Improving Hospital Mortality Prediction with Medical Named Entities and Multimodal Learning

摘要: Clinical text provides essential information to estimate the acuity of a patient during hospital stays in addition structured clinical data. In this study, we explore how can complement predictive learning task. We leverage an internal medical natural language processing service perform named entity extraction and negation detection on notes compose selected entities into new corpus train document representations. then propose multimodal neural network jointly time series signals unstructured representations predict in-hospital mortality risk for ICU patients. Our model outperforms benchmark by 2% AUC.