Explainable Clinical Decision Support from Text

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1478–1489, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 1478 Explainable Clinical Decision Support from Text Jinyue Feng Unity Health Toronto, Toronto ON jinyue@cs.toronto.edu Chantal Shaib Unity Health Toronto, Toronto ON chantal.shaib@unityhealth.to Frank Rudzicz Unity Health Toronto, Toronto ON Department of Computer Science, University of Toronto, Toronto ON Vector Institute for Artificial Intelligence, Toronto ON Surgical Safety Technologies, Toronto ON frank@cs.toronto.edu Abstract Clinical prediction models often use structured variables and provide outcomes that are not readily interpretable by clinicians. Further, free-text medical notes may contain informa- tion not immediately available in structured variables. We propose a hierarchical CNN- transformer model with explicit attention as an interpretable, multi-task clinical language model, which achieves an AUROC of 0.75 and 0.78 on sepsis and mortality prediction on the English MIMIC-III dataset, respectively. We also explore the relationships between learned features from structured and unstructured vari- ables using projection-weighted canonical cor- relation analysis. Finally, we outline a proto- col to evaluate model usability in a clinical de- cision support context. From domain-expert evaluations, our model generates informative rationales that have promising real-life appli- cations. 1 Introduction Electronic medical records (EMRs) store both structured data (e.g., vitals and laboratory mea- surements) and unstructured data (e.g., nursing and physician notes). Previous clinical prediction tasks have focused on structured data (e.g., Desautels et al., 2016; Gultepe et al., 2013; Ghassemi et al., 2014) which, despite their utility, may not capture all of the useful information in associated text. Clin- ical decision support systems rarely take advantage of free-text notes due to the complex nature of clin- ical language and interpretation. Rules and special- ized grammars can be applied to circumvent issues around clinical language; however, these methods rely on the presence of certain phrases and spelling, and do not account for the highly variable note structures across departments and hospitals (Yao et al., 2019; Mykowiecka et al., 2009; Assale et al., 2019). Further, opaque models without explain- ability are often met with resistance in medical contexts (Challen et al., 2019; Ahmad et al., 2018; Gordon et al., 2019). To address these challenges, we propose a novel multi-task language model that also provides rationales for decisions in medicine. Our multi-task model leverages ClinicalBERT (Alsentzer et al., 2019), which is a transformer- based model pre-trained on clinical corpora. Given the uniqueness of medical text, we introduce a com- bination of CNN and transformer encoders to cap- ture phrase-level patterns and global contextual relationships. Additionally, we explore latent atten- tion layers to generate rationales. Based on availability, we use the MIMIC-III database (Johnson et al., 2016) to predict two out- comes: sepsis and mortality in the intensive care unit (ICU). All experiments are conducted on notes written in English. We define the task of sepsis prediction more rigorously than previous work due both to using textual data only, and to emphasize the practicality of this model in real-world appli- cations. Moreover, we use canonical correlation analysis (CCA; …

Content truncated to 500 words