José Alberto Benítez-Andrades, Ph.D.

Suicides have increased in recent years in many countries around the world. According to the World Health Organization (WHO), suicide is the second leading cause of death among people aged 15 to 29. The WHO estimates that approximately 800,000 people die by suicide each year, equivalent to a suicide rate of 11.4 per 100,000 inhabitants. The WHO recognizes suicide as a public health priority. Suicide attempts have been shown to be a significant predictive factor for death by suicide, and most suicides are preceded by an average of four previous suicide attempts. Therefore, identifying suicide attempts among at-risk patients and preventing their suicidal behaviors is an urgent problem. There are different strategies to try to identify people at risk of suicide. For years, the most accurate have been based on complex screening based on information contained in the electronic health record (EHR). In the medical field, the use of artificial intelligence (AI) is increasingly being used for the diagnosis of different pathologies. The aim of this project, SUICIDETECT, is to investigate new methods of detecting suicide attempts in at-risk populations by combining natural language processing (NLP) techniques, semantic enrichment, and Large Language Models (LLM), applying prompt engineering to optimize the interpretation and analysis of Electronic Health Records (EHR). The initial hypothesis is that, through the use of these technologies, more effective supervised machine learning classification models can be developed to predict suicide attempts in at-risk patients. We currently have a dataset of 6,785 patient records referred to the psychiatry service, from 2017 to 2023. In addition, there is an expression of interest from the Junta de Castilla y León to collaborate in this research. To validate this hypothesis, work will be carried out in different lines that will provide the following main contributions: (i) an updated systematic review in the field of semantic enrichment techniques, LLMs, and information extraction in EHR; (ii) generation of approaches beyond the state of the art in the field of NLP, semantic enrichment, and LLMs; and (iii) work on the creation of machine learning models that make use of the new vectors based on a complete context within the health environment. In the processing of the EHR, new methods of obtaining knowledge about the concepts detected in the EHR will be added. For this, collaborative knowledge graphs such as Wikidata and DBpedia will be used. The final goal is to provide classification and prediction models of suicide attempts to help prevent them. Two approaches will be used for the machine learning models: the use of tabular data and graph-based data. We hope to discover new features available in the EHR that help predict possible suicide attempts in patients who have already attempted suicide at least once.

Turkey’s Bursa Uludağ University (BUU), a public research university in Bursa, will receive a big boost through a twinning partnership with Spain’s University of León (ULE) and Ireland’s Atlantic Technological University (ATU) in business model innovation (BMI). Under the EU-funded REMODEL project, BUU staff will enhance their management and administration skills in the hospitality sector. A new BMI laboratory will be created leveraging and combining the strengths of the ULE consumer behaviour analysis laboratory and ATU DiceLabs. It will be instrumental for the execution of the research and innovation part of the project by applying acquired knowledge and skills in consumer behavioural analysis, neuromarketing, research management, business innovation, leadership and entrepreneurship, developing business models for 10 Turkish SMEs in the hospitality sector in the context of COVID-19 pandemic.

Smart, Sustainable and coheSive Digitalization conceived as a Digital Innovation Hub
Digital Innovation Hubs (DIHs) are entities that help companies and public administration in a given territory access the information, services and facilities they need to successfully address their digital transformation processes.

It is one of the main tools of the European Commission to achieve the objectives set out in its digitalization strategy.

La Universidad de Leoben (Austria), la Universidad Técnica de Freiberg (Alemania), la Universidad de Petrosani (Rumanía), la Universidad de León (España), la Universidad Técnica de Creta (Grecia), la Universidad Tecnológica de Silesia (Polonia) y la Universidad de Ciencias Aplicadas de Mittweida (Alemania) son siete instituciones de enseñanza superior situadas en seis Estados miembros de la UE diferentes que han unido sus fuerzas para crear una universidad europea fuerte y única en el ámbito del consumo y la producción responsables (PCR): EURECA-PRO.

Tiene una doble misión social y planetaria. Por un lado, contribuye de manera holística a la cuestión de gran actualidad del PCR en el marco del Objetivo de Desarrollo Sostenible (ODS) 12 y, por otro, contribuye eficazmente a la transformación del Espacio Europeo de Educación Superior de manera complementaria al ODS 4.

La innovación es la clave para alcanzar el objetivo de reducción de CO2 y las prácticas de sostenibilidad asociadas del Pacto Verde de la UE hasta 2050. Su realización se basa en nuevas tecnologías y procesos que integren los flujos de materiales y recursos en el sentido de la Economía Circular y en comportamientos de consumo responsable que estén alineados con las expectativas de la sociedad respecto a la lucha contra el cambio climático. Una educación inclusiva, sin fronteras e integrada apoya a los titulados más competentes y cualificados que pueden contribuir a este importante reto de la sociedad europea.

Robots are part of our daily lives, and their operation and evolution are determined by a simple yet elegant and effective concept: the divide and conquer strategy. Each mobile robotic platform is composed of a large number of interconnected individual systems, each responsible for processing a small amount of information or performing a specific task. The integration of these systems results in complex robotic platforms capable of perceiving their environment and acting accordingly. This modularization in robotics not only allows the design of highly customized systems but also enables the development of new functionalities in isolation, without the need to modify the rest of the robot’s systems.

This doctoral thesis leverages this design strategy to propose research covering three of the fundamental components required for the proper functioning of a robotic platform.

First, a roadway detection system for a vehicle is proposed. This system is designed for use in an autonomous vehicle and provides information about the region of the environment free of obstacles—that is, the area where the vehicle can safely circulate. Using environmental data from a monocular camera and a three-dimensional LiDAR sensor, the system estimates the position, size, and morphology of the roadway through a Convolutional Neural Network specifically developed for this work. The proposed neural architecture, named TEDNet, has been evaluated on the well-known Kitti dataset, ranking among the top ten models achieving the best performance when using both types of sensors.

The second part of this work focuses on the analysis of road networks and the proposal of a route that allows all streets within them to be traversed. The developed solution involves representing the road network as a weighted graph and applying optimization techniques to find the route that completes the task while covering the shortest possible distance. The evaluation of several real road networks has validated the proposed system for generating routes for an inspection vehicle.

Finally, this research addresses the integration of different environmental representations, whether pre-existing or generated by different types of robots, and the definition of a common reference system for all of them. The CAD2SLAM system is presented, which enables adaptive transitions between reference systems while minimizing the effect of errors present in each robot’s maps. The evaluation of this system in various real-world scenarios has demonstrated its utility and accuracy in combining multiple maps.

The conducted research has been presented in articles published in JCR-indexed journals, as well as in national and international conferences. Additionally, the implementation of the developed systems has been made available to the community for verification of the obtained results and for the reuse of these systems in other projects.

This doctoral thesis examines the application of Artificial Intelligence (AI) to the detection and prediction of risky sexual behaviors among Men who have Sex with Men (MSM) who use Dating and Contact Applications (DCAs). Machine learning techniques are applied to semantic, sociodemographic, and network variables, combining Social Network Analysis (SNA), k-means clustering, and mixed qualitative–quantitative methods. The study is structured into three sections: (1) analysis of language and symbolism in DCAs, (2) SNA of sexual networks and risk behaviors, and (3) application of AI to predict risk patterns. The findings emphasize the interplay between individual behavior, network dynamics, and risks associated with chemsex.

The linguistic and symbolic analysis reveals a wide diversity of terms and emojis on platforms such as Grindr®, reflecting sexual preferences, substance use, and risk practices. Some terms are linked to GHB or cocaine consumption, while others denote group encounters. Results also show a normalization of practices such as gagging and the evolution of terms like “chill”, underlining the importance of contextualized language analysis for prevention strategies.

The typical DCA user profile corresponds to cisgender men aged 31–40, single, highly educated, and urban. Data highlight a high prevalence of unprotected oral (89.6%) and anal sex (31%), with chemsex reaching 43.7%. Ketamine users display particularly risky patterns. STI incidence is elevated—chlamydia/gonorrhea (27.1%), syphilis (14.6%)—as is HIV prevalence (8.3%, all undetectable with treatment). Despite this, health literacy is high, with most users regularly tested for HIV.

SNA shows diverse, mostly casual networks averaging 12 partners in six months. Grindr® emerges as the most central application, with others playing secondary roles. Predictive models based on k-means clustering and Girvan-Newman analysis provide complementary insights into structural and individual risk factors.

Overall, the results confirm that MSM sexual networks are dynamic and heterogeneous, shaped by sociodemographics and substance use. The thesis underscores the value of integrating SNA, AI, and qualitative approaches to design targeted interventions, focusing on key nodes, influential platforms, and culturally embedded language. These findings provide a foundation for more effective and inclusive public health strategies for the MSM population.

Robotic systems, and any system with a certain degree of intelligence that interacts with the environment, need to receive information from the surrounding medium. This information comes from a set of sensors and provides the most accurate possible knowledge of the environment in which they operate, which is usually unknown. The mobile agent must generate the most accurate map based on the data provided by the sensors and simultaneously estimate its position within it. This process is known as Simultaneous Localization and Mapping (SLAM). The quality of the generated maps, as well as the mobile’s position and the associated uncertainty, are conditioned by the accuracy of the data acquired by the sensors.

Optical remote sensing systems such as Light Detection and Ranging (LiDAR) perform polar measurements of distance and angle, characterizing the elements of their environment. Distance measurement is carried out by evaluating the reflected infrared radiation. The accuracy of their measurements, compared to other positional measurement systems, along with the high density of data they provide, makes them increasingly common sensors in robotic systems, autonomous vehicles, or consumer electronic devices. The information they provide enables the development of detailed maps and the estimation of the robotic system’s position on such maps. However, the large volume of information, while offering a high level of detail, significantly increases the computational load of the SLAM process.

Two research approaches are identified in the processing of LiDAR data within the SLAM domain. Some researchers work directly with the full volume of data, seeking the best match between scans by evaluating the distance between point clouds and minimizing it using mathematical optimization techniques. Others simplify the sensor information to a reduced set of geometric features that uniquely define the mapped profile.

This doctoral thesis is framed within the mapping stage of the SLAM process based on information provided by two-dimensional LiDAR sensors. The first general objective focuses on simplifying the environment into a polygonal set of straight segments and virtual intersections capable of defining it in the most efficient and accurate way possible. To this end, two methods are proposed for extracting geometric primitives that uniquely define the environment. The proposed methodologies are Weighted Conformal LiDAR-Mapping (WCLM) and Conditional Weighted Linear Fitting (CWLF). These are compared, in computational and uncertainty terms, with two classical reference methods. The WCLM methodology estimates the parameters that form the straight segments of the profile and their intersections in the inverse complex domain. This method assigns greater weight to points closer to the line compared to distant ones, following a bivariate distribution. As for the CWLF methodology, it obtains the same characteristic elements as its WCLM counterpart, but in this case through a conditional linear regression, assuming zero error in the independent variable. The probability distribution followed by CWLF is conditional, simplifying the methodology compared to WCLM.

The second general objective of this research is to evaluate the influence of the scanning frequency in 2D LiDAR sensors during the feature extraction process of the environment. The quality of the detected significant points, as well as the computational load of the process, are affected by the value of this frequency. This parameter, which is variable in some commercial sensors, must be adjusted according to application requirements.

The analysis of the integration of Robot Operating System (ROS) with MATLAB in the teleoperation of a robotic system equipped with a camera and 2D LiDAR sensor constitutes the last general objective. The remote control of the system is carried out entirely from MATLAB. To this end, a human–machine interface has been developed where data acquired by the sensors can be visualized, along with the data processed by ROS or MATLAB, as well as the implementation of a virtual joystick to control the robot remotely. Among the data processed in MATLAB are the features extracted with the WCLM methodology, represented in real time on the environment map.