PharmBERT: a Fine-tuned Model for Pharmaceutical Error Prediction

Abstract

Every year, billions of prescriptions are dispensed in North America. Shockingly, medication errors result in up to 9,000 deaths annually in the United States alone. However, the current system for tracking service quality during the medication dispensation process is severely limited. It is essential to identify and understand the patterns of these errors to effectively prevent them. In this study, we employ a deep learning model called Bidirectional Encoder Representations from Transformers (BERT) to predict medication errors related to pharmacy services. Our preliminary experimental results demonstrate that our fine-tuned model achieves an impressive accuracy of approximately 88+%, accurately predicting whether a dispensation procedure will result in a near-miss (caught beforehand) or an incident (caught afterward) error. The attention scores generated by the model parameters offer valuable insights into the data features. We believe that the proposed approach can serve as a vital initial step in uncovering error patterns and ultimately contribute to reducing medication errors.