Automatic Annotation of Training Data for Deep Learning Based De-identification of Narrative Clinical Text

Electronic health records contain information about patients’ medical history which is important for research but the text must be de-identified before use. This study utilized dictionaries constructed from publicly available lists of identifiers to automatically annotate a training dataset for a named entity recognition model to de-identify names, streets, and locations in Danish narrative clinical text. Ambiguous identifiers were not annotated if they occurred more than expected for an identifier. The model had recall 93.43%, precision 86.10%, and F1 89.62%. We found that the model generalized from the training data to achieve better performance than simply using the dictionaries to directly annotate text.