Remedi : a medical information extraction system

Abstract

Novel articles containing modern advancements and discoveries in the medical industry are being published online daily. Due to privacy issues, there is a lack of data available to build knowledge extraction tools specialized in the medical industry. This study presents a biomedical information extraction tool called Remedi, specialized in identifying ‘medical problem’ and ‘treatment’ entities from unstructured text in the medical domain, as well as the relations between them. The tool consists of a Biomedical Named Entity Recognition (BM-NER) model which employs a Bidirectional-Long Short-Term Memory Conditional Random Field (Bi-LSTM-CRF) model, as well as a Biomedical Relation Extraction (BMRE) model which consists of a Bi-LSTM model. A subset of the i2b2 2010 challenge dataset has been acquired to train the BM-NER and BM-RE components. The dataset includes annotated de-identified medical reports that can be used for concept extraction and relation classification tasks. We evaluated the performance of the Bi-LSTM and Bi-LSTM-CRF model when given additional external features such as the UMLS and the GENIA tagger. Results showed that with the addition of external features, both models outperformed similar models that did not leverage on such features. The BM-NER system achieved a micro-average F1 score of 0.846 which can be ranked second amongst top-performing models presented in the i2b2 2010 concept extraction challenge. On the other hand, the BM-RE system achieved a micro-average F1 score of 0.652 and needs further improvements to reach the performance of the best models in the i2b2 2010 relation extraction challenge.