Joeky Senders

145 Deep Learning and NLP learning curves Introduction Clinically-derived patient information is generally increasing in volume and granularity with the expansion and improvement of online medical record systems. 1 Although analysis of this information allows for the generation of knowledge to improve future patient care, its full potential remains unrealized because most of it is stored in an unstructured, free-text format unsuitable for traditional methods of analysis. Manual review is necessary to extract and structure relevant patient information. As data sets continue to grow, however, manual chart review becomes increasingly inefficient, inconsistent, and prone to error. 2 Various studies have already demonstrated the utility of automated methods for the processing and analysis of free-text clinical reports. 3–14 These algorithms have the potential to assist in structuring the immense stream of free-text clinical information produced on a day-to-day basis. However, they are generally evaluated on a static text corpus of clinical reports with a fixed number of training examples, thereby lacking evidence on their learning capacity (i.e., the required number of examples to train high-performing models). 3–14 Yet, learning efficiency is instrumental, as the availability of training examples can be limited, and their labelling can be time consuming or expensive. In the current study, we aimed to compare the learning curves of various natural language processing approaches and develop an open-source framework for clinical text mining. Therefore, we have developed models that determine the histological diagnosis of brain tumor patients based on free-text pathology reports. Furthermore, we used various training samples to compare the efficiency of each algorithm’s learning curve. Methods Participants This study was conducted and reported according to the Transparent Reporting of a Multivariable Prediction Model of Individual Prognosis Or Diagnosis (TRIPOD) statement. 15 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. We included all patients who underwent an operation at our institution for a histopathologically confirmed diagnosis of glioma, meningioma, or brain metastasis between January 2002 and July 2018. Patients were