Acute hypotension is a sudden drop in blood pressure that reduces adequate blood flow to vital organs. If acute hypotensive episodes in intensive care unit (ICU) patients are not promptly recognised and treated, they are linked to negative consequences. Using regularly obtained vital signs from the eICU Collaborative Research Database, this study suggests an explainable machine learning framework for the early prediction of acute hypotensive episodes. Simple temporal summary features were derived by resampling minute-level time series of mean arterial pressure, systolic and diastolic blood pressure, heart rate, respiratory rate, and oxygen saturation to one-minute resolution and then using sliding 4-hour windows. Models were trained to predict whether an episode would occur within the next hour, and episodes were defined using a modified PhysioNet criterion based on sustained mean arterial pressure ≤ 60 mmHg. Tree-based models like random forest and XGBoost significantly outperformed a logistic regression baseline in terms of discriminative performance using data from 200 intensive care unit (ICU) stays. In order to support clinical interpretability and the framework’s potential use as a decision support tool in critical care settings, feature importance analysis identified the blood pressure and heart rate patterns that were most strongly linked to impending hypotension.
Roy et al. (Sun,) studied this question.