A Study of Missing Data Processing Method Based on Long Short Term Memory Neural Network for Electronic Health Record

Abstract

Data-driven healthcare is considered as a promising technology of health care reform and Electronic Health Record (EHR) is an important vehicle. However, EHR are characterized by high dimensionality, temporality, sparsity and so on and it is hard for traditional deep learning algorithms to directly use sparse EHR data. In this paper, we first select the medical information mart for intensive care (MIMIC-III) database to detect the test data after patient admission within 48 hours. Then we use Long Short Term Memory Neural Network (LSTM) to learn the characteristic change model of existing data and apply the learned model to generate the missing values. Finally, the performance of the missing data processing method is verified by the prediction results of the classification model on patient mortality. Experimental results demonstrate that LSTM is an effective method for filling in missing data and the filled data based on LSTM is superior to the data filled by Linear Regression (L), K Nearest Neighbor (KNN) and Forward Padding (F) in predicting patient death outcomes.