Extracting Pharmaceutical Safety Information from Nursing Records: Utilizing ChatGPT for Data Categorization

Abstract

This study explores the potential of nursing records in improving pharmaceutical safety by analyzing adverse effects (AEs) using ChatGPT-4 for data categorization. A total of 927 anonymized records from a Japanese community hospital were analyzed through a three-stage protocol: identifying AEs using ChatGPT-4, constructing a co-occurrence network to analyze relationships among key terms, and identifying central patterns using network metrics. Results revealed a high proportion of entries classified as "Unknown" for both AE status (674 cases) and patient conditions (865 cases), highlighting documentation gaps. The most frequently documented medicines were Red Blood Cells (123 instances) and Fresh Frozen Plasma (87 instances), while Thermal Therapy (30 instances) and CT Scans (23 instances) were the most used medical devices. However, AE documentation was limited, with only 6 cases for medicines and 2 for devices. The co-occurrence network emphasized the importance of informed consent (IC) in communicating treatment risks, particularly in chemotherapy and antibiotic use. Despite challenges, this study underscores the critical role of nursing records in AE management and calls for standardized documentation and integration. Future efforts could explore combining large language models with human-in-the-loop feedback to further improve AE extraction accuracy from incomplete nursing notes. into clinical workflows to enhance patient safety.