve as highly beneficial forms of therapy and exercise for people with disabilities, yet they are often inaccessible due to their control schemes. In particular, driving games offer experiences that cannot be replicated in real life, but the steering wheels, gamepads or keyboards typically used are not usable by individuals with very limited mobility. This paper examines the accessibility challenges faced by such users when using both specialized and generic hardware and software—including an analysis of the games themselves and the existing assistive solutions. Based on this analysis, a solution is designed and proposed that relies exclusively on the mouse, without support from other devices or third parties.
-
Automated Code Generation for Tabular Question Answering in Zero-Shot Settings
Our work presents an effective and robust approach to Tabular Question Answering based on a zero-shot pipeline that uses a Large Language Model to generate executable code for extracting information from tables. Our system includes a modular architecture: a code generation module supported by column selection and type inference components to guide accurate code synthesis, together with an iterative repair mechanism for runtime errors into revised prompts. Our approach is successfully tested both in English and Spanish.
-
Immersive Visualization of Home Renovations for Accessibility Using Virtual Reality
Designing accessible spaces is a key challenge in contemporary architecture, especially when adapting existing housing. People with reduced mobility face physical barriers in their homes that affect their quality of life, and a personalized renovation helps address this. However, 2D floor plans are hard to interpret for non-experts, making it difficult to envision changes. Immersive technologies like Virtual Reality (VR) enable first-person exploration of architectural spaces, improving spatial understanding and decision-making. This work presents a VR tool that reconstructs and visualizes an apartment from CAD drawings, allowing users to move through the space, collide with different objects and obstacles, and switch between the current state and a proposed post-renovation layout.
-
Morphological Classification of Galaxies From the SDSS Using Machine Learning Techniques
Given the vast amount of labeled data from surveys like SDSS and Galaxy Zoo 2, AI has become essential for galaxy morphology classification. Challenges such as redshift, low resolution, and similar shapes make even expert labeling difficult. We developed a hierarchical two-stage CNN model to classify galaxy images. The first model categorized them into four classes, merging the underrepresented and visually similar cigar-shaped and edge-on types. A second submodel was then trained to distinguish between these two classes, using data augmentation to address class imbalance. Using these approaches, we achieved 95\% accuracy, outperforming other research models using the same classes and demonstrating better classification accuracy and generalization on imbalanced datasets.
Descargas
Publicado
Licenza
Esta obra está baixo unha licenza internacional Creative Commons Atribución-NoComercial-SinDerivadas 4.0.
