Deep learning model to identify and validate hypotension endotypes in surgical and critically ill patients

Background: Hypotension is associated with organ injury and death in surgical and critically ill patients. In clinical practice, treating hypotension remains challenging because it can be caused by various underlying haemodynamic alterations. We aimed to identify and independently validate endotypes of hypotension in big datasets of surgical and critically ill patients using unsupervised deep learning.

Methods: We developed an unsupervised deep learning algorithm, specifically a deep learning autoencoder model combined with a Gaussian mixture model, to identify endotypes of hypotension based on stroke volume index, heart rate, systemic vascular resistance index, and stroke volume variation observed during episodes of hypotension. The algorithm was developed with data from 871 surgical patients who had 6962 hypotensive events and validated in two independent datasets, one including 1000 surgical patients who had 7904 hypotensive events and another including 1000 critically ill patients who had 53 821 hypotensive events. We defined hypotension as a mean arterial pressure <65 mm Hg for at least 1 min.

Results: In the development dataset, we identified four hypotension endotypes. Based on their physiological and clinical characteristics, we labelled them as: vasodilation, hypovolaemia, myocardial depression, and bradycardia. The same four hypotension endotypes were identified in the two independent validation datasets of surgical and critically ill patients.

Conclusions: Unsupervised deep learning identified four endotypes of hypotension in surgical and critically ill patients: vasodilation, hypovolaemia, myocardial depression, and bradycardia. The algorithm provides the probability of each endotype for each hypotensive data point. Identifying hypotensive endotypes could guide clinicians to causal treatments for hypotension.