Background: cardiovascular diseases (CVDs) have
become the leading causes of death worldwide. Arterial stiffness and
elasticity are important indicators of cardiovascular health. Pulse
wave analysis (PWA) is essential for analyzing arterial stiffness
and elasticity, which are highly dependent on the tidal peak
(P
2). P
2 is one of the four key physiological points,
which also include percussion peaks (P
1), diastolic notches
(P
3), and diastolic peaks (P
4). P
1, P
3, and
P
4 are often local maxima or minima, facilitating their
identification via the second derivatives method, a classic
localization method for key physiological points. Classic methods
such as the second derivative method, Empirical Mode Decomposition
(EMD), and Wavelet Transform (WT), have been employed for the
extraction and analysis of the P
2. Due to individual variation
and arterial stiffness, locating the P
2 using classic methods
is particularly challenging.
Methods: we propose a hybrid neural network based on Residual
Networks (ResNet) and bidirectional Long Short-Term Memory Networks
(Bi-LSTM), successfully achieving high-precision localization of the
P
2 in radial artery pulse signals. Meanwhile, we compared our
method with the second derivative method, EMD, WT, Convolutional
Neural Networks (CNN) and the hybrid model with ResNet and LSTM.
Results: the results indicate that our proposed model
exhibits significantly higher accuracy compared to other
algorithms. Overall, MAEs and RMSEs for our proposed method are
62.60% and 58.84% on average less than those for other
algorithms. The average R
Adj
2 is 29.20% higher. The
outcomes of the efficiency evaluation suggest that the hybrid model
performs more balancedly without any significant shortcomings, which
indicates that the Bi-LSTM structure upgrades the performances of
LSTM.
Significance: our hybrid model can provide the medical field
with improved diagnostic tools and promote the development of
clinical practice and research.