PhysioJEPA: Joint Embedding Representations of Physiological Signals for Real Time Risk Estimation in the Intensive...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Critical Care Medicine

Background:

• Self-supervised learning for multi-modal physiological signals is underexplored in critical care.
• Existing methods may not fully leverage complex, high-frequency bedside monitoring data.

Purpose of the Study:

• To introduce PhysioJEPA, a Joint Embedding Predictive Architecture (JEPA) for self-supervised representation learning from multi-modal physiological signals.
• To evaluate PhysioJEPA's performance in predicting clinical outcomes like hypotension and shock index.

Main Methods:

• PhysioJEPA was trained on multi-modal physiological signals (arterial blood pressure, ECG, photoplethysmography) from the MIMIC-III Waveform Database.
• Learned representations were used to predict 5-minute risk of hypotension and shock index.
• Performance was compared against Patch Time Series Transformer, ECG-JEPA, and a supervised convolutional model.

Main Results:

• PhysioJEPA achieved strong performance in predicting hypotension (AUROC = 0.83) and shock index (AUROC = 0.95).
• It demonstrated comparable or superior performance to existing self-supervised and supervised models.
• PhysioJEPA exhibited superior generalization to an independent healthcare system compared to all other models.

Conclusions:

• Self-supervised JEPA representation learning is a promising approach for multi-modal bedside monitoring signals.
• PhysioJEPA offers a robust method for extracting clinically relevant information from critical care data.
• The model's ability to generalize suggests potential for wider clinical application.