A major hurdle in the development of natural language processing (NLP) methods for Electronic Health Records (EHRs) is the lack of large, annotated datasets. Privacy concerns prevent the distribution of EHRs, and the annotation of data is known to be costly and cumbersome. Synthetic data presents a promising solution to the privacy concern, if synthetic data has comparable utility to real data and if it preserves the privacy of patients. However, the generation of synthetic text alone is not useful for NLP because of the lack of annotations. In this work, we propose the use of neural language models (LSTM and GPT-2) for generating artificial EHR text jointly with annotations for named-entity recognition. Our experiments show that artificial documents can be used to train a supervised named-entity recognition model for de-identification, which outperforms a state-of-the-art rule-based baseline. Moreover, we show that combining real data with synthetic data improves the recall of the method, without manual annotation effort. We conduct a user study to gain insights on the privacy of artificial text. We highlight privacy risks associated with language models to inform future research on privacy-preserving automated text generation and metrics for evaluating privacy-preservation during text generation.

Unstructured information in electronic health records provide an invaluable resource for medical research. To protect the confidentiality of patients and to conform to privacy regulations, de-identification methods automatically remove personally identifying information from these medical records. However, due to the unavailability of labeled data, most existing research is constrained to English medical text and little is known about the generalizability of de-identification methods across languages and domains. In this study, we construct a varied dataset consisting of the medical records of 1260 patients by sampling data from 9 institutes and three domains of Dutch healthcare. We test the generalizability of three de-identification methods across languages and domains. Our experiments show that an existing rule-based method specifically developed for the Dutch language fails to generalize to this new data. Furthermore, a state-of-the-art neural architecture performs strongly across languages and domains, even with limited training data. Compared to feature-based and rule-based methods the neural method requires significantly less configuration effort and domain-knowledge. We make all code and pre-trained de-identification models available to the research community, allowing practitioners to apply them to their datasets and to enable future benchmarks.