Thesis Project Form

Early differentiation between candidemia and bacteremia is crucial in patients presenting with sepsis or septic shock, as delays in initiating appropriate antifungal therapy can be fatal. Deep learning models trained on laboratory data may help identify candidemia earlier, complementing existing diagnostic markers and enhancing clinical decision support.

This thesis builds upon recent work on a deep learning model by focusing or improving its architecture and explore the training potential through synthetic data to differentiate candidemia from bacteremia using routinely collected laboratory features. The student will:
• Review the literature on AI applications in the early diagnosis of candidemia and bacteremia;
• Explore and preprocess a large dataset of automatically extracted laboratory data;
• Train and validate the improved deep learning model for candidemia prediction;
• Compare performance against baseline classifiers and specific biomarkers such as ?-D-glucan (BDG) and procalcitonin (PCT);
• Investigate model explainability to identify patterns in nonspecific laboratory features;
• Assess the potential of integrating laboratory-based models into clinical workflows as complementary tools for rapid infection diagnosis.

The student will acquire hands-on experience in applying deep learning to biomedical datasets, handling clinical laboratory data, and evaluating predictive models. They will develop skills in feature engineering, model validation, and explainability, as well as in interdisciplinary collaboration at the interface between data science and clinical infectious diseases.