Abstract TP437: Machine Learning Based Approach for Predicting High In-Hospital Systolic Blood Pressure Variability Among Intracerebral Hemorrhage Patients During Early Hospitalization

Abstract

Introduction: Recent data suggest that early high systolic blood pressure variability (HSPBV) is associated with poor long term outcomes in intracerebral hemorrhage (ICH) patients. We employed a Machine Learning (ML) based approach to predict HSBPV during hospitalization. Methods: Adult radiologically confirmed ICH patients were enrolled in a multisite cohort. A semi-automated algorithm extracted systolic blood pressure (SBP) data from electronic medical records (EMR) and linkage to database with demographic, clinical, and outcomes information was created. Pre-hospital and early admission variables were used to develop predictive algorithms. Patients who expired during hospitalization were excluded. Generalized estimating equations quantified inter and intra-patient SBPV. An 80/20 data split was used to train and test predictive models for HSBPV. Two supervised ML models were trained using a repeated cross-validation bagged classification and regression tree, the first with a complete set of records and the second utilizing imputation (Figure). Results: A total of 455 patients were included with mean(SD) age 63.6 (14.98), females 36.7%, African American 34.5%, and median ICH score 2.0. The average per-patient observation time was 9.7 days resulting in 152,691 SBP readings. A 25 variable imputed model was parameterized (accuracy: 72.2%, sensitivity: 76.5%, specificity: 66.7%, and kappa: 0.43) (Figure). The top influential variables were age, hemorrhage volume, glucose, LDL, and platelets. Methodological and output contrasts with traditional likelihood-based modeling approaches will be presented. Conclusions: We demonstrate that ML-based pipelines can be developed for predicting HSBPV by utilizing SBP values generated from routine monitoring of ICH patients. Future development will incorporate streaming EMR SBP data points during patient stay, and evaluating high-dimensionality reduction methods such as principal component analysis.