Are synthetic clinical notes useful for real natural language processing tasks: A case study on clinical entity...

Area of Science:

• Computational Linguistics
• Medical Informatics
• Artificial Intelligence

Background:

• Clinical natural language processing (NLP) development requires extensive clinical data.
• Privacy and security concerns limit public access to clinical documents.
• Generating synthetic clinical data is a potential solution to data scarcity.

Purpose of the Study:

• To develop and evaluate methods for generating synthetic clinical notes.
• To assess the utility of synthetic notes in training clinical NLP models, specifically Named Entity Recognition (NER).

Main Methods:

• Implemented four text generation models: CharRNN, SegGAN, GPT-2, and CTRL.
• Generated synthetic clinical text for the History and Present Illness section.
• Trained NER models on natural, synthetic (GPT-2), and combined corpora.
• Evaluated NER model performance on independent natural clinical corpora.

Main Results:

• GPT-2 achieved the highest BLEU score for text generation.
• NER models trained on GPT-2 synthetic data showed comparable or slightly superior performance to those trained on natural data (F1 scores up to 0.748).
• Combining natural and synthetic corpora further improved NER model performance.

Conclusions:

• Synthetic clinical notes generated by advanced models like GPT-2 are valuable for training clinical NER models.
• This approach can mitigate privacy concerns and increase data accessibility for NLP research.
• Further research is needed to integrate synthetic data generation into practical clinical NLP applications.