Detecting Adverse Drug Event Safety Signals from MEDLINE Reports: Challenges in Employing Cross-Terminology Mapping of MeSH to MedDRA Abhivyakti Sawarkar, MD, MMSc1; Alfred Sorbello, DO, MPH1; Anna Ripple, MLS2; Olivier Bodenreider MD, PhD2 1Center for Drug Evaluation and Research, U.S. Food and Drug Administration, DHHS; 2Lister Hill National Center for Biomedical Communications, U.S. National Library of Medicine, National Institutes of Health, DHHS FDA recently developed a web-based, prototype for prospective detection of emerging adverse drug event (ADE) safety signals from scientific literature reports by quantitative data mining of Medical Subject Heading (MeSH) indexing terms1. The MeSH terms are leveraged to extract ADE relations from MEDLINE reports and to convey ADE content2. We aim to facilitate the interoperability of pharmacovigilance resources at FDA by investigating cross-terminology mapping from MeSH to the Medical Dictionary for Regulatory Affairs (MedDRA), the medical term dictionary used in ADE safety signal generation from the FDA Adverse Event Reporting System. To optimize the interpretation and information value of prototype-generated ADE safety signals and their MedDRA mappings for the FDA regulatory review community, it is important that the MeSH descriptors unambiguously convey the highly specific ADE content of the MEDLINE citation. Introduction We leveraged the Unified Medical Language System (UMLS) and exploited hierarchical relationships to render preliminary, uncurated mappings of MeSH descriptors to equivalent or finer grained MedDRA preferred terms (PT). We manually evaluated the cross-terminology mapping of disease concepts in the context of MEDLINE citations retrieved for an arbitrarily selected marketed drug, ciprofloxacin. Methods For 70% of the MEDLINE reports (382/542) for the drug ciprofloxacin, only one ADE pair was generated to represent the specific ADE content of the entire citation. Analysis of the MeSH descriptor-to-MedDRA PT crosswalk evidenced three general mapping categories (see Table 1). Manual assessment of the 83 one to one MeSH:MedDRA exact mappings confirmed the linked terms conveyed similar meaning in translating the clinical ADE content of the source citations. Results Table 1: MeSH:MedDRA Cross-terminology Table for Ciprofloxacin MeSH:MedDRA Mapping Category n (%) Interpretation ambiguity of MeSH descriptor information loss? MeSH descriptor (MeSH Code) MedDRA PTs (MedDRA Code) mapped to the MeSH descriptor 1:1 exact or closely related mappings 83 (38) No Nausea (D009325) Nausea (10028813) 1:many mappings with broad ‘equivalence’ grouping of related clinical concepts 123 (56) Possible Stevens Johnson Syndrome (D013262) Stevens Johnson Syndrome (10042033) and Toxic epidermal necrolysis (10044223) 1:many mappings with ‘granularity mismatches’ 14 (6) Yes See Table 2 Table 2: Top five 1:many MeSH:MedDRA ‘granularity mismatch’ mappings for Ciprofloxacin MeSH descriptor MeSH code # of unique candidate MedDRA PTs* Examples Musculoskeletal diseases D009140 276 Injection site erosion, Application site pruritus, Tendon calcification Gastrointestinal diseases D005767 219 Obstruction gastric, Diverticular perforation, Hepatic lesion Eye diseases D005128 157 Eye swelling, Macular fibrosis, Ocular fistula Joint diseases D007592 66 Arthrotoxicity, Administration site joint discomfort, Lateral patellar compression syndrome Blood coagulation disorders D001778 41 Fibrin increased, Hypocoagulable state, Plasminogen activator inhibitor polymorphism *The MedDRA PTs span various clinical entities, laboratory findings, and genetic  polymorphisms. Additional details about the  specific ADE content of the citation are  needed to resolve the MeSH term mapping to the single, best fit MedDRA PT. Conclusions MeSH descriptors mapped to single MedDRA PTs provide the best links to translate clinical ADE content between the terminologies. Underlying differences in domain and scope between MeSH (biomedical literature; coarser) and MedDRA (ADEs; fine-grained) impact how ADE content is conveyed. Ambiguity in interpretation of some conceptually broad MeSH descriptors risks information loss and increases the complexity of cross-terminology mapping to MedDRA. Future work will include further curation of the UMLS-rendered cross-terminology mappings. Secondary exploration involving text mining with machine learning applied to abstracts and full text citations will seek to extract the ADE content so as to disambiguate between MeSH-MedDRA granularity mismatch instances rendered from the UMLS cross-terminology mappings. References Sorbello A, Ripple A, Tonning J, Munoz M, Hasan R, Ly T, Francis H, Bodenreider O. Harnessing scientific literature reports for pharmacovigilance. Prototype software analytical tool development and usability testing. Appl Clin Inform. 2017 Mar 22;8(1):291-305 Winnenburg R, Sorbello A, Ripple A, Harpaz R, Tonning J, Szarfman A, Francis H, Bodenreider O. Leveraging MEDLINE indexing for pharmacovigilance - Inherent limitations and mitigation strategies. J Biomed Inform. 2015 Oct;57:425-35. 22; 8(1):291-305. Acknowledgements: This project was supported in part by appointment to the research participation program at CDER administered by the Oak Ridge Institute for Science and Education (ORISE) for the FDA. Funding support received from the FDA/CDER/Office of Translational Sciences and the Intramural Research Program, NIH, National Library of Medicine. Disclaimer: The views expressed are those of the authors and do not necessarily represent the views of the US FDA, the NIH, or the US Government.