Category Classification for Lung Computed Tomography of COVID-19 by Natural Language Processing in Japanese Radiology Report

Purpose: We screened patients admitted for coronavirus disease 2019 (COVID-19) via lung computed tomography (CT) using our own five-level categorization of imaging findings. We postulated that natural language processing (NLP) and machine learning (ML) could predict categorization using Japanese radiology reports. Methods: We screened 528 patients, including 40 polymerase chain reaction (PCR) test-positive patients. We built ML models to predict these categories and the results of PCR tests using a CoreML 3 framework. Results: When categories 1-3 were considered positive predictions, the precision of the probability of PCR results predicted by radiologists was 0.24 with recall of 0.65; specificity of 0.83; accuracy of 0.82; and F1 score of 0.35. The precision of the ML models was 0.62 with recall if 0.53; specificity of 0.88; accuracy of 0.78; and F1 score of 0.57. The macro-averaged accuracy of the reproducibility of the ML models for classification was 0.47. The area under the curve of receiver operating curve for PCR tests was 0.644, whereas that for categories 1-3 was 0.680. Conclusion: Although the understanding of Japanese radiology reports by NLP is still limited, the use of categorization may increase its usefulness in screening for COVID-19