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

Summary

Using an unsupervised autoencoder plus Gaussian mixture model, the authors identified four reproducible hypotension endotypes (vasodilation, hypovolaemia, myocardial depression, bradycardia) across surgical and ICU populations. The model outputs endotype probabilities at each hypotensive timepoint, enabling causal, physiology-directed therapy rather than treating blood pressure alone.

Key Findings

• Identified four physiologic hypotension endotypes: vasodilation, hypovolaemia, myocardial depression, and bradycardia.
• Independent validation across two large datasets (1,000 surgical; 1,000 ICU) reproduced the same endotypes.
• Algorithm uses stroke volume index, heart rate, systemic vascular resistance index, and stroke volume variation during MAP <65 mm Hg episodes.
• Outputs endotype probabilities for each hypotensive data point to inform causal therapy.

Clinical Implications

Endotype probabilities can guide targeted therapies: vasopressors for vasodilation, fluids for hypovolaemia, inotropes for myocardial depression, and chronotropic/pacing strategies for bradycardia. Integration into monitors may standardize causal treatment of hypotension.

Limitations

• No interventional trial to show clinical outcome improvement when using endotypes
• Potential site/device variability in hemodynamic measurements

Future Directions

Prospective trials integrating endotype classification into real-time decision support to test outcome benefits; assessment of generalizability across monitoring platforms; development of clinician-facing interfaces.