Natural language generation for electronic health records

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Health Data Science

Background:

• Generating fully-synthetic, representative electronic health records (EHRs) is crucial for data sharing and research.
• Existing synthetic EHR methods struggle to generate unstructured clinical text, such as chief complaints.

Purpose of the Study:

• To develop a deep learning model capable of generating synthetic chief complaints from discrete EHR variables.
• To assess the model's ability to preserve epidemiological information and de-identify patient data.

Main Methods:

• Utilized an encoder-decoder deep learning model, common in machine translation.
• Trained the model end-to-end on authentic EHR records linking discrete variables to chief complaints.

Main Results:

• The model successfully generated realistic synthetic chief complaint text.
• Generated text preserved epidemiological information and was free of abbreviations, misspellings, and personally identifiable information (PII).

Conclusions:

• The encoder-decoder model shows promise for generating de-identified, synthetic clinical text.
• This approach, combined with methods like GANs, can facilitate the creation of fully-synthetic EHRs for improved data sharing and privacy.