Automatic Classification of Valve Diseases Through Natural Language Processing in Spanish and Active Learning

Abstract

A correct label classification of data allows improving healthcare processes and research. However, labeling is a difficult and expensive process, which limits its use and quality. We propose a proof of concept based on Natural Language Processing and active learning, in order to automatically structure information from a text in Spanish in the field of echocardiography.

Echocardiographic reports from a Health National System Cardiology Department were analyzed. Reports were divided into a training corpus (26,699 reports) and a validation corpus (2,881 reports). The design of the model was focused on the automatic labeling of aortic and mitral valve disease (stenosis/insufficiency) and their valve nature (native/prosthetic). The following steps were followed to build the models: data preparation, vectorization, and model fitting and validation. Results were compared with the manually labeled ground truth data from the physicians reporting the echocardiographic studies.

Four machine learning algorithms were compared: logistic regression, naïve bayes, random forest, and support vector machine; obtaining the last our best results with areas under the ROC curve of 0.92 and 0.93 for aortic and mitral stenosis, 0.87 and 0.89 for aortic and mitral insufficiency, and 0.97 and 0.96 for native aortic and mitral valve, respectively. Natural Language processing tools are useful to automatically structure and label echocardiographic information in Spanish text format. The developed models combined with active learning are capable of performing a correct prospective labeling.