Data mining and knowledge discovery from emergency department clinical data

Clinical pathway analysis determines which medical behaviors are essential, where the temporal ordering of these behaviors is characterized with numerical bounds. While such predefined pathways are less applicable in the Emergency Department, their analytical principles can still be used to study patient journeys, identify good practices, and improve decision support in acute care settings. Emergency Departments provide immediate care; thus, timely risk stratification is crucial to improve patient journeys. However, reliable tools for comorbidity profiling and identifying patients at risk for in-hospital mortality and healthcare-associated infections are still lacking. This thesis develops a comorbidity extraction algorithm for Italian free-text patient reports, investigates the association between spending a night in the emergency department and in-hospital mortality with its key predictors, and identifies the main predictors of developing healthcare-associated infections. To pursue these goals, we evaluated current tools for comorbidity extraction and predictors affecting undesirable outcomes in the emergency department, such as in-hospital mortality and healthcare-associated infections. We conducted experiments using large language models and regular expressions for comorbidity extraction, and Machine Learning methods including logistic regression, random forest, extreme gradient boosting, and light gradient boosting with balancing techniques. This thesis investigates approaches that both extract comorbidities and predict critical outcomes for patients in the emergency department. By identifying high-risk patients early, these methods could give clinicians valuable support in timely decisions, improving clinical trajectories and reducing preventable complications.