Paper 1058-2021
Abstract
Patient narratives reported in clinical study reports (CSRs) provide clinical evidence of adverse events that occurred to a patient and help scientific reviewers during pharmacovigilance activities. The manual review of these narratives is a daunting task for safety reviewers as it is time consuming and resource intensive. How can we improve the efficiency of identifying safety signals from patient narratives? Can deep learning technology help to overcome the review challenges in an automated way? This paper suggests an implementation to accurately categorize one adverse event term, Serotonin Syndrome, as an example of what SAS deep learning technology is capable of. We first generate sentence level embeddings from terms contained in patient narratives. Following this, we generate term embeddings within a SAS deep learning framework. Subsequently, we obtain a category decision on whether or not the narrative text relates to Serotonin syndrome as an output. Finally, we compare this method to other deep learning and machine learning methods, including the SAS supervised Boolean rule builder algorithm, which provide a layer of interpretability. We expect that use of a Deep Learning methodology shall improve the accuracy of the medical coding (example MedDRA coding) process for adverse events. It will also help in identifying drug-event pairs, drug interactions, and clinical evidence from narratives benefiting safety reviewers during the safety review process.
PRESENTATION
Watch Developing a SAS Deep Learning MedDRA encoder (MedDRA-DeepCoder) for Patient Narratives as presented by the author on the SAS Users YouTube channel.
