Automatic Classification of Thyroid Findings Using Static and Contextualized Ensemble Natural Language Processing Systems: Development Study

Abstract

BackgroundIn the case of Korean institutions and enterprises that collect nonstandardized and nonunified formats of electronic medical examination results from multiple medical institutions, a group of experienced nurses who can understand the results and related contexts initially classified the reports manually. The classification guidelines were established by years of workers’ clinical experiences and there were attempts to automate the classification work. However, there have been problems in which rule-based algorithms or human labor–intensive efforts can be time-consuming or limited owing to high potential errors. We investigated natural language processing (NLP) architectures and proposed ensemble models to create automated classifiers. ObjectiveThis study aimed to develop practical deep learning models with electronic medical records from 284 health care institutions and open-source corpus data sets for automatically classifying 3 thyroid conditions: healthy, caution required, and critical. The primary goal is to increase the overall accuracy of the classification, yet there are practical and industrial needs to correctly predict healthy (negative) thyroid condition data, which are mostly medical examination results, and minimize false-negative rates under the prediction of healthy thyroid conditions. MethodsThe data sets included thyroid and comprehensive medical examination reports. The textual data are not only documented in fully complete sentences but also written in lists of words or phrases. Therefore, we propose static and contextualized ensemble NLP network (SCENT) systems to successfully reflect static and contextual information and handle incomplete sentences. We prepared each convolution neural network (CNN)-, long short-term memory (LSTM)-, and efficiently learning an encoder that classifies token replacements accurately (ELECTRA)-based ensemble model by training or fine-tuning them multiple times. Through comprehensive experiments, we propose 2 versions of ensemble models, SCENT-v1 and SCENT-v2, with the single-architecture–based CNN, LSTM, and ELECTRA ensemble models for the best classification performance and practical use, respectively. SCENT-v1 is an ensemble of CNN and ELECTRA ensemble models, and SCENT-v2 is a hierarchical ensemble of CNN, LSTM, and ELECTRA ensemble models. SCENT-v2 first classifies the 3 labels using an ELECTRA ensemble model and then reclassifies them using an ensemble model of CNN and LSTM if the ELECTRA ensemble model predicted them as “healthy” labels. ResultsSCENT-v1 outperformed all the suggested models, with the highest F1 score (92.56%). SCENT-v2 had the second-highest recall value (94.44%) and the fewest misclassifications for caution-required thyroid condition while maintaining 0 classification error for the critical thyroid condition under the prediction of the healthy thyroid condition. ConclusionsThe proposed SCENT demonstrates good classification performance despite the unique characteristics of the Korean language and problems of data lack and imbalance, especially for the extremely low amount of critical condition data. The result of SCENT-v1 indicates that different perspectives of static and contextual input token representations can enhance classification performance. SCENT-v2 has a strong impact on the prediction of healthy thyroid conditions.