Frontiers in Drug Safety and Regulation (Feb 2023)
Natural language processing for automated triage and prioritization of individual case safety reports for case-by-case assessment
Abstract
Objective: To improve a previously developed prediction model that could assist in the triage of individual case safety reports using the addition of features designed from free text fields using natural language processing.Methods: Structured features and natural language processing (NLP) features were used to train a bagging classifier model. NLP features were extracted from free text fields. A bag-of-words model was applied. Stop words were deleted and words that were significantly differently distributed among the case and non-case reports were used for the training data. Besides NLP features from free-text fields, the data also consisted of a list of signal words deemed important by expert report assessors. Lastly, variables with multiple categories were transformed to numerical variables using the weight of evidence method.Results: the model, a bagging classifier of decision trees had an AUC of 0.921 (95% CI = 0.918–0.925). Generic drug name, info text length, ATC code, BMI and patient age. were most important features in classification.Conclusion: this predictive model using Natural Language Processing could be used to assist assessors in prioritizing which future ICSRs to assess first, based on the probability that it is a case which requires clinical review.
Keywords