Concurrent Imputation and Prediction on EHR data using Bi-Directional GANs: Bi-GANs for EHR imputation and prediction

Area of Science:

• Health Informatics
• Machine Learning
• Biomedical Data Science

Background:

• Electronic Health Records (EHRs) present significant challenges for predictive modeling due to variable lengths, observation counts, and missing entries.
• These data irregularities hinder the accuracy and reliability of models trained on EHR data.
• Existing methods often struggle to address both data imputation and prediction tasks simultaneously for complex EHR datasets.

Purpose of the Study:

• To develop a unified model capable of both imputing missing values and predicting future outcomes within irregularly observed, variable-length EHR data.
• To address the limitations of current approaches by offering a single, integrated solution for EHR data challenges.
• To enhance the performance of predictive models by effectively handling missing data and temporal variations in EHRs.

Main Methods:

• Proposed a novel model, Bi-GAN (Bidirectional Generative Adversarial Network), integrating a bidirectional recurrent network within a generative adversarial framework.
• The generator network imputes missing EHR values, while the discriminator network distinguishes between real and generated data.
• The model learns to impute and predict using the complete input data, accommodating varying time steps and missing entries.

Main Results:

• Bi-GAN demonstrated superior performance in imputing and predicting Body Mass Index (BMI) values across two large EHR datasets.
• The model successfully handled time-series data of varying lengths with missing entries, outperforming state-of-the-art methods.
• Evaluations confirmed the model's effectiveness for both short-term and long-term predictions without requiring predefined window lengths during training.

Conclusions:

• Bi-GAN offers a significant advancement by performing both imputation and prediction tasks within a single model for EHR data.
• The model's ability to handle variable-length time-series with missing data makes it highly adaptable for diverse clinical applications.
• This approach provides a robust solution for improving the utility of EHR data in developing accurate predictive health models.