TY - JOUR T1 - Using Machine Learning to Parse Breast Pathology Reports JF - bioRxiv DO - 10.1101/079913 SP - 079913 AU - Adam Yala AU - Regina Barzilay AU - Laura Salama AU - Molly Griffin AU - Grace Sollender AU - Aditya Bardia AU - Constance Lehman AU - Julliette M. Buckley AU - Suzanne B. Coopey AU - Fernanda Polubriaginof AU - Judy E. Garber AU - Barbara L. Smith AU - Michele A. Gadd AU - Michelle C. Specht AU - Thomas M. Gudewicz AU - Anthony Guidi AU - Alphonse Taghian AU - Kevin S. Hughes Y1 - 2016/01/01 UR - http://biorxiv.org/content/early/2016/10/10/079913.abstract N2 - Purpose Extracting information from Electronic Medical Record is a time-consuming and expensive process when done manually. Rule-based and machine learning techniques are two approaches to solving this problem. In this study, we trained a machine learning model on pathology reports to extract pertinent tumor characteristics, which enabled us to create a large database of attribute searchable pathology reports. This database can be used to identify cohorts of patients with characteristics of interest.Methods We collected a total of 91,505 breast pathology reports from three Partners hospitals: Massachusetts General Hospital (MGH), Brigham and Womens Hospital (BWH), and Newton-Wellesley Hospital (NWH), covering the period from 1978 to 2016. We trained our system with annotations from two datasets, consisting of 6,295 and 10,841 manually annotated reports. The system extracts 20 separate categories of information, including atypia types and various tumor characteristics such as receptors. We also report a learning curve analysis to show how much annotation our model needs to perform reasonably.Results The model accuracy was tested on 500 reports that did not overlap with the training set. The model achieved accuracy of 90% for correctly parsing all carcinoma and atypia categories for a given patient. The average accuracy for individual categories was 97%. Using this classifier, we created a database of 91,505 parsed pathology reports.Conclusions Our learning curve analysis shows that the model can achieve reasonable results even when trained on a few annotations. We developed a user-friendly interface to the database that allows physicians to easily identify patients with target characteristics and export the matching cohort. This model has the potential to reduce the effort required for analyzing large amounts of data from medical records, and to minimize the cost and time required to glean scientific insight from this data. ER -