1259: CHARACTERIZATION AND PREDICTION OF NOREPINEPHRINE RESPONSE IN CRITICALLY ILL ADULTS

Abstract

Introduction: Norepinephrine (NE) is a widely used vasopressor to control blood pressure in hypotensive patients. NE titrations are performed manually based upon clinician interpretations of physiologic status, trajectory, and goals. Characterization of short-term NE response phenotypes can introduce objectivity to NE use and facilitate future automated titration systems. Methods: We used retrospective data from adults who were managed exclusively with NE and had arterial lines in ICUs at a single center between July 2017 and October 2020. Considering each NE dose change as t=0, response was assessed by comparing physiologic features from -15 min to 0 vs 0 to +15 min. Response was defined as a significant increase in mean arterial pressure (MAP) trajectory after a NE dose change, and we employed interrupted time series analysis to screen for 3 response phenotypes: (I) immediate, sustained; (II) delayed, sustained; and (III) immediate, not sustained. Next, a light gradient boosting machine was developed to predict response after any given NE dose change based on demographics and physiological data prior to this change. Shapley additive explanation (SHAP) was used to explain the predictions. Results: 2436 NE dose changes were included from 208 ICU patients (mean age of 61.2 +/-13.7 years, 63.5% male, 48.6% Black). Of these, 64 (2.6%) resulted in response I, 426 (17.5%) in response II, and 171 (7%) in response III; the remaining 1775 (72.9%) had no response. Response events had significantly lower physiologic feature slopes before NE dose changes (p< 0.01). Our model classified responses with a sensitivity of 76.3%, specificity of 75.7%, and AUC of 0.85 (CI95% 0.81-0.88). Immediate, sustained responses to NE were classified with a sensitivity of 81.8%, specificity of 87.0%, and AUC of 0.93 (CI95% 0.87-0.98). SHAP analysis indicated that a response was more likely if blood pressure was decreasing before a NE dose change. The likelihood of response was higher in lower weight-adjusted NE doses (< 0.1 mcg/kg/min), and dose increases were positively associated with the response. Conclusions: Responses to NE can be phenotyped and predicted using machine learning. Delayed, sustained responses are most commonly observed. Responses are more likely at low NE doses and when blood pressure is falling.