Robust modelling of arterial blood pressure reconstruction from photoplethysmography Authors: Jiating Pan, Lishi Liang, Yongbo Liang, Qunfeng Tang, Zhencheng Chen, Jianming Zhu Published: 2024 ABSTRACT Blood pressure is a crucial indicator of cardiovascular disease, and arterial blood pressure (ABP) waveforms contain information that reflects the cardiovascular status. We propose a novel deep-learning method that converts photoplethysmogram (PPG) signals into ABP waveforms. We used \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net as a feature extractor and designed a Bi-block to capture individualised time information in encoder feature extraction. We further enhanced the prediction accuracy of the ABP waveforms by applying a combined loss function to each layer of deep supervision. We also propose a total error index (TEI) to measure overall performance. Furthermore, we extended our method from the UCI dataset to the VitalDB dataset, achieving mean absolute error ± standard deviation (MAE ± STD) values of 2.48 ± 1.95, 1.42 ± 1.42, and 1.48 ± 1.36 mmHg for systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean arterial pressure (MAP) in UCI dataset, and 2.16 ± 1.53, 1.12 ± 0.59, and 1.35 ± 0.84 mmHg in VitalDB dataset, respectively. The mean ± STD values of the TEI index are 0.29 ± 0.10 in UCI dataset and 0.29 ± 0.15 in VitalDB dataset. These results demonstrate the superiority of the proposed method over existing methods and its robustness to different sampling frequencies and devices. Introduction Continuous blood pressure monitoring is essential for maintaining overall health and preventing cardiovascular diseases, such as coronary heart disease, stroke, and kidney failure, as hypertension is a significant risk factor for these conditions1. Research indicates that individuals with prehypertension have a 3.2/1.74 times higher risk of developing cardiovascular diseases than those with normal blood pressure2. In 2015, approximately 10.7 million deaths (19.2% of all deaths) were attributed to all causes with SBP between 110 and 115 mmHg, and 7.8 million deaths (14.0% of all deaths) were associated with SBP of 140 mmHg3. Consequently, blood pressure is a critical indicator of overall health and requires timely intervention when abnormal blood pressure is detected. Furthermore, postoperative blood pressure monitoring is crucial for patient’ recovery following interventions such as percutaneous coronary intervention, as intraoperative blood pressure waveforms can provide a model for postoperative monitoring, ensuring a safe and effective return to work and predicting patients’ quality of life4. Blood pressure represents the force exerted by the blood on the arterial wall per unit area and is the sum of the haemodynamic, kinetic, and hydrostatic energies5. Blood pressure not only reflects the presence of hypertension but also indicates arterial and cardiac functions. Blood pressure measurement methods include numerical and waveform measurements. The clinical measurement of blood pressure values typically involves using a cuff to measure SBP and DBP. Invasive catheterisation allows continuous blood pressure waveform measurement. Cuff-based blood pressure monitors provide only two parameters of the arterial pressure waveform, SBP, and DBP, which are suitable for the routine diagnosis of hypertension. However, waveform morphology and beat-to-beat variability can reflect a wide range of information, including aortic valve disease and myocardial functional disorders6. For acute patients, interpreting their condition solely based on changes in SBP and DBP can lead to misdiagnosis because of the complexity of various factors that cause changes in ABP and their interrelationships. Blood pressure waveform analysis can provide more physiological information, which is beneficial for diagnosis. In addition to providing beat-by-beat estimates of SBP, DBP, and MAP, blood pressure waveforms can also provide left ventricular function assessment by evaluating the heart rate, systolic time intervals, and other information7. Analysis of ABP waveforms is a key method for obtaining arterial function-related parameters such as peripheral resistance, arterial compliance, and wave reflection8. Therefore, analysing blood pressure waveforms can provide a more in-depth understanding of the haemodynamic state6,9. Although ABP waveforms can offer valuable physiological insights, the invasiveness of the gold-standard method for measuring ABP limits its utility in daily health monitoring and postoperative recovery10. Consequently, the non-invasive measurement of the ABP waveform is considerably important. Therefore, the development and application of PPG are promising alternatives. PPG is a non-invasive technique that indirectly measures changes in blood volume using a light source and a photodetector placed on the skin. For decades, this cost-effective method has been acknowledged for its potential in assessing cardiovascular diseases and is widely employed in areas such as blood pressure measurement and vascular assessment11. The rationale for using PPG to predict ABP is as follows. First, as shown in Fig.1, PPG and ABP have a remarkable similarity in morphology, with similar upstrokes, downstrokes, and dicrotic notches. Moreover, ABP and PPG in individuals with normal blood pressure, prehypertension, and hypertension displayed a strong resemblance in both the time and frequency domains, suggesting a consistent and significant causal relationship. These similarities suggest a consistent and significant causal relationship, thereby enabling the prediction of ABP from PPG signals12.Fig. 1Synchronised ABP and PPG signals. Synchronised ABP and PPG signals. In recent years, substantial advancements have been made in the field of non-invasive blood pressure monitoring through the reconstruction of ABP waveforms from PPG signals. A variety of methods have been proposed to address this challenge. Ibtehaz et al. proposed PPG2ABP, which is an approach for predicting the ABP from PPG signals. This method uses a one-dimensional U-Net approximation network after filtering, and a one-dimensional MultiResUNet refinement network to refine the results13. Cheng et al. introduced ABP-Net, a Wave-U-Net-based architecture that reconstructs ABP waveforms using the PPG, its first-order derivative (VPG), and its second-order derivative (APG) as inputs14. Tang et al. presented a deep learning model with a W-Net architecture comprising of two cascaded U-Net structures for reconstructing ABP from PPG signals15. Additionally, Qin et al. developed a deep autoencoder (DAE) model for reconstructing ABP pressure waveforms from PPG signals16. Although their results were superior to other related experiments, the need for both PPG and ECG signals complicates its application. Athaya et al. used a U-NET to reconstruct ABP signals from PPG signals and predict SBP and DBP, but their data was only a small subset of the MIMIC dataset17.However, they conducted experiments on a single dataset and evaluated the prediction results using the predicted values of SBP, DBP, and MAP. A comprehensive metric is required to measure the numerical prediction accuracy and similarity of the reconstructed blood pressure waveforms. This study established an end-to-end model for predicting ABP using a one-dimensional signal transformation of PPG signals. This model is based on an improved version of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net, a deep convolutional neural network with a U-Net structure primarily used for salient object detection (SOD)18. In \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net, the encoder and decoder have multi-scale feature maps that capture signal features at different scales. To obtain long-term dependencies in the sequence, we integrated the concept of bidirectional Long Short-Term into our model. This integration significantly boosted the capability of the model to predict the ABP from PPG signals. We designed a comprehensive loss function based on existing experiments under deep supervision. We propose a novel integrated evaluation metric that measures the accuracy of the predicted values and assesses the waveform similarity. We also collected data from two different datasets to verify the universality of the model and the experiment. The results showed that the reconstructed ABP waveforms in both datasets were satisfactory in terms of waveform similarity and achieved satisfactory results for SBP, DBP, and MAP indicators. Data sources The data used in this experiment were obtained from the UCI dataset19and VitalDB20 dataset. The MIMIC II dataset contained signals sampled at 125 Hz. In this dataset, the ABP signals were derived from invasive radial artery measurements, whereas the PPG signals were obtained from fingertip measurements. Due to the large size of the MIMIC II dataset and the variability in the quality of signals, some of which contain distortions and artifacts, we opted for the well-organized UCI dataset21, which is also derived from the MIMIC II dataset and has been preprocessed by Mohamad Kachuee and others. UCI dataset comprises three types of signals: PPG, ABP, and electrocardiogram (ECG), totalling 12000 records. For this experiment, only the PPG and ABP signals were required. Considering related studies, we extracted records with durations exceeding 8 mins and excluded records with ABP values greater than 200 mmHg14,15. Ultimately, we selected the first 150 records from the filtered data as the research subjects. Fig.2 shows the distributions of SBP and DBP in the training and prediction sets after segmenting 150 records. After segmentation, 38829 signal segments were obtained. The VitalDB dataset contained data sampled at 500 Hz. The PPG and ABP signals in this dataset were obtained from a GE Solar8000 monitor, with PPG signals measured from the fingertip and ABP signals derived from invasive radial artery measurements. Initially, we acquired cases containing PPG and ABP signals from a Web Monitoring dataset and downloaded the data for the first twenty cases for experimentation. Specific information regarding the experimental subjects is presented in Table 1. We also extracted two synchronised signals with more than 8 min duration after signal stabilisation and ABP values below 200 mmHg. Fig.2 illustrates the distributions of the SBP and DBP following signal segmentation. In total, 7172 signal segments were obtained.Table 1Individual Information in VitalDB dataset.CaseidAgeSexOpname177MLow anterior resection474MDistal gastrectomy566MAneurysmal repair752FLung lobectomy1072MDistal gastrectomy1367FMetastasectomy1657MLiver segmentectomy1785MAneurysmal repair1974MPylorus preserving pancreaticoduodenectomy2253FTotal thyroidectomy2566MTotal gastrectomy2664MLung lobectomy2750FDistal gastrectomy2867FThymectomy2953FDonor hepatectomy3158MLung wedge resection3469FBreast-conserving surgery3866MLung lobectomy4145MNephrectomy4364MLow anterior resection Table 1Individual Information in VitalDB dataset.CaseidAgeSexOpname177MLow anterior resection474MDistal gastrectomy566MAneurysmal repair752FLung lobectomy1072MDistal gastrectomy1367FMetastasectomy1657MLiver segmentectomy1785MAneurysmal repair1974MPylorus preserving pancreaticoduodenectomy2253FTotal thyroidectomy2566MTotal gastrectomy2664MLung lobectomy2750FDistal gastrectomy2867FThymectomy2953FDonor hepatectomy3158MLung wedge resection3469FBreast-conserving surgery3866MLung lobectomy4145MNephrectomy4364MLow anterior resection Individual Information in VitalDB dataset. The distributions of the training and prediction sets were similar for the individual experiments. This could be inferred from the distribution plots. Moreover, the blood pressure distribution encompassed normal, high, and low blood pressure, which enhanced the adaptability of the model.Fig. 2Distribution of SBP and DBP in the training and prediction sets for two datasets. Distribution of SBP and DBP in the training and prediction sets for two datasets. Input sigals Studies have revealed that PPG derivatives, specifically VPG and APG, provide additional heart cycle-related information closely associated with blood pressure22. Moreover, experiments verified that VPG and APG can improve pressure prediction23. Consequently, this study incorporated PPG, VPG, and APG as model inputs to enhance the ABP waveform prediction. In this study, we utilized two different datasets: UCI dataset and VitalDB, with sampling rates of 125 Hz and 500 Hz, respectively. The difference in sampling rates necessitated the use of different mathematical methods for calculating the APG and VPG. We calculated APG and VPG for both datasets using the 2-point method. Suppose each segment of the PPG signal is denoted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{S}_{{ppg}}}^{(n) }$$\end{document}, with the PPG signal segment containing N sample points from the two datasets, each with a data length of 8.192 s. The calculation of VPG (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{S}_{{vpg}}}^{(n) }$$\end{document} ) is then as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {S_{{vpg}}}^{(n)}(n) = {\left\{ \begin{array}{ll} {S_{{ppg}}}^{(n)}(n)-{S_{{ppg}}}^{(n)}(n-1) & \text {, } 1 < n \le N \\ {S_{{ppg}}}^{(n)}(n+1)-{S_{{ppg}}}^{(n)}(n) & \text {, } n = 1 \end{array}\right. } \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{{ppg}}^{(n)}(n)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{{vpg}}^{(n)}(n)$$\end{document} are the n-th sample points of the PPG and VPG signals, respectively. The calculation of APG(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{S}_{{apg}}}^{(n) }$$\end{document}) is as follows:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {S_{{apg}}}^{(n)}(n) = {\left\{ \begin{array}{ll} {S_{{vpg}}}^{(n)}(n)-{S_{{vpg}}}^{(n)}(n-1) & \text {, } 1 < n \le N \\ {S_{{vpg}}}^{(n)}(n+1)-{S_{{vpg}}}^{(n)}(n) & \text {, } n = 1 \end{array}\right. } \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{S}_{{apg}}}^{(n) }(n)$$\end{document} denotes the n-th sample point of the APG signal. Preprocessing of datasets The preprocessing operations for both the UCI and VitalDB datasets shared many similarities, with some adjustments due to differences in signal acquisition and sampling rates. Both datasets underwent the following steps:Detrending: Linear detrending was applied to PPG signals to remove baseline drift, device-related interference, and linear trends, thus improving signal analysis accuracy.Segmentation: For both datasets, synchronized PPG and ABP signals were segmented into 8.192-second-long segments. However, due to the different sampling rates, each segment contained 1024 sample points for the UCI dataset (125 Hz) and 4096 sample points for the VitalDB dataset (500 Hz). A 75% overlap was maintained between consecutive segments. The corresponding VPG and APG signals were calculated using equations 1 and 2, respectively. For both datasets, 80% of the total data segments (containing PPG, VPG, APG, and ABP signals) were used for training, while the remaining 20% were used for testing.Normalisation: Min-max normalisation was applied to the PPG, VPG, and APG signals in both datasets, with the normalised signals defined as: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} X_{\text {norm}} = \frac{X - X_{\text {min}}}{X_{\text {max}} - X_{\text {min}}} \end{aligned}$$\end{document} where X represents the original PPG, VPG, or APG signals, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{\text {norm}}$$\end{document} is the normalised signal, ranging from 0 to 1. ABP signals were normalised by dividing the signal by 200.Despite these common steps, the two datasets required some additional dataset-specific operations:UCI Dataset: As the dataset had already undergone preprocessing operations (removal of unacceptable signals and smoothing using a simple average filter21), we applied only detrending, segmentation, and normalisation.VitalDB Dataset: Because this dataset represented continuous signals from the entire surgical procedure, a stable measurement interval was selected after signal stabilisation. In addition to detrending, segmentation, and normalisation, the following steps were also performed:Missing-value Imputation: Linear interpolation was applied to NaN values.Smoothing: A third-order Savitzky-Golay filter with a window size of 101 was used to smooth the signals, ensuring consistent preprocessing across datasets. Detrending: Linear detrending was applied to PPG signals to remove baseline drift, device-related interference, and linear trends, thus improving signal analysis accuracy. Detrending: Linear detrending was applied to PPG signals to remove baseline drift, device-related interference, and linear trends, thus improving signal analysis accuracy. Segmentation: For both datasets, synchronized PPG and ABP signals were segmented into 8.192-second-long segments. However, due to the different sampling rates, each segment contained 1024 sample points for the UCI dataset (125 Hz) and 4096 sample points for the VitalDB dataset (500 Hz). A 75% overlap was maintained between consecutive segments. The corresponding VPG and APG signals were calculated using equations 1 and 2, respectively. For both datasets, 80% of the total data segments (containing PPG, VPG, APG, and ABP signals) were used for training, while the remaining 20% were used for testing. Segmentation: For both datasets, synchronized PPG and ABP signals were segmented into 8.192-second-long segments. However, due to the different sampling rates, each segment contained 1024 sample points for the UCI dataset (125 Hz) and 4096 sample points for the VitalDB dataset (500 Hz). A 75% overlap was maintained between consecutive segments. The corresponding VPG and APG signals were calculated using equations 1 and 2, respectively. For both datasets, 80% of the total data segments (containing PPG, VPG, APG, and ABP signals) were used for training, while the remaining 20% were used for testing. Normalisation: Min-max normalisation was applied to the PPG, VPG, and APG signals in both datasets, with the normalised signals defined as: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} X_{\text {norm}} = \frac{X - X_{\text {min}}}{X_{\text {max}} - X_{\text {min}}} \end{aligned}$$\end{document} where X represents the original PPG, VPG, or APG signals, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{\text {norm}}$$\end{document} is the normalised signal, ranging from 0 to 1. ABP signals were normalised by dividing the signal by 200. Normalisation: Min-max normalisation was applied to the PPG, VPG, and APG signals in both datasets, with the normalised signals defined as: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} X_{\text {norm}} = \frac{X - X_{\text {min}}}{X_{\text {max}} - X_{\text {min}}} \end{aligned}$$\end{document} where X represents the original PPG, VPG, or APG signals, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{\text {norm}}$$\end{document} is the normalised signal, ranging from 0 to 1. ABP signals were normalised by dividing the signal by 200. UCI Dataset: As the dataset had already undergone preprocessing operations (removal of unacceptable signals and smoothing using a simple average filter21), we applied only detrending, segmentation, and normalisation. UCI Dataset: As the dataset had already undergone preprocessing operations (removal of unacceptable signals and smoothing using a simple average filter21), we applied only detrending, segmentation, and normalisation. VitalDB Dataset: Because this dataset represented continuous signals from the entire surgical procedure, a stable measurement interval was selected after signal stabilisation. In addition to detrending, segmentation, and normalisation, the following steps were also performed:Missing-value Imputation: Linear interpolation was applied to NaN values.Smoothing: A third-order Savitzky-Golay filter with a window size of 101 was used to smooth the signals, ensuring consistent preprocessing across datasets. VitalDB Dataset: Because this dataset represented continuous signals from the entire surgical procedure, a stable measurement interval was selected after signal stabilisation. In addition to detrending, segmentation, and normalisation, the following steps were also performed:Missing-value Imputation: Linear interpolation was applied to NaN values.Smoothing: A third-order Savitzky-Golay filter with a window size of 101 was used to smooth the signals, ensuring consistent preprocessing across datasets. Missing-value Imputation: Linear interpolation was applied to NaN values. Missing-value Imputation: Linear interpolation was applied to NaN values. Smoothing: A third-order Savitzky-Golay filter with a window size of 101 was used to smooth the signals, ensuring consistent preprocessing across datasets. Smoothing: A third-order Savitzky-Golay filter with a window size of 101 was used to smooth the signals, ensuring consistent preprocessing across datasets. Method Model As shown in Fig.3, the proposed signal conversion model is based on a one-dimensional \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net with deep supervision. The Residual U (RSU) module follows the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net model structure, and the Bi-block includes two tensor-dimensional transformations: a BiLSTM layer, and a 1 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 1 one-dimensional convolution layer. ‘Maxpool’ represents the max-pooling operation, ‘Conv1d’ is the one-dimensional convolution operation, ‘Upsample’ refers to the upsampling operation using linear interpolation, and ‘±’ denotes the concatenate operation. We uses \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net as a feature extractor to obtain spatial features from the raw input data and then input these features into the Bi-block to capture temporal dependencies. This approach combines the advantages of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net for spatial feature extraction and BiLSTM for handling long- and short-term temporal dependencies. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net was initially designed for SOD and is a nested U-Net model structure. The RSU module proposed in this model mixes receptive fields of different sizes, thereby enabling it to obtain additional information at different scales. The encoder output is concatenated with the features of the current level at each decoder stage. This fusion process leads to the loss or blurring of temporal information. To improve the model’s efficiency of the model in utilising temporal information and better leverage the temporal features extracted by the encoder during decoding, we used Bi-blocks to process the temporal information of these fused features.Fig. 3Detailed Architecture of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net-Based Signal Conversion Model with Deep Supervision and Bi-block Integration. Detailed Architecture of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net-Based Signal Conversion Model with Deep Supervision and Bi-block Integration. Loss function Because deep supervision was used in this experiment, supervision signals were introduced at every network level, and the loss was calculated as the sum of the loss functions at each level and the final output loss. Studies have shown that using the maximum absolute difference and Pearson correlation coefficient as loss functions can improve prediction accuracy14,15. Therefore, we adopted a combined loss function incorporating the mean squared error (MSE), maximum absolute difference, and Pearson correlation coefficient.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} L = \alpha \cdot MSE + \beta \cdot Max + \gamma \cdot (1 - r) \end{aligned}$$\end{document}where L represents the combined loss, MSE is the mean squared error between the predictions and targets, Max is the maximum absolute difference between the predictions and targets, and r is the Pearson correlation coefficient between predictions and targets. The weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document} were set to 1, 0.1 and 1. Subsequently, the final multi-level loss is calculated as the sum of the combined losses at each level:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} L_{sum} = L_0+L_1+L_2+L_3+L_4+L_5+L_6 \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_2$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_3$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_4$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_5$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_6$$\end{document} represent the loss functions of the supervision signals at the intermediate level before the final concatenation operation, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_0$$\end{document} denotes the loss function of the final output. Evaluation metrics We assessed the accuracy of important numerical values using two metrics: MAE and STD. In addition, we introduced a comprehensive index called TEI to measure the overall accuracy of the waveform predictions. The most important numerical values were SBP, DBP, and MAP. The formulae are as follows:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} SBP=max({S_{ABP}}^{(n)}) \end{aligned}$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} DBP=min({S_{ABP}}^{(n)}) \end{aligned}$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} MBP=\frac{1}{N}\sum _{n=1}^{n}({S_{ABP}}^{(n)}(n)) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S_{ABP}}^{(n)}$$\end{document} is an ABP segment with N points, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S_{ABP}}^{(n)}(n)$$\end{document} represents the n-th point.MAE of SBP, DBP, and MAP The MAE of SBP, DBP, and MAP are defined as: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} MAE_s=\frac{1}{M}\sum _{m=1}^{M}\left| {BP_{pre}}^{(m)}(m)-{BP_{gt}}^{(m)}(m) \right| \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${BP_{pre}}^{(m)}(m)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${BP_{gt}}^{(m)}(m)$$\end{document} represent the m-th predicted and ground truth values of the blood pressure, respectively.STD of MAE for SBP, DBP, and MAP The STD of the MAE for SBP, DBP, and MAP is defined as: 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} STD_s = \sqrt{\frac{1}{M}\sum _{m=1}^{M}\left( {\left( \left| {BP_{\text {diff}}}^{(m)}(m)\right| -avg({\left| BP_{\text {diff}}\right| })\right) }^2 \right) } \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${BP_{\text {diff}}}^{(m)}(m)$$\end{document} represents the difference between the ground truth and the predicted blood pressure values.TEI We propose a TEI metric to evaluate the overall reconstruction performance of the reconstructed waveform, encompassing both numerical and waveform similarities. This metric incorporates both the normalised root mean squared error (NRMSE) numerical prediction indicator and normalised dynamic time warping (NDTW) waveform prediction indicator. We normalised this index to facilitate comparisons among reconstructed signals with different lengths and amplitudes. The formula for this index is as follows: 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\left\{ \begin{array}{ll} NRMSE_{SBP}=\frac{\sqrt{\frac{1}{M}\sum _{m=1}^{M}\left( {SBP_{gt}}^{(m)}(m)-{SBP_{pre}}^{(m)}(m)\right) ^2}}{max({SBP_{gt}}^{(m)})-min({SBP_{gt}}^{(m)})} \\ NRMSE_{DBP}=\frac{\sqrt{\frac{1}{M}\sum _{m=1}^{M}\left( {DBP_{gt}}^{(m)}(m)-{DBP_{pre}}^{(m)}(m)\right) ^2}}{max({DBP_{gt}}^{(m)})-min({DBP_{gt}}^{(m)})} \\ NDTW=\frac{dtw}{N\times \left| max({ABP_{gt}}^{n})-min({ABP_{gt}}^{n})\right| } \\ TEI=\sqrt{\frac{{NRMSE_{SBP}}^2+{NRMSE_{DBP}}^2+\left( \frac{1}{M}NDTW\right) ^2}{3}} \\ \end{array}\right. } \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${ABP_{gt}}^{(n)}$$\end{document} represents a ground-truth ABP segment with N points and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SBP_{pre}}^{(m)} (m)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${DBP_{pre}}^{(m)}(m)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SBP_{gt}}^{(m)}(m)$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${DBP_{gt}}^{(m)}(m)$$\end{document} represent the m-th predicted value and the ground truth value of M segments, respectively. MAE of SBP, DBP, and MAP MAE of SBP, DBP, and MAP The MAE of SBP, DBP, and MAP are defined as: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} MAE_s=\frac{1}{M}\sum _{m=1}^{M}\left| {BP_{pre}}^{(m)}(m)-{BP_{gt}}^{(m)}(m) \right| \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${BP_{pre}}^{(m)}(m)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${BP_{gt}}^{(m)}(m)$$\end{document} represent the m-th predicted and ground truth values of the blood pressure, respectively. The MAE of SBP, DBP, and MAP are defined as: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} MAE_s=\frac{1}{M}\sum _{m=1}^{M}\left| {BP_{pre}}^{(m)}(m)-{BP_{gt}}^{(m)}(m) \right| \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${BP_{pre}}^{(m)}(m)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${BP_{gt}}^{(m)}(m)$$\end{document} represent the m-th predicted and ground truth values of the blood pressure, respectively. STD of MAE for SBP, DBP, and MAP STD of MAE for SBP, DBP, and MAP The STD of the MAE for SBP, DBP, and MAP is defined as: 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} STD_s = \sqrt{\frac{1}{M}\sum _{m=1}^{M}\left( {\left( \left| {BP_{\text {diff}}}^{(m)}(m)\right| -avg({\left| BP_{\text {diff}}\right| })\right) }^2 \right) } \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${BP_{\text {diff}}}^{(m)}(m)$$\end{document} represents the difference between the ground truth and the predicted blood pressure values. The STD of the MAE for SBP, DBP, and MAP is defined as: 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} STD_s = \sqrt{\frac{1}{M}\sum _{m=1}^{M}\left( {\left( \left| {BP_{\text {diff}}}^{(m)}(m)\right| -avg({\left| BP_{\text {diff}}\right| })\right) }^2 \right) } \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${BP_{\text {diff}}}^{(m)}(m)$$\end{document} represents the difference between the ground truth and the predicted blood pressure values. TEI We propose a TEI metric to evaluate the overall reconstruction performance of the reconstructed waveform, encompassing both numerical and waveform similarities. This metric incorporates both the normalised root mean squared error (NRMSE) numerical prediction indicator and normalised dynamic time warping (NDTW) waveform prediction indicator. We normalised this index to facilitate comparisons among reconstructed signals with different lengths and amplitudes. The formula for this index is as follows: 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\left\{ \begin{array}{ll} NRMSE_{SBP}=\frac{\sqrt{\frac{1}{M}\sum _{m=1}^{M}\left( {SBP_{gt}}^{(m)}(m)-{SBP_{pre}}^{(m)}(m)\right) ^2}}{max({SBP_{gt}}^{(m)})-min({SBP_{gt}}^{(m)})} \\ NRMSE_{DBP}=\frac{\sqrt{\frac{1}{M}\sum _{m=1}^{M}\left( {DBP_{gt}}^{(m)}(m)-{DBP_{pre}}^{(m)}(m)\right) ^2}}{max({DBP_{gt}}^{(m)})-min({DBP_{gt}}^{(m)})} \\ NDTW=\frac{dtw}{N\times \left| max({ABP_{gt}}^{n})-min({ABP_{gt}}^{n})\right| } \\ TEI=\sqrt{\frac{{NRMSE_{SBP}}^2+{NRMSE_{DBP}}^2+\left( \frac{1}{M}NDTW\right) ^2}{3}} \\ \end{array}\right. } \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${ABP_{gt}}^{(n)}$$\end{document} represents a ground-truth ABP segment with N points and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SBP_{pre}}^{(m)} (m)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${DBP_{pre}}^{(m)}(m)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SBP_{gt}}^{(m)}(m)$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${DBP_{gt}}^{(m)}(m)$$\end{document} represent the m-th predicted value and the ground truth value of M segments, respectively. TEI We propose a TEI metric to evaluate the overall reconstruction performance of the reconstructed waveform, encompassing both numerical and waveform similarities. This metric incorporates both the normalised root mean squared error (NRMSE) numerical prediction indicator and normalised dynamic time warping (NDTW) waveform prediction indicator. We normalised this index to facilitate comparisons among reconstructed signals with different lengths and amplitudes. The formula for this index is as follows: 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\left\{ \begin{array}{ll} NRMSE_{SBP}=\frac{\sqrt{\frac{1}{M}\sum _{m=1}^{M}\left( {SBP_{gt}}^{(m)}(m)-{SBP_{pre}}^{(m)}(m)\right) ^2}}{max({SBP_{gt}}^{(m)})-min({SBP_{gt}}^{(m)})} \\ NRMSE_{DBP}=\frac{\sqrt{\frac{1}{M}\sum _{m=1}^{M}\left( {DBP_{gt}}^{(m)}(m)-{DBP_{pre}}^{(m)}(m)\right) ^2}}{max({DBP_{gt}}^{(m)})-min({DBP_{gt}}^{(m)})} \\ NDTW=\frac{dtw}{N\times \left| max({ABP_{gt}}^{n})-min({ABP_{gt}}^{n})\right| } \\ TEI=\sqrt{\frac{{NRMSE_{SBP}}^2+{NRMSE_{DBP}}^2+\left( \frac{1}{M}NDTW\right) ^2}{3}} \\ \end{array}\right. } \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${ABP_{gt}}^{(n)}$$\end{document} represents a ground-truth ABP segment with N points and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SBP_{pre}}^{(m)} (m)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${DBP_{pre}}^{(m)}(m)$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${SBP_{gt}}^{(m)}(m)$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${DBP_{gt}}^{(m)}(m)$$\end{document} represent the m-th predicted value and the ground truth value of M segments, respectively. Results We performed five experiments to assess the performance of our enhanced model, the efficacy of the loss function, and the applicability of the model to other dataset.Table 2Comparison of experimental results.ExperimentDatasetModelLoss function\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text {MAE}}_{\text {SBP}} \, (\text {mmHg})$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text {MAE}}_{\text {DBP}} \, (\text {mmHg})$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text {MAE}}_{\text {MAP}} \, (\text {mmHg})$$\end{document}TEI (mean ± STD)Experiment IUCI\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-NetMSE3.07 ± 2.282.10 ± 2.131.38 ± 1.330.51 ± 0.57Experiment IIUCI\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-NetOur loss2.91 ± 2.501.99 ± 1.951.47 ± 1.410.45 ± 0.38Experiment IIIUCIOur modelMSE2.52 ± 2.041.43 ± 1.451.44 ± 1.420.29 ± 0.12Experiment IVUCIOur modelOur loss2.48 ± 1.951.42 ± 1.421.48 ± 1.360.29 ± 0.10Experiment VVitalDBOur modelOur loss2.16 ± 1.531.12 ± 0.591.35 ± 0.840.29 ± 0.15 Table 2Comparison of experimental results.ExperimentDatasetModelLoss function\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text {MAE}}_{\text {SBP}} \, (\text {mmHg})$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text {MAE}}_{\text {DBP}} \, (\text {mmHg})$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text {MAE}}_{\text {MAP}} \, (\text {mmHg})$$\end{document}TEI (mean ± STD)Experiment IUCI\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-NetMSE3.07 ± 2.282.10 ± 2.131.38 ± 1.330.51 ± 0.57Experiment IIUCI\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-NetOur loss2.91 ± 2.501.99 ± 1.951.47 ± 1.410.45 ± 0.38Experiment IIIUCIOur modelMSE2.52 ± 2.041.43 ± 1.451.44 ± 1.420.29 ± 0.12Experiment IVUCIOur modelOur loss2.48 ± 1.951.42 ± 1.421.48 ± 1.360.29 ± 0.10Experiment VVitalDBOur modelOur loss2.16 ± 1.531.12 ± 0.591.35 ± 0.840.29 ± 0.15 Comparison of experimental results. Table 2 shows that the MAE ± STD values for SBP and DBP in Experiment I are 3.07 ± 2.28 mmHg and 2.10 ± 2.13 mmHg, respectively. In Experiment II, these values improve to 2.91 ± 2.50 mmHg for SBP and 1.99 ± 1.95 mmHg for DBP, respectively. Additionally, the progression observed from Experiment III to Experiment IV, in terms of the growth in SBP, DBP, and TEI indices, indicates that the application of the combined loss function in each layer of our deep supervision has a positive impact. From the above, we conclude that this approach, whether applied in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net or in our model, underscores the effectiveness of this methodology in refining the accuracy of ABP estimation. However, the MAE ± STD value for MAP in Experiment I is 1.38 ± 1.33 mmHg, while in Experiment II, it is slightly increased to 1.47 ± 1.41 mmHg. In our model, from Experiment III to Experiment IV, a similar increase in MAE is observed, though the magnitude of this increase is smaller. This is because the MAX loss function introduced in the combined loss function focuses more on minimizing the maximum error, which sacrifices the accuracy of the waveform’s mean value to some extent. However, according to the TEI index, the combined loss function achieves a better balance between the two evaluation metrics of numerical and waveform morphology predictions. In Experiments I and III, under the MSE loss function, the proposed model achieved better prediction results than \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net. Similarly, in Experiments II and IV, using our optimised loss function, our model outperformed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net in terms of important numerical predictions. Additionally, the TEI index’s mean ± STD in Experiment IV has been significantly improved to 0.29 ± 0.10. These results indicate that the proposed model exhibits superior accuracy for ABP waveform prediction tasks. In our extended dataset, the VitalDB dataset, which has a different sampling frequency and PPG gain compared to the UCI dataset, we observed that the MAE ± STD for SBP, DBP, and MAP were 2.16 ± 1.53 mmHg, 1.12 ± 0.59 mmHg, and 1.35 ± 0.84 mmHg, respectively. The mean ± STD for the TEI index reached 0.29 ± 0.15. These results not only demonstrate the outstanding predictive accuracy of our model but also confirm its robustness when applied to data collected with different sampling frequencies and from various devices.Fig. 4Comparison of reconstructed and ground truth waveforms for different TEI indices. Original ABP signal in red and predicted ABP signal in blue.Fig. 5Error distributions of SBP, MAP, and DBP for two datasets. (a), (c), (e) The error distributions of SBP, MAP, and DBP in the UCI dataset; (b), (d), (f) The error distributions of SBP, MAP, and DBP in the VitalDB dataset.Table 3Performance Comparison of SBP, DBP, and MAP Using MAE with Existing Studies.MethodDatasetInputSBP (mmHg)DBP (mmHg)MAP (mmHg)Lin et al.23MIMIC II/MIMIC IIIPPG, VPG and APG4.592.47NRWang et al.24MIMICPPG3.952.14NRIbtehaz et al.13UCIPPG5.733.452.31Cheng et al.14UCIPPG, VPG and APG3.271.901.49Tang et al.15UCIPPG2.601.45NRAthaya et al.17MIMIC/MIMIC IIIPPG3.681.972.17Zhang et al.25MIMIC I/VitalDBPPG and ECG7.734.30NROursUCIPPG, VPG and APG2.481.421.48OursVitalDBPPG, VPG and APG2.161.121.35 Comparison of reconstructed and ground truth waveforms for different TEI indices. Original ABP signal in red and predicted ABP signal in blue. Error distributions of SBP, MAP, and DBP for two datasets. (a), (c), (e) The error distributions of SBP, MAP, and DBP in the UCI dataset; (b), (d), (f) The error distributions of SBP, MAP, and DBP in the VitalDB dataset. Table 3Performance Comparison of SBP, DBP, and MAP Using MAE with Existing Studies.MethodDatasetInputSBP (mmHg)DBP (mmHg)MAP (mmHg)Lin et al.23MIMIC II/MIMIC IIIPPG, VPG and APG4.592.47NRWang et al.24MIMICPPG3.952.14NRIbtehaz et al.13UCIPPG5.733.452.31Cheng et al.14UCIPPG, VPG and APG3.271.901.49Tang et al.15UCIPPG2.601.45NRAthaya et al.17MIMIC/MIMIC IIIPPG3.681.972.17Zhang et al.25MIMIC I/VitalDBPPG and ECG7.734.30NROursUCIPPG, VPG and APG2.481.421.48OursVitalDBPPG, VPG and APG2.161.121.35 Performance Comparison of SBP, DBP, and MAP Using MAE with Existing Studies. Fig. 6Bland-Altman plots of SBP, MAP, and DBP for two datasets. (a), (c), (e) Bland-Altman plots of SBP, MAP, and DBP in the UCI dataset; (b), (d), (f) Bland-Altman plots of SBP, MAP, and DBP in the VitalDB dataset. Bland-Altman plots of SBP, MAP, and DBP for two datasets. (a), (c), (e) Bland-Altman plots of SBP, MAP, and DBP in the UCI dataset; (b), (d), (f) Bland-Altman plots of SBP, MAP, and DBP in the VitalDB dataset. Figure 4 shows the reconstructed and original ABP waveforms for different TEI indices. For the TEI in (a) and (b), the results show high similarity in waveform morphology and comprehensive prediction of SBP and DBP. The abnormal blood pressure waveform was also predicted accurately. As the TEI indices increases in (c), (d), (e), and (f), a deviation in the overall predictive performance for SBP, DBP, and waveform morphology was observed. The error distributions of SBP, DBP, and MAP across all samples are shown in Fig.5. For both datasets, the error distribution resembles a Gaussian distribution. It is observable that the vast majority of error values for SBP, DBP, and MAP in both datasets are concentrated within the −5 to 5 range. Figure 6 displays the Bland-Altman analysis of the SBP, DBP, and MAP data for the two datasets. Specifically, (a) shows that the mean bias of SBP in the UCI dataset is −0.68, with a 95% limits of agreement ranges from −7.94 to 6.58 mmHg. (e) indicates that the mean bias of MAP in the UCI dataset is −0.28, with a 95% limits of agreement ranges from −5.21 to 4.66 mmHg. (c) reveals that the mean bias of DBP in the UCI dataset is 0.17, with a 95% limits of agreement ranges from −4.51 to 4.85 mmHg. Furthermore, (b) shows that the mean bias of SBP in the VitalDB dataset is −0.20, with a 95% limits of agreement ranges from −6.39 to 5.98 mmHg. (f) indicates that the mean bias of MAP in the VitalDB dataset is 0.03, with a 95% limits of agreement ranges from −3.82 to 3.87 mmHg. Finally, (f) reveals that the mean bias of DBP in the VitalDB dataset is 0.19, with a 95% limits of agreement ranges from −2.91 to 3.29 mmHg. Comparison with existing studies To validate the effectiveness of our model, we extended our dataset to include a VitalDB dataset and compared it with existing related studies. We not only compared individual experiments related to waveform reconstruction but also compared individual experiments related to the numerical prediction of SBP, DBP, and MAP from PPG. The comparison results are shown in Table 3, including the dataset and the MAE ± STD of SBP, DBP, and MAP. From the table, we can see that in the waveform reconstruction work, Ibtehaz et al., who only used PPG as input, achieved SBP, DBP, and MAP values of 5.73 ± 9.16, 3.45 ± 6.15, and 2.31 ± 4.44 mmHg respectively. Cheng et al., using PPG, VPG, and APG as inputs and employing Wave-U-Net, achieved SBP, DBP, and MAP values of 3.27 ± 3.92, 1.90 ± 2.44, and 1.49 ± 2.10 mmHg respectively. The results of Tang et al., who used W-Net, were the closest to ours. Their results only measured SBP and DBP, which reached 2.60 ± 1.89 and 1.45 ± 1.33 mmHg. In the numerical prediction experiments we compared, Lin et al. used features from PPG, VPG, and APG as input, with the prediction results only measuring SBP and DBP, which reached 4.59 ± 1.78 and 2.47 ± 1.09 mmHg respectively. Wang et al. only used PPG as input, with their results measuring SBP and DBP at 3.95 ± 4.38 and 2.14 ± 2.40 mmHg. Athaya et al. used a U-Net based model with PPG as input, achieving results of 3.68 ± 4.42 mmHg for SBP, 1.97 ± 2.92 mmHg for DBP, and 2.17 ± 3.06 mmHg for MAP. Zhang et al. proposed a BP estimation method based on multiple physiological parameters (including HRV, PTT, and pulse wave morphology features) extracted from ECG and PPG signals, and after individual calibration, achieved MAE 7.73 mmHg for SBP and 4.30 mmHg for DBP. In our predictions on the UCI dataset, we achieved MAE ± STD values of 2.48 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm$$\end{document} 1.95, 1.42 ± 1.42, and 1.48 ± 1.36 mmHg for SBP, DBP, and MAP, respectively, demonstrating enhanced accuracy in ABP waveform numerical prediction. Despite the differing source of the VitalDB dataset, we achieved SBP, DBP, and MAP prediction values of 2.16 ± 1.53, 1.12 ± 0.59, and 1.35 ± 0.84 mmHg, respectively, which also yielded comparably favorable results. These comparison results show that our model achieved a higher precision in the numerical prediction of SBP, DBP, and MAP. Furthermore, it demonstrated robustness and performed well across different datasets. Comparison with AAMI & BHS standards To ensure the accuracy and reliability of the blood pressure estimation method, we applied the standards set by the British Hypertension Society (BHS)26and Association for the Advancement of Medical Instrumentation (AAMI)27 for evaluation. The BHS standard was used to evaluate the accuracy of blood pressure measurements. This standard categorises accuracy into four grades, A, B, C, and D, based on the deviation of the measurement results from the gold standard. Grade A required at least 60% of the measurement errors to be within 5 mmHg, 85% within 10 mmHg, and 95% within 15 mmHg. Grade B required at least 50% of the measurement errors to be within 5 mmHg, 75% within 10 mmHg, and 90% within 15 mmHg. Grade C required at least 40% of the measurement errors to be within 5 mmHg, 65% within 10 mmHg, and 85% within 15 mmHg. The AAMI standard is also a benchmark of blood pressure measurement accuracy. This standard stipulates that the mean measurement error must be within 5 mmHg, with a standard deviation not exceeding 8 mmHg. These criteria ensure reliable and precise data for cuff-less blood pressure estimation studies. Table 4 and Table 5 present the evaluation results in accordance with the AAMI and BHS standards. The data demonstrate that our method conforms to the stipulated error metrics and standard deviation limits of AAMI, and it achieves a Grade A classification under the criteria set by the BHS.Table 4BHS Standard Compliance Assessment.Percentage±5 mmHg±10 mmHg±15 mmHgBHS standardGrade A60%85%95%Grade B50%75%90%Grade C40%65%85%OursSBP87.2%96.9%99.6%DBP95.9%98.8%99.9%MAP95.2%99.1%99.8%Table 5AAMI Standard Compliance Assessment.ME(mmHg)STD(mmHg)AAMI standard\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\le$$\end{document} 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\le$$\end{document} 8OursSBP2.453.63DBP1.392.29MAP1.482.47 Table 4BHS Standard Compliance Assessment.Percentage±5 mmHg±10 mmHg±15 mmHgBHS standardGrade A60%85%95%Grade B50%75%90%Grade C40%65%85%OursSBP87.2%96.9%99.6%DBP95.9%98.8%99.9%MAP95.2%99.1%99.8% BHS Standard Compliance Assessment. Table 5AAMI Standard Compliance Assessment.ME(mmHg)STD(mmHg)AAMI standard\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\le$$\end{document} 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\le$$\end{document} 8OursSBP2.453.63DBP1.392.29MAP1.482.47 AAMI Standard Compliance Assessment. Discussion In recent years, studies on blood pressure measurements based on PPG signals have mainly included numerical measurement methods for SBP, DBP, and MAP, and methods for reconstructing the ABP waveform through PPG. Current numerical measurement methods for SBP, DBP, and MAP primarily involve blood pressure estimation methods based on pulse wave velocity (PWV), pulse arriving time (PAT), and pulse transit time (PTT)28–30. Some methods predict blood pressure by acquiring features from the PPG, its derivatives, and the ECG31,32, whereas others apply deep learning techniques to predict blood pressure values directly from raw PPG and its derivatives33,34. Existing methods for obtaining the ABP waveform from PPG are primarily based on deep learning. This approach extracts important parameters, such as SBP, DBP, and MAP, from the ABP waveform derived from PPG and provides more physiological information. However, a comprehensive consideration of the numerical values and waveform morphology is lacking in the current waveform reconstruction methods, and related deep learning methods have not yet been extended to other dataset to test the robustness of their models under different sampling frequencies and devices. The main contribution of this study is the design of a feature extractor based on the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net. A Bi-block was designed to capture the temporal information that emerged after extracting features at different levels from the bidirectional information flow in individual experiments. A Bi-block is used to process the temporal information of the fused features in the deep supervision outputs to avoid information loss and blurring during the fusion process. In addition, this study optimised the loss function for each layer in deep supervision based on existing research and employed a combined loss function. This design enhances feature extraction capabilities, improves model generalisation, and increases the performance and accuracy of the model on unprocessed data. Our results for the two models confirm this, as the models demonstrate good robustness on the two datasets and achieve high accuracy and improved metrics for important parameters such as SBP, DBP, MAP, and the composite index TEI. One limitation of this study is that our model was tailored for individual modelling. The primary reason for this is that PPG and ABP signals vary with the test subject’s age and other variable conditions, and individual differences are expressed differently in the PPG and ABP signals, often leading to larger prediction errors. Despite established healthcare standards prohibiting calibration of testing devices during evaluation16,35, in practical applications, individual blood pressure waveforms are continuously measured and recorded, offering the potential to leverage personal historical data for calibrating generalized models. Our future studies will focus on applying this model to individual devices and extending it to group studies. Conclusion In this study, we introduced an innovative end-to-end network designed to predict ABP waveforms from PPG signals. Building on existing research, our approach incorporates a combined loss function at each level of our deeply supervised architecture. A comprehensive TEI was proposed to assess the accuracy of the vital numerical predictions and waveform shape fidelity. To verify the robustness of our model, we expanded our experiment to include a VitalDB dataset from UCI dataset. An important aspect to note is the inherent complexity of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {U}^{2}$$\end{document}-Net used for feature extraction in our model, which comprises 14,698,765 parameters. In the complete architecture of our model, this number increases to 17,629,901, reflecting its enhanced complexity and capabilities. This augmentation inevitably extends the training time of the model. Although this increase in parameters and the associated training duration presents a challenge, there is a trade-off for the enhanced ability of our model to capture and process complex temporal relationships in the data, leading to improved accuracy in ABP waveform prediction. This enhancement was evident in the high accuracy of the key numerical predictions across both the UCI dataset and VitalDB dataset, along with a commendable overall performance in waveform prediction. We plan to deploy our model on medical devices to provide more accurate numerical and waveform information. Future studies will involve optimising the model to balance the increased complexity with efficient training, thereby ensuring practical applicability in real-world settings.