November 29, 2025
Background: Hypotension is associated with organ injury and mortality in both surgical and critically ill patients. Management remains challenging because hypotension can arise from diverse underlying haemodynamic mechanisms. This study aimed to identify and independently validate distinct hypotension endotypes using unsupervised deep learning applied to large datasets of surgical and critically ill patients.
Methods: An unsupervised deep learning algorithm, consisting of a deep learning autoencoder combined with a Gaussian mixture model, was developed to classify hypotensive episodes based on stroke volume index, heart rate, systemic vascular resistance index, and stroke volume variation. The development dataset included 871 surgical patients with 6,962 hypotensive episodes. External validation was performed using two independent cohorts: 1,000 surgical patients with 7,904 hypotensive episodes and 1,000 critically ill patients with 53,821 hypotensive episodes. Hypotension was defined as a mean arterial pressure below 65 mmHg for at least 1 minute.
Results: Four distinct hypotension endotypes were identified in the development cohort. Based on their physiological and clinical characteristics, these were classified as vasodilation, hypovolaemia, myocardial depression, and bradycardia. The same four endotypes were consistently identified in both independent validation cohorts of surgical and critically ill patients.
Conclusions: Unsupervised deep learning successfully identified four reproducible hypotension endotypes across surgical and critically ill populations: vasodilation, hypovolaemia, myocardial depression, and bradycardia. The algorithm estimates the probability of each endotype for individual hypotensive episodes, potentially enabling more individualized diagnosis and haemodynamic management.
©2026 OR. All Rights Reserved
