What type of study is this?

September 10, 2025Open Access

Machine Learning-based Prediction of Hypotension During Anesthesia Induction

Key Points

Machine learning models can predict hypotension within 10 minutes after anesthetic induction, improving patient outcomes.
The best-performing model is the random forest, achieving an accuracy of 88.3% without feature selection.
Feature selection methods enhance model performance, with an AUC-ROC of 0.947 retained by the random forest model.
Clinical interventions such as patient positioning are significant in enhancing the accuracy of hypotension predictions.

Abstract

Background: Hypotension is a common complication during the induction of anesthesia, leading to adverse outcomes such as acute kidney injury (AKI), myocardial injury, and, in high-risk patients, death. Objectives: This study aimed to predict post-induction hypotension (PIH) by considering clinical interventions using machine learning (ML) methods. Methods: Prior to the induction phase of anesthesia, patient data were collected, and cardiac monitoring was set to measure non-invasive blood pressure (NIBP) at 1-minute intervals. Afterwards, induction was performed by the anesthesiologist. Hypotension was assessed 30 minutes after induction, defined as either a 20% drop in mean arterial pressure (MAP), an absolute MAP below 65 mmHg, or a systolic blood pressure (SBP) below 90 mmHg. The ML techniques were employed to develop a real-time hypotension predictor. These models utilize data gathered from five minutes to predict occurrences of hypotension in the next 10 minutes. Feature selection methods such as dimension reduction and sequential feature selection algorithms were utilized to provide more informative inputs to the ML models. Static features such as clinical features and dynamic features like vital signs were collected from patients undergoing general anesthesia across multiple hospital centers. Among the 215 patients, 110 developed PIH. Results: Without employing feature selection methods, the best performance belongs to the random forest (RF) model, with an accuracy of 88.3%, precision of 87.6%, recall of 85%, and an area under the curve of the receiver operating characteristic (AUC-ROC) at 0.945. Moreover, when utilizing feature selection methods, the RF model retained its status as the best model, with accuracy, precision, recall, and AUC-ROC values of 88.1%, 88.1%, 85.5%, and 0.947, respectively. Conclusions: We discovered that ML models hold the potential to predict PIH within the subsequent 10 minutes by utilizing data collected five minutes prior. Furthermore, considering clinical interventions, such as the patient's position and type of anesthetic drug injection, have a positive impact on the performance of ML models.

KI fragen

Bookmark

View Full Paper