Contrastive Transformer-Driven Discovery of Temporal Hemodynamic Subphenotypes in Cardiac Surgery Patients

Area of Science:

• Cardiology
• Artificial Intelligence
• Intensive Care Medicine

Background:

• Post-operative cardiac surgery patients exhibit dynamic hemodynamic changes requiring intensive monitoring.
• Personalized management strategies can be informed by identifying distinct hemodynamic patient subgroups.

Purpose of the Study:

• To develop and validate a novel approach for identifying hemodynamic subphenotypes in early post-operative cardiac surgery patients.
• To compare the efficacy of a contrastive-transformer framework against dynamic time warping (DTW) for hemodynamic subtyping.

Main Methods:

• Utilized 24-hour high-resolution physiologic and treatment data from 6,630 MIMIC-IV and 1,963 SICdb patients.
• Trained a transformer encoder with a reconstruction-contrastive objective to derive patient-level embeddings.
• Applied spectral clustering to identify hemodynamic subphenotypes and compared results with DTW-derived embeddings.

Main Results:

• Identified three reproducible hemodynamic subphenotypes using the contrastive-transformer framework.
• Subphenotype 3 showed significantly higher in-hospital mortality, longer ICU stays, and longer hospitalizations compared to subphenotype 1.
• DTW-derived subphenotypes demonstrated weaker prognostic separation.

Conclusions:

• The contrastive-transformer framework effectively identifies clinically meaningful temporal hemodynamic subphenotypes.
• These identified subphenotypes can optimize post-operative risk stratification for cardiac surgery patients.
• This approach holds potential for informing personalized management strategies in intensive care settings.