BackgroundElderly patients are at increased risk for Adverse Drug Events (ADEs). Proactively screening elderly people visiting the emergency department for the possibility of their hospital admission being drug-related helps to improve patient care as well as prevent potential unnecessary medical costs. Existing routine ADE assessment heavily relies on a rule-based checking process. Recently, machine learning methods have been shown to be effective in automating the detection of ADEs, however, most approaches used only either structured data or free texts for their feature engineering. How to better exploit all available EHRs data for better predictive modeling remains an important question. On the other hand, automated reasoning for the preventability of ADEs is still a nascent line of research. MethodsClinical information of 714 elderly ED-visit patients with ADE preventability labels was provided as ground truth data by Jeroen Bosch Ziekenhuis hospital, the Netherlands. Methods were developed to address the challenges of applying feature engineering to heterogeneous EHRs data. A Dual Autoencoders (2AE) model was proposed to solve the problem of imbalance embedded in the existing training data. ResultsExperimental results showed that 2AE can capture the patterns of the minority class without incorporating an extra process for class balancing. 2AE yields adequate performance and outperforms other more mainstream approaches, resulting in an AUPRC score of 0.481. ConclusionsWe have demonstrated how machine learning can be employed to analyze both structured and unstructured data from electronic health records for the purpose of preventable ADE prediction. The developed algorithm 2AE can be used to effectively learn minority group phenotype from imbalanced data.
Read full abstract