Publications

Systematic review of AI models for predicting hospital emergency department demand, covering articles published between 2019 and July 2025 in five databases (PubMed, IEEE, Springer, ScienceDirect and ACM). Across the 11 studies included, machine learning and deep learning models outperform classical ARIMA/SARIMA, especially with non-linear, high-dimensional data, and adding external variables (weather, air quality, calendar) consistently improves results (MAPE between 3% and 18%). Limited external validation and lack of interpretability remain the main barriers to clinical adoption.

Analysis of the spatio-temporal patterns of emergency department visits at Hospital Royo Villanova (Zaragoza) between 2011 and 2024, combining clinical, demographic, environmental and socio-economic data. Using geospatial analysis and machine learning (XGBoost interpreted with SHAP), the time of day turns out to be the most influential variable for every diagnosis, while temperature, humidity and pollutants such as NO₂, SO₂ and O₃ shape respiratory and injury-related visits. Persistent high-demand areas and structural changes after COVID-19, especially in paediatrics, are identified.

Efficient planning of health transport is key to improving access and quality of care. Using seven years of health transport data from Asturias, the study applies data mining techniques to identify the most relevant variables for predicting demand. Selecting the significant variables reduces computational cost without losing accuracy, which improves the management of medical transport resources and the planning of emergency services.

Drowning is a major public health problem in childhood. This study reviews 100 paediatric drowning cases (aged 0–14) attended by the Galician emergency services between 2004 and 2020. Lack of supervision was involved in 55% of cases; younger children drowned mainly in swimming pools and adolescents on beaches. Bystander CPR was performed in 42% of cases, the mean emergency response time was 12 minutes and survival at hospital discharge reached 77%, with pneumonia as a common complication.

Simulation pilot trial with 21 participants without basic life support training, guided by a beta chatbot with speech recognition and machine learning. 91% completed the cardiac arrest recognition sequence correctly and all of them called 112; 62% placed their hands in the right spot. One third achieved high-quality CPR (mean depth of 42 mm and 100 compressions/min). The voice assistant is useful for lay people, although it needs shorter response times and better feedback on compressions.

This work covers the design and integration of a process control system based on the digital twin concept for the agri-food industry: a virtual replica of the production process, fed with plant data, that makes it possible to monitor, simulate and optimise operations on the way to a more efficient, digitalised industry.

Pilot study exploring the use of a chatbot as an assistance tool so that people without medical training can recognise a cardiac arrest and perform the first resuscitation manoeuvres. It analyses how conversational technology can guide the initial response to a cardiac emergency step by step until the emergency services arrive.