Parallel convolutional neural networks for non-invasive cardiac hemodynamic estimation: integrating uncalibrated PPG signals with nonlinear feature analysis

Abstract

Objective. Understanding cardiac hemodynamic status (CHS) is essential for accurate cardiovascular health assessment, as it is governed by key parameters such as cardiac output (CO), systemic vascular resistance (SVR), and arterial compliance (AC). This study aims to develop a non-invasive method using digital photoplethysmography (PPGD) signals and deep learning techniques to predict these biomarkers for a comprehensive CHS evaluation. Approach. A dataset of 4374 virtual subjects was used. Nonlinear features were extracted from PPGD signals to capture their inherent complexity and irregularity. A parallel convolutional neural network (PCNN) was implemented to process both raw signals and nonlinear features concurrently. Model performance was evaluated using R2, root mean squared error (RMSE), mean squared error (MSE), and mean absolute error (MAE). Main results. The PCNN demonstrated satisfactory predictive performance with R2, RMSE, MSE, and MAE values of 0.872, 0.086, 0.008, and 0.068 for CO; 0.851, 0.074, 0.006, and 0.058 for SVR; and 0.938, 0.049, 0.003, and 0.038 for AC. The proposed PCNN-based method offers a novel, non-invasive approach for predicting key cardiovascular biomarkers, providing an accurate CHS assessment. Significance. This method advances non-invasive cardiovascular diagnostics by combining PPGD signals and deep learning. Future work will focus on validating this findings in real-world settings for improved clinical applicability.