Sequence-Model-Based Medication Extraction from Clinical Narratives in German

Clinical narratives are a rich resource of patient-based information, where the automatic extraction of specific types of entities is still challenging due to various idiosyncrasies of non-standardized routine documentation. Much of the literature in clinical natural language processing (NLP) has been focused on English language. In this work, we focus on de-identified German-language clinical narratives. Predominant supervised approaches in the area of NLP, specifically named entity recognition (NER), need an expressive set of annotations for robust sequence modeling via a machine learning approach. This work investigates to what extent three main model types, conditional random fields (CRFs), bidirectional long short-term memory (BiLSTM) and Transformer models (BERT), perform on a limited set of annotations for medication information extraction in a specific clinical domain. The top performing feature, an optimized CRF model exploiting embedding representations out of a natively pre-trained language model using fastText, performed best with an overall F1 score 0.91. A multilingual BERT model incorporating German language resources reached an overall F1 score of 0.88, followed by the BiLSTM approach with an F1 score of 0.81. BERT based end-to-end systems nearly reach the performance of an optimized CRF approach. The results highlight the importance of pre-trained clinical language models available in languages other than English for fast, high performing, problem domain adaption.