We present an intelligent system designed for the environment of a healthcare facility, such as a hospital or a nursing home. The aim of the system is to log the services carried out by the service provider (a nurse) without the need to choose individual items from a series of menus, which is currently the standard procedure. Instead, we used a voice recording of the nurses reporting their actions in a natural language, i.e. without particular voice commands, coupled with a fine-tuned Large Language Model (LLM). We discuss legal aspects related to recording and processing of conversations in a healthcare facility. We outline the implementation and present the results of tests using different types of LLMs. We focus on the Slovenian language, which can be seen as a medium-sized language, instead of English which is widely supported by LLMs. Our approach produces up to 97% both in precision and recall.
With the increasing adoption of AI in safety-critical applications within urban environments, the interpretability of these systems is paramount. This study explores the application of Explainable Artificial Intelligence (XAI) techniques to enhance transparency in audio-based detection of emergency vehicle sirens, a crucial component in urban sound management. Adopting methods such as SHAP (SHapley Additive exPlanations) values, Permutation Feature Importance, and model-specific feature scores, this research identifies key audio features, including mid-frequency spectral contrasts and targeted chroma components, which significantly help in distinguishing siren sounds among urban noise. The study examines various machine learning models, identifying K-Nearest Neighbors (KNN) and XGBoost as top performers; KNN excelled in class-specific precision, while XGBoost demonstrated strong cross-class discrimination. The findings highlight the potential of XAI in improving both accuracy and accountability for sound detection systems in safety-critical urban applications, advancing the deployment of transparent AI within smart city infrastructures.
Cyclist safety in urban areas is a growing concern that requires innovative planning and citizen participation approaches. This research introduces SPUR, a serious gaming application designed to evaluate cyclist safety on a city scale through a data-driven traffic infrastructure analysis and immersive evaluation of the current state of the traffic network and planned alternatives. The concept of SPUR introduces a digital city twin that uses data from city models, road infrastructure, point clouds, geo-referenced statistics, and citizen feedback to identify vulnerabilities and guide towards safer planning alternatives. Users are given assistance for interactively generating planning scenarios in an editor and can create simulations that are made tangible with a VR setup with a functional bicycle and a traffic simulation based on SUMO, ensuring realistic scenarios for the evaluation. While this approach focuses on cyclist safety, the workflow is generally applicable to other use cases in urban and traffic infrastructure planning.
Our research proposes a novel methodology for the design of immersive learning environments that prioritize accessibility and inclusion of adults with various disabilities. In contrast to existing hyper-specialized approaches that lack generalizability, it integrates Universal Design for Learning and participatory design to balance scalability and personalization. The iterative process promotes adaptable and universally accessible solutions through multidisciplinary collaboration and empirical validation. Preliminary results confirm the potential of Extended Reality to overcome pedagogical, spatial and social barriers and underline the need for standardized, flexible design guidelines.
Artificial Intelligence is playing an increasingly prominent role in healthcare. For this integration to continue evolving effectively, professionals in the field, including doctors, psychologists, and nurses, must trust the technology. In the domain of Large Language Models, this trust requires transparency in the sources of information, ensuring they are verifiable and reviewed. This paper presents an intelligent diagnostic assistant designed to support differential diagnosis and disease comparison, leveraging an LLM enriched with the Retrieval-Augmented Generation technique to address trust-related challenges. The knowledge base for our system is PubMed, a biomedical article aggregator, which supplies relevant articles on the considered disorders, as well as evidence supporting or refuting potential links between them. Retrieval-Augmented Generation empowers the model to incorporate external knowledge tailored to the specific question while providing citations for its statements. These citations not only enhance the trustworthiness of the answers but also enable practitioners to explore the referenced articles in greater detail. We developed a prototype of our system, and conducted an evaluation through questionnaires administered to 12 professional psychologists. Results show that our system compares favorably with state of the art Large Language Models in providing the relevant information to distinguish between two diseases, and that the provided details are useful for supporting differential diagnosis.
