Context-Aware Time Series Imputation for Multi-Analyte Clinical Data.

Area of Science:

• Biomedical Informatics
• Data Science
• Machine Learning

Background:

• Clinical time series imputation is crucial for data analytics but often relies on limiting assumptions or local data properties.
• Existing models struggle to effectively incorporate global dependencies within patient health trajectories.
• Advancing imputation techniques is vital for reliable clinical data interpretation.

Purpose of the Study:

• To develop a novel imputation model for clinical time series data that addresses limitations of current methods.
• To improve the accuracy and reliability of missing data imputation in healthcare.
• To leverage global dependencies within patient data for enhanced imputation.

Main Methods:

• Proposed Context-Aware Time Series Imputation (CATSI), a bidirectional LSTM framework.
• Introduced a "global context vector" to capture patients' overall health states from entire time series.
• Incorporated a cross-feature imputation component to model complex feature correlations.

Main Results:

• Achieved a normalized root mean square deviation (nRMSD) of 0.1998, outperforming state-of-the-art models by 10.6%.
• Demonstrated the effectiveness of the global context vector in generating accurate imputations, particularly for consecutive missing data.
• Validated the model's performance in the 2019 ICHI Data Analytics Challenge on Missing Data Imputation (DACMI).

Conclusions:

• CATSI offers a significant advancement in clinical time series imputation by effectively utilizing global context.
• The model's ability to capture long-range dependencies and feature correlations leads to superior imputation accuracy.
• This approach holds promise for more robust clinical data analytics and improved patient outcome prediction.