Comparison of text processing methods in social media–based signal detection. Issue 10 (7th August 2019)

• Record Type: Journal Article
• Authors: Gavrielov‐Yusim, Natalie; Kürzinger, Marie‐Laure; Nishikawa, Chihiro; Pan, Chunshen; Pouget, Julie; Epstein, Limor BH; Golant, Yan; Tcherny‐Lessenot, Stephanie; Lin, Stephen; Hamelin, Bernard; Juhaeri, Juhaeri
• Is Part Of: Pharmacoepidemiology and drug safety. Volume 28:Issue 10(2019)
• Page Start: 1309
• Page End: 1317
• Publication Date: 2019-08-07
• Subjects: adverse event (AE) -- lift -- machine learning (ML) -- natural language processing (NLP) -- pharmacoepidemiology -- signal of disproportionate reporting (SDR) -- social media (SM); Pharmacoepidemiology -- Periodicals; Chemotherapy -- Periodicals; Epidemiology -- Periodicals; 615.705
• DOI: 10.1002/pds.4857
• Languages: English
• ISSNs: 1053-8569

Abstract

Abstract: Purpose: Adverse event (AE) identification in social media (SM) can be performed using various types of natural language processing (NLP) and machine learning (ML). These methods can be categorized by complexity and precision level. Co‐occurrence‐based ML methods are rather basic, as they identify simultaneous appearance of drugs and clinical events in a single post. In contrast, statistical learning methods involve more complex NLP and identify drugs, events, and associations between them. We aimed to compare the ability of co‐occurrence and NLP to identify AEs and signals of disproportionate reporting (SDR) in patient‐generated SM. We also examined the performance of lift in SM‐based signal detection (SD). Methods: Our examination was performed in a corpus of SM posts crawled from open online patient forums and communities, using the spontaneously reported VigiBase data as reference data set. Results: We found that co‐occurrence and NLP produce AEs, which are 57% and 93% consistent with VigiBase AEs, respectively. Among the SDRs identified both in SM and in VigiBase, up to 55.3% were identified earlier in co‐occurrence, and up to 32.1% were identified earlier in NLP‐processed SM. Using lift in SM SD provided performance similar to frequentist methods, both in co‐occurrence and in NLP‐processed AEs. Conclusion: Our results indicate that using SM as a data source complementary to traditional pharmacovigilance sources should be considered further. Various levels of SM processing may be considered, depending on the preferred policies and tolerance for false‐positive to false‐negative balance in routine pharmacovigilance processes.