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

Area of Science:

• Clinical data analysis
• Biomedical informatics
• Machine learning for healthcare

Background:

• Missing values are prevalent in clinical datasets, hindering accurate data analysis.
• Multivariate time series (MTS) clinical data presents unique challenges for imputation.
• Effective missing value estimation is critical for downstream clinical data utilization.

Purpose of the Study:

• To develop a highly accurate methodology for estimating missing values in MTS clinical data.
• To introduce the MD-MTS approach, leveraging multi-directional features for improved imputation.
• To address the limitations of existing methods in handling complex clinical time series data.

Main Methods:

• The MD-MTS methodology utilizes multi-directional features, combining temporal and cross-variable information.
• Temporal information captures sequential dependencies within a single variable.
• Cross-variable information models correlations among different variables at fixed time-stamps.
• An efficient gradient boosting decision tree (LightGBM) algorithm is employed.

Main Results:

• MD-MTS demonstrated superior performance in estimating missing values compared to baseline methods (3D-MICE, Amelia II, BRITS).
• The method achieved the lowest root-mean-square error (RMSE) on both offline (0.1717) and online (0.1720) test datasets.
• MD-MTS secured first place in the ICHI challenge 2019, surpassing numerous competing models.

Conclusions:

• MD-MTS offers an accurate and robust solution for missing value estimation in MTS clinical data.
• The approach effectively integrates temporal and cross-variable insights for enhanced imputation.
• MD-MTS can serve as a valuable preprocessing step for various clinical data analysis tasks.