Attention-based Imputation of Missing Values in Electronic Health Records Tabular Data.

Area of Science:

• Machine Learning
• Biostatistics
• Data Science

Background:

• Imputing missing values (IMV) in electronic health records (EHR) tabular data is essential for accurate patient-specific predictive modeling.
• Existing deep learning methods struggle with tabular data, limiting their success in IMV.
• Current statistical and machine learning imputation methods have limitations in handling complex data patterns.

Purpose of the Study:

• To propose a novel attention-based imputation framework for reconstructing missing values in tabular EHR data.
• To enhance the generalization of imputation models using contrastive learning techniques.
• To evaluate the performance of the proposed method against state-of-the-art imputation techniques.

Main Methods:

• Developed a novel attention-based missing value imputation framework leveraging self-attention or between-sample attentions.
• Employed data manipulation strategies from contrastive learning to improve model generalization.
• Utilized five diverse tabular datasets and two EHR datasets for comprehensive evaluation.

Main Results:

• The proposed self-attention imputation method significantly outperformed existing statistical and decision-tree-based imputation methods.
• Achieved substantial reductions in normalized root mean squared error (RMSE): 18.4%–74.7% on general tabular data and 52.6%–82.6% on EHR data.
• Demonstrated superior performance across various missingness levels (10%–50%) under a missing completely at random assumption.

Conclusions:

• The attention-based imputation framework offers a superior approach for handling missing values in tabular EHR data.
• The method shows significant potential for improving patient-specific predictive modeling through more accurate data reconstruction.
• This work advances the field of IMV for tabular data, particularly in healthcare applications.