1. Introduction

Biopotentials such as electrocardiogram (ECG), photoplethysmogram (PPG), and ballistocardiogram (BCG) signals are widely used for cuffless blood pressure (BP) estimation due to their non-invasive approach ^[1]. Biosignals are used to calculate pulse transit time (PTT)—the time taken by a pulse wave between two arterial sites—which has an inverse correlation with BP ^[2].

Several deep neural network (DNNs) models have been proposed to predict PTT-based blood pressure ^[3-6]. In particular, the recurrent neural network (RNNs) models is the most widely used architecture, since it could extract the sequential information from time series data ^[7]. However, one of the major obstacles of the RNNs approach is a long-term dependency when processing the sequential data. As the layers in networks become deeper, initial weights could not be considered due to the vanishing gradient problem ^[8]. Additionally, the training time for RNNs increases significantly when training data have long sequences ^[9]. Thus, an RNNs might not be a suitable model for edge devices like mobile phones or smart watches owing to its high computational costs.

Therefore, this paper investigates an attention mechanism particularly known as Luong attention to estimate the BP using the data of three channels: ECG, PPG, and BCG signals. The proposed algorithm comprises four layers of a 1D convolutional neural network (CNNs) followed by a self-attention algorithm to extract correlations among segments of the input data sequences. The attention mechanism is faster than the RNNs due to its parallel computations for all input data ^[10]. Moreover, the attention mechanism is explainable, making it more interpretable than conventional DNNs.

The remaining sections cover the data acquisition process, a demonstration of the algorithm, and the performance results.

2. Data Acquisition

The data used for model training and testing are as follows. Five subjects participated in the experiment, and 30-minute long data from ECG and PPG signals were recorded using a biosignal amplifier (Biopac MP36, BIOPAC Systems Inc., Goleta, CA, USA). A BCG signal was simultaneously measured using a data acquisition module (NI 6225, National Instrument Corp, Austin, TX, USA). At the same time, continuous ABP was measured using a beat-to-beat BP monitor (Finometer PRO, Netherlands).

This experiment was approved by the Institutional Review Board of Kwangwoon University (IRB No.7001546-20200823-HR(SB)-008-07). During the experiment, subjects were introduced to a cold pressor condition three times, where the subject’s right foot was put into cold water for one minute to elevate blood pressure.

2.1 Data Preprocessing

After the data were obtained at the sampling frequency (1KHz), band-pass filtering was performed to remove noise and baseline wandering. Fig. 1 illustrates the signals before and after band-pass filtering. Table 1 shows the cutoff frequency of the band-pass filter for each signal. For the input data, five-second windows with corresponding SBP and DBP were used to train and test the model. Fig. 2 displays the continuous BP values utilized for labels to train the model. Outlier BP values beyond the range of the mean ${\pm}$1.96 standard deviation (SD) were removed from the dataset.

Fig. 1. The preprocessing of input signals (Top: raw signals; Bottom: filtered signals).

Fig. 2. Continuous BP used as labels to train and test the model (Top: continuous BP for 30 mins; Middle: arterial BP from Finometer; Bottom: distribution of BP).

Table 1. Cutoff frequencies of band-pass filtered input.

Signal	High-pass filter	Low-pass filter
ECG	0.5 Hz	35 Hz
PPG	4 Hz	15 Hz
BCG	0.5 Hz	15 Hz

2.2 Deep Learning Architecture

The proposed model to estimate BP consists of a 1D CNN and the Luong attention mechanism. The overall architecture is illustrated in Fig. 3. First, the preprocessed input data are fed into the 1D CNN for feature extraction. The dimensions of the initial data are 1000 $\times 3$, and the CNN layers are composed of 64, 128, and 256 layers, followed by batch normalization and max pooling to downsample the features.

Fig. 3. The overall architecture of the proposed DNN model.

3. Results

The performance of the model was validated based on mean absolute error (MAE), root mean square error (RMSE), and the $R^{2}$ coefficient. The metrics are shown in Eqs. (2) and (3).

In Eqs. (2) and (3), the evaluation metrics are described in which $y$ denotes ground truth, and $\hat{y}$ is the reference BP. MAE and RMSE represent the difference between the predicted values and ground truth. $R^{2}$ provides the correlation coefficient between the prediction and the reference BP:

Table 2 shows that the MAE was 3.299${\pm}$2.419 for SBP and 2.69${\pm}$1.821 for DBP. The results for MAE and RMSE outperformed models that connect the CNN, RNN, and attention algorithms ^[13].

4. Conclusion

In this paper, we proposed a cuffless BP estimation model using an attention mechanism that does not require an RNN and that yields a highly accurate BP prediction. This approach could be an alternative to the conventional cuff-based BP monitoring model, and is more accessible and comfortable when applied on a daily basis.