First Experiences With An Ai-Assisted Clinical Evidence System To Evaluate Clinical Consensus Among Clinical Trial Publications.

e18583 Background: The ability to review the most current evidence is pressured by the volume and rapid growth of new and sometimes inconsistent medical information. IBM Watson Evidence Service (WES) is being developed to provide a level of confidence in published findings based on patient cohorts and treatment outcome consensus. Methods: A set of 60,000 articles for 12 cancer types are defined using PubMed query language for WES. The resulting corpus is processed for extraction by natural language processing (NLP) using deep learning and term disambiguation as well as extraction of Medical Subject Headings (MeSH) from the Medline record. Extraction of comparative statements related to efficacy is done for survival, response, toxicity and quality of life along with 18 oncology relevant attributes. Then a set of abstracts that report relevant outcomes for a therapy and for a patient cohort are extracted. The number of positive, negative and neutral outcomes are aggregated over the set of articles relevant to a given treatment. A summary is produced outlining the degree of concordance seen in each type of reported outcome over the set of studies involving the therapy of interest. Results: The WES runs though the corpus and measures both the volume and degree of consensus across all studies reported for each patient cohort/treatment combination and each type of reported outcome. The system measures the degree to which specific treatments have been evaluated for efficacy and how consistent the findings are for each type of outcome across the identified set of published study results. For example, fora stage IV, NSCLC patient cohort, 69 articles were found evaluating use of carboplatin + docetaxel vs other treatments. Over 51% of the studies reported relatively less favorable survival outcomes for this treatment while over 77% of these studies reported favorable toxicities for this treatment regimen. Conclusions: Here we present a novel approach that can read and clinically interpret vast amounts of unstructured and structured data, extract clinical outcome statements from clinical research publications and classify the degree to which observed outcomes are consistent across a set of clinical studies.