Abstract
OBJECTIVE: Large volume radiological text data have been accumulated since the incorporation of electronic health record (EHR) systems in clinical practice. We aimed to determine whether deep natural language processing algorithms could aid radiologists in improving thyroid cancer diagnosis. METHODS: Sonographic EHR data were obtained from the EHR database. Pathological reports were used as the gold standard for diagnosing thyroid cancer. We developed thyroid cancer diagnosis based on natural language processing (THCaDxNLP) to interpret unstructured sonographic text reports for thyroid cancer diagnosis. We used the area under the receiver operating characteristic curve (AUROC) as the primary metric to measure the performance of the THCaDxNLP. We compared the performance of thyroid ultrasound radiologists aided with THCaDxNLP vs. those without THCaDxNLP using 5 independent test sets. RESULTS: We obtained a total number of 788,129 sonographic radiological reports. The number of thyroid sonographic data points was 132,277, 18,400 of which were thyroid cancer patients. Among the 5 test sets, the numbers of patients per set were 439, 186, 82, 343, and 171. THCaDxNLP achieved high performance in identifying thyroid cancer patients (the AUROC ranged from 0.857-0.932). Thyroid ultrasound radiologists aided with THCaDxNLP achieved significantly higher performances than those without THCaDxNLP in terms of accuracy (93.8% vs. 87.2%; one-sided t-test, adjusted P = 0.003), precision (92.5% vs. 86.0%; P = 0.018), and F1 metric (94.2% vs. 86.4%; P = 0.007). CONCLUSIONS: THCaDxNLP achieved a high AUROC for the identification of thyroid cancer, and improved the accuracy, sensitivity, and precision of thyroid ultrasound radiologists. This warrants further investigation of THCaDxNLP in prospective clinical trials.