Noninvasive hemodynamics monitoring method and system based on deep learning

Abstract

The invention discloses a non-invasive hemodynamics monitoring method and system based on deep learning, and relates to the technical field of medical instruments, and the method comprises the steps: obtaining invasive blood flow data and non-invasive radial artery waveform data of a corresponding time point; time reference alignment is carried out on invasive blood flow data and non-invasive radial artery waveform data through time points, time drift errors between invasive and non-invasive equipment are eliminated, an initial non-invasive detection model is constructed, the method better fits the actual state of non-invasive detection, non-invasive feature vectors are used as input features after vascular elasticity compensation processing, and the non-invasive detection accuracy is improved. The initial non-invasive detection model is trained by taking blood flow parameters in the invasive data set as supervision annotation data, so that the consistency of model output and real invasive blood flow parameters is improved, 'waveform fragment-blood flow parameter 'space-time mapping is established, the running dynamic state of the non-invasive detection model is evaluated and analyzed, and the detection accuracy is improved. The performance change of the model in the training and running process can be monitored in real time, and convergence, over-fitting and under-fitting problems can be found in time.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and in particular to a non-invasive hemodynamic monitoring method and system based on deep learning. Background Technology

[0002] In the field of medical diagnosis and monitoring, the accurate acquisition of blood pressure and hemodynamic parameters is crucial for assessing an individual's physiological state and health status. Precise, continuous, and non-invasive monitoring of hemodynamic parameters is key to assessing an individual's physiological state, diagnosing cardiovascular diseases, and guiding clinical treatment. As one of the core parameters, blood pressure contains rich information about the cardiovascular system through its dynamic changes. In recent years, blood pressure monitoring methods based on photoplethysmography (PPG) signals have received widespread attention. By analyzing the morphological characteristics of PPG signals, physiological information closely related to blood pressure fluctuations can be extracted.

[0003] Traditional cuff-based blood pressure measurement methods can only provide intermittent blood pressure values, failing to achieve continuous monitoring and unable to obtain more comprehensive hemodynamic parameters (such as cardiac output, peripheral resistance, and vascular compliance). Current non-invasive hemodynamic monitoring methods primarily rely on the radial artery flattening tension method for indirect measurement, which presents the following problems during the measurement process:

[0004] In data acquisition, individual differences in vascular elasticity lead to large deviations in measurement parameters. The sensor array design is not optimized enough, resulting in insufficient signal acquisition sensitivity and stability. When acquiring electrocardiogram and heart sound signals, it is impossible to effectively achieve synchronous monitoring and calibration with hemodynamic parameters, resulting in a lack of high-quality labeled data, affecting the accuracy of measurement results, and making it difficult to meet the clinical standards for invasive monitoring.

[0005] To address the aforementioned technical shortcomings, a solution is proposed. Summary of the Invention

[0006] The purpose of this invention is to correct non-invasive parameters with invasive parameters during non-invasive testing, thereby achieving non-invasive and accurate detection of clinically equivalent blood flow parameters.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a non-invasive hemodynamic monitoring method and system based on deep learning, comprising the following steps:

[0008] Step 1: Acquire invasive blood flow data and non-invasive radial artery waveform data at corresponding time points. Based on the time points, align the invasive blood flow data and non-invasive radial artery waveform data with a time reference and perform data preprocessing. Based on the processed non-invasive radial artery waveform data, extract data features related to invasive blood flow parameters to generate non-invasive feature vectors. Integrate the non-invasive feature vectors with the corresponding invasive blood flow parameters to construct invasive and non-invasive datasets.

[0009] Step 2: Based on the non-invasive dataset, construct an initial non-invasive detection model. After performing vascular elasticity compensation processing on the non-invasive feature vectors in the non-invasive dataset, use them as the input features of the initial non-invasive detection model. Use the blood flow parameters in the invasive dataset as supervised annotation data to train the initial non-invasive detection model and generate the trained non-invasive detection model.

[0010] Step 3: Based on the non-invasive detection model, a dynamic convergence detection mechanism is adopted, and combined with clinical error standards, the dynamic operation of the non-invasive detection model is evaluated and analyzed, dynamic test results of the non-invasive detection model are generated, standard judgments are made on the non-invasive detection model based on the dynamic test results, standard judgment instructions for the non-invasive detection model are generated, and the non-invasive detection model is operated.

[0011] Step 4: When performing detection based on the non-invasive detection model, acquire real-time non-invasive radial artery waveform data and input it into the non-invasive detection model to output clinically equivalent blood flow parameters.

[0012] Furthermore, the acquisition of invasive blood flow data and non-invasive radial artery waveform data at corresponding time points in step one specifically includes the following:

[0013] S100: A catheter is inserted via radial artery puncture through an arterial catheter, and a high-precision pressure sensor is connected to the end of the catheter. The pressure sensor acquires ABP waveform, cardiac output, and stroke volume in real time to form invasive blood flow data.

[0014] S101. Using a matrix pressure sensor array, identify the sensor channel with the largest signal amplitude and the highest signal-to-noise ratio in the array, extract the continuous tension waveform of the channel, form radial artery tension waveform data, and normalize the waveform amplitude.

[0015] S102. Through FPGA hardware-level signal alignment, the FPGA generates a synchronization trigger signal at the start of data acquisition, simultaneously initiating the acquisition of invasive and non-invasive data, and synchronizing the time of invasive ABP waveform, cardiac output, stroke volume and non-invasive radial artery tension waveform data.

[0016] Furthermore, after vascular elasticity compensation processing, the input features used as the initial non-invasive detection model specifically include the following:

[0017] S200. Based on the non-invasive radial artery pressure waveform and the invasive waveform, perform a vascular elasticity difference analysis and establish a vascular elasticity compensation model. The objective function is:

[0018]

[0019] In the above formula, VII is the compensation value. The rate of change of pressure over time. Let k1 be the rate of change of the tube cross-sectional area with pressure, k2 be the individualized dynamic elastic coefficient for the patient, and k2 be the individualized static elastic coefficient for the patient. The values ​​of the two are [0.1, 0.9] and [0.01, 0.1], respectively. ΔT is the pulse wave conduction time or the delay time in signal processing.

[0020] S201. The non-invasive radial artery pressure waveform is preprocessed using a vascular elasticity compensation model. The vascular elasticity compensation value is superimposed on the non-invasive radial artery pressure waveform to obtain the preprocessed non-invasive radial artery pressure waveform.

[0021] S202. Use the preprocessed non-invasive radial artery pressure waveform as the input feature of the initial non-invasive detection model.

[0022] Furthermore, the trained non-invasive detection model is generated, specifically including the following:

[0023] S300, Input Layer: Obtain the preprocessed non-invasive radial artery pressure waveform and use it as the input feature of the initial non-invasive detection model;

[0024] S301, Hidden Layer: Based on a convolutional neural network, the non-invasive feature vector received from the input layer after vascular elasticity compensation is processed to extract temporal features. The blood flow parameters in the invasive dataset are used as supervised annotation data to establish a mapping from the non-invasive feature vector to the invasive blood flow data. The initial non-invasive detection model is trained. Based on the linear activation function, the non-invasive feature vector and the invasive blood flow data are fused to generate a high-dimensional feature vector.

[0025] S302, Output Layer: Set the number of neurons in the output layer. Each neuron corresponds to a prediction parameter. Based on a high-dimensional feature vector, output the target blood flow prediction parameters.

[0026] Furthermore, step three also includes establishing a clinical comparison database, specifically comprising the following steps:

[0027] S400: Acquire raw data from the non-invasive blood flow monitoring device, preprocess and select data based on clinical reference standard data, and obtain a dataset for comparison.

[0028] S401. By analyzing the dataset, define the acceptable range of difference between non-invasive testing results and reference standards;

[0029] S402. The analytical method is based on absolute error, relative error, and correlation analysis to evaluate non-invasive testing and reference standards, and to set clinical error standards.

[0030] Furthermore, the evaluation and analysis of the operational dynamics of the non-invasive detection model specifically includes the following:

[0031] During the training of the S500 non-invasive detection model, a loss function for the non-invasive detection model is defined. Based on the defined loss function L, the loss between the non-invasive detection prediction value and the invasive labeled data is calculated.

[0032] The loss function L is:

[0033] In the above formula, L is the loss value. λ represents the prediction parameters of the non-invasive detection model, Y represents the mapped invasive blood flow data, and λ is the penalty factor with a value of [0.1, 10].

[0034] S501. Using the backpropagation algorithm, the loss value is propagated back along the network starting from the loss function. The gradient of each weight parameter with respect to the loss is calculated. The weight parameters are updated using the stochastic gradient descent of the optimizer. Multiple rounds of iterative training are performed on the training data until the loss function converges to a smaller value.

[0035] S502. The dynamic convergence detection mechanism judges the trend of the loss function. If the loss function continues to decrease and tends to stabilize, and belongs to the clinical error standard, then the non-invasive detection model is determined to have converged. The dynamic test result of the non-invasive detection model is generated as converged, and the standard instructions of the non-invasive detection model are generated. Then the current non-invasive detection model construction is completed. Otherwise, it has not converged, and the non-standard instructions of the non-invasive detection model are generated. Then the non-invasive detection model structure needs to be adjusted and the non-invasive detection model needs to be retrained.

[0036] Furthermore, the output of clinically equivalent blood flow parameters in step four includes the following steps:

[0037] S600: Through a non-invasive matrix pressure sensor, it collects radial artery pulse wave signals in real time, extracts relevant non-invasive feature vectors, performs vascular elasticity compensation, generates processed feature vectors, and inputs them into the input layer of the non-invasive detection model.

[0038] S601. Non-invasive feature vector mapping analysis and comparison are performed through the hidden layer of the non-invasive detection model to output target blood flow prediction parameters;

[0039] S602. A dynamic calibration mechanism is adopted. The calibration process is automatically triggered when the first test is started each day. The updated vascular elasticity parameters k1 and k2 are sent to the vascular elasticity compensation model for calibration.

[0040] S603 also includes a dynamic channel selection strategy, which involves simultaneously acquiring radial artery waveforms from multiple sensor channels, calculating the SNR of each channel and applying a specified standard threshold; calculating the difference between the maximum and minimum SNR of all channels, and if the difference is greater than the standard threshold, then the signal quality difference between the channels is considered significant and fusion processing is required; otherwise, the channel with the largest SNR is selected.

[0041] Weighted fusion strategy: Dynamically allocate weights based on the SNR values ​​of each channel, perform weighted averaging on the multi-channel signals, and generate the final output waveform;

[0042] Single-channel selection strategy: directly select the channel with the highest SNR as the output.

[0043] A non-invasive hemodynamic monitoring system based on deep learning includes a data acquisition module, a model building module, a model detection module, and a real-time detection module;

[0044] The data acquisition module acquires invasive blood flow data and non-invasive radial artery waveform data at corresponding time points. Based on the time points, it aligns the invasive blood flow data and non-invasive radial artery waveform data with a time reference and performs data preprocessing. Based on the processed non-invasive radial artery waveform data, it extracts data features related to invasive blood flow parameters and generates non-invasive feature vectors. After integrating the non-invasive feature vectors with the corresponding invasive blood flow parameters, it constructs invasive and non-invasive datasets.

[0045] The model building module constructs an initial non-invasive detection model based on the non-invasive dataset. The non-invasive feature vectors in the non-invasive dataset are processed by vascular elasticity compensation and used as the input features of the initial non-invasive detection model. Blood flow parameters in the invasive dataset are used as supervised annotation data to train the initial non-invasive detection model and generate a trained non-invasive detection model.

[0046] The model testing module, based on the non-invasive testing model, adopts a dynamic convergence testing mechanism and combines clinical error standards to evaluate and analyze the operational dynamics of the non-invasive testing model, generate dynamic test results of the non-invasive testing model, make standard judgments on the non-invasive testing model based on the dynamic test results, generate standard judgment instructions for the non-invasive testing model, and perform non-invasive testing model operation.

[0047] The real-time detection module acquires real-time non-invasive radial artery waveform data when performing detection based on the non-invasive detection model, and inputs it into the non-invasive detection model to output clinically equivalent blood flow parameters.

[0048] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0049] This deep learning-based non-invasive hemodynamic monitoring method and system aligns invasive blood flow data and non-invasive radial artery waveform data at specific time points to eliminate time drift errors between invasive and non-invasive devices. An initial non-invasive detection model is constructed based on the non-invasive dataset, better reflecting the actual state of non-invasive detection. The non-invasive feature vectors are processed with vascular elasticity compensation before being used as input features. The influence of physiological factors on blood flow parameters is considered, improving the completeness and accuracy of the input data. Blood flow parameters from the invasive dataset are used as supervised annotation data to train the initial non-invasive detection model, enabling the model to learn under the guidance of known accurate results. By continuously adjusting the model parameters, the consistency between the model output and real invasive blood flow parameters is improved. A spatiotemporal mapping of "waveform segment-blood flow parameter" is established, and the dynamic operation of the non-invasive detection model is evaluated and analyzed. This allows for real-time monitoring of model performance changes during training and operation, timely detection of convergence, overfitting, and underfitting issues, and improved clinical reliability of non-invasive detection results. Attached Figure Description

[0050] Figure 1 A schematic diagram of the process flow of the method steps of the present invention is shown;

[0051] Figure 2 A schematic diagram of the overall system structure of the present invention is shown. Detailed Implementation

[0052] 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.

[0053] Example 1:

[0054] like Figure 1 As shown, a non-invasive hemodynamic monitoring method based on deep learning includes the following steps:

[0055] Step 1: Acquire invasive blood flow data and non-invasive radial artery waveform data at corresponding time points. Based on the time points, align the invasive blood flow data and non-invasive radial artery waveform data with a time reference and perform data preprocessing. Based on the processed non-invasive radial artery waveform data, extract data features related to invasive blood flow parameters to generate non-invasive feature vectors. Integrate the non-invasive feature vectors with the corresponding invasive blood flow parameters to construct invasive and non-invasive datasets.

[0056] Step 2: Based on the non-invasive dataset, construct an initial non-invasive detection model. After performing vascular elasticity compensation processing on the non-invasive feature vectors in the non-invasive dataset, use them as the input features of the initial non-invasive detection model. Use the blood flow parameters in the invasive dataset as supervised annotation data to train the initial non-invasive detection model and generate the trained non-invasive detection model.

[0057] Step 3: Based on the non-invasive detection model, a dynamic convergence detection mechanism is adopted, and combined with clinical error standards, the dynamic operation of the non-invasive detection model is evaluated and analyzed, dynamic test results of the non-invasive detection model are generated, standard judgments are made on the non-invasive detection model based on the dynamic test results, standard judgment instructions for the non-invasive detection model are generated, and the non-invasive detection model is operated.

[0058] Step 4: When performing detection based on the non-invasive detection model, acquire real-time non-invasive radial artery waveform data and input it into the non-invasive detection model to output clinically equivalent blood flow parameters.

[0059] Step one, which involves acquiring invasive blood flow data and corresponding non-invasive radial artery waveform data at the same time point, specifically includes the following:

[0060] S100: A catheter is inserted via radial artery puncture through an arterial catheter, and a high-precision pressure sensor is connected to the end of the catheter. The pressure sensor acquires ABP waveform, cardiac output, and stroke volume in real time to form invasive blood flow data.

[0061] S101. Using a matrix pressure sensor array, identify the sensor channel with the largest signal amplitude and the highest signal-to-noise ratio in the array, extract the continuous tension waveform of the channel, form radial artery tension waveform data, and normalize the waveform amplitude.

[0062] S102. Through FPGA hardware-level signal alignment, the FPGA generates a synchronization trigger signal at the start of data acquisition, and simultaneously starts the acquisition of invasive and non-invasive data. The time synchronization of invasive ABP waveform, cardiac output, stroke volume and non-invasive radial artery tension waveform data is controlled within 10ms.

[0063] After vascular elasticity compensation processing, the input features used as the initial non-invasive detection model include the following:

[0064] S200. Based on the non-invasive radial artery pressure waveform and the invasive waveform, perform a vascular elasticity difference analysis and establish a vascular elasticity compensation model. The objective function is:

[0065]

[0066] In the above formula, VII is the compensation value. The rate of change of pressure over time. Let k1 be the rate of change of the tube cross-sectional area with pressure, k2 be the individualized dynamic elastic coefficient of the patient, and k2 be the individualized static elastic coefficient of the patient. The values ​​of the two are [0.1, 0.9] and [0.01, 0.1], respectively. ΔT is the pulse wave conduction time or the delay time in signal processing.

[0067] S201. The non-invasive radial artery pressure waveform is preprocessed using a vascular elasticity compensation model. The vascular elasticity compensation value is superimposed on the non-invasive radial artery pressure waveform to obtain the preprocessed non-invasive radial artery pressure waveform.

[0068] S202. Use the preprocessed non-invasive radial artery pressure waveform as the input feature of the initial non-invasive detection model.

[0069] The trained non-invasive detection model is generated, specifically including the following:

[0070] S300, Input Layer: Obtain the preprocessed non-invasive radial artery pressure waveform and use it as the input feature of the initial non-invasive detection model;

[0071] S301, Hidden Layer: Based on a convolutional neural network, the non-invasive feature vector received from the input layer after vascular elasticity compensation is processed to extract temporal features. The blood flow parameters in the invasive dataset are used as supervised annotation data to establish a mapping from the non-invasive feature vector to the invasive blood flow data. The initial non-invasive detection model is trained. Based on the linear activation function, the non-invasive feature vector and the invasive blood flow data are fused to generate a high-dimensional feature vector.

[0072] S302, Output Layer: Set the number of neurons in the output layer. Each neuron corresponds to a prediction parameter. Based on a high-dimensional feature vector, output the target blood flow prediction parameters.

[0073] Step three also includes establishing a clinical comparison database, which specifically includes the following steps:

[0074] S400: Acquire raw data from the non-invasive blood flow monitoring device, preprocess and select data based on clinical reference standard data, and obtain a dataset for comparison.

[0075] S401. By analyzing the dataset, define the acceptable range of difference between non-invasive testing results and reference standards;

[0076] S402. The analytical method is based on absolute error, relative error, and correlation analysis to evaluate non-invasive testing and reference standards, and to set clinical error standards.

[0077] The evaluation and analysis of the operational dynamics of the non-invasive detection model specifically includes the following:

[0078] During the training of the S500 non-invasive detection model, a loss function for the non-invasive detection model is defined. Based on the defined loss function L, the loss between the non-invasive detection prediction value and the invasive labeled data is calculated.

[0079] The loss function L is:

[0080] In the above formula, L is the loss value. λ represents the prediction parameters of the non-invasive detection model, Y represents the mapped invasive blood flow data, and λ is the penalty factor with a value of [0.1, 10].

[0081] S501. Using the backpropagation algorithm, the loss value is propagated back along the network starting from the loss function. The learning rate is set to [0.001, 0.01]. The gradient of each weight parameter with respect to the loss is calculated. The weight parameters are updated using the stochastic gradient descent of the optimizer. Multiple rounds of iterative training are performed on the training data until the loss function converges to a smaller value.

[0082] S502. The dynamic convergence detection mechanism judges the trend of the loss function. If the loss function continues to decrease and tends to stabilize, and belongs to the clinical error standard, then the non-invasive detection model is determined to have converged. The dynamic test result of the non-invasive detection model is generated as converged, and the standard instructions of the non-invasive detection model are generated. Then the current non-invasive detection model construction is completed. Otherwise, it has not converged, and the non-standard instructions of the non-invasive detection model are generated. Then the non-invasive detection model structure needs to be adjusted and the non-invasive detection model needs to be retrained.

[0083] Step four, which involves outputting clinically equivalent blood flow parameters, includes the following steps:

[0084] S600: Through a non-invasive matrix pressure sensor, it collects radial artery pulse wave signals in real time, extracts relevant non-invasive feature vectors, performs vascular elasticity compensation, generates processed feature vectors, and inputs them into the input layer of the non-invasive detection model.

[0085] S601. Non-invasive feature vector mapping analysis and comparison are performed through the hidden layer of the non-invasive detection model to output target blood flow prediction parameters;

[0086] S602. A dynamic calibration mechanism is adopted. The calibration process is automatically triggered when the first test is started each day. The updated vascular elasticity parameters k1 and k2 are sent to the vascular elasticity compensation model for calibration.

[0087] S603 also includes a dynamic channel selection strategy, which involves simultaneously acquiring radial artery waveforms from multiple sensor channels, calculating the SNR of each channel and applying a specified standard threshold; calculating the difference between the maximum and minimum SNR of all channels, and if the difference is greater than the standard threshold, then the signal quality difference between the channels is considered significant and fusion processing is required; otherwise, the channel with the largest SNR is selected.

[0088] Weighted fusion strategy: Dynamically allocate weights based on the SNR values ​​of each channel, perform weighted averaging on the multi-channel signals, and generate the final output waveform;

[0089] Single-channel selection strategy: directly select the channel with the highest SNR as the output.

[0090] Example 2:

[0091] like Figure 2 As shown, a non-invasive hemodynamic monitoring system based on deep learning includes a data acquisition module, a model building module, a model detection module, and a real-time detection module.

[0092] The data acquisition module acquires invasive blood flow data and non-invasive radial artery waveform data at corresponding time points. Based on the time points, it aligns the invasive blood flow data and non-invasive radial artery waveform data with a time reference and performs data preprocessing. Based on the processed non-invasive radial artery waveform data, it extracts data features related to invasive blood flow parameters and generates non-invasive feature vectors. After integrating the non-invasive feature vectors with the corresponding invasive blood flow parameters, it constructs invasive and non-invasive datasets.

[0093] The model building module constructs an initial non-invasive detection model based on the non-invasive dataset. The non-invasive feature vectors in the non-invasive dataset are processed by vascular elasticity compensation and used as the input features of the initial non-invasive detection model. Blood flow parameters in the invasive dataset are used as supervised annotation data to train the initial non-invasive detection model and generate a trained non-invasive detection model.

[0094] The model testing module, based on the non-invasive testing model, adopts a dynamic convergence testing mechanism and combines clinical error standards to evaluate and analyze the operational dynamics of the non-invasive testing model, generate dynamic test results of the non-invasive testing model, make standard judgments on the non-invasive testing model based on the dynamic test results, generate standard judgment instructions for the non-invasive testing model, and perform non-invasive testing model operation.

[0095] The real-time detection module acquires real-time non-invasive radial artery waveform data when performing detection based on the non-invasive detection model, and inputs it into the non-invasive detection model to output clinically equivalent blood flow parameters.

[0096] Acquire invasive blood flow data and corresponding non-invasive radial artery waveform data at the following time points:

[0097] S100: A catheter is inserted via radial artery puncture through an arterial catheter, and a high-precision pressure sensor is connected to the end of the catheter. The pressure sensor acquires ABP waveform, cardiac output, and stroke volume in real time to form invasive blood flow data.

[0098] S101. Using a matrix pressure sensor array, identify the sensor channel with the largest signal amplitude and the highest signal-to-noise ratio in the array, extract the continuous tension waveform of the channel, form radial artery tension waveform data, and normalize the waveform amplitude.

[0099] S102. Through FPGA hardware-level signal alignment, the FPGA generates a synchronization trigger signal at the start of data acquisition, and simultaneously starts the acquisition of invasive and non-invasive data. The time synchronization of invasive ABP waveform, cardiac output, stroke volume and non-invasive radial artery tension waveform data is controlled within 5ms.

[0100] After vascular elasticity compensation processing, the input features used as the initial non-invasive detection model include the following:

[0101] S200. Based on the non-invasive radial artery pressure waveform and the invasive waveform, perform a vascular elasticity difference analysis and establish a vascular elasticity compensation model. The objective function is:

[0102]

[0103] In the above formula, VII is the compensation value. The rate of change of pressure over time. Let k1 be the rate of change of the tube cross-sectional area with pressure, k2 be the individualized dynamic elastic coefficient for the patient, and k2 be the individualized static elastic coefficient for the patient. The values ​​of the two are [0.1, 0.9] and [0.01, 0.1], respectively. ΔT is the pulse wave conduction time or the delay time in signal processing.

[0104] S201. The non-invasive radial artery pressure waveform is preprocessed using a vascular elasticity compensation model. The vascular elasticity compensation value is superimposed on the non-invasive radial artery pressure waveform to obtain the preprocessed non-invasive radial artery pressure waveform.

[0105] S202. Use the preprocessed non-invasive radial artery pressure waveform as the input feature of the initial non-invasive detection model.

[0106] The trained non-invasive detection model is generated, specifically including the following:

[0107] S300, Input Layer: Obtain the preprocessed non-invasive radial artery pressure waveform and use it as the input feature of the initial non-invasive detection model;

[0108] S301, Hidden Layer: Based on a convolutional neural network, the non-invasive feature vector received from the input layer after vascular elasticity compensation is processed to extract temporal features. The blood flow parameters in the invasive dataset are used as supervised annotation data to establish a mapping from the non-invasive feature vector to the invasive blood flow data. The initial non-invasive detection model is trained. Based on the linear activation function, the non-invasive feature vector and the invasive blood flow data are fused to generate a high-dimensional feature vector.

[0109] S302, Output Layer: Set the number of neurons in the output layer. Each neuron corresponds to a prediction parameter. Based on a high-dimensional feature vector, output the target blood flow prediction parameters.

[0110] This also includes establishing a clinical comparative database, specifically including the following steps:

[0111] S400: Acquire raw data from the non-invasive blood flow monitoring device, preprocess and select data based on clinical reference standard data, and obtain a dataset for comparison.

[0112] S401. By analyzing the dataset, define the acceptable range of difference between non-invasive testing results and reference standards;

[0113] S402. The analytical method is based on absolute error, relative error, and correlation analysis to evaluate non-invasive testing and reference standards, and to set clinical error standards.

[0114] The evaluation and analysis of the operational dynamics of the non-invasive detection model specifically includes the following:

[0115] During the training of the S500 non-invasive detection model, a loss function for the non-invasive detection model is defined. Based on the defined loss function L, the loss between the non-invasive detection prediction value and the invasive labeled data is calculated.

[0116] The loss function L is:

[0117] In the above formula, L is the loss value. Here are the prediction parameters for the non-invasive detection model, Y represents the mapped invasive blood flow data, and λ is the penalty factor with a value of [0.5, 5].

[0118] S501. Using the backpropagation algorithm, the loss value is propagated back along the network starting from the loss function. The learning rate is set to [0.005, 0.05]. The gradient of each weight parameter with respect to the loss is calculated. The weight parameters are updated using the stochastic gradient descent of the optimizer. Multiple rounds of iterative training are performed on the training data until the loss function converges to a smaller value.

[0119] S502. The dynamic convergence detection mechanism judges the trend of the loss function. If the loss function continues to decrease and tends to stabilize, and belongs to the clinical error standard, then the non-invasive detection model is determined to have converged. The dynamic test result of the non-invasive detection model is generated as converged, and the standard instructions of the non-invasive detection model are generated. Then the current non-invasive detection model construction is completed. Otherwise, it has not converged, and the non-standard instructions of the non-invasive detection model are generated. Then the non-invasive detection model structure needs to be adjusted and the non-invasive detection model needs to be retrained.

[0120] Outputting clinically equivalent blood flow parameters includes the following steps:

[0121] S600: Through a non-invasive matrix pressure sensor, it collects radial artery pulse wave signals in real time, extracts relevant non-invasive feature vectors, performs vascular elasticity compensation, generates processed feature vectors, and inputs them into the input layer of the non-invasive detection model.

[0122] S601. Non-invasive feature vector mapping analysis and comparison are performed through the hidden layer of the non-invasive detection model to output target blood flow prediction parameters;

[0123] S602. A dynamic calibration mechanism is adopted. The calibration process is automatically triggered when the first test is started each day. The updated vascular elasticity parameters k1 and k2 are sent to the vascular elasticity compensation model for calibration.

[0124] S603 also includes a dynamic channel selection strategy, which involves simultaneously acquiring radial artery waveforms from multiple sensor channels, calculating the SNR of each channel and applying a specified standard threshold; calculating the difference between the maximum and minimum SNR of all channels, and if the difference is greater than the standard threshold, then the signal quality difference between the channels is considered significant and fusion processing is required; otherwise, the channel with the largest SNR is selected.

[0125] Weighted fusion strategy: Dynamically allocate weights based on the SNR values ​​of each channel, perform weighted averaging on the multi-channel signals, and generate the final output waveform;

[0126] Single-channel selection strategy: directly select the channel with the highest SNR as the output.

[0127] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0128] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0129] In the two embodiments provided in this application, it should be understood that the disclosed apparatus and system can be implemented in other ways; for example, the apparatus embodiments described above are merely illustrative, and the division of modules is merely a logical functional division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; furthermore, the coupling or direct coupling or communication connection between the shown or discussed mutuals can be through some interfaces, and the indirect coupling or communication connection between the apparatus or modules can be electrical, mechanical or other forms.

[0130] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A non-invasive hemodynamic monitoring method based on deep learning, characterized in that, Includes the following steps: Step 1: Acquire invasive blood flow data and non-invasive radial artery waveform data at corresponding time points. Based on the time points, align the invasive blood flow data and non-invasive radial artery waveform data with a time reference and perform data preprocessing. Based on the processed non-invasive radial artery waveform data, extract data features related to invasive blood flow parameters to generate non-invasive feature vectors. Integrate the non-invasive feature vectors with the corresponding invasive blood flow parameters to construct invasive and non-invasive datasets. Step 2: Based on the non-invasive dataset, construct an initial non-invasive detection model. After performing vascular elasticity compensation processing on the non-invasive feature vectors in the non-invasive dataset, use them as the input features of the initial non-invasive detection model. Use the blood flow parameters in the invasive dataset as supervised annotation data to train the initial non-invasive detection model and generate the trained non-invasive detection model. Step 3: Based on the non-invasive detection model, a dynamic convergence detection mechanism is adopted, and combined with clinical error standards, the dynamic operation of the non-invasive detection model is evaluated and analyzed, dynamic test results of the non-invasive detection model are generated, standard judgments are made on the non-invasive detection model based on the dynamic test results, standard judgment instructions for the non-invasive detection model are generated, and the non-invasive detection model is operated. Step 4: When performing detection based on the non-invasive detection model, acquire real-time non-invasive radial artery waveform data and input it into the non-invasive detection model to output clinically equivalent blood flow parameters.

2. The non-invasive hemodynamic monitoring method and system based on deep learning according to claim 1, characterized in that, Step one, which involves acquiring invasive blood flow data and corresponding non-invasive radial artery waveform data at the same time point, specifically includes the following: S100: A catheter is inserted via radial artery puncture through an arterial catheter, and a high-precision pressure sensor is connected to the end of the catheter. The pressure sensor acquires ABP waveform, cardiac output, and stroke volume in real time to form invasive blood flow data. S101. Using a matrix pressure sensor array, identify the sensor channel with the largest signal amplitude and the highest signal-to-noise ratio in the array, extract the continuous tension waveform of the channel, form radial artery tension waveform data, and normalize the waveform amplitude. S102. Through FPGA hardware-level signal alignment, the FPGA generates a synchronization trigger signal at the start of data acquisition, simultaneously initiating the acquisition of invasive and non-invasive data, and synchronizing the time of invasive ABP waveform, cardiac output, stroke volume and non-invasive radial artery tension waveform data.

3. The non-invasive hemodynamic monitoring method and system based on deep learning according to claim 1, characterized in that, After vascular elasticity compensation processing, the input features used as the initial non-invasive detection model include the following: S200. Based on the non-invasive radial artery pressure waveform and the invasive waveform, perform a vascular elasticity difference analysis and establish a vascular elasticity compensation model. The objective function is: In the above formula, VII is the compensation value. The rate of change of pressure over time. Let k1 be the rate of change of the tube cross-sectional area with pressure, k2 be the individualized dynamic elastic coefficient for the patient, and k2 be the individualized static elastic coefficient for the patient. The values ​​of the two are [0.1, 0.9] and [0.01, 0.1], respectively. ΔT is the pulse wave conduction time or the delay time in signal processing. S201. The non-invasive radial artery pressure waveform is preprocessed using a vascular elasticity compensation model. The vascular elasticity compensation value is superimposed on the non-invasive radial artery pressure waveform to obtain the preprocessed non-invasive radial artery pressure waveform. S202. Use the preprocessed non-invasive radial artery pressure waveform as the input feature of the initial non-invasive detection model.

4. The non-invasive hemodynamic monitoring method and system based on deep learning according to claim 1, characterized in that, The trained non-invasive detection model is generated, specifically including the following: S300, Input Layer: Obtain the preprocessed non-invasive radial artery pressure waveform and use it as the input feature of the initial non-invasive detection model; S301, Hidden Layer: Based on a convolutional neural network, the non-invasive feature vector received from the input layer after vascular elasticity compensation is processed to extract temporal features. The blood flow parameters in the invasive dataset are used as supervised annotation data to establish a mapping from the non-invasive feature vector to the invasive blood flow data. The initial non-invasive detection model is trained. Based on the linear activation function, the non-invasive feature vector and the invasive blood flow data are fused to generate a high-dimensional feature vector. S302, Output Layer: Set the number of neurons in the output layer. Each neuron corresponds to a prediction parameter. Based on a high-dimensional feature vector, output the target blood flow prediction parameters.

5. The non-invasive hemodynamic monitoring method and system based on deep learning according to claim 1, characterized in that, Step three also includes establishing a clinical comparison database, which specifically includes the following steps: S400: Acquire raw data from the non-invasive blood flow monitoring device, preprocess and select data based on clinical reference standard data, and obtain a dataset for comparison. S401. By analyzing the dataset, define the acceptable range of difference between non-invasive testing results and reference standards; S402. The analytical method is based on absolute error, relative error, and correlation analysis to evaluate non-invasive testing and reference standards, and to set clinical error standards.

6. The non-invasive hemodynamic monitoring method and system based on deep learning according to claim 1, characterized in that, The evaluation and analysis of the operational dynamics of the non-invasive detection model specifically includes the following: During the training of the S500 non-invasive detection model, a loss function for the non-invasive detection model is defined. Based on the defined loss function L, the loss between the non-invasive detection prediction value and the invasive labeled data is calculated. The loss function L is: In the above formula, L is the loss value. λ represents the prediction parameters of the non-invasive detection model, Y represents the mapped invasive blood flow data, and λ is the penalty factor with a value of [0.1, 10]. S501. Using the backpropagation algorithm, the loss value is propagated back along the network starting from the loss function. The gradient of each weight parameter with respect to the loss is calculated. The weight parameters are updated using the stochastic gradient descent of the optimizer. Multiple rounds of iterative training are performed on the training data until the loss function converges to a smaller value. S502. The dynamic convergence detection mechanism judges the trend of the loss function. If the loss function continues to decrease and tends to stabilize, and belongs to the clinical error standard, then the non-invasive detection model is determined to have converged. The dynamic test result of the non-invasive detection model is generated as converged, and the standard instructions of the non-invasive detection model are generated. Then the current non-invasive detection model construction is completed. Otherwise, it has not converged, and the non-standard instructions of the non-invasive detection model are generated. Then the non-invasive detection model structure needs to be adjusted and the non-invasive detection model needs to be retrained.

7. The non-invasive hemodynamic monitoring method based on deep learning according to claim 1, characterized in that, Step four, which involves outputting clinically equivalent blood flow parameters, includes the following steps: S600: Through a non-invasive matrix pressure sensor, it collects radial artery pulse wave signals in real time, extracts relevant non-invasive feature vectors, performs vascular elasticity compensation, generates processed feature vectors, and inputs them into the input layer of the non-invasive detection model. S601. Non-invasive feature vector mapping analysis and comparison are performed through the hidden layer of the non-invasive detection model to output target blood flow prediction parameters; S602. A dynamic calibration mechanism is adopted. The calibration process is automatically triggered when the first test is started each day. The updated vascular elasticity parameters k1 and k2 are sent to the vascular elasticity compensation model for calibration. S603 also includes a dynamic channel selection strategy, which involves simultaneously acquiring radial artery waveforms from multiple sensor channels, calculating the SNR of each channel and applying a specified standard threshold; calculating the difference between the maximum and minimum SNR of all channels, and if the difference is greater than the standard threshold, then the signal quality difference between the channels is considered significant and fusion processing is required; otherwise, the channel with the largest SNR is selected. Weighted fusion strategy: Dynamically allocate weights based on the SNR values ​​of each channel, perform weighted averaging on the multi-channel signals, and generate the final output waveform; Single-channel selection strategy: directly select the channel with the highest SNR as the output.

8. A non-invasive hemodynamic monitoring system based on deep learning, characterized in that, It includes a data acquisition module, a model building module, a model detection module, and a real-time detection module; The data acquisition module acquires invasive blood flow data and non-invasive radial artery waveform data at corresponding time points. Based on the time points, it aligns the invasive blood flow data and non-invasive radial artery waveform data with a time reference and performs data preprocessing. Based on the processed non-invasive radial artery waveform data, it extracts data features related to invasive blood flow parameters and generates non-invasive feature vectors. After integrating the non-invasive feature vectors with the corresponding invasive blood flow parameters, it constructs invasive and non-invasive datasets. The model building module constructs an initial non-invasive detection model based on the non-invasive dataset. The non-invasive feature vectors in the non-invasive dataset are processed by vascular elasticity compensation and used as the input features of the initial non-invasive detection model. Blood flow parameters in the invasive dataset are used as supervised annotation data to train the initial non-invasive detection model and generate a trained non-invasive detection model. The model testing module, based on the non-invasive testing model, adopts a dynamic convergence testing mechanism and combines clinical error standards to evaluate and analyze the operational dynamics of the non-invasive testing model, generate dynamic test results of the non-invasive testing model, make standard judgments on the non-invasive testing model based on the dynamic test results, generate standard judgment instructions for the non-invasive testing model, and perform non-invasive testing model operation. The real-time detection module acquires real-time non-invasive radial artery waveform data when performing detection based on the non-invasive detection model, and inputs it into the non-invasive detection model to output clinically equivalent blood flow parameters.

Images