Improving clinical outcome predictions using convolution over medical entities with multimodal learning

Abstract

Early prediction of mortality and length of stay (LOS) of a patient is vital for saving a patient's life and management of hospital resources. Availability of Electronic Health Records (EHR) makes a huge impact on the healthcare domain and there are several works on predicting clinical problems. However, many studies did not benefit from the clinical notes because of the sparse, and high dimensional nature. In this work, we extract medical entities from clinical notes and use them as additional features besides time-series features to improve proposed model predictions. The proposed convolution based multimodal architecture, which not only learns effectively combining medical entities and time-series Intensive Care Unit (ICU) signals of patients but also allows to compare the effect of different embedding techniques such as Word2vec and FastText on medical entities. Results show that the proposed deep multimodal method outperforms all other baseline models including multimodal architectures and improves the mortality prediction performance for Area Under the Receiver Operating Characteristics (AUROC) and Area Under Precision-Recall Curve (AUPRC) by around 3%. For LOS predictions, there is an improvement of around 2.5% over the time-series baseline. The code for the proposed method is available at https://github.com/tanlab/ConvolutionMedicalNer.