A Multi-directional Approach for Missing Value Estimation in Multivariate Time Series Clinical Data

Abstract

Missing values are common in clinical datasets which bring obstacles for clinical data analysis. Correctly estimating the missing parts plays a critical role in utilizing these analysis approaches. However, only limited works focus on the missing value estimation of multivariate time series (MTS) clinical data, which is one of the most challenge data types in this area. We attempt to develop a methodology (MD-MTS) with high accuracy for the missing value estimation in MTS clinical data. In MD-MTS, temporal and cross-variable information are constructed as multi-directional features for an efficient gradient boosting decision tree (LightGBM). For each patient, temporal information represents the sequential relations among the values of one variable in different time-stamps, and cross-variable information refers to the correlations among the values of different variables in a fixed time-stamp. We evaluated the estimation method performance based on the gap between the true values and the estimated values on the randomly masked parts. MD-MTS outperformed three baseline methods (3D-MICE, Amelia II and BRITS) on the ICHI challenge 2019 datasets that containing 13 time series variables. The root-mean-square error of MD-MTS, 3D-MICE, Amelia II and BRITS on offline-test dataset are 0.1717, 0.2247, 0.1900, and 0.1862, respectively. On online-test dataset, the performance for the former three methods is 0.1720, 0.2235, and 0.1927, respectively. Furthermore, MD-MTS got the first in ICHI challenge 2019 among dozens of competition models. MD-MTS provides an accurate and robust approach for estimating the missing values in MTS clinical data, which can be easily used as a preprocessing step for the downstream clinical data analysis.