A Method for Extracting Dissatisfaction Entities in the Pharmaceutical Sector Using Large Language Model

Abstract

In today’s VUCA environment, pharmaceutical companies face mounting pressure to enhance both innovation and patient satisfaction amid shrinking domestic markets and intensifying global competition. Despite the growing emphasis on Patient Centricity—the integration of patients' voices into the drug development process—Japan still lags in the systematic incorporation of patient feedback. This study proposes a novel, patient-centered approach that leverages large language models (LLMs) and prompt engineering to extract dissatisfaction entities from patient-generated content, automatically generate a Dissatisfaction Dictionary, and visualize the interrelationships between these entities. Using the “Medical & Welfare_Pharmaceuticals” category from the Dissatisfaction Survey Dataset provided by the National Institute of Informatics, we analysed 300 patient comments with a Conversation Chain built on the GPT-4o API and LangChain. As a result, dissatisfaction entities were extracted and vectorized using TF-IDF and cosine similarity to form thematic clusters. This analysis revealed interconnected concerns related to pricing, efficacy, usability, and medical service processes. For example, discrepancies in pricing and perceived ineffectiveness of generics frequently co-occurred with complaints about pharmacist communication. Our findings offer pharmaceutical firms a systematic, scalable framework to reflect patient dissatisfaction in drug development, thereby enhancing patient engagement, satisfaction, and strategic alignment with Patient Centricity principles.