Zero-Shot Clinical Concept Normalization Using Prompt-Based Language Models

Abstract

This paper explores the domain of zero-shot clinical concept normalization, leveraging prompt-based language models to align unstructured clinical text with standardized medical terminologies. The approach circumvents traditional data-hungry methods by framing the normalization task as a conditional text inference problem, invoking the model's latent conceptual understanding. We address the complexity of heterogeneous medical vocabulary by prompting an underlying model to infer the most probable canonical label, given minimal or no explicitly labeled training samples. The method is grounded in the principle that each concept, represented by a textual descriptor, can be mapped onto a structured taxonomy through a contextual prompt. By directly prompting large language models with carefully designed prompts, the system capitalizes on the model’s prior knowledge, thereby enabling on-the-fly resolution of diverse clinical expressions. We propose a rigorous formal framework and employ advanced mathematical concepts to enhance interpretability, offering insights into the underlying reasoning within the model. With experiments on widely used clinical corpora, results highlight competitive performance in normalizing unseen or minimally sampled expressions. Notably, the technique addresses lexical variation and out-of-vocabulary issues by exploiting prompt-driven cross-lingual and cross-domain transfer abilities. Our findings advance the state of the art in zero-shot clinical concept normalization and pave the way for broader medical natural language processing applications.