MP76-18 NATURAL LANGUAGE PROCESSING ALLOWS FOR ACCURATE AND AUTOMATED EXTRACTION OF DATA FROM PROSTATE BIOPSY PATHOLOGY REPORTS

Abstract

You have accessJournal of UrologyGeneral & Epidemiological Trends & Socioeconomics: Value of Care: Cost & Outcomes Measures I1 Apr 2018MP76-18 NATURAL LANGUAGE PROCESSING ALLOWS FOR ACCURATE AND AUTOMATED EXTRACTION OF DATA FROM PROSTATE BIOPSY PATHOLOGY REPORTS Gregory Joice, Brant Chee, Natasha Gupta, and Michael Johnson Gregory JoiceGregory Joice More articles by this author , Brant CheeBrant Chee More articles by this author , Natasha GuptaNatasha Gupta More articles by this author , and Michael JohnsonMichael Johnson More articles by this author View All Author Informationhttps://doi.org/10.1016/j.juro.2018.02.2586AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookTwitterLinked InEmail INTRODUCTION AND OBJECTIVES Evaluation of prostate biopsy data is a critical aspect of prostate cancer research. Non-standardized formats and syntax in pathology reports make large-scale analyses challenging. Additionally, significant manpower and cost are needed to manually review electronic medical records. Natural language processing (NLP) allows for rapid and automated extraction of important data, but it is unclear how this compares to manual data extraction in accuracy and recall. METHODS We randomly selected 1,000 systematic prostate biopsy reports from men diagnosed with prostate cancer at our institution. All pathology reports were annotated for Gleason score, anatomical location, total number of cores, cancerous cores, percent cancer per core, and benign pathology. We utilized these annotations to train and develop the NLP engine to identify these variables from de novo pathology reports. Next, two independent reviewers annotated an additional 150 prostate biopsy reports to use as a validation set. We evaluated inter-reviewer reliably by utilizing Cohen's kappa statistic. The NLP engine performance was assessed by calculating the precision, recall, and F1 score. RESULTS Inter-reviewer reliability was good between the two reviewers with kappa of 0.84. Overall, the NLP engine had excellent performance with an aggregate precision of 0.978, recall of 0.998, and F1 statistic of 0.987 (Table). NLP performed best at identifying benign cores (F1 = 0.997), percent cancer per core (F1 = 0.993), and Gleason score (F1 = 0.991). The most challenging variables were location (F1 = 0.984), total cores (F1 = 0.982), and cancerous cores (F1 = 0.978) but all still demonstrated excellent overall precision and recall. CONCLUSIONS There is good inter-reviewer reliability for manual data extraction from prostate biopsy pathology reports. We developed an NLP engine that can efficiently extract important data from biopsy reports with excellent precision and recall. NLP can be utilized to review and extract data from medical records to assist in performing large-scale population based analyses. © 2018FiguresReferencesRelatedDetails Volume 199Issue 4SApril 2018Page: e1025-e1026 Advertisement Copyright & Permissions© 2018MetricsAuthor Information Gregory Joice More articles by this author Brant Chee More articles by this author Natasha Gupta More articles by this author Michael Johnson More articles by this author Expand All Advertisement Advertisement PDF downloadLoading ...