Automated Concept Indexing for Health Measurement Scale Items through Prompt Learning with Pre-trained Language Models

Health measurement scales are fundamental in clinical practice and health sciences research, enabling the quantification of subjective health states. The complexities involved in the discovery and organization of these scales highlight the need for advanced semantic retrieval and medical concept indexing techniques. To address the challenge, this study presents an automated concept indexing approach for health measurement scale items, utilizing prompt learning with pre-trained large language models. We constructed a specialized corpus dataset for lung cancer, comprising 347 items annotated with two-dimensional index terms to capture the measurement purpose and domain. Employing prompt learning with PubMedBERT, our approach achieved up to 94% effectiveness in low-resource data environments. The ablation study underscored the critical contribution of key components within our approach, significantly boosting concept indexing performance. Notably, our error analysis revealed the unique challenges posed by the colloquial and flexible language prevalent in scale items, differing from the conventional medical terms. This pioneering study not only improves the management and retrieval of health measurement scale resources but also sets the stage in knowledge discovery and organization at the item level, potentially transforming the utilization of these measurement tools by clinicians and researchers. Moreover, the practical implications of this work extend to health care, where enhanced accessibility and utility of scale items could lead to improved patient care and more impactful health research.