Wearable ECG-PPG Deep Learning Model for Cardiac Index-Based Noninvasive Cardiac Output Estimation in Cardiac Surgery...

Area of Science:

• Biomedical Engineering
• Cardiovascular Physiology
• Artificial Intelligence in Medicine

Background:

• Accurate cardiac output (CO) measurement is crucial for hemodynamic management but typically requires invasive monitoring.
• Continuous, out-of-hospital CO monitoring is limited by current invasive methods.
• Wearable sensors combined with deep learning present a promising noninvasive alternative.

Purpose of the Study:

• To develop and validate a lightweight deep learning model using wearable electrocardiography (ECG) and photoplethysmography (PPG) signals for noninvasive CO prediction.
• To assess if cardiac index (CI)-based normalization improves the performance of CO prediction models.
• To enable continuous, catheter-free hemodynamic monitoring.

Main Methods:

• A deep learning model was developed using simultaneous ECG and PPG signals from 27 cardiac surgery patients.
• Three models were trained: direct CO prediction, indirect CO prediction, and CI prediction.
• Model performance was evaluated against pulmonary artery catheter thermodilution reference values.

Main Results:

• The cardiac index (CI) model demonstrated the best performance.
• Indirect CO prediction models showed significant reductions in error metrics (MAE, RMSE, bias) (p < 0.0001).
• Indirect CO estimates achieved a Pearson correlation coefficient of 0.904 and a percentage error of 23.75%, meeting the <30% benchmark.

Conclusions:

• Wearable ECG-PPG fusion deep learning enables accurate, noninvasive cardiac output estimation.
• Cardiac index normalization enhances model agreement with invasive measurements.
• This technology supports continuous, catheter-free hemodynamic monitoring.