HealthPrompt: A Zero-shot Learning Paradigm for Clinical Natural Language Processing

Area of Science:

• Computational linguistics
• Medical informatics
• Artificial intelligence in healthcare

Background:

• Clinical natural language processing (NLP) development requires large annotated datasets, which are scarce and costly to produce.
• The absence of accessible annotated clinical text is a significant barrier to advancing clinical NLP systems.
• Zero-Shot Learning (ZSL) offers a potential solution by enabling models to classify unseen data.

Purpose of the Study:

• To introduce HealthPrompt, a novel prompt-based framework for clinical NLP.
• To evaluate the efficacy of prompt-based learning in a zero-shot setting for clinical text analysis.
• To assess HealthPrompt's performance across various pre-trained language models (PLMs) without requiring training data.

Main Methods:

• Development of the HealthPrompt framework utilizing prompt-based learning for NLP tasks.
• Application of prompt-based learning to clinical text data, focusing on task definition tuning via prompt templates.
• In-depth analysis and evaluation of HealthPrompt on six different PLMs in a no-training-data scenario.

Main Results:

• HealthPrompt effectively captures contextual information within clinical texts.
• The framework demonstrates strong performance on clinical NLP tasks despite the absence of task-specific training data.
• Prompt-based learning proved viable for ZSL in the clinical domain.

Conclusions:

• HealthPrompt offers a promising approach to overcome data limitations in clinical NLP.
• Prompt-based learning can be successfully applied to clinical text for zero-shot classification tasks.
• This framework facilitates the development of clinical NLP tools without the need for extensive annotated datasets.