Named Entity Recognition for Case Narratives of Adverse Event Reports

In the field of pharmacovigilance (PV), signal detection and assessment activities play a crucial role. They require a PV assessor to read through countless adverse event reports which is manual labor-intensive work. To ease the reading process, visual highlighters can be provided by leveraging natural language processing techniques. These can help in focusing on keyinformation in free-text narratives of adverse event reports. This thesis project is an attempt to address how can the existing information extraction tools be fined-tuned to support signal assessment of these adverse event report narratives. To accomplish this, an annotation guideline and gold-standard dataset were created. Information extraction tools namely MedspaCy, CLAMP and Stanza were explored. Named entity recognition models were developed using these tools to extract five entities of interest namely Adverse Event, Drug, Negation, Date and Problem. Subsequently, the models were evaluated on common performance metrics resulting in the highest scoring model with 81.98% F1 score. Additionally, an interactive user interface was also developed for these named entity recognition models and the assessment of its impact on PV assessors was identified as a potential avenue for future work