Noninvasive blood pressure waveform reconstruction method based on 3D blood vessel modeling and depth sequential network

Abstract

The invention provides a non-invasive blood pressure waveform reconstruction method based on 3D blood vessel modeling and a depth sequential network, and belongs to the field of non-invasive physiological signal processing and cardiovascular function evaluation. The method comprises the following steps: constructing a physiological simulation model containing a vascular network from a radial artery to a fingertip by adopting VTK; the method comprises the following steps: integrating a Moens-Korteweg wave equation, a vascular compliance model and a Beer-Lambert optical law, and establishing a signal conversion model from ABP to PPG; designing a deep sequential network architecture comprising a U2-Net nested U-shaped structure and a bidirectional time processor (BTP); aBP waveform reconstruction is realized through a depth supervision strategy and a combined loss function; and finally, closed-loop application is realized in a hardware system integrated with the MAX30102 sensor and the STM32F103 processor. Experiments show that according to the method, the prediction precision of SBP 2.48 + / -1.95 mmHg and DBP 1.42 + / -1.42 mmHg is achieved in an MIMICII data set, and the AAMI standard is achieved. Through combination of theoretical modeling and an intelligent algorithm and implementation of a hardware system, a novel noninvasive monitoring scheme is provided for family health monitoring.

Description

Technical Field

[0001] This invention relates to a non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks, belonging to the field of non-invasive physiological signal processing and cardiovascular function assessment. Background Technology

[0002] Cardiovascular disease has become the leading threat to global health, with hypertension being the most important controllable risk factor. Arterial blood pressure waveform monitoring offers significant advantages over traditional cuff measurements, providing 17 hemodynamic parameters such as peripheral resistance and arterial compliance, which are crucial for early disease diagnosis. However, current technologies face significant challenges: invasive monitoring carries a high risk of infection, limiting its application to intensive care; non-invasive methods for converting photoplethysmography (PPG) to arterial blood pressure (ABP) suffer from unclear physiological mechanisms and insufficient algorithmic accuracy. Furthermore, traditional methods currently suffer from three major bottlenecks: a lack of comprehensive theoretical modeling of vascular network transmission characteristics; insufficient algorithmic accuracy (systolic blood pressure error approximately 5.3 ± 3.8 mmHg); and neglect of individual vascular characteristics. In deep learning applications, architectures like U-Net often perform poorly in waveform detail reconstruction and handling individual differences, while multimodal inputs further increase system complexity.

[0003] To address the aforementioned issues, this patent proposes an innovative solution. By constructing a "coordinated positive and negative problem" research paradigm, it establishes for the first time a complete signal conversion model from ABP to PPG, integrating three major physical processes: pressure wave propagation, vascular mechanical response, and optical detection. In terms of algorithm implementation, a hybrid architecture combining U²-Net and a bidirectional time processor (BTP) was developed to achieve multi-scale feature fusion and personalized temporal feature capture. Validated using the MIMIC-II and VitalDB datasets, this solution reduces systolic blood pressure prediction error to 2.48±1.95 mmHg and diastolic blood pressure error to 1.42±1.42 mmHg, meeting the AAMI standard and BHS protocol Level A requirements. In terms of hardware implementation, an embedded system based on the MAX30102 sensor and STM32F103 processor provides a reliable non-invasive continuous monitoring solution for clinical use. Summary of the Invention

[0004] Objective: This invention addresses key issues in existing non-invasive blood pressure monitoring technologies, such as unclear physiological mechanisms, insufficient waveform reconstruction accuracy, and poor individual adaptability. It proposes a non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks. This method constructs a multi-physics coupling model integrating hemodynamics, vascular mechanics, and optical properties to reveal the conversion mechanism between ABP and PPG signals from a first-principles perspective. It designs an innovative deep temporal network architecture, combining the multi-scale feature extraction capabilities of U²-Net and the personalized temporal modeling advantages of a bidirectional time processor (BTP). Ultimately, it achieves high-precision end-to-end reconstruction from PPG signals to ABP waveforms, providing a safe and convenient continuous blood pressure monitoring solution for clinical monitoring and home health management.

[0005] Technical Solution: To address the above problems, this invention proposes a non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks. This method includes the following steps:

[0006] 1. A non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks, characterized in that the method includes the following steps:

[0007] Step 1: Construct a physiological simulation model from the radial artery to the fingertip using VTK;

[0008] Step 2: Integrate the Moens-Korteweg wave equation, vascular compliance model and Beer-Lambert optical law to establish a signal conversion model from arterial blood pressure (ABP) to photoplethysmography (PPG);

[0009] Step 3, design including U 2 -A deep temporal network architecture with a nested U-shaped structure and a bidirectional time processor (BTP);

[0010] Step 4: Achieve high-precision arterial blood pressure waveform reconstruction through a deep supervision strategy and a combined loss function;

[0011] Step 5: Validate the reconstructed arterial blood pressure waveform using multi-source data;

[0012] Step 6: Perform PPG acquisition and ABP real-time acquisition and reconstruction in the hardware system integrating the MAX30102 sensor and STM32F103 processor.

[0013] 2. The non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks according to claim 1, characterized in that, in step 1, a Visualization Toolkit is used for vascular geometric modeling, realizing a complete arterial tree structure including the radial artery trunk, palmar arch branches, and distal digit vessels. Each vascular path generates a smooth centerline through spline interpolation, and a gradually changing diameter is set according to Murray's law to ensure compliance with the physiological laws of vascular branching. The specific formula is as follows:

[0014]

[0015] in For the first Segmental vessel diameter, The initial vessel diameter, The cumulative path length is calculated. The constructed model covers a multi-level vascular system from the radial artery of the wrist (2.5 mm in diameter) to the arterioles of the fingertips (0.4-0.5 mm in diameter), including a total of 11 major vascular segments. At the same time, the physiological parameters of the vascular network are configured.

[0016] 3. The non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks according to claim 1, characterized in that the specific method of step 2 is as follows:

[0017] Step 2.1: Construct a hemodynamic propagation model based on the Moens-Korteweg equation, expressed as:

[0018]

[0019] in This refers to the elastic modulus of the blood vessel wall. For wall thickness, Blood density, Let be the radius of the blood vessel. The pulse wave velocity for different vessel diameters can be calculated from the above formula. At the same time, the influence of factors such as vascular elasticity and branch reflection on the waveform must also be considered, and the conservation law must be applied at each branch node of the blood vessel to ensure the continuity of pressure and flow.

[0020] Step 2.2, establish a vascular volume conversion model, which includes the nonlinear relationship between vascular pressure and volume:

[0021]

[0022] in Baseline blood volume, For reference pressure, diastolic pressure is usually used. The compliance coefficient. The value is a nonlinear exponent. Here, considering the characteristics of peripheral blood vessels, the model introduces frequency-dependent attenuation and peripheral amplification effects, where the attenuation coefficient for high-frequency components above 25Hz is -3dB.

[0023] Step 2.3: Design the optical signal conversion system and apply the modified Beer-Lambert law:

[0024]

[0025] in The tissue absorption coefficient, The blood absorption coefficient, As the DC blood volume integral, the peripheral amplification factor is set to 1.15 here to convert the fingertip pressure waveform into a PPG signal, realizing a complete conversion from mechanical signal to optical signal.

[0026] 4. The non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks according to claim 1, characterized in that the specific implementation steps of step 3 are as follows:

[0027] Step 3.1, construct U 2 The -Net multi-scale feature extraction architecture employs a 5-level nested U-shaped structure design, with each level containing multiple levels of Residual U-block (RSU) modules. It captures the spatial topological information of the vascular network contained in the PPG signal, corresponding to the vascular branching structure from the radial artery to the fingertip in the forward problem. The downsampling process uses max pooling operation, and the upsampling operation using linear interpolation achieves resolution restoration. This architecture captures multi-scale features from local details to global trends in the PPG signal through the feature reuse mechanism between levels.

[0028] Step 3.2: Design and integrate a bidirectional time processor module. Embed BTP units consisting of bidirectional LSTM layers and 1×1 convolutional layers at the jump connections of each U-shaped structure. The dynamic update mechanism satisfied by its temporal feature processing is as follows:

[0029]

[0030] in for The hidden state at all times for The mechanism can simultaneously model the physiological temporal dependence of forward and backward directions, effectively capturing the differences in pulse wave conduction caused by individual vascular characteristics.

[0031] Step 3.3: Construct a 5-level nested U-shaped convolutional network. An 8-second PPG signal segment with a sampling rate of 125Hz is used as the standard input window. This duration setting ensures that the complete cardiac cycle features are included. The number of channels in the network convolutional kernels is configured as 64, 128, 256, 512 and 1024 channels respectively. The first-level RSU module uses a 5-layer depthwise separable convolutional structure to process the original PPG signal, and multi-scale feature fusion is achieved between each layer through skip connections.

[0032] 5. The non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks according to claim 1, characterized in that the specific implementation steps of step 4 are as follows:

[0033] Step 4.1: Configure training optimization parameters. The batch size is fixed at 32 to balance memory usage and gradient stability. The initial learning rate is set to 0.001 and dynamically adjusted using a cosine decay strategy. The mathematical expression for the learning rate update is as follows:

[0034]

[0035] The maximum learning rate =0.001, minimum learning rate =0.0001, T is the total training period set to 200 rounds, thus achieving stable convergence of model parameters;

[0036] Step 4.2, construct a hierarchical supervision network system in U 2 - Each decoder output of the Net architecture is equipped with an independent supervision node, forming a full-stack supervision network from high-resolution low-level features to low-resolution high-level features. The low-level supervision nodes focus on the extraction of local morphological features of the pulse wave, while the high-level supervision nodes strengthen the modeling of the vascular network topology. Gradient propagation is achieved through 5-level skip connections.

[0037] Step 4.3: Design a multi-objective combined optimization system, establish a combined loss function consisting of waveform fidelity, extreme point accuracy, and physiological correlation; and implement an adaptive weight scheduling mechanism, setting a hierarchical weight reduction strategy from 1.0 at the bottom layer to 0.2 at the top layer to ensure that each layer of the network can learn effective feature representations.

[0038] 6. The non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal network according to claim 1, characterized in that the specific implementation steps of step 5 are as follows: performance evaluation is performed on the MIMIC II, Vital DB dataset.

[0039] Step 5.1: Select sampling data of different frequencies for verification. Extract records from the MIMIC II dataset that last for more than 8 minutes and have an ABP value of less than 200 mmHg, and select the first 150 records as the research subjects. In the Vital DB dataset, obtain cases containing PPG and ABP signals from the Web monitoring dataset, and download the data of the first 10 cases for experiments.

[0040] Step 5.2 uses two metrics to assess the accuracy of key values: MAE and STD. The specific formulas are as follows:

[0041]

[0042] in and These represent the first and second blood pressure values, respectively. One predicted value and one actual value. STD is defined as:

[0043]

[0044] in This indicates the difference between the actual blood pressure value and the predicted blood pressure value.

[0045] 7. A non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks according to claim 1, characterized in that, in step 6, a blood pressure waveform monitoring hardware system is built, which integrates a minimum system circuit, a reflective photoelectric biosensor circuit, and a temperature and humidity sensor circuit through an STM32F103RET6 microcontroller to form a core control module; a MAX30102 PPG sensor is used to collect blood pressure waveform signals, and an SHT35 high-precision temperature and humidity sensor is used to monitor environmental parameters; an LED display unit, a communication interface, and a power management circuit are connected synchronously to form a finger clip wearable structure to realize real-time detection and data transmission of blood pressure waveforms.

[0046] The beneficial effects of this invention after the above technical processing are mainly manifested in:

[0047] 1. By constructing a complete 3D vascular network model from the radial artery to the fingertip, this study innovatively integrates three major physical principles: the Moens-Korteweg wave equation, the vascular nonlinear compliance model, and the modified Beer-Lambert optical law. For the first time, a complete conversion mechanism from ABP to PPG signals is established at the first-principles level. This multi-physics coupled modeling method not only overcomes the limitations of traditional data-driven technologies but also provides a solid theoretical foundation for non-invasive blood pressure monitoring.

[0048] 2. Employing a nested U-shaped structure of U²-Net combined with a bidirectional time processor (BTP), this approach achieves effective fusion of multi-scale features and accurately captures the temporal characteristics of individual blood vessels. The innovatively designed combined loss function achieves an optimal balance between waveform fidelity and key parameter accuracy through scientific proportioning. Furthermore, this algorithm architecture significantly improves reconstruction accuracy while maintaining a lightweight model.

[0049] 3. The hardware system employs an optimized combination of a medical-grade MAX30102 sensor and an STM32F103 processor, ensuring both high-precision signal acquisition and real-time processing capabilities. The specially designed finger clip structure achieves extremely low power consumption while ensuring wearing comfort, enabling long-term continuous monitoring. The system's stable performance on multiple clinical datasets fully validates its clinical applicability. Attached Figure Description

[0050] Figure 1 This is a technical roadmap of the non-invasive blood pressure waveform reconstruction method of the present invention.

[0051] Figure 2 This is a diagram showing the pressure propagation and optical signal conversion results of the present invention.

[0052] Figure 3 This is a block diagram of the photoplethysmography (PPG) acquisition system of the present invention. Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the invention.

[0054] To more clearly illustrate the objectives, technical solutions, and advantages of this invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] like Figure 1 As shown in the figure, an embodiment of the present invention provides a non-invasive blood pressure waveform reconstruction method based on 3D vascular modeling and deep temporal networks, which includes the following steps:

[0056] Step 1: Construct a physiological simulation model from the radial artery to the fingertip using VTK;

[0057] Step 2: Integrate the Moens-Korteweg wave equation, vascular compliance model and Beer-Lambert optical law to establish a signal conversion model from arterial blood pressure (ABP) to photoplethysmography (PPG);

[0058] Step 3, design including U 2 -A deep temporal network architecture with a nested U-shaped structure and a bidirectional time processor (BTP);

[0059] Step 4: Achieve high-precision arterial blood pressure waveform reconstruction through a deep supervision strategy and a combined loss function;

[0060] Step 5: Validate the reconstructed arterial blood pressure waveform using multi-source data;

[0061] Step 6: Perform PPG acquisition and ABP real-time acquisition and reconstruction in the hardware system integrating the MAX30102 sensor and STM32F103 processor.

[0062] Furthermore, in step 1, the Visualization Toolkit was used for vascular geometric modeling, realizing a complete arterial tree structure including the radial artery trunk, palmar arch branches, and distal digits. Each vascular path was used to generate a smooth centerline through spline interpolation, and a gradually changing diameter was set according to Murray's law to ensure compliance with the physiological laws of vascular branching. The specific formula is as follows:

[0063]

[0064] in For the first Segmental vessel diameter, The initial vessel diameter, The cumulative path length is calculated. The constructed model covers a multi-level vascular system from the radial artery of the wrist (2.5 mm in diameter) to the arterioles of the fingertips (0.4-0.5 mm in diameter), including a total of 11 major vascular segments. At the same time, the physiological parameters of the vascular network are configured.

[0065] Furthermore, the specific implementation steps of step 2 are as follows:

[0066] Step 2.1: Construct a hemodynamic propagation model based on the Moens-Korteweg equation, expressed as:

[0067]

[0068] in This refers to the elastic modulus of the blood vessel wall. For wall thickness, Blood density, Let be the radius of the blood vessel. The pulse wave velocity for different vessel diameters can be calculated from the above formula. At the same time, the influence of factors such as vascular elasticity and branch reflection on the waveform must also be considered, and the conservation law must be applied at each branch node of the blood vessel to ensure the continuity of pressure and flow.

[0069] Step 2.2, establish a vascular volume conversion model, which includes the nonlinear relationship between vascular pressure and volume:

[0070]

[0071] in Baseline blood volume, For reference pressure, diastolic pressure is usually used. The compliance coefficient. The value is a nonlinear exponent. Here, considering the characteristics of peripheral blood vessels, the model introduces frequency-dependent attenuation and peripheral amplification effects, where the attenuation coefficient for high-frequency components above 25Hz is -3dB.

[0072] Step 2.3: Design the optical signal conversion system and apply the modified Beer-Lambert law:

[0073]

[0074] in The tissue absorption coefficient, The blood absorption coefficient, For direct current blood volume integration, the peripheral amplification factor is set to 1.15 here. Figure 2 As shown, the simulation ultimately converts the fingertip pressure waveform into a PPG signal using a nonlinear vascular compliance model and a modified Beer-Lambert law, achieving a complete conversion from mechanical to optical signals.

[0075] Furthermore, the specific implementation steps of step 3 are as follows:

[0076] Step 3.1, construct U 2 The -Net multi-scale feature extraction architecture employs a 5-level nested U-shaped structure design, with each level containing multiple levels of Residual U-block (RSU) modules. It captures the spatial topological information of the vascular network contained in the PPG signal, corresponding to the vascular branching structure from the radial artery to the fingertip in the forward problem. The downsampling process uses max pooling operation, and the upsampling operation using linear interpolation achieves resolution restoration. This architecture captures multi-scale features from local details to global trends in the PPG signal through the feature reuse mechanism between levels.

[0077] Step 3.2: Design and integrate a bidirectional time processor module. Embed BTP units consisting of bidirectional LSTM layers and 1×1 convolutional layers at the jump connections of each U-shaped structure. The dynamic update mechanism satisfied by its temporal feature processing is as follows:

[0078]

[0079] in for The hidden state at all times for The mechanism can simultaneously model the physiological temporal dependence of forward and backward directions, effectively capturing the differences in pulse wave conduction caused by individual vascular characteristics.

[0080] Step 3.3: Construct a 5-level nested U-shaped convolutional network. An 8-second PPG signal segment with a sampling rate of 125Hz is used as the standard input window. This duration setting ensures that the complete cardiac cycle features are included. The number of channels in the network convolutional kernels is configured as 64, 128, 256, 512 and 1024 channels respectively. The first-level RSU module uses a 5-layer depthwise separable convolutional structure to process the original PPG signal, and multi-scale feature fusion is achieved between each layer through skip connections.

[0081] Furthermore, the specific implementation steps of step 4 are as follows:

[0082] Step 4.1: Configure training optimization parameters. The batch size is fixed at 32 to balance memory usage and gradient stability. The initial learning rate is set to 0.001 and dynamically adjusted using a cosine decay strategy. The mathematical expression for the learning rate update is as follows:

[0083]

[0084] The maximum learning rate =0.001, minimum learning rate =0.0001, T is the total training period set to 200 rounds, thus achieving stable convergence of model parameters;

[0085] Step 4.2, construct a hierarchical supervision network system in U 2 - Each decoder output of the Net architecture is equipped with an independent supervision node, forming a full-stack supervision network from high-resolution low-level features to low-resolution high-level features. The low-level supervision nodes focus on the extraction of local morphological features of the pulse wave, while the high-level supervision nodes strengthen the modeling of the vascular network topology. Gradient propagation is achieved through 5-level skip connections.

[0086] Step 4.3: Design a multi-objective combined optimization system, establish a combined loss function consisting of waveform fidelity, extreme point accuracy, and physiological correlation; and implement an adaptive weight scheduling mechanism, setting a hierarchical weight reduction strategy from 1.0 at the bottom layer to 0.2 at the top layer to ensure that each layer of the network can learn effective feature representations.

[0087] Furthermore, the specific implementation steps of step 5 are as follows:

[0088] Step 5.1: Select sampling data of different frequencies for verification. Extract records from the MIMIC II dataset that last for more than 8 minutes and have an ABP value of less than 200 mmHg, and select the first 150 records as the research subjects. In the Vital DB dataset, obtain cases containing PPG and ABP signals from the Web monitoring dataset, and download the data of the first 10 cases for experiments.

[0089] Step 5.2 uses two metrics to assess the accuracy of key values: MAE and STD. The specific formulas are as follows:

[0090]

[0091] in and These represent the first and second blood pressure values, respectively. One predicted value and one actual value. STD is defined as:

[0092]

[0093] in This indicates the difference between the actual blood pressure value and the predicted blood pressure value.

[0094] Step 5.3: Construct an evaluation index to assess the overall reconstruction performance of the reconstructed waveform, including numerical and waveform similarity. This index combines a normalized root mean square error numerical prediction index and a normalized dynamic time warping waveform prediction index. The index is normalized to facilitate comparison of reconstructed signals of different lengths and amplitudes.

[0095] Furthermore, the specific method for step 6 is as follows:

[0096] A blood pressure waveform monitoring hardware system was built, integrating a minimum system circuit, a reflective photoelectric biosensor circuit, and a temperature and humidity sensor circuit using an STM32F103RET6 microcontroller to form the core control module. A MAX30102 PPG sensor was used to acquire blood pressure waveform signals, in conjunction with a SHT35 high-precision temperature and humidity sensor to monitor environmental parameters. An LED display unit, communication interface, and power management circuit were simultaneously connected to form a finger-clip wearable structure, enabling real-time detection and data transmission of blood pressure waveforms. Figure 3 The diagram shown is a block diagram of a photoplethysmography (PPG) acquisition system.

[0097] This invention achieves high-precision reconstruction of blood pressure waveforms through collaborative modeling of forward and inverse problems. The forward problem, starting from physiological mechanisms, constructs a 3D vascular network simulation platform based on VTK, fully modeling the vascular branching structure from the radial artery to the fingertip, and integrating multiphysics equations to reveal the conversion mechanism from ABP to PPG from first-principles calculations; the inverse problem utilizes U... 2- A deep temporal network model using Net and BTP is used to reconstruct blood pressure waveforms. A combined loss function is designed, and deep learning is used to mine physiological processes. Positive and negative problems are mutually verified and optimized on a dataset.

Claims

1. A non-invasive blood pressure waveform reconstruction method based on 3D blood vessel modeling and deep time series network, characterized in that, The method comprises the following steps: Step 1: Construct a physiological simulation model containing radial artery to fingertips using VTK; Step 2: Integrate Moens-Korteweg wave equation, vascular compliance model and Beer-Lambert optical law to establish a signal conversion model from arterial blood pressure (ABP) to photoplethysmography (PPG); Step 3, design containing U 2 - Deep temporal network architecture with Net-in-Net U-shaped structure and bidirectional temporal processor (BTP); Step 4: Realize high-precision arterial blood pressure waveform reconstruction through deep supervision strategy and combined loss function; Step 5: Verify the reconstructed arterial blood pressure waveform using multi-source data; Step 6: Perform PPG acquisition and ABP real-time acquisition reconstruction in a hardware system integrating MAX30102 sensor and STM32F103 processor.

2. The non-invasive blood pressure waveform reconstruction method based on 3D blood vessel modeling and deep time series network according to claim 1, characterized in that, In step 1, Visualization Toolkit is used for vascular geometry modeling to realize a complete arterial tree structure containing radial artery trunk, palmar arch branches and five-finger end vessels. Each blood vessel path generates a smooth centerline through spline interpolation, and the diameter is gradually set according to Murray's law to ensure compliance with the physiological laws of blood vessel branching. The specific formula is as follows: wherein is the first segment blood vessel diameter, is the initial blood vessel diameter, is the cumulative path length. The constructed model encompasses a multi-level blood vessel system from the wrist radial artery (diameter 2.5 mm) to the fingertip arteriole (diameter 0.4-0.5 mm), including a total of 11 major vessel segments, while the physiological parameters of the vessel network are configured.

3. The non-invasive blood pressure waveform reconstruction method based on 3D blood vessel modeling and deep time series network according to claim 1, characterized in that, The specific method of step 2 is as follows: Step 2.1: Construct a blood flow dynamics propagation model according to the Moens-Korteweg equation, which is expressed as: wherein is the vessel wall elastic modulus, is the wall thickness, is the blood density, is the vessel radius. The pulse wave velocity for different vessel diameters can be calculated from the above equation, while also taking into account the effects of vessel elasticity, branch reflections, etc. on the waveforms, and applying conservation laws at each vessel branch junction to ensure continuity of pressure and flow. Step 2.2: Establish a blood vessel volume conversion model, which includes the nonlinear relationship between blood vessel pressure and volume: wherein is the baseline blood volume, is the reference pressure, usually taken as the diastolic pressure, is the compliance coefficient, is the nonlinearity index. Here for the characteristics of the peripheral vessels, the model introduces a frequency-dependent attenuation and peripheral amplification effect, where the high-frequency component greater than 25 Hz attenuation coefficient is -3 dB; Step 2.3: Design an optical signal conversion system and apply the modified Beer-Lambert law: wherein is the tissue absorption coefficient, is the blood absorption coefficient, is the direct current blood volume component, where the peripheral amplification factor is set to 1.15 to convert the finger tip pressure waveform to a PPG signal, achieving a complete conversion from mechanical to optical signals.

4. The non-invasive blood pressure waveform reconstruction method based on 3D blood vessel modeling and deep time series network according to claim 1, characterized in that, The specific implementation steps of step 3 are as follows: Step 3.1, construction of U 2 -Net multi-scale feature extraction architecture, using a 5-level nested U-shaped structure design, each level contains multiple levels of Residual U-block (RSU) modules, capturing the spatial topological information of the vascular network contained in the PPG signal, corresponding to the vascular branch structure from the radial artery to the fingertip in the forward problem; the down-sampling process uses the maximum pooling operation, and the up-sampling operation using linear interpolation is used to realize the resolution recovery, and the architecture captures the multi-scale features from local details to global trends in the PPG signal through the feature reuse mechanism between levels; Step 3.2: Design and integrate a bidirectional time processor module, embed a BTP unit composed of a bidirectional LSTM layer and a 1×1 convolution layer at the jump connection of each level of U-shaped structure, and the dynamic update mechanism of its time sequence feature processing process is: wherein is the hidden state at time is the PPG signal input at time. This mechanism can model both forward and backward physiological temporal dependencies, effectively capturing the pulse wave conduction differences caused by individual vascular characteristics; Step 3.3: Construct a 5-level nested U-shaped convolution network, and an 8-second PPG signal segment with a sampling rate of 125Hz is used as the standard input window. The time length setting ensures that the complete cardiac cycle characteristics are included; The number of network convolution kernel channels is configured as 64, 128, 256, 512 and 1024 channels respectively, and the first level RSU module uses a 5-layer depth separable convolution structure to process the original PPG signal, and realizes multi-scale feature fusion through jump connection between levels.

5. The non-invasive blood pressure waveform reconstruction method based on 3D blood vessel modeling and deep time series network according to claim 1, characterized in that, The specific implementation steps of step 4 are as follows: Step 4.1: Configure training optimization parameters, fix the batch size to 32 to balance memory usage and gradient stability, set the initial learning rate to 0.001 and use the cosine decay strategy for dynamic adjustment, and the learning rate update follows the mathematical law expression: where the maximum learning rate = 0.001, the minimum learning rate = 0.0001, T is the total training period set to 200 rounds, so as to realize the stable convergence of the model parameters; Step 4.2, construct hierarchical supervision network system, in U 2 - Set up independent supervision nodes at the output end of each decoder of the Net architecture to form a full-stack supervision network from high-resolution bottom-layer features to low-resolution high-layer features, wherein the bottom-layer supervision nodes focus on the extraction of pulse wave local morphological features, and the high-layer supervision nodes strengthen the modeling of the blood vessel network topology, and gradient propagation is achieved through 5-level jump connections; Step 4.3: Design a multi-objective combined optimization system, establish a combined loss function composed of waveform fidelity, extreme point accuracy and physiological correlation; and implement an adaptive weight scheduling mechanism, set a hierarchical weight decreasing strategy from bottom 1.0 to top 0.2, to ensure that each level of the network can learn effective feature representation.

6. The non-invasive blood pressure waveform reconstruction method based on 3D blood vessel modeling and deep time series network according to claim 1, characterized in that, The specific implementation steps of step 5 are as follows: Step 5.1, select different frequency of sampling data for verification, extract records with duration more than 8 minutes and ABP value less than 200mmHg from MIMIC II dataset, and select the first 150 records as the research object; in Vital DB dataset, get the cases containing PPG and ABP signals from the web monitoring dataset, and download the data of the first 10 cases for experiment; Step 5.2, use two indicators to evaluate the accuracy of important values: MAE and STD. The specific formula is as follows: wherein and respectively represent the first predicted and true values of the blood pressure values. STD is defined as: wherein represents the difference between the true value and the predicted blood pressure value. Step 5.3, construct evaluation index to evaluate the overall reconstruction performance of reconstructed waveform, including numerical value and waveform similarity. The index combines normalized root mean square error value prediction index and normalized dynamic time warping waveform prediction index. Normalize the index to facilitate comparison of reconstructed signals of different lengths and amplitudes.

7. The method of claim 1, wherein, In step 6, a blood pressure waveform monitoring hardware system is built, which integrates the minimum system circuit, reflective photoelectric biosensor circuit and temperature and humidity sensor circuit through STM32F103RET6 microcontroller to form the core control module; MAX30102 PPG sensor is used to collect blood pressure waveform signals, and SHT35 high-precision temperature and humidity sensor is used to monitor environmental parameters; synchronously connect LED display unit, communication interface and power management circuit to form a finger clip wearable structure, realize real-time detection and data transmission of blood pressure waveform.

Images