Non-invasive blood pressure reconstruction system and method based on physical decomposition and blood flow drift modeling
By decoupling the blood pressure reconstruction task into morphological reconstruction and drift calibration stages, and combining deep learning and a trust window mechanism, the problem of insufficient blood pressure reconstruction accuracy in existing technologies is solved, achieving high-precision and stable blood pressure monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING QICHENG MEDICAL TECHNOLOGY CO LTD
- Filing Date
- 2026-05-20
- Publication Date
- 2026-06-26
Smart Images
Figure CN122286267A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biomedical signal processing technology, specifically to a non-invasive blood pressure reconstruction system and method based on physical decomposition and blood flow drift modeling. Background Technology
[0002] Continuous arterial blood pressure waveform monitoring is a core clinical requirement for perioperative management, intensive care, and postoperative ward monitoring, directly impacting the assessment of patients' hemodynamic status and the timing of emergency interventions. Current mainstream clinical monitoring methods have significant limitations. Non-invasive continuous blood pressure waveform reconstruction technology based on photoplethysmography (PPG) signals, with its dual advantages of non-invasive safety and continuous monitoring, has become an ideal alternative to traditional methods.
[0003] Existing end-to-end models require a single network to simultaneously learn two conflicting tasks: reconstructing the fine morphology of blood pressure waveforms and regressing absolute values such as systolic and diastolic blood pressure. These two objectives compete with each other in the loss function, leading to waveform distortion or insufficient accuracy in absolute values. In long-term clinical monitoring, time-varying factors such as peripheral vascular tension, depth of anesthesia, vasoactive drugs, and changes in body position can alter the mapping relationship between PPG and blood pressure. Existing two-stage cascaded methods, with highly consistent learning objectives in both stages, only refine the accuracy progressively, failing to decouple the tasks and unable to incorporate calibration information to combat drift. Essentially, they do not overcome the inherent limitations of end-to-end models. Summary of the Invention
[0004] The purpose of this invention is to provide a non-invasive blood pressure reconstruction system and method based on physical decomposition and blood flow drift modeling, so as to solve the problems raised in the prior art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: In the first aspect, this application provides a non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling, including the following steps: The absolute arterial blood pressure waveform reconstruction task is decoupled into a waveform morphology reconstruction stage and a blood pressure drift calibration stage. The waveform morphology reconstruction stage outputs a normalized arterial blood pressure waveform based on photoplethysmography pulse wave signals and individual characteristic information. The blood pressure drift calibration stage uses intermittent non-invasive blood pressure measurements as calibration anchors and combines calibration timeliness information to predict the amount of blood pressure parameter drift. The normalized arterial blood pressure waveform is combined with the blood pressure parameter drift to form the absolute arterial blood pressure waveform; Train the waveform morphology reconstruction stage and the blood pressure drift calibration stage separately; Trust window management is performed based on the calibration time interval, and safety output control is triggered when calibration information exceeds the validity period; when a new intermittent measurement value arrives, the calibration anchor point is updated and the calibration time interval is reset.
[0006] In conjunction with the first aspect, in a first embodiment of the first aspect of this application, the waveform morphology reconstruction stage outputs a normalized arterial blood pressure waveform based on photoplethysmography (PPG) signals and individual characteristic information, including: The input to the waveform morphology reconstruction stage is a multi-channel signal tensor, including the preprocessed original photoplethysmography (PPG) signal and its first and second derivatives, as well as the demographic features of the multi-channel feature map; the multi-channel signal tensor is spliced in the channel dimension; the multi-channel feature map is obtained by multilayer perceptron encoding and is the same length as the PPG segment.
[0007] In conjunction with the first aspect, in a second embodiment of the first aspect of this application, the waveform morphology reconstruction stage includes: The waveform reconstruction stage uses a deep learning codec network as the backbone architecture. Side outputs are generated in each decoding stage. After being upsampled to a uniform length, each side output participates in the loss calculation together with the main output to achieve deep supervision. The supervision weight of the side output decreases from shallow to deep layers.
[0008] In conjunction with the first aspect, in a third embodiment of the first aspect of this application, the waveform morphology reconstruction stage includes: The waveform morphology reconstruction stage adopts a multi-objective weighted loss function with waveform morphology fidelity as the objective. Specifically, it is a weighted combination of point-by-point average absolute error loss, Pearson correlation coefficient loss, frequency domain loss, boundary loss and deep supervision loss, wherein the frequency domain loss is composed of a weighted sum of amplitude spectrum loss and phase spectrum loss.
[0009] Multi-objective weighted loss function in waveform morphology reconstruction stage The formula is as follows: ; in, The weighting coefficients for the point-by-point average absolute error loss are: The weighting coefficients for the Pearson correlation coefficient loss are... These are the weighting coefficients for the frequency domain loss. These are the weighting coefficients for the boundary loss. The point-by-point average absolute error loss represents the point-by-point amplitude matching accuracy between the constrained predicted waveform and the true normalized waveform. To constrain the overall morphological similarity between the predicted waveform and the true waveform, the Pearson correlation coefficient loss is used. For frequency domain loss, To mitigate boundary loss, additional constraints are imposed on key physiological inflection points such as the contraction peak and diastolic trough of the waveform. The deep supervision loss is the sum of the weighted losses of the outputs of each decoding stage of the network, which alleviates the gradient vanishing phenomenon in deep networks and forces each layer to learn effective morphological features.
[0010] In conjunction with the first aspect, in the fourth embodiment of the first aspect of this application, the blood pressure drift calibration stage uses intermittent non-invasive blood pressure measurements as calibration anchors and predicts the amount of blood pressure parameter drift by combining calibration timeliness information, including: The blood pressure drift calibration stage receives a fixed-dimensional context vector as input. The context vector is composed of the most recent intermittent non-invasive blood pressure measurement value after normalization, the calibration time interval, and demographic features. The calibration time interval is obtained by normalizing the maximum trust duration.
[0011] Normalized calibration time interval The calculation formula is as follows: ; in, The absolute time interval between the current moment and the most recent effective intermittent noninvasive blood pressure measurement. The time of the most recent intermittent measurement. This is the preset maximum trust duration. This encoding enables the network to perceive the time distance between the current inference moment and the most recent calibration, thereby automatically learning drift modeling strategies for different calibration ages during training.
[0012] In conjunction with the first aspect, in the fifth embodiment of the first aspect of this application, the blood pressure drift calibration stage includes: The blood pressure drift calibration stage predicts two scalars: pulse pressure drift and diastolic pressure drift. It employs a loss function targeting the absolute accuracy of blood pressure parameters. The loss function is a weighted combination of smoothed L1 losses of systolic pressure, diastolic pressure, and mean arterial pressure. The systolic pressure loss term introduces an asymmetric weighting mechanism to resist regression to the mean. Using the preset population blood pressure mean as the boundary, a larger penalty weight is applied when the predicted systolic pressure is biased towards the population mean, and a smaller penalty weight is applied when it deviates from the population mean.
[0013] The blood pressure drift calibration stage uses a completely different loss function than the waveform morphology reconstruction stage, with the absolute accuracy of the blood pressure parameters as the optimization objective. The formula is as follows: ; in, The total loss function of the drift calibration network during the blood pressure drift calibration phase is given. The weighting coefficient for the systolic pressure loss term. The weighting coefficient for the diastolic pressure loss term. The weighting coefficient for the mean arterial pressure loss term. For the smooth L1 loss of systolic pressure, For smooth L1 loss of diastolic pressure, Smooth L1 loss of mean arterial pressure; In conjunction with the first aspect, in the sixth embodiment of the first aspect of this application, the step of synthesizing the normalized arterial blood pressure waveform and the blood pressure parameter drift into an absolute arterial blood pressure waveform includes: The normalized arterial blood pressure waveform and the blood pressure parameter drift are synthesized into an absolute arterial blood pressure waveform using a physical decomposition reconstruction equation. The physical decomposition reconstruction equation is as follows: ; in, Here is the absolute arterial blood pressure waveform at time t, where t is the time variable. Here is the normalized arterial blood pressure waveform at time t. This represents the lower bound clamping function. This is the reference value for pulse pressure from the most recent intermittent measurement. This is the pulse pressure drift. This is the physiological lower limit constraint value for pulse pressure. This is the reference value for diastolic blood pressure from the most recent intermittent measurement. The diastolic blood pressure drift is quantified as follows: the trust window threshold is set to A to B minutes and can be configured according to the clinical scenario. When the calibration time interval exceeds the threshold, the system refuses to output the absolute blood pressure value and prompts for a new intermittent measurement, where A and B are positive integers set by the user.
[0014] The clinical significance of this trust window mechanism lies in its acknowledgment of the inherent limitations of the model when there is no calibration reference for extended periods, proactively avoiding providing potentially unreliable blood pressure values to clinical decision-making. This "knowing what you don't know" safety design is crucial for meeting the risk management requirements of medical device regulations.
[0015] The values for A and B can be configured according to the clinical scenario, with typical values ranging from 5 to 30 minutes. In perioperative scenarios, a longer trust duration can be used during the maintenance of anesthesia, while a shorter trust duration is preferable during the drastic changes in blood pressure during the induction and recovery phases of anesthesia.
[0016] In conjunction with the first aspect, in the seventh embodiment of the first aspect of this application, the step of training the waveform morphology reconstruction stage and the blood pressure drift calibration stage separately includes: First, the waveform morphology reconstruction network is trained independently until convergence and all its network parameters are frozen. Then, the blood pressure drift calibration network is trained independently. When training the blood pressure drift calibration network, controlled noise is injected into the real normalized arterial blood pressure waveform to simulate the prediction error distribution in the waveform morphology reconstruction stage, thereby enhancing the robustness of the drift calibration network to the error in the waveform morphology reconstruction stage.
[0017] Specifically, for the independent training waveform morphology reconstruction stage, the waveform morphology reconstruction network is trained to convergence using paired data of photoplethysmography pulse wave and normalized arterial blood pressure waveform. After training, all network parameters for the waveform morphology reconstruction stage are frozen.
[0018] For the independent training phase of blood pressure drift calibration, a context vector is constructed containing intermittent measurement reference values, calibration time intervals, and demographic features. The corresponding real blood pressure parameters are paired as supervision labels to train the drift calibration network. These real blood pressure parameters include systolic blood pressure, diastolic blood pressure, and mean arterial pressure. The output of the waveform morphology reconstruction phase in the training data of the blood pressure drift calibration phase can be obtained in the following ways: directly using the frozen waveform morphology reconstruction phase network to perform inference on the training data to obtain the predicted output; or using a real normalized arterial blood pressure waveform with controlled noise injection to simulate the prediction error distribution of the waveform morphology reconstruction phase, making the blood pressure drift calibration phase robust to the errors of the waveform morphology reconstruction phase.
[0019] If the two phases are trained end-to-end, the gradients will flow freely between the two phases. This can lead to the waveform morphology reconstruction phase implicitly learning absolute value information and the blood pressure drift calibration phase implicitly learning morphological information, thus blurring the functional division between the two. This strategy forces the preservation of the functional purity of each phase.
[0020] In conjunction with the first aspect, the eighth embodiment of the first aspect of this application also includes a clinical adaptation enhancement mechanism, specifically: The context vector extension includes perioperative stage labels and clinical event vectors, which include one or more of vasoactive drug injection, volume infusion, and body position adjustment. The blood pressure drift calibration stage adopts a dual-output structure to obtain the expected drift amount and time-varying prediction variance, and further outputs the reconstructed waveform and time-varying confidence band. When the calibration time interval exceeds the trust window threshold, it enters a degraded mode and only outputs the normalized waveform for morphological trend reference.
[0021] Secondly, this application provides a non-invasive blood pressure reconstruction system based on physical decomposition and blood flow drift modeling, including: The photoplethysmography (PPG) signal acquisition module is used to continuously acquire and preprocess the PPG signal of the subject. An intermittent measurement receiving module is used to receive non-invasive blood pressure measurements and record timestamps; The demographic information input module is used to receive the demographic characteristic information of the subjects; The waveform morphology reconstruction module includes a deep learning encoder-decoder network, which is used to fuse the preprocessed photoplethysmography pulse wave signal with encoded demographic features and output a normalized arterial blood pressure waveform that retains only the relative morphological features. The drift calibration module includes a context encoder network for constructing a context vector containing the most recent intermittent noninvasive blood pressure measurement, the time interval since the last measurement, and demographic features, to predict the pulse pressure drift and diastolic blood pressure drift relative to the measurement. The physical reconstruction module is used to synthesize the absolute arterial blood pressure waveform according to the physical decomposition reconstruction equation; The trust window determination module is used to reject output and issue a measurement prompt when the calibration time interval exceeds a threshold. The calibration update module is used to overwrite the previous pulse pressure reference value and diastolic pressure reference value with the latest measurement value when a new intermittent non-invasive blood pressure measurement value is received, and to reset the calibration time interval to zero. The interface module integrates with the perioperative anesthesia information system, automatically acquiring intermittent measurements, surgical stage labels, and clinical event information through this interface.
[0022] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention decomposes the blood pressure waveform into three independent physical components: normalized morphology, pulse pressure, and diastolic baseline. The waveform morphology reconstruction stage focuses on learning the waveform morphology, and the blood pressure drift calibration stage focuses on calibrating the absolute value, thus completely resolving the contradiction of coupled learning. The waveform restoration is more accurate and the absolute value is more reliable.
[0023] 2. This invention reduces the high-dimensional waveform drift to two scalars: pulse pressure drift and diastolic pressure drift. It uses intermittent measurement values as anchor points to learn the drift amount, which greatly reduces the difficulty of drift learning and effectively offsets the systematic bias caused by factors such as vascular status, drugs, and body position, resulting in stable long-term monitoring accuracy.
[0024] 3. The two stages of this invention use different networks, loss functions and training processes. The waveform morphology reconstruction stage is trained first and the parameters are frozen. Then the blood pressure drift calibration stage is trained to avoid functional cross-interference. The model converges faster and is more robust. Attached Figure Description
[0025] Figure 1 This is a schematic diagram of the steps of the non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling of the present invention; Figure 2 This is a system structure diagram of the non-invasive blood pressure reconstruction system based on physical decomposition and blood flow drift modeling of the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] Example: Figures 1-2 As shown, the present invention provides a technical solution: like Figure 1 As shown, this application provides a non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling, including the following steps: Step S100: Decouple the absolute arterial blood pressure waveform reconstruction task into a waveform morphology reconstruction stage and a blood pressure drift calibration stage; the waveform morphology reconstruction stage outputs a normalized arterial blood pressure waveform based on photoplethysmography pulse wave signals and individual characteristic information; the blood pressure drift calibration stage uses intermittent non-invasive blood pressure measurements as calibration anchors and combines calibration timeliness information to predict the amount of blood pressure parameter drift. Specifically, the input to the waveform morphology reconstruction stage is a multi-channel signal tensor, including the preprocessed original photoplethysmography (PPG) signal and its first and second derivatives, as well as the demographic features of the multi-channel feature map; the multi-channel signal tensor is spliced in the channel dimension; the multi-channel feature map is obtained by multilayer perceptron encoding and is the same length as the PPG segment.
[0028] Furthermore, the waveform morphology reconstruction stage adopts a deep learning codec network as the backbone architecture, and side outputs are extracted in each decoding stage. After being upsampled to a uniform length, each side output participates in the loss calculation together with the main output to achieve deep supervision. Moreover, the supervision weight of the side output decreases from shallow to deep layers.
[0029] Furthermore, the waveform morphology reconstruction stage adopts a multi-objective weighted loss function with waveform morphology fidelity as the objective. Specifically, it is a weighted combination of point-by-point average absolute error loss, Pearson correlation coefficient loss, frequency domain loss, boundary loss and deep supervision loss, wherein the frequency domain loss is composed of a weighted sum of amplitude spectrum loss and phase spectrum loss.
[0030] Furthermore, the blood pressure drift calibration stage receives a fixed-dimensional context vector as input, which is composed of the normalized most recent intermittent non-invasive blood pressure measurement, the calibration time interval, and demographic features; the calibration time interval is obtained by normalizing the maximum trust duration.
[0031] Furthermore, in the blood pressure drift calibration stage, the predicted pulse pressure drift and diastolic pressure drift are calculated using a loss function that targets the absolute accuracy of the blood pressure parameters. This loss function is a weighted combination of the smoothed L1 loss of systolic pressure, diastolic pressure, and mean arterial pressure. The systolic pressure loss term introduces an asymmetric weighting mechanism to resist regression to the mean. Using the preset population blood pressure mean as a boundary, a larger penalty weight is applied when the predicted systolic pressure is biased towards the population mean, and a smaller penalty weight is applied when it deviates from the population mean.
[0032] In one specific embodiment, 10 patients with ASA I-II undergoing elective abdominal surgery were selected. During the maintenance anesthesia period, the gold standard data of fingertip photoplethysmography (PPG) and invasive arterial blood pressure were simultaneously collected, and intermittent cuff blood pressure measurements were connected to the operating room monitor. The maximum trust duration was set to 10 minutes, and the physiological lower limit of pulse pressure was set to 5 mmHg. The sampling rate of the fingertip PPG was 125 Hz, and the intermittent cuff blood pressure measurement was automatically taken once every 10 minutes.
[0033] In the waveform morphology reconstruction stage, a nested residual encoder-decoder network with 6 encoders and 5 decoders was used. The loss function weights were set as follows: λ_MAE=5.0, λ_corr=2.0, λ_freq=0.2, λ_bnd=0.5. The side output supervision weights from shallow to deep layers were 0.3, 0.25, 0.2, 0.15, and 0.1, respectively. In the blood pressure drift calibration stage, a 3-layer sensor was used. The loss function weights were: λ_SBP=1.5, λ_DBP=1.0, λ_MAP=1.0. The asymmetric weighting ratio of systolic blood pressure was set to 1.5, and the population reference mean was set to 110 mmHg.
[0034] Experimental results show that, during continuous monitoring for 60 minutes, the average absolute error of systolic blood pressure remained stable at 2.8±1.1 mmHg, and the average absolute error of diastolic blood pressure was 2.1±0.8 mmHg, with a waveform Pearson correlation coefficient of 0.96±0.02. In comparison, the mainstream end-to-end U-Net method had an initial error of 3.2±1.3 mmHg, which drifted to 8.7±2.5 mmHg after 30 minutes, with a waveform correlation coefficient of only 0.88±0.05. In the hypotensive range (systolic blood pressure below 90 mmHg), the average absolute error of this invention was 2.5±0.9 mmHg, significantly better than the existing method's 5.1±1.8 mmHg, without significant regression to the mean bias. No physical abnormalities (pulse pressure below 5 mmHg) were observed in any of the test samples, and the system accurately triggered measurement prompts after calibration timeouts.
[0035] Step S200: Combine the normalized arterial blood pressure waveform with the blood pressure parameter drift to form an absolute arterial blood pressure waveform; Specifically, the normalized arterial blood pressure waveform and the blood pressure parameter drift are synthesized into an absolute arterial blood pressure waveform using a physical decomposition reconstruction equation. The physical decomposition reconstruction equation is as follows: ; in, Here is the absolute arterial blood pressure waveform at time t, where t is the time variable. Here is the normalized arterial blood pressure waveform at time t. This represents the lower bound clamping function. This is the reference value for pulse pressure from the most recent intermittent measurement. This is the pulse pressure drift. This is the physiological lower limit constraint value for pulse pressure. This is the reference value for diastolic blood pressure from the most recent intermittent measurement. The diastolic blood pressure drift is quantified as follows: the trust window threshold is set to A to B minutes and can be configured according to the clinical scenario. When the calibration time interval exceeds the threshold, the system refuses to output the absolute blood pressure value and prompts for a new intermittent measurement, where A and B are positive integers set by the user.
[0036] In one specific embodiment, monitoring data of 30 minutes during the anesthesia maintenance period of one ASA II laparoscopic cholecystectomy patient were selected, with the maximum trust duration set at 10 minutes and the lower physiological limit of pulse pressure PP_min = 5 mmHg, where A = 5 minutes and B = 30 minutes.
[0037] The initial intermittent calibration values were systolic blood pressure 122 mmHg and diastolic blood pressure 76 mmHg, corresponding to a pulse pressure reference value PP_ref = 46 mmHg. At 5 minutes, the predicted pulse pressure drift ΔPP = +3 mmHg and diastolic blood pressure drift ΔDBP = +2 mmHg during the blood pressure drift calibration phase. Substituting these values into the reconstruction equation, we get: systolic blood pressure = 1*(46+3)+(76+2) = 127 mmHg, diastolic blood pressure = 0*(46+3)+(76+2) = 78 mmHg, with errors of 1.2 mmHg and 0.8 mmHg respectively compared to the invasive gold standard. At 10 minutes, the system automatically triggered a new calibration, updating the reference values to systolic blood pressure 125 mmHg and diastolic blood pressure 79 mmHg, and the calibration time interval was reset to zero. At 15 minutes, the predicted ΔPP = -2 mmHg and ΔDBP = +1 mmHg, with reconstructed blood pressure errors of 2.3 mmHg and 1.5 mmHg compared to the gold standard. Meanwhile, the error of the non-drift compensation scheme had risen to 6.8 mmHg and 4.2 mmHg, with ΔPP = 0 and ΔDBP = 0.
[0038] When the calibration interval was extended to 11 minutes, the system automatically refused to output the absolute blood pressure value and displayed a measurement prompt. During the test, an extreme predicted value of ΔPP = -43 mmHg was observed. After correction by the clamping function, the pulse pressure was 5 mmHg, and no physiological abnormalities were output.
[0039] Step S300: Train the waveform morphology reconstruction stage and the blood pressure drift calibration stage separately; Specifically, the waveform morphology reconstruction network is first trained independently until convergence and all its network parameters are frozen. Then, the blood pressure drift calibration network is trained independently. When training the blood pressure drift calibration network, controlled noise is injected into the real normalized arterial blood pressure waveform to simulate the prediction error distribution in the waveform morphology reconstruction stage, thereby enhancing the robustness of the drift calibration network to errors in the waveform morphology reconstruction stage.
[0040] In one specific embodiment, synchronous PPG and invasive ABP paired data from 800 critically ill patients in the MIMIC-IV database were used. The data were divided into training and testing sets according to patient severity, with a uniform sampling rate of 125 Hz. The training set consisted of 640 patients, and the testing set consisted of 160 patients.
[0041] The waveform morphology reconstruction stage involved training a nested residual encoder-decoder network for 100 epochs until convergence. The validation set showed a mean absolute error of 0.032 for the normalized waveforms and a Pearson correlation coefficient of 0.958. All parameters were frozen after training. The blood pressure drift calibration stage employed a 3-layer perceptron. The training set contained 12,000 timestamped intermittent measurement samples, divided into two groups: Group A directly used the output from the frozen waveform morphology reconstruction stage, while Group B added Gaussian noise with a standard deviation of 0.04 to the real normalized waveforms to simulate the prediction error in the waveform morphology reconstruction stage.
[0042] Test results show that the drift calibration network trained in group B, with an output amplitude fluctuation of ±0.05 during the waveform morphology reconstruction stage, achieved a mean absolute error of 2.7±0.9 mmHg in systolic blood pressure, representing a 13% improvement in robustness compared to group A's 3.1±1.2 mmHg. In contrast, the end-to-end joint training scheme reduced the waveform correlation coefficient to 0.912 during the waveform morphology reconstruction stage, and the systolic blood pressure error increased to 7.5±2.3 mmHg after 30 minutes of long-term monitoring. The separate training scheme of this invention, however, maintained a stable long-term error of 2.9±1.1 mmHg, showing no significant accuracy degradation.
[0043] Step S400: Perform trust window management based on the calibration time interval, triggering safety output control when calibration information exceeds the valid time limit; when a new intermittent measurement value arrives, update the calibration anchor point and reset the calibration time interval.
[0044] This invention also includes a clinical adaptation enhancement mechanism, specifically: The context vector extension includes perioperative stage labels and clinical event vectors, which include one or more of vasoactive drug injection, volume infusion, and body position adjustment. The blood pressure drift calibration stage adopts a dual-output structure to obtain the expected drift amount and time-varying prediction variance, and further outputs the reconstructed waveform and time-varying confidence band. When the calibration time interval exceeds the trust window threshold, it enters a degraded mode and only outputs the normalized waveform for morphological trend reference.
[0045] In one specific embodiment, a patient was selected during the induction of general anesthesia, in a scenario of severe blood pressure fluctuations. The trust window threshold was set to 5 minutes, and all clinical adaptation enhancement mechanisms were enabled. A short window was configured for the induction period, and the standard maintenance period was set to 10-30 minutes.
[0046] The initial calibration values were systolic blood pressure of 118 mmHg and diastolic blood pressure of 65 mmHg. 0.1 mg of norepinephrine was injected at the 3rd minute. The model incorporating the vasoactive drug event vector predicted a systolic blood pressure of 136 mmHg 2 minutes after the injection, with an error of 1.8 mmHg compared to the invasive gold standard. The model without the event vector predicted a systolic blood pressure of 124 mmHg, with an error of 10.2 mmHg. The dual-output structure initially showed a 95% confidence band for systolic blood pressure of ±3.2 mmHg. As the calibration time increased to 4.5 minutes, the confidence band naturally broadened to ±4.7 mmHg.
[0047] At 5 minutes and 10 seconds, the calibration interval exceeded the 5-minute threshold, and the system automatically triggered safety control, entering degraded mode: stopping the output of absolute blood pressure values, retaining only the normalized waveform for heart rate and heart rhythm trend monitoring, and displaying a red prompt requesting recalibration. A new cuff calibration was completed at 6 minutes, the calibration interval was reset to zero, the confidence band returned to ±2.5 mmHg, and absolute blood pressure output automatically resumed.
[0048] like Figure 2 As shown, this application provides a non-invasive blood pressure reconstruction system based on physical decomposition and blood flow drift modeling, including: The photoplethysmography (PPG) signal acquisition module is used to continuously acquire and preprocess the PPG signal of the subject. An intermittent measurement receiving module is used to receive non-invasive blood pressure measurements and record timestamps; The demographic information input module is used to receive the demographic characteristic information of the subjects; The waveform morphology reconstruction module includes a deep learning encoder-decoder network, which is used to fuse the preprocessed photoplethysmography pulse wave signal with encoded demographic features and output a normalized arterial blood pressure waveform that retains only the relative morphological features. The drift calibration module includes a context encoder network for constructing a context vector containing the most recent intermittent noninvasive blood pressure measurement, the time interval since the last measurement, and demographic features, to predict the pulse pressure drift and diastolic blood pressure drift relative to the measurement. The physical reconstruction module is used to synthesize the absolute arterial blood pressure waveform according to the physical decomposition reconstruction equation; The trust window determination module is used to reject output and issue a measurement prompt when the calibration time interval exceeds a threshold. The calibration update module is used to overwrite the previous pulse pressure reference value and diastolic pressure reference value with the latest measurement value when a new intermittent non-invasive blood pressure measurement value is received, and to reset the calibration time interval to zero. The interface module integrates with the perioperative anesthesia information system, automatically acquiring intermittent measurements, surgical stage labels, and clinical event information through this interface.
[0049] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling, characterized in that, Includes the following steps: The absolute arterial blood pressure waveform reconstruction task is decoupled into a waveform morphology reconstruction stage and a blood pressure drift calibration stage. The waveform morphology reconstruction stage outputs a normalized arterial blood pressure waveform based on photoplethysmography pulse wave signals and individual characteristic information. The blood pressure drift calibration stage uses intermittent non-invasive blood pressure measurements as calibration anchors and combines calibration timeliness information to predict the amount of blood pressure parameter drift. The normalized arterial blood pressure waveform is combined with the blood pressure parameter drift to form the absolute arterial blood pressure waveform; Train the waveform morphology reconstruction stage and the blood pressure drift calibration stage separately; Trust window management is performed based on the calibration time interval, and safety output control is triggered when calibration information exceeds the validity period; when a new intermittent measurement value arrives, the calibration anchor point is updated and the calibration time interval is reset.
2. The non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling according to claim 1, characterized in that, The waveform morphology reconstruction stage outputs a normalized arterial blood pressure waveform based on photoplethysmography (PPG) pulse wave signals and individual characteristic information, including: The input to the waveform morphology reconstruction stage is a multi-channel signal tensor, including the preprocessed original photoplethysmography (PPG) signal and its first and second derivatives, as well as the demographic features of the multi-channel feature map; the multi-channel signal tensor is spliced in the channel dimension; the multi-channel feature map is obtained by multilayer perceptron encoding and is the same length as the PPG segment.
3. The non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling according to claim 2, characterized in that, The waveform morphology reconstruction stage includes: The waveform reconstruction stage uses a deep learning codec network as the backbone architecture. Side outputs are generated in each decoding stage. After being upsampled to a uniform length, each side output participates in the loss calculation together with the main output to achieve deep supervision. The supervision weight of the side output decreases from shallow to deep layers.
4. The non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling according to claim 3, characterized in that, The waveform morphology reconstruction stage includes: The waveform morphology reconstruction stage adopts a multi-objective weighted loss function with waveform morphology fidelity as the objective. Specifically, it is a weighted combination of point-by-point average absolute error loss, Pearson correlation coefficient loss, frequency domain loss, boundary loss and deep supervision loss, wherein the frequency domain loss is composed of a weighted sum of amplitude spectrum loss and phase spectrum loss.
5. The non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling according to claim 1, characterized in that, The blood pressure drift calibration phase uses intermittent non-invasive blood pressure measurements as calibration anchors and combines calibration timeliness information to predict the amount of blood pressure parameter drift, including: The blood pressure drift calibration stage receives a fixed-dimensional context vector as input. The context vector is composed of the most recent intermittent non-invasive blood pressure measurement value after normalization, the calibration time interval, and demographic features. The calibration time interval is obtained by normalizing the maximum trust duration.
6. The non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling according to claim 5, characterized in that, The blood pressure drift calibration phase includes: The blood pressure drift calibration stage predicts two scalars: pulse pressure drift and diastolic pressure drift. It employs a loss function targeting the absolute accuracy of blood pressure parameters. The loss function is a weighted combination of smoothed L1 losses of systolic pressure, diastolic pressure, and mean arterial pressure. The systolic pressure loss term introduces an asymmetric weighting mechanism to resist regression to the mean. Using the preset population blood pressure mean as the boundary, a larger penalty weight is applied when the predicted systolic pressure is biased towards the population mean, and a smaller penalty weight is applied when it deviates from the population mean.
7. The non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling according to claim 1, characterized in that, The process of synthesizing the normalized arterial blood pressure waveform and the blood pressure parameter drift into an absolute arterial blood pressure waveform includes: The normalized arterial blood pressure waveform and the blood pressure parameter drift are synthesized into an absolute arterial blood pressure waveform using a physical decomposition reconstruction equation. The physical decomposition reconstruction equation is as follows: ； in, Here is the absolute arterial blood pressure waveform at time t, where t is the time variable. Here is the normalized arterial blood pressure waveform at time t. This represents the lower bound clamping function. This is the reference value for pulse pressure from the most recent intermittent measurement. This is the pulse pressure drift. This is the physiological lower limit constraint value for pulse pressure. This is the reference value for diastolic blood pressure from the most recent intermittent measurement. The diastolic blood pressure drift is quantified as follows: the trust window threshold is set to A to B minutes and can be configured according to the clinical scenario. When the calibration time interval exceeds the threshold, the system refuses to output the absolute blood pressure value and prompts for a new intermittent measurement, where A and B are positive integers set by the user.
8. The non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling according to claim 1, characterized in that, The separate training of the waveform morphology reconstruction stage and the blood pressure drift calibration stage includes: First, the waveform morphology reconstruction network is trained independently until convergence and all its network parameters are frozen. Then, the blood pressure drift calibration network is trained independently. When training the blood pressure drift calibration network, controlled noise is injected into the real normalized arterial blood pressure waveform to simulate the prediction error distribution in the waveform morphology reconstruction stage, thereby enhancing the robustness of the drift calibration network to the error in the waveform morphology reconstruction stage.
9. The non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling according to claim 1, characterized in that, It also includes clinical adaptation enhancement mechanisms, specifically: The context vector extension includes perioperative stage labels and clinical event vectors, which include one or more of vasoactive drug injection, volume infusion, and body position adjustment. The blood pressure drift calibration stage adopts a dual-output structure to obtain the expected drift amount and time-varying prediction variance, and further outputs the reconstructed waveform and time-varying confidence band. When the calibration time interval exceeds the trust window threshold, it enters a degraded mode and only outputs the normalized waveform for morphological trend reference.
10. A non-invasive blood pressure reconstruction system based on physical decomposition and blood flow drift modeling, using the non-invasive blood pressure reconstruction method based on physical decomposition and blood flow drift modeling as described in any one of claims 1-9, characterized in that, include: The photoplethysmography (PPG) signal acquisition module is used to continuously acquire and preprocess the PPG signal of the subject. An intermittent measurement receiving module is used to receive non-invasive blood pressure measurements and record timestamps; The demographic information input module is used to receive the demographic characteristic information of the subjects; The waveform morphology reconstruction module includes a deep learning encoder-decoder network, which is used to fuse the preprocessed photoplethysmography pulse wave signal with encoded demographic features and output a normalized arterial blood pressure waveform that retains only the relative morphological features. The drift calibration module includes a context encoder network for constructing a context vector containing the most recent intermittent noninvasive blood pressure measurement, the time interval since the last measurement, and demographic features, to predict the pulse pressure drift and diastolic blood pressure drift relative to the measurement. The physical reconstruction module is used to synthesize the absolute arterial blood pressure waveform according to the physical decomposition reconstruction equation; The trust window determination module is used to reject output and issue a measurement prompt when the calibration time interval exceeds a threshold. The calibration update module is used to overwrite the previous pulse pressure reference value and diastolic pressure reference value with the latest measurement value when a new intermittent non-invasive blood pressure measurement value is received, and to reset the calibration time interval to zero. The interface module integrates with the perioperative anesthesia information system, automatically acquiring intermittent measurements, surgical stage labels, and clinical event information through this interface.