Joeky Senders

111 Automating clinical chart review Introduction Analyzing patient characteristics and outcomes can be instrumental for optimizing clinical decision-making. However, most medical information is confined in narratively- written reports, which precludes efficient data gathering and analysis. 1 Manual chart review not only poses substantial costs in terms of time and human resources, variation between and within clinical reviewers can even lead to inconsistencies in data collection and consequently to erroneous inferences from biased study results. 2 Natural language processing (NLP) provides an automatic and deterministic alternative for the extraction of medical information from free-text clinical reports. It therefore has the potential to accelerate the speed and scale at which clinical research can be performed. 3 Although various pipelines have been developed to automate the extraction of medical information, external validation and optimization of these frameworks is impeded as only a few study groups have released their code on a publicly-accessible repository. Furthermore, the current medical literature on NLP predominantly focusses on the reporting of model performance, whereas it lacks insights into the methodological characteristics that drive model performance and help identify clinical variables eligible for text mining. In this study, we aimed to develop an open-source NLP pipeline for automated variable extraction using a corpus of free-text radiology reports of glioblastoma patients. Our secondary aim was to provide insight into the feasibility of NLP by studying the statistical properties of the variables of interest. Therefore, we examined how model performance was associated with the frequency distribution of the variables of interest, as well as the interrater agreement of the manually provided consensus labels. These insights can help identify variables eligible, as well as methodological boundaries, for text mining in clinical research. Methods Retrieval of the free-text radiology reports This study was conducted and reported according to the Transparent Reporting of a Multivariable Prediction Model of Individual Prognosis Or Diagnosis (TRIPOD) statement. 4 The Institutional Review Board of Brigham and Women’s Hospital approved the current study and waived the need for informed consent due to its retrospective, observational design. All patients with a histopathologically confirmed diagnosis of a glioblastoma operated at our institution between January 2005 and May 2018 were