Learning Patient Representations from Text

Proceedings of the 7th Joint Conference on Lexical and Computational Semantics (*SEM), pages 119–123 New Orleans, June 5-6, 2018. c⃝2018 Association for Computational Linguistics Learning Patient Representations from Text Dmitriy Dligach1 and Timothy Miller2 1Loyola University Chicago 2Boston Children’s Hospital and Harvard Medical School 1ddligach@luc.edu 2timothy.miller@childrens.harvard.edu Abstract Mining electronic health records for patients who satisfy a set of predefined criteria is known in medical informatics as phenotyping. Phenotyping has numerous applications such as outcome prediction, clinical trial recruit- ment, and retrospective studies. Supervised machine learning for phenotyping typically re- lies on sparse patient representations such as bag-of-words. We consider an alternative that involves learning patient representations. We develop a neural network model for learn- ing patient representations and show that the learned representations are general enough to obtain state-of-the-art performance on a stan- dard comorbidity detection task. 1 Introduction Mining electronic health records for patients who satisfy a set of predefined criteria is known in medical informatics as phenotyping. Phenotyping has numerous applications such as outcome pre- diction, clinical trial recruitment, and retrospec- tive studies. Supervised machine learning is cur- rently the predominant approach to automatic phe- notyping and it typically relies on sparse patient representations such as bag-of-words and bag-of- concepts (Shivade et al., 2013). We consider an alternative that involves learning patient represen- tations. Our goal is to develop a conceptually sim- ple method for learning lower dimensional dense patient representations that succinctly capture the information about a patient and are suitable for downstream machine learning tasks. Our method uses cheap supervision in the form of billing codes and thus has representational power of a large dataset. The learned representations can be used to train phenotyping classifiers with much smaller datasets. Recent trends in machine learning have used neural networks for representation learning, and these ideas have propagated into the clinical in- formatics literature, using information from elec- tronic health records to learn dense patient repre- sentations (Choi et al., 2016, 2017; Lipton et al., 2016; Miotto et al., 2016; Nguyen et al., 2017; Pham et al., 2016). Most of this work to date has used only codified variables, including ICD (Inter- national Classification of Diseases) codes, proce- dure codes, and medication orders, often reduced to smaller subsets. Recurrent neural networks are commonly used to represent temporality (Choi et al., 2016, 2017; Lipton et al., 2016; Pham et al., 2016), and many methods map from code vocab- ularies to dense embedding input spaces (Choi et al., 2016, 2017; Nguyen et al., 2017; Pham et al., 2016). One of the few patient representation learning systems to incorporate electronic medical record (EMR) text is DeepPatient (Miotto et al., 2016). This system takes as input a variety of features, including coded diagnoses as the above systems, but also uses topic modeling on the text to get topic features, and applies a tool that maps text spans to clinical concepts in standard vocabularies (SNOMED and RxNorm). To learn the represen- tations they use a model consisting of stacked de- noising autoencoders. In …

Content truncated to 500 words