Continuous blood pressure prediction system using Conv-LSTM network on hybrid latent features of photoplethysmogram (PPG) and electrocardiogram (ECG) signals

Abstract

Continuous blood pressure (BP) monitoring is essential for managing cardiovascular disease. However, existing devices often require expert handling, highlighting the need for alternative methods to simplify the process. Researchers have developed various methods using physiological signals to address this issue. Yet, many of these methods either fall short in accuracy according to the BHS, AAMI, and IEEE standards for BP measurement devices or suffer from low computational efficiency due to the complexity of their models. To solve this problem, we developed a BP prediction system that merges extracted features of PPG and ECG from two pulses of both signals using convolutional and LSTM layers, followed by incorporating the R-to-R interval durations as additional features for predicting systolic (SBP) and diastolic (DBP) blood pressure. Our findings indicate that the prediction accuracies for SBP and DBP were 5.306 ± 7.248 mmHg with a 0.877 correlation coefficient and 3.296 ± 4.764 mmHg with a 0.918 correlation coefficient, respectively. We found that our proposed model achieved a robust performance on the MIMIC III dataset with a minimum architectural design and high-level accuracy compared to existing methods. Thus, our method not only meets the passing category for BHS, AAMI, and IEEE guidelines but also stands out as the most rapidly accurate deep-learning-based BP measurement device currently available.

Keywords: Blood pressure; Deep learning; Electrocardiography; LSTM; Photoplethysmography.

Figure 1

The flow diagram of our proposed method. The figure shows the process of predicting the systolic and diastolic blood pressure. Record files from the MIMIC-III dataset are passed to the preprocessing module. The output of the preprocessing module is three signals: PPG, ECG, and ABP. PPG and ECG signals are inputted to the DBP Prediction Network and SBP Prediction Network, whereas the ABP signal determines the target SBP and DBP values for training the Network.

Figure 2

Preprocessing of the MIMIC III dataset used in our experiments. The figure shows the inside of the preprocessing module for extracting PPG and ECG signals. It consists of 7 consecutive blocks, with each block depicting a particular function.

Figure 3

The SBP and DBP Prediction Networks. For predicting SBP, real SBP data was used as the target value. For predicting DBP, real DBP data was used. The network consists of Conv1D layers that process the ECG signal, Conv1D layers that process the PPG signal, a concatenation layer that mixes the features of PPG and ECG, an LSTM layer, and lastly, a fully connected layer to predict either SBP or DBP.

Figure 4

Distribution of subjects based on the total signal duration for the Training and Testing Datasets. A clustered column plot with a y-axis showing total the number of subjects belonging to a time duration category and an x-axis showing the duration length categories.

Figure 5

Distribution of blood pressure values in the Training and Testing datasets. The SBP and DBP distributions of the Training and Testing datasets are shown in panel (a), and the distributions of calculated PP for the Training and Testing datasets are shown in panel (b). The distributions are shown in a vertically stacking bar plot for both panels (a) and (b) with the y-axis showing the number of samples and the x-axis showing the blood pressure value in mmHg.

Figure 6

Error distributions and the Bland–Altman plots of the predicted SBP and DBP by the proposed model: (a) Error histogram for SBP, (b) Error histogram for DBP, (c) Bland–Altman plot for SBP, and (d) Bland–Altman plot for DBP. Figure panel (a) and (b) show an unimodal distribution of prediction error with the mean shown in bold-dashed lines. Panel (c) and (d) show scattered points of prediction data, shown in red color for SBP and blue color for DBP in the Bland–Altman plot, which shows the difference between the error of the predicted measurements against the averages of the error.

Figure 7

Distribution of subjects based on the total signal duration for the Training and Testing Datasets. Panel (a) shows a scatterplot of actual versus predicted SBP values from Conv1D-LSTM without R-to-R intervals. Panel (b) shows a scatterplot of actual versus predicted DBP values from Conv1D-LSTM with R-to-R intervals. Each plot has a regression line, orange for SBP and purple for DBP. This line overlays a dashed black line indicating perfect predictions.