CSAI: Conditional Self-Attention Imputation for Healthcare Time-series.

Area of Science:

• Machine Learning
• Biomedical Informatics
• Data Science

Background:

• Electronic Health Records (EHRs) generate complex multivariate time series data.
• Missing data is a significant challenge in EHR time series, impacting analysis and model performance.
• Existing imputation methods often fail to capture the unique characteristics of clinical data.

Purpose of the Study:

• To introduce the Conditional Self-Attention Imputation (CSAI) model, a novel recurrent neural network for EHR time series imputation.
• To address limitations of current imputation techniques by incorporating EHR-specific features.
• To improve the accuracy and reliability of data restoration in partially observed EHR datasets.

Main Methods:

• Developed CSAI, a recurrent neural network incorporating attention-based hidden state initialization.
• Implemented domain-informed temporal decay to reflect clinical recording patterns.
• Utilized a non-uniform masking strategy to model non-random missingness in EHR data.
• Evaluated CSAI on four benchmark EHR datasets.

Main Results:

• CSAI demonstrated superior performance in data restoration compared to state-of-the-art architectures.
• The model showed effectiveness in downstream tasks utilizing imputed EHR data.
• Evaluations confirmed CSAI's ability to handle long- and short-range temporal dependencies and non-random missingness.

Conclusions:

• CSAI offers a significant advancement in neural network-based imputation for EHR data.
• The model's design aligns algorithmic imputation more closely with clinical realities.
• CSAI is available in the open-source PyPOTS toolbox, facilitating its application in time series analysis.