Synthetic4Health

Synthetic4Health generates annotated synthetic clinical letters using pre-trained language models to provide reliable, diverse, and de-identified clinical text for training models and supporting medical research.

Key Features:

• Pre-trained language model investigation: Multiple pre-trained language models were investigated for text masking and generation, with focus on Bio_ClinicalBERT.
• Masking strategies: Different masking strategies were applied and evaluated, including masking stopwords, nouns, and verbs.
• Qualitative and quantitative evaluation: Evaluation included qualitative and quantitative assessments using BERTScore as the primary quantitative metric with other metrics as supplementary references.
• Downstream NER evaluation: Downstream named entity recognition tasks were performed to assess substitution of synthetic letters for real ones.
• Clinical-focused evaluation: Clinically focused evaluations were conducted using BioGPT and GPT-3.5-turbo.
• Entity and structure preservation: Clinical entities and document structure are preserved during generation to align with task objectives.

Scientific Applications:

• Medical model training: Synthetic clinical letters support training of machine learning models without exposing sensitive patient information.
• Medical research: De-identified synthetic letters enable research applications that are limited by privacy constraints on real clinical data.
• Named entity recognition: Synthetic letters can effectively substitute real clinical letters in downstream NER tasks.
• Dataset expansion: The framework helps expand datasets in the clinical domain by generating diverse clinical letters.

Methodology:

Multiple pre-trained language models were investigated for text masking and generation, with focus on Bio_ClinicalBERT and different masking strategies. Encoder-only and encoder-decoder models were compared. Evaluation included qualitative and quantitative assessments, downstream named entity recognition tasks, and clinically focused evaluations using BioGPT and GPT-3.5-turbo.