Abstract
BACKGROUND: Identifying specific causes of intraoperative hypotension (IOH) is a challenge in clinical practice. Improving the causal treatment of hypotension requires a more detailed understanding of the underlying hemodynamic alterations during hypotension. This study aims to identify distinct hemodynamic endotypes of IOH by applying a deep learning model to high-resolution intraoperative hemodynamic data. METHODS: We conducted a retrospective analysis for surgical patients who had undergone continuous intraoperative monitoring of systemic vascular resistance index (SVRI), stroke volume index (SVI), stroke volume variation (SVV), cardiac index (CI), and heart rate (HR). IOH was defined as a mean arterial pressure (MAP) < 65 mmHg sustained for at least 1 min. A long short-term memory (LSTM)-based autoencoder was developed to compress multivariate time-series data into a two-dimensional latent space. Unsupervised clustering, k-means, was performed on the two-dimensional latent representations, and the optimal number of clusters was determined by Calinski-Harabasz (CH) and Davies-Bouldin (DB) index. RESULTS: A total of 184 patients experienced at least one episode of IOH, and 1304 hypotensive episodes with 253,380 data points were included for analysis. K-means identified five distinct IOH endotypes. Based on the characteristic hemodynamic profiles of each cluster, we labeled the five endotypes: (1) severe vasodilation with high CI, (2) hypovolemia, (3) myocardial depression, (4) bradycardia, and (5) mild vasodilation with preserved CI. Mild vasodilation with preserved CI occurred mainly in the first and second quartiles of the procedure. Bradycardia occurred throughout the procedure, but more frequently in the fourth quartile. Myocardial depression occurred primarily in the second quartile, and hypovolemia occurred frequently throughout the procedure. It was also observed that intra-event transitions between endotypes within the same hypotensive episode occurred in five events. CONCLUSIONS: Five endotypes of IOH were identified, and this may support the development of causal treatment strategies of intraoperative hypotension, pending future validation.