Ascle-A Python Natural Language Processing Toolkit for Medical Text Generation: Development and Evaluation Study.

Area of Science:

• Biomedical Informatics
• Natural Language Processing

Background:

• Medical texts present unique challenges for manual curation.
• Existing natural language processing (NLP) toolkits lack text generation capabilities.
• There is a need for integrated, user-friendly NLP solutions in the biomedical domain.

Purpose of the Study:

• Develop and evaluate Ascle, an all-in-one NLP toolkit for biomedical researchers and clinical staff.
• Introduce novel generative functions including question-answering, summarization, simplification, and machine translation.
• Integrate essential NLP functions and clinical database query capabilities into a single platform.

Main Methods:

• Fine-tuned 32 domain-specific language models evaluated on 27 benchmarks.
• Developed a retrieval-augmented generation (RAG) framework with a medical knowledge graph for question-answering.
• Conducted physician validation to assess the quality of generated content.

Main Results:

• Fine-tuned models improved machine translation by 20.27 BLEU score.
• RAG framework increased ROUGE-L score by 18% for question-answering.
• Physician validation yielded high scores for readability (4.95/5) and relevancy (4.43/5).

Conclusions:

• Ascle is a user-friendly NLP toolkit for medical text generation.
• The toolkit offers advanced generative and essential NLP functions.
• All code and models are publicly available, promoting accessibility and further research.