An Imputation Approach to Electronic Medical Records Based on Time Series and Feature Association

Due to the interruption of network transmission, the collected electronic medical records are usually incomplete data. Therefore, the imputation of missing values is of great significance, whose main challenge is to accurately predict the missing data. Current state-of-the-art methods often employ time relationships to predict the missing data, while ignoring the correlation of similar features. This motivates us to explore the association relationships of similar features, and fuse them with time series data mining. In this paper, we propose a novel time series and feature association approach, GRU-RMF, which can effectively improve the accuracy of missing value imputation in electronic medical records. In our GRU-RMF, we utilize the matrix factorization principle to decompose electronic medical records into time feature matrix and spatial feature matrix; we design a nonlinear regularized GRU deep neural network, which can learn the time relationship in the time feature matrix; we design a time-space feature fusion method with alternate solving to organically fuse the feature correlation and time correlation, which can improve the comprehensive accuracy of missing value imputation. The experimental results on several public datasets show that the proposed GRU-RMF has the advantages of higher imputation accuracy and better scalability.