Estimation of Beat-by-Beat Blood Pressure and Heart Rate From ECG and PPG Using a Fine-Tuned Deep CNN Model

Given that current cuffless blood pressure (BP) measurement technologies feature acceptable overall accuracy, this paper proposed a sufficiently accurate cuffless BP estimation method based on photoplethysmography (PPG) and electrocardiography (ECG) signals. This study used single-channel PPG and ECG signals to estimate heart rate (HR), diastolic BP (DBP), and systolic BP (SBP). A modified long-term recurrent convolutional network comprising a multi-scale convolution network and a long short-term memory (LSTM) network was used to develop a deep learning model for accurately estimating BP and HR. The PPG and ECG signal data of 1551 patients were obtained from the Data Sets-UCI Machine Learning Repository of the University of California, Irvine. The study dataset comprised ECG, PPG, and arterial BP (ABP) signals from the PhysioNet MIMIC II dataset. The original signals were processed by removing noise and artifacts. The aforementioned dataset contains 12,000 records in a hierarchical data format, with each record containing three signals, namely 125-Hz ECG signals from channel II (ECG lead II), 125-Hz PPG signals from the fingertip, and 125-Hz invasive ABP signals. To validate the stability and performance of the developed model, ten-fold cross-validation was conducted. The mean absolute error (MAE) (standard deviation (SD)) values of the developed model for predicting SBP, DBP, and HR were 2.24 mmHg (3.59 mmHg), 1.40 mmHg (2.56 mmHg), and 0.84 bpm (2.23 bpm), respectively. In addition, the estimated SBP and DBP values satisfied the standards of the British Hypertension Society and the Association for the Advancement of Medical Instrumentation. Compared with the methods proposed in other studies, the deep learning model developed in this study required a lower number of layers to provide accurate SBP, DBP, and HR estimations. The results of this study confirmed the effectiveness of the proposed deep learning architecture.


I. INTRODUCTION
Cardiovascular diseases (CVDs) are a major cause of death worldwide. Common CVDs include heart diseases, cerebrovascular diseases, hypertension, nephritis, nephrotic syndrome, and nephropathy. The effective prevention of CVDs The associate editor coordinating the review of this manuscript and approving it for publication was Venkata Rajesh Pamula . requires appropriate health management (i.e., via the periodic measurement of blood pressure (BP)). However, BP measurements in medical settings might be inaccurate because the BP of some patients is higher in medical settings than in routine settings due to the ''white coat effect.'' To address this problem, researchers have used cuffless methods based on photoplethysmography (PPG) and electrocardiography (ECG) signals for home-based BP measurements. VOLUME 10, 2022 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ PPG involves the use of optical sensors to monitor the pulse and detect blood volume changes in the capillary bed, thus, the PPG signal amplitude reflects the blood volume in blood vessels. ECG is the most commonly adopted heart examination method. In this method, electrodes are attached to the chest to collect weak currents generated during heart systole and diastole that reflect potential changes in the peripheral conductive tissues of the heart. When currents flow through the entire body, the electrodes of the ECG instrument convert them into fluctuating waves that can be used to estimate heart rate (HR).
Deep learning methods have been successfully used for addressing various medical problems and have become the main methods adopted in many biomedical applications [1], [2], [3], [4], [5], [6], especially BP estimations based on PPG and ECG signals. Recent studies have used original signals to estimate BP without manual feature extraction. In these studies, features were extracted automatically from signals using deep learning methods. A deep learning model that combines feature extraction and regression analysis for predicting BP can effectively overcome the drawbacks of conducting manual feature extraction on complex data. In addition, deep learning improves the accuracy of BP and HR estimations based on PPG and ECG signals because it enables cardiovascular features to be extracted from these signals.
A recurrent neural network (RNN) is used to perform natural language interpretation and sequence modeling in video processing but if a sequence is long, the traditional RNN creates the exploding or vanishing gradient problem [7]. To solve this long-term dependency problem, studies have proposed two modified RNN-based models, the long short-term memory (LSTM) [8] and gated recurrent unit (GRU) [9] models. Although these models outperform a traditional RNN [10], the LSTM model is more accurate than the GRU model because the LSTM model contains more learning parameters. Thus, the LSTM model was adopted in the present study.
This study used PPG and ECG sensors and the combination of a convolutional neural network (CNN) with an LSTM network. The long-term recurrent convolutional networks (LRCNs) proposed in previous studies were modified to develop new architectures [11], [12]. The PP-Net proposed in [11] is an LRCN and consists of a CNN and an LSTM network. The data of 1,557 participants were collected from the Multi-parameter Intelligent Monitoring in Intensive Care (MIMIC II) database to calibrate and test PP-Net. The standard deviations (SDs) obtained for the predictions of SBP and DBP with the aforementioned model were 5.65 and 5.41, respectively. This network is lightweight and also has good estimation accuracy. A BP estimation model was developed using the receptive field parallel attention shrinkage network to increase the effectiveness of feature extraction from PPG signals. The mean absolute errors (MAEs) ± SDs for the estimated SBP and DBP were 1.63 ± 2.43 and 2.26 ± 4.04 mmHg, respectively [13]. Li et al. proposed a blood pressure estimation method, which can be calibrated with reference inputs rather than with retraining.
Many researchers have explored the relationship between BP and PPG signals in cuffless BP estimations. In literature [15], it proposed framework estimates the BP values through processing vital signals and extracting two types of features, which are based on either physiological parameters or whole-based representation of vital signals. And then the regression algorithms are employed for the BP estimation. A pulse arrival time (PAT)-based algorithm facilitated the continual cuffless SBP, DBP, and mean arterial pressure estimations, with relatively high errors in SBP and DBP of 10.09 and 6.14 mmHg, respectively. Li et al. developed a wavelet neural network algorithm that considers the relationship between PPG signals and BP. This algorithm constructs complete BP waveforms based on PPG signals to extract SBP and diastolic BP (DBP). Most relevant studies have used different machine learning algorithms and manual feature extraction, which is a complicated task that requires considerable computation, to estimate BP [16].
Different machine learning or deep learning technologies have been used to predict BP based on the measured PPG waveforms. Due to the inadequacy of PATs and PPTs to characterize BP, some machine learning works were reported to employ more features in PPG waveforms to achieve better accuracy. The deep learning networks were used to handle feature extraction and establish models integrally, with the aim to cope with the inadequacy of pre-determined (hand-crafted) features to characterize BP and the difficulty of extracting pre-determined features by a designed algorithm. Therefore, the deep learning achieved much better BP accuracy by using a single PPG sensor. Deep learning technologies can, to some extent, adapt to the changes in bio-optical characteristics between individuals and increase the accuracy of BP estimations [17]. An artificial neural network (ANN) has been proposed to predict BP continually using PPG waveforms. The prediction method involved manually extracting a total of 21 features. Although the BP prediction results boasted low SDs, the mean errors (ME) were high [18]. Duan et al. identified 57 candidate features, and three feature sets were eventually proposed, each of which comprised 11 features. The BP was predicted through a support vector regression (SVR) algorithm [19]. Zhang and Feng extracted many detailed features from PPG signals for input into an SVM algorithm to predict BP [20]. PPG signals and the second derivative of a photoplethysmogram have been used to extract many detailed features for BP estimations. In addition to 21 traditional time-domain PPG features, 14 second-derivative-based features were extracted to develop an SVR-based BP measurement instrument [21]. The five predetermined PPG features (i.e., pulse area, pulse rise time, width of 25%, width of 50%, and width of 75%) were extracted and multiple regression analysis, SVM, and decision tree regression were used to calculate BP [22].
Another study used a multitaper method to extract features from signals, and an ANN was adopted to predict SBP and DBP [23]. A deep RNN for BP prediction used seven features as the inputs of a bidirectional LSTM network [24]. A number of new indicators were extracted from PPG recordings and a linear regression method was used to construct BP estimation models based on the PPG indicators and pulse transit time (PTT). results showed that the best PPG-based BP estimation model could achieve a decrease of 0.31 ± 0.08 mmHg in systolic BP (SBP) and 0.33 ± 0.01 mmHg in diastolic BP (DBP) on estimation errors of grand absolute mean (GAM) and standard deviation (GSD) [25]. A waveform-based hierarchical artificial neural network-long short-term memory (ANN-LSTM) model was proposed for BP estimation. The model consists of two hierarchy levels, where the lower hierarchy level uses ANNs to extract necessary morphological features from ECG and PPG waveforms and the upper hierarchy level uses LSTM layers to account for the time domain variation of the features extracted by the lower hierarchy level. The proposed model can automatically extract the necessary features and their time-domain variations to estimate BP reliably in a noninvasive continuous manner [26].
Yan et al. carried out experiments with and without calibration procedure in training stage to evaluate the performance of new method in different application scenarios. The experiment results show that the Deep-BP model achieves high accuracy and outperforms existing methods, in the experiments both with and without calibration [27]. Shimazaki et al. proposed a three-layer automatic coding machine for the automatic generation of features from PPG waveforms to estimate BP [28]. A method to estimate blood pressure and to automatically generate features from pulse wave of PPG used the CNN method. For the input layer of the neural network, pulse wave of one beat, the velocity plethysmography (VPG), the APG, the third derivative pulse wave and the fourth derivative pulse wave are used. Then, age, height, weight, sex, presence or absence of drug, pulse rate are also entered to fully connected layer [29]. Slapničar et al. used the PPG alongside its first and second derivative as inputs for a novel spectro-temporal deep neural network with residual connections. In a leave-onesubject-out experiment, the network could model the dependency between PPG and BP, achieving mean absolute errors of 9.43 for systolic and 6.88 for diastolic BP. Additionally, the personalization of models is important and substantially improves the results, while deriving a good general predictive model is difficult [30].In literature [31], it proposed two schemes: extraction through multiple dilated convolution, and concentration through strided convolution with a large kernel, to process sequential ECG and PPG signals through CNN. The results shown that the BP prediction performance was the best when both ECG and PPG signals were used together. Another study estimated BP using a model consisting of a CNN, a bidirectional GRU (BiGRU), and an attention mechanism. Although the aforementioned model exhibited low errors in SBP and DBP predictions, the number of participants (n = 15) was too low to prove the robustness of the model [32]. An ANN was used to develop a nonlinear dynamic autoregression model with a nonlinear autoregressive exogenous network, with the feature extraction performed on ECG and PPG waveforms to predict BP [33]. A Siamese network was used to estimate BP from PPG signals, and the estimated BP approached the actual BP [34]. A U-Net structure was used to convert PPG waveforms into arterial BP (ABP) waveforms, with SBP and DBP predicted according to the peaks and troughs of the ABP waveforms. The SDs for the predictions of SBP and DBP with the aforementioned waveforms were 4.42 and 2.92, respectively [35]. Aguirre et al. proposed a structure comprising a GRU and an attention mechanism to estimate BP but the developed structure had large estimation errors, with the SDs for DBP and SBP predictions being 7.32 and 15.67, respectively [36].
The results obtained from these references are illustrated. For estimation accuracy, the depth of the network must be increased or complex signal processing techniques must be involved. Because the limitations of hardware devices, the network of wearable devices need to be lightweight. Therefore, a lightweight network was designed to accurately estimate BP and HR in the study. The proposed multi-scale convolution CNN-LSTM model combines the advantages of an LSTM model, a CNN model, and the feature extraction function of the multi-scale convolution models. The MAEs of the proposed network to predict SBP, DBP, and HR were 2.24, 1.40, and 0.84, respectively, with an R 2 score of 0.99. Since the objective of the present study was to estimate BP and HR accurately to aid the prevention of CVDs, verification in patients with CVDs is critical and necessary for clinical applications, especially for monitoring of heart and stroke in critically ill postoperative patients.
The remainder of this paper is structured as follows. Section II describes the adopted methods. Section III presents the obtained results and a discussion of the results. Section IV concludes this study and offers future recommendations.

II. METHODS
During the conducted experiment, a cuffless BP estimation dataset was used as the source of ECG, PPG, and ABP signals. The proposed deep learning framework is tested on University of California, Irvine (UCI) Machine learning Repository dataset, derived from publically available largest database 'Multi-parameter Intelligent Monitoring in Intensive Care (MIMIC-II)' which is available at Physionet repository. This database consists simultaneous recordings of multi-parameters of Intensive care unit (ICU) patients which include physiological signals as well as physiological parameters. This dataset was generated by Kachuee et al. [37], who collected ECG, PPG, and ABP signals from the Phy-sioNet MIMIC II dataset [38]. The collected original signals were processed by removing noises and artifacts. The aforementioned dataset contains 12000 records in a hierarchical data format. Each record contains three signals, VOLUME 10, 2022 namely 125-Hz ECG signals from channel II (ECG lead II), 125-Hz PPG signals from the fingertip, and 125-Hz invasive ABP signals. The training and testing values of SBP and DBP were obtained by extracting the peaks and troughs of ABP signals. A complete heartbeat is the complete cycle of each contraction and relaxation of the heart, so we calculate the number of peaks and refer to the ABP signal to estimate the HR value. The ground truth scores for SBP, DBP and HR are calculated using the ABP signal, applying the approach used in the existing studies. The proposed framework simultaneously estimates SBP, DBP, and HR after obtaining PPG and ECG signals, which facilitates the realization of extensive medical and health-care monitoring.

A. PREPOSSINMG
In the preprocessing stage of this study, data with a duration shorter than 8 min were removed. Data with missing values (Nan) and poor signal quality such as extremely high or low mean BP and HR over 8 min (SBP ≤ 80, SBP ≥ 180, DBP ≤ 60, DBP ≥ 130, HR < 40, or HR > 220) were also removed. After the removal of the aforementioned data, the total number of records reduced from 12 000 to 1551.
The ABP signals were not processed. They were simply used to train, validate, and test the model developed in this study for predicting SBP, DBP, and HR.
The PPG signals were processed using a bandpass filter, namely the fourth-order Chebyshev II filter, in a frequency range of 0.5-10 Hz. The low-and high-frequency noises in the PPG signals were filtered and removed. Polynomial fitting was then conducted to process the baseline drift of the PPG signals. Subsequently, the data were divided into 8-s-long segments, and 75% of each data segments overlapped with the preceding and following data segments.
ECG signals were processed using a 0.1-Hz, eighth-order, low-pass filter to remove low-frequency noises and baseline drift from these signals. The discrete wavelet transform was then used to filter out noises with a frequency higher than 90 Hz. Subsequently, the data were divided into 8-s segments, and 75% of each data segment overlapped with the preceding and following data segments.
After the aforementioned preprocessing, ECG and PPG signals without baseline drift and high-frequency noises were obtained. Subsequently, signals with extremely high or low BP or HR values in the 8-s segments (SBP ≤ 80, SBP ≥ 180, DBP ≤ 60, DBP ≥ 130, HR < 40, or HR > 220) were filtered out to eliminate poor signals. ECG and PPG signals with identical numbers of P-peaks were retained, as shown in Fig. 1. Downsampling was then conducted to convert the obtained 1000 data segments with a duration of 8 s into 250 data segments. This approach reduced the data quantity and enabled the developed model to be implemented on platforms with limited computing resources. Subsequently, the processed PPG data; processed ECG data; and HR, SBP, and DBP data obtained from the ABP signals were processed using an algorithm to facilitate their use in the developed deep learning network.

B. DEEP LEARNING MODEL
A CNN extracts key features from large quantities of data, such as images, and exhibits remarkable performance in processing signals in the biomedical field [39], [40], [41], [42], [43]. Most of the previous studies used single-scale convolution for feature extraction of BP value estimation. It focusing on the information of the waveform itself, while ignoring the potentially useful information at different scales. Therefore, in order to obtain the characteristics of different PPG/ECG distributions at different scales. The deep learning model proposed in this study comprises two convolutional layers of two different sizes and an LSTM layer. It has simple and lightweight network characteristics and is suitable for wearable devices. This model automatically extracts features from the input signals. The adopted deep learning network was designed to provide multiple outputs simultaneously (i.e., DBP, SBP, and HR) on the basis of PPG and ECG signals. Fig. 2 illustrates the overall architecture of the proposed model, which has a blended architecture. PPG and ECG data are input into convolutional layers of two different sizes to extract feature vectors [44]. The symbol ''+'' refers to concatenated operation. The two-scale convolution model used in this study is made up of two convolutional layers with different sizes, activation functions, pooling layers, and a dropout layer. The activation functions are accelerated and trained using the rectified linear unit activation function. The outputs of the two blocks of the adopted CNN are transmitted to two LSTM models to extract temporal feature vectors. These outputs are then transferred to the output layer of the developed model to calculate SBP, DBP, and HR.

C. TRAINING AND EVALUATION
A total of 150 batches containing 100 pieces of data each were used to train the developed deep learning model. The Adam optimizer was used to conduct optimization during model training. MAE and SD were used as loss functions to evaluate the performance of the proposed model. The proposed model was also evaluated using k-fold cross-validation because simple train-test splitting methods could validate the generalizability of the model. In this study, k was set This section presents the BP results estimated by the proposed model when using the PPG + ECG data as input. The performance of the proposed model and that of models proposed in previous studies were compared. This study verified whether the results obtained with the proposed model met the standards of the British Hypertension Society (BHS) and the Association for the Advancement of Medical Instrumentation.

A. PERFORMANCE EVALUATION
This study designed a deep learning model with a limited number of layers and favorable accuracy to enable the real-time monitoring of BP and HR on mobile platforms. Thus, the architecture of an LRCN was modified by replacing its convolutional layer with a two-scale convolution model. This model enables the simultaneous estimations of DBP, SBP, and HR. In addition, downsampling was conducted to compress the input PPG and ECG signals, thereby reducing the data processing complexity, directly decreasing the computational load of the developed deep learning model.
In the present study, only PPG data, only ECG data, and PPG + ECG data were used as inputs. Moreover, after using the two layers of two-scale convolution model, the LSTM and GRU models were used to identify which input led to the optimal results. When using PPG signals as inputs, a single-layer LSTM model, and the two layers of two-scale CNN. The ten-fold cross-validation method was used to estimate the MAEs ± SDs of the model. The results revealed that under the aforementioned settings, the MAEs ± SDs values of the proposed model for predictions of SBP, DBP, and HR were 3.46 ± 5.90 mmHg, 2.01 ± 3.65 mmHg, and 1.13 ± 2.49 bpm, respectively. When using ECG signals as inputs, a singlelayer LSTM model, and the two layers of two-scale CNN. The ten-fold cross-validation method was used to estimate the MAEs ± SDs of the model. The results revealed that under the aforementioned settings, the MAEs ± SDs values of the proposed model for predictions of SBP, DBP, and HR were 2.63 ± 4.20 mmHg, 1.62 ± 2.84 mmHg, and 1.20 ± 2.40 bpm, respectively. When using PPG + ECG signals as inputs, a single-layer LSTM model, and the two layers of two-scale CNN. The ten-fold cross-validation method was used to estimate the MAEs ± SDs of the model. The results revealed that under the aforementioned settings, the MAEs ± SDs values of the proposed model for predictions of SBP, DBP, and HR were 2.44 ± 3.59 mmHg, 1.40 ± 2.56 mmHg, and 0.84 ± 2.23 bpm, respectively.
When using PPG signals as inputs, a single-layer GRU model, and the two layers of two-scale CNN. The ten-fold cross-validation method was used to estimate the MAEs ± SDs of the model. The results revealed that under the aforementioned settings, the MAEs ± SDs values of the proposed model for predictions of SBP, DBP, and HR were 3.65 ± 5.38 mmHg, 2.10 ± 3.45 mmHg, and 1.11 ± 2.41 bpm, respectively. When using ECG signals as inputs, a singlelayer GRU model, and the two layers of two-scale CNN. The ten-fold cross-validation method was used to estimate the MAEs ± SDs of the model. The results revealed that under the aforementioned settings, the MAEs ± SDs values of the proposed model for predictions of SBP, DBP, and HR were 3.03 ± 4.71 mmHg, 1.85 ± 3.17 mmHg, and 1.11 ± 2.43 bpm, respectively. When using PPG + ECG signals as inputs, a single-layer GRU model, and the two layers of twoscale CNN. The ten-fold cross-validation method was used to estimate the MAEs ± SDs of the model. The results revealed that under the aforementioned settings, the MAEs ± SDs values of the proposed model for predictions of SBP, DBP, and HR were 2.85 ± 4.37 mmHg, 1.73 ± 3.05 mmHg, and 0.90 ± 2.31 bpm, respectively.
From the experiment results, the optimal MAEs of 2.24 mmHg, 1.40 mmHg, and 0.84 bpm for SBP, DBP, and HR, respectively, were achieved when using PPG + ECG data as inputs, a single-layer LSTM model, and the two layers of two-scale CNN. Fig. 3 displays the loss curve of the proposed model under the aforementioned settings. The ten-fold validation method was used to validate this loss curve. The results indicated that the proposed model converged quickly within a few layers and attained favorable performance under VOLUME 10, 2022    the aforementioned conditions. Fig. 4 displays the normal distributions of the errors in SBP, DBP, and HR estimation.
From Tables 1, it can be found that two layers of two-scale CNN + LSTM model is better in SBP and DBP estimations under single PPG, ECG and PPG+ECG signals. However, in terms of HR estimation, two layers of two-scale CNN + GRU model is better in PPG or ECG signal inputs, which may be caused by the different distribution range of blood pressure and heart rate. If we only concern of the balance design between lightweight network and estimation accuracy, the two layers of two-scale CNN + GRU model is being a good choice.

B. ASSESSMENT WITH RESPECT TO THE BHS AND AAMI STANDARDS
The BP prediction results obtained with the proposed model were assessed with respect to the BHS and AAMI standards to validate the effectiveness of using PPG + ECG signals as inputs, a single-layer LSTM model, and the two-scale CNN. Cumulative error percentage was used to assess the model accuracy with respect to the BHS standard. This percentage is classified into three categories, as presented in Table 2. The BP prediction results of the proposed model were validated using the actual data for 1551 individuals ( Table 2). According to BHS standard, the DBP and SBP results of the proposed model were classified into Grade A, which is within the defined standard scopes. According to the AAMI standards, the ME and SD of BP measurement instruments should be lower than 5 and 8 mmHg, respectively. Table 3 presents the performance of different models with respect to the AAMI standards. The proposed model satisfied the AAMI standards. Fig. 5 depicts the Bland-Altman plots of ten-fold validation for the DBP, SBP, and HR predictions. The correlation coefficients are presented in Fig. 6.

C. COMPARISON OF THE PROPOSED MODEL WITH MODELS PROPOSED IN PREVIOUS STUDIES
To verify the effectiveness of the proposed model, its performance was compared with that of models proposed in previous studies. There are two main ways to improve the accuracy of estimating blood pressure and heart rate. The first is to improve the signal clarity to increase the network feature extraction capability or via a feature extraction method to identify useful features. The second is to improve the network model, for example, a lightweight network designed for use in wearable devices. An attention mechanism network improves the accuracy of numerical estimation and a regression method is used to increase the accuracy of numerical estimation. Also, more features can be used to form a residual network VOLUME 10, 2022 (ResNet) [30]. For this study, it was decided to focus on improving the network's ability to extract features from the input signal. Table 4 presents a comparison of the BP estimation results obtained with the proposed model and other models. It lists the feature selection and extraction steps of the models proposed in previous studies. The proposed model exhibited MAE ± SD values of 2.44 ± 3.59 mmHg and 1.40 ± 2.56 mmHg for predictions of the SBP and DBP, respectively, of 1551 individuals. The BP prediction results obtained with the proposed model were superior to those obtained using the models proposed in previous studies. Thus, the proposed model is effective in real-time clinical applications. Table 5 presents a comparison of the HR values predicted by the proposed model and other models. The datasets used for the proposed model differed from those used in most previous relevant studies. Moreover, the proposed model uses more data for estimation than do the other models and thus has higher accuracy and generalizability than the other models do. It indicates that many of the considered studies estimated HR separately to reduce the load during model training. The proposed model estimates DBP, SBP, and HR simultaneously, which increases the load of weight training for the model. However, the results of this study revealed that the load of the proposed model differed marginally from those of the other models. Moreover, the HR prediction results of the proposed model were superior to those of most of the other models. The proposed model achieved an MAE ± SD of 0.84 ± 2.23 bpm in HR estimation.
In addition to the need for accurate blood pressure and heart rate estimation for the proposed technology, lightweight networking is also an important consideration in order to be 85466 VOLUME 10, 2022 used in wearable devices. Therefore, the balance between accuracy and lightweight is the criterion for our network design. If more accurate estimates are required in the future, the advanced data processing technology and depth of the network can be further increased. Because the proposed model simultaneously estimates DBP, SBP, and HR separate training and feature extraction need not be conducted for each parameter. Therefore, in real-time system analysis, the computation load of the proposed model is lower than those of models in which DBP, SBP, and HR are estimated separately.

IV. CONCLUSION
The results of this study confirmed the effectiveness of the proposed deep learning model in the simultaneous estimations of three physiological parameters: SBP, DBP, and HR. In the present study, the physiological data of 1551 patients were examined. The optimal prediction results with the proposed model were achieved when using PPG + ECG data as inputs, the combination of an LSTM model, and the two layers of two-scale convolution model. The optimal MAE ± SD values for the predictions of SBP, DBP, and HR were 2.24 ± 3.59 mmHg, 1.40 ± 2.56 mmHg, and 0.84 ± 2.23 bpm, respectively. Moreover, the optimal R 2 score was 0.984. The SBP, DBP, and HR estimation accuracy performance of the proposed model outperformed in previous studies.
The SBP and DBP estimation results that ranked as Grade A according to the BHS standard, which validated the effectiveness of the proposed model in estimating physiological information. It indicates that the proposed model has potential for use in general health-care monitoring for patients with diabetes, patients rehabilitating from stroke, older patients, postoperative patients, and can also be feasibly used in telemedicine.