Predicting hypotension episode with numerical vital sign signals in the intensive care unit

Abstract

Introduction: Even short periods of hypotension are associated with increased morbidity and mortality. Using high-density numerical physiologic data, we developed a machine learning (ML) model to predict hypotension episodes, and further characterized risk trajectories leading to hypotension.

Methods: A subset of subjects with 1/60Hz physiological data was extracted from MIMIC2, a richly annotated multigranular database. Hypotension was defined as >5 measurements of systolic blood pressure ≤ 90 mmHg and mean arterial pressure ≤ 60 mmHg, within a 10-minute window. Derived features using raw measurements of heart rate, respiratory rate, oxygen saturation, and blood pressure were computed. Random Forest (RF), K-Nearest Neighbors (KNN), and Logistic Regression models were trained with 10-fold cross validation to predict instantaneous risk of hypotension using features extracted from the data leading to the first episode of hypotension (cases) or ICU discharge in subjects never experiencing hypotension (controls). For a given subject, risk trajectory was computed from the collation of instantaneous risks.

Results: From a source population of 2808 subjects, 442 subjects met our definition of hypotension, and 724 subjects without hypotension comprised the control group. 204 features were generated from the four vital signs. The area under the curve (AUC) for Random Forest classifier was 0.829, out-performing Logistic Regression (AUC 0.826) or K-Nearest Neighbors (AUC 0.783) (Fig 1). Risk trajectories analysis showed average controls risk scores <0.3 (<30% risk of future hypotension), while the hypotension group had a rising risk score (0.45 to 0.7) in the 8 hours leading to the first hypotension episode, and significantly higher scores leading into subsequent episodes (Fig 2). Conclusions: Hypotension episodes can be predicted from vital sign time series using supervised ML. Subjects developed hypotension have an increased risk compared to controls at least 8 hours prior to the episode.