Advancing PPG-based cf-PWV estimation with an integrated CNN-BiLSTM-Attention model

Carotid-to-femoral pulse wave velocity (cf-PWV) is essential for assessing arterial stiffness and managing cardiovascular diseases (CVDs). Traditional cf-PWV measurement methods are cumbersome and error-prone, requiring expert intervention. To overcome these challenges, we introduce a novel non-invasive technique using photoplethysmography (PPG) signals, simplifying the process, increasing accuracy, and enhancing accessibility. Our deep learning framework integrates a convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and an attention mechanism to automatically extract features, capture temporal dependencies, and focus on critical parts of the sequence, respectively. This architecture adapts to both 1D PPG signals and their spectrogram representations, automating feature extraction and improving estimation accuracy. For 1D signal inputs, our model achieved a mean absolute percentage error (MAPE) of 1.674%, a root mean squared error (RMSE) of 0.188, and an R-squared ( \(\:{\text{R}}^{2}\) ) value of 0.992. Spectrogram inputs further improved results, yielding a MAPE of 0.991%, RMSE of 0.128, and an \(\:{\text{R}}^{2}\) of 0.996. These significant advancements suggest that our method has the potential to streamline routine cardiovascular health assessments, facilitating earlier detection and more efficient management of CVDs with simpler and more reliable cf-PWV measurements. Future work will focus on clinical validation to confirm its effectiveness and practicality in diverse real-world settings.