In radiology, natural language processing (NLP) allows the extraction of valuable information from radiology reports. It can be used for various downstream tasks such as quality improvement, epidemiological research, and monitoring guideline adherence. Class imbalance, variation in dataset size, variation in report complexity, and algorithm type all influence NLP performance but have not yet been systematically and interrelatedly evaluated. In this study, we investigate these factors on the performance of four types [a fully connected neural network (Dense), a long short-term memory recurrent neural network (LSTM), a convolutional neural network (CNN), and a Bidirectional Encoder Representations from Transformers (BERT)] of deep learning-based NLP. Two datasets consisting of radiologist-annotated reports of both trauma radiographs (n = 2469) and chest radiographs and computer tomography (CT) studies (n = 2255) were split into training sets (80%) and testing sets (20%). The training data was used as a source to train all four model types in 84 experiments (Fracture-data) and 45 experiments (Chest-data) with variation in size and prevalence. The performance was evaluated on sensitivity, specificity, positive predictive value, negative predictive value, area under the curve, and F score. After the NLP of radiology reports, all four model-architectures demonstrated high performance with metrics up to > 0.90. CNN, LSTM, and Dense were outperformed by the BERT algorithm because of its stable results despite variation in training size and prevalence. Awareness of variation in prevalence is warranted because it impacts sensitivity and specificity in opposite directions.
Read full abstract