Vrednotenje metod strojnega učenja na enotni bazi podatkov o možganski kapi

Stroke is one of the leading causes of death and disability, but the development of predictive models using machine learning (ML) has the potential to reduce the number of fatalities. However, progress is hampered by a lack of high-quality public datasets and challenges in research reproducibility. This thesis presents a framework for evaluating ML models on public stroke data. We collected several public datasets and used nested cross-validation to evaluate different ML algorithms. We created a semantic model based on public ontologies (OntoExp, Schema.org) to document the entire experimental process, making the data and results FAIR (findable, accessible, interoperable, reusable). The annotated data is stored in a public knowledge graph, accessible via SPARQL endpoint. For easier access we developed an interactive online catalog (http://semantichub.ijs.si/StrokeBench/), which allows data to be explored without technical knowledge. The framework enables the construction of reliable artificial intelligence for predicting stroke. Future work will add new datasets, advanced models, and a natural language interface powered by large language models for easier data querying.