Deconstructing Intensive Care Unit Triage Identifying factors influencing

Deconstructing Intensive Care Unit Triage Identifying factors influencing admission decisions using natural language processing and semantic technology Yue Liu, 2 MS; Evan Patton, 2 MS; Heng Ji, 2 Ph. D; Deborah L. Mc. Guinness, 2, 3 Ph. D Critical Care, & Sleep Medicine, Dept. of Medicine; Dept. of Emergency Medicine, Icahn School of Medicine at Mount Sinai, New York, NY; 2 Tetherless World Constellation, Dept. of Computer Science, Rensselaer Polytechnic Institute (RPI), Troy, NY, 3 Dept. of Cognitive Science, RPI, Troy, NY RESEARCH OBJECTIVES Figure 1: Semantic analysis & Acronym expansion “ 61 y. o. M pt with hx of COPD, HTN, … etc. ” Online resource Domain Ontology Input Distributional Semantic Computation Text Normalization Natural Language Processing BACKGROUND • ICU admission decisions can hinge on rapid clinical assessments, often involving factors that are not easily captured within electronic health record (EHR) data. • Ontology mapping has primarily focused on clinical data and diagnoses, but the semantics of “clinical gestalt” require more customized approaches. • NLP and semantic technology can be used to analyze various data formats and can be customized to qualitatively and quantitatively study subjective elements of the ICU triage process. METHODS • Setting: Academic tertiary care hospital, with a 14 -bed Medical ICU, operating at 91% average occupancy • Study population: All adult patients referred for Medical ICU admission over a 12 -month period • Data source: Prospectively collected Medical ICU physician semi-structured consult documentation, augmented with clinical data (patient characteristics/clinical course, University Health. Systems Consortium (UHC) expected mortality risk) • Computer NLP model: Fact extraction training and evaluation on 1160 patient consults, for pattern recognition within free-text sections of consult logs. (Figures 1 -3) • Analysis: Chi-square testing and multivariate regression, incorporating a priori selected clinical elements and NLP themes to predict (1) odds of ICU “accept” decision, and (2) in-hospital mortality or hospice discharge • Domain-specific resources combined with a trained model can expand acronyms, interpret jargon, and extract semantic relations between entities. DISCUSSION RESULTS Table 1: Cohort characteristics Accept (N=703) Reject (N=457) Age, years (mean ± SD)*** 59. 3 ± 16. 6 64. 0 ± 16. 2 Male gender (%)* 361 (51. 4) 209 (45. 7) UHC Expected Mortality (median, IQR) 0. 54 (0. 01 -0. 23) 0. 45 (0. 00 -0. 19) Pre-consult LOS, days (median, IQR) 0. 8 (0. 3 -4. 9) 0. 8 (0. 3 -6. 4) ED origin (%)** 226 (54. 6) 188 (45. 4) Discharge disposition (%) • In-hospital mortality • Hospice 204 (29. 0) 55 (7. 8) 126 (27. 6) 45 (9. 8) • Illness severity, as captured by UHC expected mortality risk index, was similar for both Accept and Reject cases. *p<0. 05 **p<0. 01 ***p<0. 001. Human correction & validation Output “ 61 year old Male patient with history of chronic obstructive pulmonary disease, hypertension, … etc. ” Figure 2: Sample “Accept” case annotation† 84 year old woman with hypertension, hyperlipidemia, colon cancer status post colectomy and with remote pulmonary metastasis status/post resection. Good functional status until a recent hospitalization for aspiration pneumonia. Was discharged to rehabilitation center about a week ago then had witnessed emesis with aspiration, and in septic shock and respiratory failure. Just extubated yesterday. Figure 3: Sample “Reject” case annotation† 83 year old female with history of metastatic colon cancer, extensive liver metastases, peritoneal carcinomatosis, hemorrhagic CVA, advanced dementia, TPN-dependent, minimal functional status, admitted for pneumonia and possible initiation of chemotherapy. Treated with antibiotics, lasix, now with worsening hypoxic respiratory failure. Patient was DNR/DNI, but reversed when patient had decompensation. Now intubated, listed to step down. Only partial clinical entities and associated relations displayed for illustrative purposes. † Entity Recognition Relation Extraction Automatic Profiling Figure 4: Word clouds for emergent descriptors, filtered against a priori selected themes, presented by weight REJECT • To develop a semantically-enabled natural learning processing (NLP) model to capture subjective elements of Intensive Care Unit (ICU) triage decisions • To determine and characterize effect of subjective assessments on odds of ICU admission and in-patient mortality. NATURAL LANGUAGE MODEL DEVELOPMENT ACCEPT 1 Pulmonary, Kusum S. Mathews, MD, 1 MPH; • Significant differences were seen for the following terms: cognitive limitations, ** family, *** limited goals of care, ** metastatic, *** poor prognosis, *** trach/PEG, * and transplant. ** Table 2: Logit models for ICU “accept” decision and in-hospital mortality or hospice discharge Model: Triage‡ Model: Mortality‡ • Analysis of ICU consult documentation, written in “real time, ” provides insight into the perceptions shaping triage decisions, including disease severity, pre-existing care limitations, and previous receipt of life-sustaining care. • Subjective assessments of overall prognosis and likelihood to benefit significantly predicted odds of receiving an ICU “accept” decision. • Although the emergent descriptors were similar in both Accept and Reject cases, the prevalence varied greatly with larger differences seen in more subjective terms. • Next steps include additional analysis on emergent themes, incorporating background knowledge to help produce more refined and discriminating descriptors (e. g. , contextualizing poor prognosis with past clinical data). CONCLUSIONS • NLP modeling and semantics can be valuable tools in qualitative and quantitative ICU triage decision analysis. • Captured through NLP analysis, prognosis and “likelihood to benefit” determinations at triage significantly influence admission decisions and are predictive of mortality and hospice utilization. Predictors (n=892) Odds ratio (CI) P value UHC Expected Mortality 1. 12 (1. 04 -1. 21) 0. 004 1. 79 (1. 61 -1. 98) <0. 001 Consult originating from ED 0. 68 (0. 49 -0. 95) 0. 025 0. 43 (0. 29 -0. 64) <0. 001 Pre-existing care limitations 0. 35 (0. 16 -0. 76) 0. 008 2. 21 (0. 98 -5. 02) 0. 057 Poor prognosis per ICU team 0. 22 (0. 13 -0. 36) <0. 001 2. 99 (1. 68 -5. 33) <0. 001 ACKNOWLEDGMENTS & FUNDING Known cancer diagnosis 0. 64 (0. 45 -0. 91) 0. 012 2. 35 (1. 61 -3. 44) <0. 001 --- 0. 99 (0. 69 -1. 42) 0. 961 The study described was supported by Award Number 1 K 12 HL 109005 -01 (Mathews) from the National Heart, Lung, and Blood Institute (NHLBI) and by the Rensselaer Polytechnic Institute (RPI) Tetherless World Constellation (Ji, Liu, Mc. Guinness, Patton). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NHLBI, the National Institutes of Health, or RPI. ICU “Accept” Decision ‡ --- Model adjusted for Gender, Race, & Ethnicity; Day shift excluded (stepwise removal; p>0. 10)