Detecting MRSA Infections by Fusing Structured and Unstructured Electronic Health Record Data

Methicillin-resistant Staphylococcus aureus (MRSA), an antibiotic resistant bacteria, is a common cause of one of the more devastating hospital-acquired infections (HAI) in the United States. In this work, we study the practicality of leveraging machine learning methods for early detection of MRSA infections based on a rich variety of patient information commonly available in modern Electronic Health Records (EHR). We explore heterogeneous types of data in EHRs including on-admission demographics, throughout-stay time series and free-form clinical notes. On-admission data capture non-clinical information (e.g., age, marital status) while Throughout-stay data include vital signs, medications, laboratory studies, and other clinical assessments. Clinical notes, free-from text documents created by medical professionals, contain expert observations about patients. Our proposed system generates dense patient-level representations for each data type, extracting features from each of our data types. It then generates scores for each patient, indicating their risk of acquiring MRSA. We evaluate prediction performance achieved by core Machine Learning methods, namely Logistic Regression, Support Vector Machine, and Random Forest, when mining these different types of EHR data retrospectively to detect patterns predictive of MRSA infection. We evaluate classification performance using MIMIC III – a critical care data set comprised of 12 years of patient records from the Beth Israel Deaconess Medical Center Intensive Care Unit in Boston, MA. Our experiments show that while all types of data contain predictive signals, the fusion of all sources of data leads to the most effective prediction accuracy.