Modelling Risk of Bleeding-Related Emergency Visits and Hospitalizations Using Medication and Patient History Data

Sammanfattning

Bleeding-related adverse events are a major patient safety concern and may lead to emergency hospital admissions with serious clinical consequences. Predicting bleeding risk is challenging due to the complex interaction of patient conditions, laboratory findings, medication use, and drug–drug interactions over time. Clinical decision support systems (CDSSs) are commonly used to improve medication safety through rule-based alerts; however, generic rule-based alerts may generate warnings that are not clinically relevant for the individual patient, contributing to alert fatigue. This thesis investigates whether medication-related knowledge from the Swedish Janusmed database can be combined with patient-condition information from electronic health records to model the risk of bleeding-related emergency visits. Using a cohort of 2,032 emergency visits from the Cambio COSMIC registry in Region Kalmar County, machine learning models were developed using patient-condition features, medication-related features, and combinations of both. Both aggregated patient representations and temporal event-based representations were explored using traditional machine learning and deep learning approaches. The results showed that patient-condition features consistently outperformed medication-related features alone across all modelling approaches. Combining both feature groups did not produce a statistically significant improvement in predictive performance compared with patient-condition features alone. Furthermore, clustering analysis indicated the potential for stratified risk levels that could support more targeted clinical alerts. The findings highlight both the challenges and opportunities of integrating longitudinal medication and patient data in future CDSS applications aimed at improving medication safety and alert relevance.