The aluminum standard: using generative Artificial Intelligence tools to synthesize and annotate non-structured...

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Medical Informatics

Background:

• Medical narratives are crucial for patient diagnosis but are often vague and difficult to categorize.
• A significant challenge exists in training AI models for clinical narrative analysis due to the lack of high-quality, non-English datasets and data privacy concerns.
• Existing datasets like MIMIC are English-specific and may introduce bias when applied to different medical domains, such as oncology.

Purpose of the Study:

• To develop a method for generating high-quality synthetic clinical narratives in German.
• To overcome limitations in data availability for training AI models in specific medical fields and languages.
• To facilitate the creation of machine-readable patient data for improved interoperability and model training.

Main Methods:

• Utilized generative AI workflows to synthesize German clinical narratives, ensuring patient data privacy.
• Developed high-quality medical prompts specifying main and co-diseases, with disease frequencies derived from hospital data to mirror real patient cohorts.
• Validated the synthetic narrative quality by annotating them to train a Named Entity Recognition (NER) algorithm.

Main Results:

• A Named Entity Recognition (NER) model was trained using the synthetic data.
• Performance metrics including precision, recall, and F1 scores were reported for the NER model, considering exact and partial entity matches.
• The trained NER model demonstrated cautious performance, achieving a precision of up to 0.8 for Entity Type match and an F1 score of 0.3.

Conclusions:

• The generative AI approach shows potential for creating synthetic clinical narratives, enhancing data interoperability across languages and regions without compromising patient safety.
• This method facilitates the synthesis of unstructured patient data, accelerating the identification and training of AI models.
• The technology can be extended to generate discharge letters for various disease combinations, potentially saving healthcare professionals significant time.