Uitbraakbeheer optimaliseren met fylogenetica, ensemblemethoden en taalmodellen

During a surge of infections, testing and tracing centers serve two goals: mitigating the spread of the pathogen and gathering individual data for epidemiological surveillance. However, previous outbreaks have shown that such public health strategies do not always scale: as cases rise, some infections remain undetected, and reconstructing transmission chains becomes increasingly challenging for contact tracers. In this thesis, we sought to address these limitations through three retrospective studies analyzing individual-level data collected from the student population in Leuven during the COVID-19 pandemic, drawn from contact tracing, diagnostics, and self-reports. In the first two studies, we showed that gradient-boosting and a long short-term memory network yield promising results in diagnostic prediction. First, we developed a configurable pre-test triage relying on predictive diagnosis; our simulations suggest that its implementation could reduce testing requirements and the effective reproduction number. Then, we integrated the long short-term memory network into a conversational framework, and showed that retrieval-augmented generation techniques could help refine the risk of infection by asking relevant, personalized questions. In a third study, we developed a methodology to assess the precision of contact tracing networks in identifying transmission events; our analyses suggest that genomically consistent case-contact pairs lead to better estimation of serial intervals. In conclusion, phylogenetics and machine learning offer strong potential for improving the test-trace-isolate strategy. Provided that genomic data and diagnostic predictive models can be deployed quickly, they could help control and monitor outbreaks more effectively and support evidence-based public health policy.