BACKGROUND
In 2018, breast cancer was the second most common cancer worldwide. Pathology reports provide important information for optimal treatment decision making.
OBJECTIVE
To elucidate the deployment of deep learning data extraction methods for pathology reports in a single institute, we investigated the performance of methods between regular expression and natural language processing (NLP) in terms of accuracy.
METHODS
This was compared to the bidirectional encoder representations from transformers (BERT) model using specific vocabulary such as BERT-basic, BioBERT, and ClinicalBERT. A total of 1,215 pathology reports were used to build annotated data to develop an extraction algorithm for pathology reports. K-fold cross-validation was used to verify the performance of BioBERT and ClinicalBERT, pre-trained in the BERT model.
RESULTS
Among them, BioBERT emerged as a highly accurate (0.99901) data parsing model based on by k-fold validation. The parsing method using the NLP model could obtain data with higher overall accuracy than the existing method using regular expressions.
CONCLUSIONS
Our results showed that BioBERT has high accuracy in pathology reports, and that the NLP model can obtain data with a higher overall accuracy than regular expressions. Taken together, our findings suggest that the process of obtaining information from pathology reports should include NLP using BioBERT.