Abstract 15673: AI-Enabled Prediction of Bleeding in Patients Supported on Extracorporeal Membrane Oxygenation

Abstract

Introduction: Clinical data within the electronic health record (EHR) combined with artificial intelligence (AI) and machine learning (ML) offer the promise of rapid advancement of knowledge at the point of care. Hypothesis: We hypothesized that aggregating clinical data with preservation of relational elements will yield a more precise and less biased representation of patient data for predictive analytics. We leveraged these data to predict clinically significant bleeding in the ECMO-supported population. Methods: The data of children 0-19 years managed on ECMO for first run cardiac indications in a single center between 1/2010-12/2020 were analyzed. Bleeding Day was defined as any 24-hour period which included pulmonary, gastrointestinal, intracranial or surgical wound hemorrhage, and/or surgical intervention or Factor VIIa administration for hemorrhage control. EHR data documenting the ECMO episode and embedding relational elements were used to inform 1) a multivariable logistic regression model matching a previously- performed analysis using manually extracted data [1], and 2) a feature-limited graph neural network (GNN) model of bleeding. Results: There were 272 patients supported with ECMO for total 2,012 days which informed the analysis, with median age 0.4 [IQR 0.03, 3.30] years, 56% male and 14% of days including a bleeding event. Direct comparison of the EHR-derived logistic regression model with previously published manual results are presented in Table 1. Inclusion of the EHR-derived covariates in a GNN model achieved similar operating characteristics. Conclusions: EHR-derived AI-prediction models of bleeding in this complex patient population are at least as accurate as models with manual data and traditional statistical analysis, but with a known pathway to potentially more accurate predictions. This methodology can be replicated to other conditions, allowing rapid insights towards a learning healthcare system.

Ähnliche Arbeiten