TimEHR: Image-Based Time Series Generation for Electronic Health Records.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Machine Learning

Background:

• Electronic Health Records (EHRs) contain complex time series data.
• Challenges in EHR time series include irregular sampling, missing values, and high dimensionality.
• Existing generative models struggle to accurately represent EHR data.

Purpose of the Study:

• To propose a novel generative adversarial network (GAN) model named TimEHR.
• To address the unique challenges of generating time series data from EHRs.
• To improve the fidelity, utility, and privacy of generated EHR data.

Main Methods:

• TimEHR utilizes a novel approach by treating time series data as images.
• The model employs 2D convolutional kernels for data representation.
• It is based on two conditional GANs: one for missingness patterns and another for time series values.

Main Results:

• TimEHR demonstrates superior performance compared to state-of-the-art methods.
• Evaluations were conducted on three real-world EHR datasets.
• The model excels in fidelity, utility, and privacy metrics.

Conclusions:

• TimEHR is an effective GAN model for generating synthetic EHR time series data.
• The proposed method successfully handles missingness patterns and irregular sampling.
• TimEHR offers a promising solution for data augmentation and privacy preservation in healthcare AI.