Deep Learning-Based Cardiac Output Estimation Using Multimodal Physiological Signals.

Area of Science:

• Biomedical Engineering
• Cardiovascular Physiology
• Artificial Intelligence in Medicine

Background:

• Cardiac output (CO) is crucial for cardiovascular monitoring.
• Current gold standard (thermodilution) is invasive.
• Existing non-invasive methods lack sufficient accuracy.

Purpose of the Study:

• To apply deep learning for non-invasive CO estimation.
• To evaluate different input signals (ART, ECG, PPG) and their combinations.
• To assess model performance using MAE, RMSE, bias, and LOA.

Main Methods:

• Utilized the public VitalDB database.
• Trained and evaluated deep learning models.
• Tested combinations of arterial pressure (ART), electrocardiography (ECG), and photoplethysmography (PPG) signals.

Main Results:

• The triadic model (ART, ECG, PPG) showed the best performance.
• Achieved a mean absolute error (MAE) of 0.66 L/min.
• Demonstrated a strong correlation (R = 0.84) with reference CO values.

Conclusions:

• Deep learning shows significant promise for accurate, non-invasive CO estimation.
• The ART, ECG, and PPG combination is highly effective.
• Future work should focus on model interpretability and real-time clinical applications.