Method, device and equipment for estimating hemodynamic parameters based on non-invasive blood pressure waveform and medium

Abstract

The application provides a hemodynamic parameter estimation method and device based on a non-invasive blood pressure waveform, equipment and a medium, which can be applied to the technical field of medical information processing. The method comprises the following steps: obtaining a non-invasive continuous arterial blood pressure waveform and individual information of a target object; preprocessing the non-invasive continuous arterial blood pressure waveform to construct a standardized single cardiac cycle blood pressure waveform; inputting the standardized single cardiac cycle blood pressure waveform and the individual information into a pre-trained hemodynamic estimation model to obtain an estimation result comprising at least cardiac output and stroke volume; wherein the hemodynamic estimation model comprises a deep neural network module, and the hemodynamic estimation model is trained by constraining the output result predicted by the deep neural network module in a physiological relationship described by a three-element Windkessel model comprising arterial compliance, peripheral resistance and characteristic impedance.

Description

Technical Field

[0001] This application relates to the field of medical information processing technology, and more specifically to a method, device, equipment and medium for estimating hemodynamic parameters based on non-invasive blood pressure waveforms. Background Technology

[0002] In clinical anesthesia, intensive care unit (ICU), and perioperative management, continuous and accurate monitoring of core hemodynamic parameters such as cardiac output (CO) and stroke volume (SV) is crucial for assessing patients' circulatory function, guiding fluid resuscitation, and using vasoactive drugs.

[0003] Currently, the gold standard or commonly used methods in clinical practice mainly include the invasive thermodilution method and invasive arterial pressure monitoring. The thermodilution method, measured via a pulmonary artery catheter, is relatively accurate, but its operation is complex, requiring specialized equipment and personnel, and is an invasive procedure, carrying risks such as catheter-related infection, arrhythmia, and pulmonary artery rupture, thus limiting its application in routine and long-term monitoring. While invasive arterial pressure monitoring provides continuous blood pressure waveforms, it also requires arterial puncture and catheterization, similarly facing risks of complications such as bleeding, embolism, and infection.

[0004] To address the limitations of invasive methods, non-invasive estimation techniques have emerged. Some existing techniques attempt to estimate hemodynamic parameters from non-invasive blood pressure waveforms (such as fingertip volume pulse waves) or photoelectric signals. However, these methods generally suffer from poor model universality, insufficient adaptability to individual differences (such as age and vascular elasticity), and susceptibility to motion artifacts and signal quality interference. This results in limited accuracy, stability, and clinical reliability of the estimation results, making it difficult to meet the high-precision and high-reliability monitoring needs of critical care medicine. Summary of the Invention

[0005] In view of the above problems, this application provides a method, device, equipment and medium for estimating hemodynamic parameters based on non-invasive blood pressure waveforms.

[0006] According to the first aspect of this application, a method for estimating hemodynamic parameters based on non-invasive blood pressure waveforms is provided, comprising:

[0007] Obtain non-invasive continuous arterial blood pressure waveforms and individualized information of the target subject;

[0008] The non-invasive continuous arterial blood pressure waveform is preprocessed to construct a standardized single-cardiac cycle blood pressure waveform;

[0009] The standardized single-cardiac cycle blood pressure waveform and the individualized information are input into a pre-trained hemodynamic estimation model to obtain estimation results including at least cardiac output and stroke volume.

[0010] The hemodynamic estimation model includes a deep neural network module. The hemodynamic estimation model is obtained by training the output predicted by the deep neural network module within the physiological relationships described by a ternary Windkessel model that includes arterial compliance, peripheral resistance, and characteristic impedance.

[0011] According to an embodiment of this application, the preprocessing of the non-invasive continuous arterial blood pressure waveform to construct a standardized single-cardiac cycle blood pressure waveform includes:

[0012] Identify the diastolic pressure point and dicrotic notch from the non-invasive continuous arterial blood pressure waveform;

[0013] Candidate heart rate cycles are determined based on adjacent diastolic pressure points;

[0014] For each candidate heartbeat cycle, a target dicrotic notch is identified within the candidate heartbeat cycle. The target dicrotic notch is located between the diastolic blood pressure point and the next systolic blood pressure point within the candidate heartbeat cycle.

[0015] Candidate heartbeat cycles that include the target dicrotic notch Notch point are identified as valid single heartbeat cycles.

[0016] The waveform of the effective single cardiac cycle is resampled to a fixed length to obtain the standardized single cardiac cycle blood pressure waveform.

[0017] According to an embodiment of this application, the method further includes:

[0018] Acquire the electrocardiogram (ECG) signal synchronously acquired with the non-invasive continuous arterial blood pressure waveform;

[0019] Using the R wave in the ECG signal as a time reference, the identified diastolic pressure point and dicrotic notch are corrected.

[0020] According to an embodiment of this application, the hemodynamic estimation model processes the standardized single-cardiac cycle blood pressure waveform in the following manner:

[0021] The standardized single-cardiac cycle blood pressure waveform is divided into multiple non-overlapping time segments;

[0022] Map each time segment to a feature vector;

[0023] Add positional information representing the temporal order to each feature vector to form a sequence of feature vectors containing temporal relationships;

[0024] Based on the feature vector sequence containing temporal relationships, waveform depth features are extracted.

[0025] According to an embodiment of this application, the deep neural network module includes a first output head and a second output head, and the method further includes:

[0026] The individualized information is mapped into an individual feature vector;

[0027] The individual feature vectors are concatenated with the waveform depth features to form a fused feature;

[0028] The fusion features are input into the first output head and the second output head respectively;

[0029] The reconstructed predicted blood pressure waveform is output through the first output head.

[0030] The second output head outputs the predicted parameter values ​​of the ternary Windkessel model, the parameters of which include at least arterial compliance, peripheral resistance, and characteristic impedance.

[0031] According to an embodiment of this application, the method further includes:

[0032] Calculate the diastolic pressure-time integral based on the predicted blood pressure waveform output by the first output head;

[0033] Based on the characteristic impedance output from the second output head and the diastolic pressure-time integral, cardiac output and stroke volume are calculated.

[0034] According to an embodiment of this application, the method further includes:

[0035] Substitute the predicted parameter values ​​output by the second output head into the differential equation of the ternary Windkessel model to calculate the theoretical blood pressure waveform;

[0036] Determine the physical consistency loss between the theoretical blood pressure waveform output by the first output head and the predicted blood pressure waveform;

[0037] The hemodynamic estimation model is trained based on the aforementioned physical consistency loss.

[0038] A second aspect of this application provides a hemodynamic parameter estimation device based on non-invasive blood pressure waveforms, the device comprising:

[0039] The acquisition module is used to acquire non-invasive continuous arterial blood pressure waveforms and individualized information of the target object;

[0040] A construction module is used to preprocess the non-invasive continuous arterial blood pressure waveform to construct a standardized single-cardiac cycle blood pressure waveform.

[0041] The input module is used to input the standardized single-cardiac cycle blood pressure waveform and the individualized information into a pre-trained hemodynamic estimation model to obtain estimation results including at least cardiac output and stroke volume.

[0042] The hemodynamic estimation model includes a deep neural network module. The hemodynamic estimation model is obtained by training the output predicted by the deep neural network module within the physiological relationships described by a ternary Windkessel model that includes arterial compliance, peripheral resistance, and characteristic impedance.

[0043] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0044] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0045] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0046] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0047] Figure 1 A flowchart illustrating a hemodynamic parameter estimation method based on non-invasive blood pressure waveforms according to an embodiment of this application is shown schematically.

[0048] Figure 2 This illustration schematically shows a flowchart of multi-bioelectric signal preprocessing and model training according to an embodiment of this application;

[0049] Figure 3 The schematic diagram illustrates a structural block diagram of a hemodynamic parameter estimation device based on a non-invasive blood pressure waveform according to an embodiment of this application;

[0050] Figure 4 A block diagram of an electronic device suitable for implementing a cross-calibration-based noninvasive continuous blood pressure monitoring method is illustrated according to an embodiment of this application. Detailed Implementation

[0051] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0052] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0053] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0054] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0055] Figure 1 A flowchart illustrating a hemodynamic parameter estimation method based on non-invasive blood pressure waveforms according to an embodiment of this application is shown.

[0056] like Figure 1 As shown, the method of this embodiment includes operations S110-S130.

[0057] In operation S110, non-invasive continuous arterial blood pressure waveforms and individualized information of the target object are acquired.

[0058] According to embodiments of this application, a non-invasive continuous arterial blood pressure waveform refers to a physiological signal obtained through any non-invasive method that continuously reflects changes in arterial pressure over time. For example, it can be acquired using a cuff-type continuous blood pressure monitoring device based on the oscillometric principle, a fingertip arterial pressure waveform obtained through volumetric clamping (such as Finapres technology), or a waveform acquired at superficial arteries such as the radial artery using a tonometry plane pressure wave sensor. This waveform is typically represented as a digital sequence, with a sampling frequency generally between 100 Hz and 500 Hz.

[0059] According to embodiments of this application, personalized information can refer to static or semi-static feature data related to the physiological state of the target object. This information is used to provide the model with individual-specific prior knowledge to improve the model's personalized adaptability. Typical personalized information includes, but is not limited to, age, gender, height, and weight. Based on height and weight, body mass index (BMI) or body surface area (BSA) can be further calculated as derived feature inputs. In some embodiments, information related to cardiovascular status, such as medical history markers of hypertension or diabetes, may also be included.

[0060] In operation S120, the non-invasive continuous arterial blood pressure waveform is preprocessed to construct a standardized single cardiac cycle blood pressure waveform.

[0061] According to embodiments of this application, the preprocessing step can be to convert the original continuous waveform into high-quality, uniformly formatted data units suitable for model input. This process includes, but is not limited to, filtering and denoising, feature point identification, period segmentation, and standardization.

[0062] Filtering and noise reduction: First, the original waveform is digitally filtered, for example, by using a low-pass filter with a cutoff frequency of 15-20 Hz (such as a Butterworth filter or an FIR filter) to eliminate high-frequency noise and power frequency interference.

[0063] Feature point identification: Automatically identify key physiological landmarks on the filtered waveform. These include at least: the systolic pressure point representing the peak pressure, the diastolic pressure point representing the trough pressure, and the dicrotic notch representing aortic valve closure and the onset of the reflected wave. Identification algorithms can employ methods based on slope thresholding, waveform second derivatives, or template matching.

[0064] Period segmentation: Using two adjacent diastolic pressure points as the boundary of a period, a waveform segment of a single cardiac cycle is extracted from the continuous waveform.

[0065] Standardization: The extracted waveforms of each period are normalized in length to ensure consistent data dimensionality in the input model. A common method is to resample each period waveform to a fixed number of points (e.g., 128, 256, or 512 points) using linear or spline interpolation. Another method is to normalize the time axis to the interval between 0 and 1.

[0066] In operation S130, the standardized single-cardiac cycle blood pressure waveform and individualized information are input into the pre-trained hemodynamic estimation model to obtain estimation results including at least cardiac output and stroke volume.

[0067] According to embodiments of this application, the hemodynamic estimation model includes a deep neural network module. The hemodynamic estimation model can be a Physical Information Neural Network (PINN), which is trained by constraining the output predicted by the deep neural network module to the physiological relationships described by a ternary Windkessel model including arterial compliance C, peripheral resistance R, and characteristic impedance Zc. The hemodynamic estimation model is integrated into a multimodal physiological signal acquisition device and can be carried by a high-performance processor (such as the RK3588) to support high-speed real-time computation.

[0068] According to embodiments of this application, the deep neural network module is a computational model used to learn and extract high-level abstract features from standardized waveforms and individualized information. This module can be any deep learning architecture suitable for sequence data processing. In one implementation, the module employs an improved Transformer model based on a self-attention mechanism, such as the encoder part of a Transformer, which enhances the ability to capture key temporal features of blood pressure waveforms through multi-head self-attention. Alternatively, it can employ networks such as Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRUs).

[0069] According to the embodiments of this application, during the training phase of the model, the optimization objective is not simply to minimize the error between the predicted parameters and the label, but rather to introduce an additional physical rule defined by the ternary Windkessel model as a constraint through a composite loss function. This constraint ensures that the learning process of the neural network is guided to a solution space that conforms to physiological laws.

[0070] One way to implement this constraint during training is to design a composite loss function. This loss function can contain multiple parts. For example, a waveform reconstruction loss term L_wave based on the predicted waveform and the real waveform, a parameter label loss term L_label based on the predicted parameters and the reference label, and a key physical constraint loss term L_parameters.

[0071] The calculation logic for the physical constraint loss term is as follows: From the specific design output of the deep neural network module (such as the predicted C, R, Zc parameters and the reconstructed blood pressure waveform), a theoretical blood pressure waveform P_windkessel(t) is calculated by substituting the predicted parameters into the differential equation of the ternary Windkessel model. Then, the consistency between this theoretical waveform and the predicted waveform reconstructed by the neural network is forced, and the difference is calculated as the physical constraint loss. By minimizing this physical constraint loss term, the model is trained so that both its internal representation and final output approximately satisfy the physiological and physical model.

[0072] According to the embodiments of this application, because the physical constraints of the Windkessel model are embedded during the model training phase, the neural network ultimately used for estimation not only learns the statistical regularities in historical data but also incorporates the physiological principles of the cardiovascular system. Therefore, even under conditions of significant individual differences or signal interference, the estimation results can maintain high physiological rationality and accuracy, overcoming the shortcomings of purely data-driven models that may produce results that defy common sense.

[0073] Figure 2 The flowchart illustrating the preprocessing of multiple bioelectrical signals and model training according to an embodiment of this application is shown in the illustration.

[0074] In some embodiments of this application, diastolic pressure points and dichroic notches can be identified from non-invasive continuous arterial blood pressure waveforms; candidate heartbeat cycles are determined based on adjacent diastolic pressure points; for each candidate heartbeat cycle, a target dichroic notch is confirmed to exist within the candidate heartbeat cycle, and the target dichroic notch is located between the diastolic pressure point and the next systolic pressure point within the candidate heartbeat cycle; the candidate heartbeat cycle containing the target dichroic notch is determined to be a valid single heartbeat cycle; the waveform of the valid single heartbeat cycle is resampled to a fixed length to obtain a standardized single heartbeat cycle blood pressure waveform. To improve robustness, dynamic statistical analysis can be performed on the distance between adjacent DBP points to assist in screening.

[0075] For example, a local minimum search algorithm can be used to identify diastolic blood pressure points. On the filtered waveform, a time window is slid across, and points within the window that have the smallest amplitude and meet a certain amplitude threshold are identified as candidate diastolic blood pressure points.

[0076] For example, identifying the diabetic notch can be done by analyzing the first and second derivatives of the waveform. The diabetic notch typically corresponds to a small inflection point or plateau during a pressure drop. It can be located near points where the first derivative rapidly changes from negative to positive, and where the second derivative crosses zero. Alternatively, template matching can be used to determine the location of the diabetic notch by correlating the waveform with a template containing typical diabetic notch patterns.

[0077] According to an embodiment of this application, using two consecutive diastolic blood pressure points as time boundaries, a segment of signal is extracted from the original continuous waveform, which can be initially defined as a candidate heartbeat cycle. For example, assuming the identified diastolic blood pressure point time series is [t_dbp1, t_dbp2, t_dbp3, ...], then the data of the first candidate heartbeat cycle is the waveform segment from t_dbp1 to t_dbp2 (excluding t_dbp2).

[0078] According to the embodiments of this application, the systolic pressure point can be located within a candidate period by searching for the local maximum value to find the pressure peak point, which is the systolic pressure point of that period.

[0079] Specifically, it can be checked whether there is one and only one identified diabetic notch within the open interval of (diastolic pressure point time of this cycle, systolic pressure point time of this cycle).

[0080] According to embodiments of this application, linear interpolation or cubic spline interpolation algorithms can be used to resample the waveform of a valid single cardiac cycle to a fixed length. For example, an original periodic waveform with a length of M points can be interpolated to generate a new waveform sequence with a length of N (e.g., 256) points uniformly distributed on a new time axis. After each cardiac cycle is resampled to a fixed length, a standard waveform template can be established through multi-cycle averaging.

[0081] In some embodiments of this application, an electrocardiogram (ECG) signal can be acquired synchronously with a non-invasive continuous arterial blood pressure waveform; the R wave in the ECG signal is used as a time reference to correct the identified diastolic blood pressure point and dicrotic notch. This can be used to help eliminate erroneous DBP point identification caused by artifacts in the ABP waveform itself.

[0082] Classic QRS wave detection algorithms (such as the Pan-Tompkins algorithm) can be used to process synchronized ECG signals to accurately identify the times when R waves appear in each heartbeat cycle (t_R1, t_R2, t_R3, ...). These R wave time sequences constitute a high-confidence time reference line.

[0083] Correction methods can include time window constraint method, delay model-assisted method, period alignment and filtering, etc.

[0084] In some embodiments of this application, the hemodynamic estimation model processes the standardized single-cardiac cycle blood pressure waveform in the following manner: the standardized single-cardiac cycle blood pressure waveform is segmented into multiple non-overlapping time segments; each time segment is mapped to a feature vector; positional information representing the temporal order is added to each feature vector to form a feature vector sequence containing temporal relationships; and waveform depth features are extracted based on the feature vector sequence containing temporal relationships. For example, the above method can be implemented using a feature extraction backbone network.

[0085] For example, suppose the standardized blood pressure waveform is a one-dimensional array of length L, such as L = 256. Define a fixed segment length, also called the patch size, denoted by P. For example, P = 16, correspondingly, it is divided into N = 32 patches. Divide the entire waveform sequentially into N = L / P non-overlapping segments (assuming L is divisible by P). For example, 256 / 16 = 16 segments. Each segment contains P consecutive samples from the original waveform. For example, the first segment contains samples 1 to 16, the second segment contains samples 17 to 32, and so on.

[0086] For example, the mapping can be achieved through a fully connected layer (Patch embedding layer), where the weight matrix W of the fully connected layer has dimensions [P, D], and the bias vector b has dimensions D. For an input fragment x_i (dimension P), its embedding operation is represented as: e_i = x_i * W + b. The result e_i is a D-dimensional feature vector. Here, D is the embedding dimension, which is usually much larger than P, for example, D = 128 or 256. This fully connected layer is part of the trainable parameters of the entire model, and during training, it learns how to map the raw pressure fragments to feature representations useful for the final task (hemodynamic parameter estimation). Simultaneously, sinusoidal positional encoding is added to preserve temporal information.

[0087] For example, the fusion operation can add the positional encoding vector p_i to the corresponding embedded feature vector e_i element-wise: z_i = e_i + p_i. This yields a new feature vector z_i that contains both local waveform semantics and its positional information within the sequence. The sequence composed of all z_i is a "feature vector sequence containing temporal relationships".

[0088] For example, the pre-processed feature vector sequence [z_1, z_2, ..., z_N] can be received, and through multi-layer complex nonlinear transformations (such as the multi-head self-attention mechanism of the Transformer encoder, convolution, and recurrent connections), the long-range dependencies and contextual information between features at various positions in the sequence can be captured, and finally a representation that can comprehensively characterize the high-level abstract features of the entire cardiac cycle waveform can be output.

[0089] In some embodiments of this application, the deep neural network module includes a first output head (waveform prediction head) and a second output head (parameter prediction head), wherein individualized information can be upscaled using a multilayer perceptron (MLP). Individualized information can be mapped to individual feature vectors; these individual feature vectors are concatenated with waveform depth features to form fused features; the fused features are input into the deep neural network module; and the reconstructed predicted blood pressure waveform P_pred(t) is output through the first output head. The parameter prediction values ​​of the ternary Windkessel model are output through the second output head, where the parameters include at least arterial compliance C, peripheral resistance Rp, and characteristic impedance Zc.

[0090] Output heads can be one or more specialized neural network layers connected to the end of a shared feature extraction backbone network, responsible for mapping shared abstract features to different specific task spaces. The first and second output heads are responsible for different prediction tasks. The two output heads can be small multilayer perceptrons with identical or different structures. They are typically connected in parallel after the output of the shared feature extractor. The waveform depth features learned by the shared part serve as the common input to both heads.

[0091] Waveform depth features and individual feature vectors are fused at a certain network layer (such as after a shared feature extractor). The fused representation is then fed into subsequent network layers (possibly including some shared layers), and finally branches into two independent output heads. The first output head's task is self-supervised reconstruction. It learns to decode the fused features and map them back to the original standardized blood pressure waveform space. Its output is a sequence with the same dimension as the input standardized single-cardiac cycle blood pressure waveform (e.g., a sequence of 256 pressure values). The second output head's task is physical parameter identification. It directly learns and infers the core lumped parameters of the ternary Windkessel model describing the current state of the individual's cardiovascular system from the fused features.

[0092] The first output head can be a deconvolutional network, a multilayer perceptron followed by a reconstructing operation, or a sequence generation layer. For example, an MLP containing several fully connected layers has the number of neurons in its last layer equal to the length L of the normalized waveform, with a linear activation function, directly outputting an L-dimensional vector, i.e., the reconstructed waveform P_pred(t).

[0093] The second output head is typically a multilayer perceptron, where the number of neurons in the last layer equals the number of parameters to be predicted. The activation function is chosen to ensure the parameter values ​​are within a reasonable range (e.g., using Softplus to ensure positive values). The predicted parameters include at least arterial compliance, peripheral resistance, and characteristic impedance.

[0094] In some embodiments of this application, the diastolic pressure-time integral is calculated based on the predicted blood pressure waveform output by the first output head; the cardiac output and stroke volume are calculated based on the characteristic impedance output by the second output head and the diastolic pressure-time integral.

[0095] Diastolic phase definition can be achieved on the standardized blood pressure waveform P_pred(t) reconstructed from the first output head. First, the diastolic phase needs to be located. This can be done automatically based on waveform characteristics: the start of diastole is the dicrotic notch on the waveform (its location is known or can be re-detected after preprocessing and waveform reconstruction), and the end point is the lowest diastolic pressure point of that cycle, i.e., the end point of the standardized waveform. Stroke volume (SV) is estimated based on hemodynamic principles. Where PSA is the diastolic pressure-time integral (diastolic blood flow integral), and Q(t) is the aortic blood flow.

[0096] Cardiac output is the total amount of blood pumped out by the heart per unit time (per minute). It is calculated from stroke volume and heart rate according to its definition: CO = SV * HR.

[0097] Heart rate can be obtained directly from the preprocessing step. Since the above preprocessing process has determined the effective single cardiac cycle, the heart rate can be calculated as: HR = 60 / T, where T is the duration of the effective cycle (seconds). The cycle duration is known before waveform normalization. CO_pred = SV_pred * HR. Here, SV_pred is in ml, HR is in beats / minute, therefore CO_pred is in ml / min, usually converted to L / min.

[0098] In some embodiments of this application, the predicted parameter values ​​output by the second output head are substituted into the differential equation of the ternary Windkessel model to calculate the theoretical blood pressure waveform; the physical consistency loss between the theoretical blood pressure waveform and the predicted blood pressure waveform output by the first output head is determined; and the hemodynamic estimation model is trained based on the physical consistency loss.

[0099] The differential equation of the ternary Windkessel model: A common form of expression is:

[0100] P(t) + τ * dP(t) / dt = R * Q(t) + τ * Zc * dQ(t) / dt

[0101] Where P(t) is arterial blood pressure, Q(t) is aortic blood flow, and τ = R * C is a time constant.

[0102] Substitute the predicted C_pred, R_pred, and Zc_pred from the second output head into the above equation. To solve the equation to obtain P(t), Q(t) needs to be known or estimated. During the training phase, Q(t) can be obtained from supervised data (if the dataset contains synchronized invasive cardiac output or flow waveforms) or derived from the predicted blood pressure waveform P_pred(t) using a simplified estimation model. For example, blood flow Q(t) can be estimated by the model. Given the initial conditions and flow Q(t), solve the differential equation using numerical integration methods (such as Euler's method or Runge-Kutta's method) to obtain the theoretical blood pressure waveform P_windkessel(t) over a period of time. The solution time can be aligned with a cardiac cycle.

[0103] The physical consistency loss L_physics is typically defined as a distance metric between two waveform sequences. Before calculation, P_pred(t) and P_windkessel(t) can be aligned on the time axis and their amplitudes normalized to ensure fair comparison.

[0104] The model training employs a two-stage training process and a composite loss function: L_total = αL_wave + βL_label + γ*L_parameters. Where, L_wave: the reconstruction loss (e.g., MSE) between the reconstructed waveform P_pred(t) from the first output head and the input true normalized waveform P_true(t). L_label: the supervision loss (e.g., MSE) between the physical parameters (C_pred, R_pred, Zc_pred) predicted by the second output head and reference values ​​from invasive measurements or high-precision derivations. Note that this may be missing in some scenarios without full supervision.

[0105] L_parameters: Physical consistency loss. α, β, γ are the weighting coefficients of each loss term, used to balance the importance of different objectives. Parameter iterative updates employ optimization algorithms (such as AdamW) and schedulers (such as CosineAnnealingLR) for real-time fitting to minimize error.

[0106] During the forward propagation of the training iterations, a batch of data is input, and the model outputs P_pred(t) and physical parameters. During loss calculation, L_wave, L_label (if any), and L_parameters are calculated, resulting in L_total. In backpropagation and parameter updates, the gradient of L_total with respect to all model parameters is calculated, and optimization algorithms (such as AdamW) are used to update the parameters, reducing L_total.

[0107] By continuously minimizing L_parameters, the model is trained to increasingly satisfy the Windkessel equations in its parameter and waveform predictions. Even in weakly supervised or self-supervised scenarios without L_label supervision (i.e., non-invasive data lacking precise physical parameter labels), L_parameters provide crucial physical guidance for the model, enabling real-time and accurate parameter estimation. In multi-bioelectric data acquisition devices, a dedicated process pool can be set up to maintain this algorithm.

[0108] Figure 3 The diagram schematically illustrates a structural block diagram of a hemodynamic parameter estimation device based on a non-invasive blood pressure waveform according to an embodiment of this application.

[0109] like Figure 3 As shown, the hemodynamic parameter estimation device based on non-invasive blood pressure waveform in this embodiment includes an acquisition module 310, a construction module 320, and an input module 330.

[0110] The acquisition module 310 is used to acquire the non-invasive continuous arterial blood pressure waveform and individualized information of the target object. In one embodiment, the acquisition module 310 can be used to perform the operation S110 described above, which will not be repeated here.

[0111] The construction module 320 is used to preprocess the non-invasive continuous arterial blood pressure waveform to construct a standardized single-cardiac cycle blood pressure waveform. In one embodiment, the construction module 320 can be used to perform the operation S120 described above, which will not be repeated here.

[0112] Input module 330 is used to input standardized single-cardiac cycle blood pressure waveforms and individualized information into a pre-trained hemodynamic estimation model to obtain estimation results including at least cardiac output and stroke volume. In one embodiment, input module 330 can be used to perform the operation S130 described above, which will not be repeated here.

[0113] The hemodynamic estimation model includes a deep neural network module. The hemodynamic estimation model is obtained by training the output predicted by the deep neural network module to constrain the physiological relationships described by the ternary Windkessel model, which includes arterial compliance, peripheral resistance, and characteristic impedance.

[0114] Figure 4 A block diagram of an electronic device suitable for implementing a cross-calibration-based noninvasive continuous blood pressure monitoring method is illustrated according to an embodiment of this application.

[0115] like Figure 4 As shown, an electronic device according to an embodiment of this application includes a processor 401, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 402 or a program loaded from a storage portion 408 into a random access memory (RAM) 403. The processor 401 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 401 may also include onboard memory for caching purposes. The processor 401 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0116] RAM 403 stores various programs and data required for the operation of electronic device 400. Processor 401, ROM 402, and RAM 403 are interconnected via bus 404. Processor 401 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 402 and / or RAM 403. It should be noted that programs may also be stored in one or more memories other than ROM 402 and RAM 403. Processor 401 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in one or more memories.

[0117] According to embodiments of this application, the electronic device may further include an input / output (I / O) interface 405, which is also connected to a bus 404. The electronic device 400 may also include one or more of the following components connected to the input / output (I / O) interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN card, modem, etc. The communication section 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to the input / output (I / O) interface 405 as needed. A removable medium 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 410 as needed so that computer programs read from it can be installed into the storage section 408 as needed.

[0118] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0119] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 402 and / or RAM 403 and / or one or more memories other than ROM 402 and RAM 403 described above.

[0120] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the cross-calibration-based non-invasive continuous blood pressure monitoring method for financial trading systems provided in the embodiments of this application.

[0121] When the computer program is executed by the processor 401, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0122] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via communication section 409, and / or installed from removable medium 411. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0123] In such an embodiment, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by processor 401, it performs the functions defined in the system of this application embodiment. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0124] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0125] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0126] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A method for estimating hemodynamic parameters based on non-invasive blood pressure waveforms, characterized in that, include: Obtain non-invasive continuous arterial blood pressure waveforms and individualized information of the target subject; The non-invasive continuous arterial blood pressure waveform is preprocessed to construct a standardized single-cardiac cycle blood pressure waveform; The standardized single-cardiac cycle blood pressure waveform and the individualized information are input into a pre-trained hemodynamic estimation model to obtain estimation results including at least cardiac output and stroke volume. The hemodynamic estimation model includes a deep neural network module. The hemodynamic estimation model is obtained by training the output predicted by the deep neural network module within the physiological relationships described by a ternary Windkessel model that includes arterial compliance, peripheral resistance, and characteristic impedance.

2. The method according to claim 1, characterized in that, The preprocessing of the non-invasive continuous arterial blood pressure waveform to construct a standardized single-cardiac cycle blood pressure waveform includes: Identify the diastolic pressure point and dicrotic notch from the non-invasive continuous arterial blood pressure waveform; Candidate heart rate cycles are determined based on adjacent diastolic pressure points; For each candidate heartbeat cycle, a target dicrotic notch is identified within the candidate heartbeat cycle. The target dicrotic notch is located between the diastolic blood pressure point and the next systolic blood pressure point within the candidate heartbeat cycle. Candidate heartbeat cycles that include the target dicrotic notch Notch point are identified as valid single heartbeat cycles. The waveform of the effective single cardiac cycle is resampled to a fixed length to obtain the standardized single cardiac cycle blood pressure waveform.

3. The method according to claim 2, characterized in that, The method further includes: Acquire the electrocardiogram (ECG) signal synchronously acquired with the non-invasive continuous arterial blood pressure waveform; Using the R wave in the ECG signal as a time reference, the identified diastolic pressure point and dicrotic notch are corrected.

4. The method according to claim 1, characterized in that, The hemodynamic estimation model processes the standardized single-cardiac cycle blood pressure waveform in the following manner: The standardized single-cardiac cycle blood pressure waveform is divided into multiple non-overlapping time segments; Map each time segment to a feature vector; Add positional information representing the temporal order to each feature vector to form a sequence of feature vectors containing temporal relationships; Based on the feature vector sequence containing temporal relationships, waveform depth features are extracted.

5. The method according to claim 4, characterized in that, The deep neural network module includes a first output head and a second output head, and the method further includes: The individualized information is mapped into an individual feature vector; The individual feature vectors are concatenated with the waveform depth features to form a fused feature; The fusion features are input into the first output head and the second output head respectively; The reconstructed predicted blood pressure waveform is output through the first output head. The second output head outputs the predicted parameter values ​​of the ternary Windkessel model, the parameters of which include at least arterial compliance, peripheral resistance, and characteristic impedance.

6. The method according to claim 5, characterized in that, The method further includes: Calculate the diastolic pressure-time integral based on the predicted blood pressure waveform output by the first output head; Based on the characteristic impedance output from the second output head and the diastolic pressure-time integral, cardiac output and stroke volume are calculated.

7. The method according to claim 5, characterized in that, The method further includes: Substitute the predicted parameter values ​​output by the second output head into the differential equation of the ternary Windkessel model to calculate the theoretical blood pressure waveform; Determine the physical consistency loss between the theoretical blood pressure waveform output by the first output head and the predicted blood pressure waveform; The hemodynamic estimation model is trained based on the aforementioned physical consistency loss.

8. A hemodynamic parameter estimation device based on non-invasive blood pressure waveform, characterized in that, The device includes: The acquisition module is used to acquire non-invasive continuous arterial blood pressure waveforms and individualized information of the target object; A construction module is used to preprocess the non-invasive continuous arterial blood pressure waveform to construct a standardized single-cardiac cycle blood pressure waveform. The input module is used to input the standardized single-cardiac cycle blood pressure waveform and the individualized information into a pre-trained hemodynamic estimation model to obtain estimation results including at least cardiac output and stroke volume. The hemodynamic estimation model includes a deep neural network module. The hemodynamic estimation model is obtained by training the output predicted by the deep neural network module within the physiological relationships described by a ternary Windkessel model that includes arterial compliance, peripheral resistance, and characteristic impedance.

9. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic is that, when the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1-7.

Images