Trustworthy assertion classification through prompting.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Machine Learning

Background:

• Accurate assertion classification in clinical notes is vital for healthcare professionals.
• Existing methods face challenges with labor-intensive engineering and class bias.

Purpose of the Study:

• To develop a prompt-based learning approach for assertion classification.
• To improve accuracy, few-shot learning, and generalizability in clinical text analysis.

Main Methods:

• Treated assertion classification as a masked language auto-completion problem.
• Evaluated the prompt-based model on six diverse clinical datasets.
• Introduced rationale faithfulness metrics (comprehensiveness, sufficiency) and LIME for explainability.

Main Results:

• Achieved a micro-averaged F-1 of 0.954 on the i2b2 2010 dataset, surpassing previous works by ~1.8%.
• Demonstrated excellence in few-shot learning and strong generalizability across five external datasets.
• Showcased superior capability in identifying comprehensive (~63.93%) and sufficient (~11.75%) linguistic features.
• Exhibited better rationale agreement with human judgment (~71.93% average F-1) using LIME.

Conclusions:

• The prompt-based learning approach offers a robust and trustworthy solution for clinical assertion classification.
• This method effectively addresses limitations of traditional approaches, particularly for underrepresented classes.
• The model's enhanced explainability and generalizability pave the way for more reliable clinical NLP applications.