Coherence-based Modeling of Clinical Concepts Inferred from Heterogeneous Clinical Notes for ICU Patient Risk Stratification

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 1012–1022 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 1012 Coherence-based Modeling of Clinical Concepts Inferred from Heterogeneous Clinical Notes for ICU Patient Risk Stratification Tushaar Gangavarapu∗ Gokul S Krishnan Sowmya Kamath S tushaargvsg45@gmail.com gsk1692@gmail.com sowmyakamath@nitk.edu.in Healthcare Analytics and Language Engineering (HALE) Lab Department of Information Technology National Institute of Technology Karnataka, Surathkal, Mangaluru, India Abstract In hospitals, critical care patients are often susceptible to various complications that ad- versely affect their morbidity and mortality. Digitized patient data from Electronic Health Records (EHRs) can be utilized to facilitate risk stratification accurately and provide prior- itized care. Existing clinical decision support systems are heavily reliant on the structured nature of the EHRs. However, the valuable patient-specific data contained in unstructured clinical notes are often manually transcribed into EHRs. The prolific use of extensive med- ical jargon, heterogeneity, sparsity, rawness, inconsistent abbreviations, and complex struc- ture of the clinical notes poses significant chal- lenges, and also results in a loss of information during the manual conversion process. In this work, we employ two coherence-based topic modeling approaches to model the free-text in the unstructured clinical nursing notes and capture its semantic textual features with the emphasis on human interpretability. Further- more, we present FarSight, a long-term aggre- gation mechanism intended to detect the onset of disease with the earliest recorded symptoms and infections. We utilize the predictive capa- bilities of deep neural models for the clinical task of risk stratification through ICD-9 code group prediction. Our experimental valida- tion on MIMIC-III (v1.4) database underlined the efficacy of FarSight with coherence-based topic modeling, in extracting discriminative clinical features from the unstructured nurs- ing notes. The proposed approach achieved a superior predictive performance when bench- marked against the structured EHR data based state-of-the-art model, with an improvement of 11.50% in AUPRC and 1.16% in AUROC. ∗Corresponding author. 1 Introduction Until recently, the healthcare industry had an inclination towards conservative approaches for the treatment and diagnosis of patients, result- ing in less patient-centric and imprecise assess- ments (Mathew and Pillai, 2015). Intensive Care Units (ICUs) utilize the most advanced medical resources to treat and monitor critically ill pa- tients. However, such advanced medical interven- tions in ICUs often make patients vulnerable to several complications (To and Napolitano, 2012). Various infections, including barotrauma, short- and long-term intubation, catheter-associated uri- nary tract infection, weaning errors, ventilator- associated pneumonia, gastrointestinal tract bleed- ing, and infections from unrecognized drug inter- actions, are associated with invasive ICU devices (Wollschlager and Conrad, 1988). The lack of ac- curate knowledge of the etiology of such compli- cations leads to the inability to accurately strat- ify risk, due to which, in most cases, adequate care is provided to patients only after the devel- opment of a complication (Huddar et al., 2016). With the advent of digitization, advancement in technology, need for evidence-based medicine, in- creased population, and rising rates of chronic diseases, the utilization of ever-increasing hetero- geneous medical …

Content truncated to 500 words