Estimation of Central Aortic Pressure Waveforms by Combination of a Meta-Learning Neural Network and a Physics-Driven Method

The accurate non-invasive detection and estimation of central aortic pressure waveforms (CAPW) are crucial for reliable treatments of cardiovascular system diseases. But the accuracy and practicality of current estimation methods need to be improved. Our study combines a meta-learning neural network and a physics-driven method to accurately estimate CAPW based on personalized physiological indicators. We collected data from 260 patients who underwent catheterization surgery, using measured CAPW and personalized physiological indicators (e.g., weight, body mass index (BMI), radial mean arterial pressure (MAP), heart rate (HR), cardiac output (CO), radial systolic blood pressure (SBP), and radial diastolic blood pressure (DBP)) as input for neural network training. The output of the neural network are the Gaussian characteristic parameters of the single-period decomposed CAPW. The neural network model was constructed using the model-agnostic meta-learning (MAML) algorithm framework. Applying the physical characteristics of CAPW to the loss function, served to increase the constraints on the output and improve the accuracy of CAPW estimation. To verify the accuracy of the model, we compared measured and estimated CAPW in 52 patients. The results are consistent with a normalized root mean square error (NRMSE) of 0.0206. The predictions had low biases, namely SBP: 4.97 ± 4.42 mmHg, DBP: 4.78 ± 5.98 mmHg, and MAP: 0.35 ± 3.36 mmHg. The results demonstrate the accuracy and practicability of the approach to estimate CAPW. It can provide personalized parameters to calculate myocardial ischemia indicators (e.g., instantaneous wave-free ratio [iFR] and fractional flow reserve [FFR]) and may contribute to the early monitoring and prevention of cardiovascular diseases.