Rethinking domain adaptation for machine learning over clinical language.

Area of Science:

• Natural Language Processing
• Machine Learning
• Biomedical Informatics

Background:

• Clinical natural language processing (NLP) systems require adaptability to diverse datasets due to high data acquisition costs.
• Existing domain adaptation methods often overlook practical constraints of clinical data sharing.

Purpose of the Study:

• To categorize domain adaptation strategies based on data shareability in clinical NLP.
• To identify and emphasize under-studied, realistic clinical NLP scenarios.
• To propose recommendations for advancing clinical NLP research and practice.

Main Methods:

• Developed a taxonomy of domain adaptation techniques, parameterized by data shareability.
• Analyzed the landscape of existing domain adaptation research in the context of clinical data.
• Identified gaps in research concerning realistic clinical data sharing scenarios.

Main Results:

• The most practical settings for clinical NLP domain adaptation are significantly under-researched.
• Current widely studied domain adaptation paradigms do not adequately address clinical data sharing limitations.
• A critical need exists for research focusing on domain adaptation under restricted data sharing.

Conclusions:

• Future clinical NLP research should prioritize domain adaptation strategies suitable for real-world data sharing constraints.
• Recommendations are provided to enhance the utility of data, shared tasks, and models in clinical NLP.
• The clinical NLP community can lead advancements in broader NLP and machine learning by addressing these specific challenges.