Method and device for monitoring cerebral blood flow autoregulation function and training model

Abstract

The present application discloses a method and device for monitoring cerebral blood flow autoregulation function and training model, belonging to the field of intelligent medical technology. The method includes: based on the arterial blood pressure signal and cerebral blood flow signal of the target individual, determining the first transfer function analysis TFA variable used to represent the current state of the cerebral blood flow autoregulation CA function of the target individual, and the first TFA variable includes frequency domain features and time domain features. Based on the frequency domain features and time domain features, through the target multi-scale fusion perception model, determine whether the CA function of the target individual is damaged. The target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and time domain features corresponding to the CA function of any individual. The present application can improve the accuracy of the detection results.

Description

Technical Field

[0001] The present application relates to the field of intelligent medical technology, and in particular to a method and device for monitoring the automatic regulation function of cerebral blood flow and training models. Background Art

[0002] Cerebral autoregulation (CA) function is an important hemodynamic physiological protection mechanism that can maintain the cerebral blood flow velocity (CBFV) at a relatively stable level when the arterial blood pressure (ABP) fluctuates. Since impaired CA function increases the risk of secondary brain injury, there is usually a need to monitor CA function to determine whether it is damaged.

[0003] In the related art, the mean of the transfer function analysis (TFA) variables (generally including phase, gain, and correlation information) within each fixed frequency band is generally determined based on the ABP signal and CBFV signal corresponding to the CA function, so as to characterize the state of the CA function through the mean of the TFA variables within each fixed frequency band. Then, the mean of the TFA variables within each fixed frequency band is fitted with a logistic regression model to determine the difference of the TFA variables, so as to monitor the CA function based on the difference of the TFA variables.

[0004] However, due to the lack of complexity of the logistic regression model, it is prone to underfitting, resulting in inaccurate monitoring results. Summary of the invention

[0005] The embodiment of the present application provides a method and device for monitoring the automatic regulation function of cerebral blood flow and training a model, which can solve the problem of inaccurate monitoring results in related technologies. The technical solution is as follows:

[0006] In a first aspect, a method for monitoring cerebral blood flow autoregulation function is provided, the method comprising:

[0007] Based on the arterial blood pressure signal and the cerebral blood flow signal of the target individual, determining a first transfer function analysis (TFA) variable for representing the current state of the cerebral blood flow autoregulation (CA) function of the target individual, wherein the first TFA variable includes a frequency domain feature and a time domain feature;

[0008] Based on the frequency domain features and the time domain features, whether the CA function of the target individual is damaged is determined through a target multi-scale fusion perception model. The target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and time domain features corresponding to the CA function of any individual.

[0009] In a second aspect, a method for training a model is provided, the method comprising:

[0010] Based on a first sample data set, iteratively training a first frequency domain module and a first time domain module respectively, wherein the first sample data in the first sample data set includes a first sample frequency domain feature and a first sample time domain feature corresponding to an unlabeled sample signal, and the first frequency domain module and the first time domain module are independent modules and both have a feature extraction function;

[0011] Constructing a second frequency domain module according to the iteratively trained first frequency domain module, and constructing a second time domain module according to the iteratively trained first time domain module, wherein both the second frequency domain module and the second time domain module have feature extraction and classification functions;

[0012] Determine a target multi-scale fusion perception model based on a second sample data set, the second frequency domain module, and the second time domain module, wherein the second sample data in the second sample data set includes second sample frequency domain features and second sample time domain features corresponding to the labeled sample signal;

[0013] Among them, the target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain characteristics and time domain characteristics corresponding to the CA function of any individual.

[0014] In a third aspect, a device for detecting the automatic regulation function of cerebral blood flow is provided, the device comprising:

[0015] A first determination module is used to determine a first transfer function analysis TFA variable for representing a current state of a cerebral blood flow autoregulation (CA) function of the target individual based on an arterial blood pressure signal and a cerebral blood flow signal of the target individual, wherein the first TFA variable includes a frequency domain feature and a time domain feature;

[0016] The second determination module is used to determine whether the CA function of the target individual is damaged based on the frequency domain features and the time domain features through a target multi-scale fusion perception model. The target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and the time domain features corresponding to the CA function of any individual.

[0017] In a fourth aspect, a device for training a model is provided, the device comprising:

[0018] A first training module, used to iteratively train a first frequency domain module and a first time domain module based on a first sample data set, wherein the first sample data in the first sample data set includes a first sample frequency domain feature and a first sample time domain feature corresponding to an unlabeled sample signal, and the first frequency domain module and the first time domain module are independent modules and both have a feature extraction function;

[0019] A construction module, used to construct a second frequency domain module according to the iteratively trained first frequency domain module, and to construct a second time domain module according to the iteratively trained first time domain module, wherein both the second frequency domain module and the second time domain module have a feature extraction function and a classification function;

[0020] A second training module is used to determine a target multi-scale fusion perception model based on a second sample data set, the second frequency domain module, and the second time domain module, wherein the second sample data in the second sample data set is a second sample frequency domain feature and a second sample time domain feature corresponding to the labeled sample signal;

[0021] Among them, the target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain characteristics and time domain characteristics corresponding to the CA function of any individual.

[0022] In a fifth aspect, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the first aspect or the second aspect when executing the computer program.

[0023] In a sixth aspect, a computer-readable storage medium is provided, on which instructions are stored. When the instructions are executed by a processor, the method described in the first aspect or the second aspect is implemented.

[0024] In a seventh aspect, a computer program product comprising instructions is provided, which, when executed on a computer, enables the computer to execute the method described in the first or second aspect above.

[0025] The beneficial effects of the technical solution provided by the embodiment of the present application are:

[0026] Based on the first sample data set, the first frequency domain module and the first time domain module are iteratively trained respectively. The first sample data in the first sample data set includes the first sample frequency domain features and the first sample time domain features corresponding to the unlabeled sample signal. The first frequency domain module and the first time domain module are independent modules, and both have feature extraction functions. The second frequency domain module is constructed according to the first frequency domain module after iterative training, and the second time domain module is constructed according to the first time domain module after iterative training. Based on the second sample data set, the second frequency domain module, and the second time domain module, the target multi-scale fusion perception model is determined. The second sample data in the second sample data set is the second sample frequency domain features and the second sample time domain features corresponding to the labeled sample signal. The obtained target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and time domain features corresponding to the CA function of any individual, that is, the CA function of the target individual can be monitored by the target multi-scale fusion perception model. Compared with the logistic regression model, the target multi-scale fusion perception model is more robust and can improve the accuracy of the detection result. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 is a flow chart of a method for training a model according to an exemplary embodiment;

[0029] Figure 2 is a flow chart of a method for training a model according to another exemplary embodiment;

[0030] Figure 3 is a schematic diagram of a first frequency domain module according to an exemplary embodiment;

[0031] Figure 4 is a schematic diagram of a first time domain module according to an exemplary embodiment;

[0032] Figure 5 is a schematic diagram showing a second frequency domain module and a second time domain module according to an exemplary embodiment;

[0033] Figure 6 is a schematic diagram of a multi-scale fusion perception model according to an exemplary embodiment;

[0034] Figure 7is a schematic diagram showing a comparison of a ROC curve and an AUC value according to an exemplary embodiment;

[0035] Figure 8 is a flow chart of a method for monitoring cerebral blood flow autoregulation function according to an exemplary embodiment;

[0036] Fig. 9 is a schematic structural diagram of a device for monitoring cerebral blood flow autoregulation function according to an exemplary embodiment;

[0037] Fig.10 is a schematic structural diagram of a device for training a model according to an exemplary embodiment;

[0038] Fig.11 The diagram is a schematic structural diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0039] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0040] It should be understood that the "multiple" mentioned in this application refers to two or more. In the description of this application, unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in order to facilitate the clear description of the technical solution of this application, the words "first" and "second" are used to distinguish between the same items or similar items with basically the same functions and effects. Those skilled in the art can understand that the words "first" and "second" do not limit the quantity and execution order, and the words "first" and "second" do not limit them to be different.

[0041] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0042] Before introducing the method provided in the embodiment of the present application in detail, the execution subject involved in the embodiment of the present application is introduced. The method provided in the embodiment of the present application can be executed by an electronic device. In one example, the electronic device can be a terminal such as a laptop, a desktop, a portable computer, a tablet computer, etc., which is not limited in the embodiment of the present application.

[0043] Next, the application scenarios involved in the embodiments of the present application are introduced.

[0044] In the medical field, CA function refers to the intrinsic ability of the brain to maintain adequate cerebral perfusion when arterial blood pressure changes. In patients with severe hypertension, brain damage, or undergoing surgery, CA function plays an important role in maintaining stable cerebral perfusion. However, it is currently known that a variety of diseases can damage CA function, and impaired CA function increases the risk of secondary brain damage. Therefore, monitoring CA function has become a clinical need to provide patients with useful clinical information. To this end, the embodiments of the present application provide a CA function monitoring method based on semi-supervised learning, which can monitor in real time and accurately whether an individual's CA function is impaired.

[0045] In an embodiment of the present application, a target multi-scale fusion perception model is used to predict whether the CA function is impaired. Among them, the target multi-scale fusion perception model is obtained through training. In order to realize the training of the target multi-scale fusion perception model, sample data needs to be collected. In one example, 61 healthy individuals, 61 individuals with severe symptoms, and 2247 individuals with mild to severe stenosis and no symptoms can be included. Among them, healthy individuals are collected from healthy people with no previous history of cardiovascular and cerebrovascular diseases and no relevant risk factors. Healthy individuals are judged to have unimpaired CA function; individuals with severe symptoms refer to subjects with severe intracranial vascular stenosis and stroke symptoms. Individuals with severe symptoms are identified as impaired CA function; the other 2247 cases refer to individuals who cannot be judged whether they are CA impaired.

[0046] During the collection of sample data, each included individual may be required to rest in a supine position with the head raised for at least 10 minutes. Professionals (such as medical staff) use a 2MHz transcranial Doppler probe head frame to continuously record the left and right CBFV of the bilateral middle cerebral arteries at a depth of 45-55mm distal to the M1 stenosis. At the same time, vascular unloading technology is used to non-invasively measure ABP from the finger artery. Healthy individuals only retain unilateral CBFV, individuals with severe symptoms retain the diseased CBFV, and individuals with mild to severe symptoms retain bilateral CBFV. In one example, in order to ensure the stability of the sample data, the CBFV signal and ABP signal collected synchronously after 10 minutes can be retained. It is not difficult to understand that for individuals who have been diagnosed, the sample data collected is labeled, for example, the label is used to indicate that the CA function is impaired or the CA function is not impaired, and the collected sample data can be called a labeled sample signal; while for individuals who cannot be diagnosed, the sample data collected is unlabeled, and the collected sample data can be called an unlabeled sample signal. In this way, a total of 122 groups of labeled sample signals (including ABP signals and CBFV signals of labeled individuals) and 4494 groups of unlabeled sample signals (including ABP signals and CBFV signals of unlabeled individuals) can be collected.

[0047] Afterwards, the collected sample data can be post-processed using Matlab software. For each set of sample data, two 90-second windows are slid simultaneously with 10-second increments of the ABP signal and the CBFV signal to obtain a frequency segment with a resolution of 1 / 90 Hz, that is, each window includes a 90-second collected signal. Then, based on the sample data obtained after post-processing, a sample data set for training is determined by the TFA method, including a first sample data set and a second sample data set, wherein the first sample data set includes a plurality of first sample data determined based on the unlabeled sample signal, and the second sample data set includes a plurality of second sample data determined based on the labeled sample signal.

[0048] In implementation, the TFA method is applied to these frequency segments to determine sample data (which may be referred to as TFA variables). For all data, the average value of the TFA variables extracted from all frequency segments represents the state of individual CA. The process of determining the TFA variables may include the following:

[0049] A finite-impulse-response (FIR) filter (N=12) is used to estimate the impulse in each window, as well as the frequency and step response, where N is the number of taps of the FIR filter. In implementation, the finite impulse response can be estimated using the least mean square method, such as calculated by formula (1):

[0050]

[0051] Where h(n) represents the finite impulse response, p represents the ABP signal, v represents the CBFV signal, represents the inverse matrix of the ABP signal autocorrelation matrix, R pv (n) represents the cross-correlation matrix between ABP signal and CBFV signal.

[0052] Afterwards, discrete time Fourier transform (DTFT) can be performed by the following formula (2), specifically:

[0053]

[0054] Wherein, fs represents the sampling frequency.

[0055] Afterwards, the gain and phase of the DTFT are determined by formula (3) and formula (4) respectively:

[0056]

[0057]

[0058] Among them, H R (f) represents the real part of H(f), H I (f) represents the imaginary part of H(f). The phase value range is [-180°, 180°].

[0059] Then, DTFT is applied to autocorrelation and cross-correlation, and correlation information is calculated through auto-spectra and cross-spectrum, as shown in formula (5):

[0060]

[0061] Among them, MSC(f) represents the correlation information, S pp (f) and S vv (f) represents the self-spectrum, which can be based on R pp (n) and R vv (n) get, S pv (f) represents the cross spectrum, which can be based on R pv (n) Determine.

[0062] Since the length of the frequency segment with a sampling frequency of 1 Hz is 90s, according to the Nyquist theorem, there are 45 frequency points in the range of 0-0.5 Hz containing phase, gain and correlation information for representing frequency domain features. In the embodiment of the present application, the phase and correlation information are used as sample frequency domain features, that is, the sample frequency domain features are two-dimensional.

[0063] In addition, the step response is determined according to the convolution of the finite impulse response and the step function. Specifically, it is calculated by formula (6):

[0064] s(n)=u(n)*h(n) (6)

[0065] Where s(n) represents a step response, u(n) represents a step function, and * represents a convolution operation.

[0066] In an example, when the value of N is 12, step responses at 12 time points can be obtained, and the obtained step responses are used as sample time domain features.

[0067] As an example of the present application, in order to make each dimension of the same feature (such as phase, correlation information or step response) comparable, the sample frequency domain feature and the sample time domain feature can be preprocessed separately. In implementation, each feature can be standardized separately by the following formula (7):

[0068]

[0069] Among them, x represents the original value of the feature, x * represents the standardized value, μ represents the mean of the feature, and σ represents the variance of the feature.

[0070] After that, the standardized features can be normalized to within a preset numerical range. In one example, the extreme value normalization method can be used for processing. Among them, the preset numerical range can be set according to actual needs, usually referring to a data range near the zero point, such as the preset numerical range of [-0.5, 0.5]. In this way, since the activation function will obtain the maximum activation value near the zero point during the subsequent training process, and the larger the activation value, the faster the model learning speed can be, so normalizing the standardized features to within the preset numerical range can speed up the learning of the model.

[0071] In this way, based on each group of unlabeled sample signals, the corresponding sample frequency domain features and sample time domain features can be determined. For the sake of distinction, they are referred to as first sample frequency domain features and first sample time domain features. The first sample frequency domain features and first sample time domain features are used as first sample data to obtain multiple first sample data. In addition, based on each group of labeled sample signals, the corresponding sample frequency domain features and sample time domain features can be determined, which are referred to as second sample frequency domain features and second sample time domain features. The second sample frequency domain features and second sample time domain features are used as second sample data to obtain multiple second sample data.

[0072] Afterwards, model training can be implemented based on the obtained multiple first sample data and multiple second sample data. The specific implementation thereof can be found in the following embodiment.

[0073] Please refer to Figure 1 , Figure 1 This is a flow chart of a method for training a model according to an exemplary embodiment. As an example but not a limitation, the method is applied to the above electronic device. The method may include the following steps:

[0074] Step 101: Based on a first sample data set, iteratively train a first frequency domain module and a first time domain module respectively, wherein the first frequency domain module and the first time domain module are independent modules and both have at least a feature extraction function.

[0075] As mentioned above, the first sample data in the first sample data set includes the first sample frequency domain features and the first sample time domain features corresponding to the unlabeled sample signal.

[0076] In the examples of this application, please refer to Figure 2 The model training process includes multiple stages. In the first stage, the first frequency domain module is trained using the first sample frequency domain features corresponding to the unlabeled sample signal, and the first time domain module is trained using the first sample time domain features corresponding to the unlabeled sample signal.

[0077] In one embodiment, the first frequency domain module includes mutually independent first K-branch convolutional autoencoders, multiple first single-branch convolutional autoencoders, and second K-branch convolutional autoencoders, the model structure of the second K-branch convolutional autoencoder is the same as the model structure after the first K-branch convolutional autoencoder and the K first single-branch convolutional autoencoders are combined, the K first single-branch convolutional autoencoders correspond one-to-one to the K branches in the first K-branch convolutional autoencoder, and K is an integer greater than 1. In one example, the value of K is 3.

[0078] Exemplarily, the first K branch convolutional autoencoder is as follows Figure 3 As shown in Figure (a), it includes three branches, bra11, bra12 and bra13. Each branch includes a convolutional layer and a deconvolutional layer, where the convolutional layer is used for encoding and the deconvolutional layer is used for decoding. The convolutional kernel sizes of the K branches are different, for example, the convolutional kernel sizes are 3, 7, and 15 respectively.

[0079] Exemplarily, a plurality of first single-branch convolutional autoencoders are as follows: Figure 3As shown in Figure (b), there are three independent first single-branch convolutional autoencoders, namely AE1, AE2 and AE3, each of which includes a pooling layer, a convolution layer, a deconvolution layer and an upsampling layer. In one example, AE1 corresponds to the bra11 branch, and the convolution kernel size of AE1 is consistent with the convolution kernel size of the bra11 branch; AE2 corresponds to the bra12 branch, and the convolution kernel size of AE2 is consistent with the convolution kernel size of the bra12 branch; AE3 corresponds to the bra13 branch, and the convolution kernel size of AE3 is consistent with the convolution kernel size of the bra13 branch.

[0080] In one example, the second K-branch convolutional autoencoder is as follows Figure 3 As shown in Figure (c), it includes three branches: bra21, bra22 and bra23. Each branch includes a convolutional layer, a pooling layer, a convolutional layer, a deconvolution layer, an upsampling layer, and a deconvolution layer.

[0081] It is not difficult to see that the model structure of the second K-branch convolutional autoencoder is the same as the model structure of the first K-branch convolutional autoencoder and the first K-branch convolutional autoencoder combined. The number of channels in each branch of the second K-branch convolutional autoencoder changes to [2, 8, 2, 8, 2], the size and step size of the pooling layer are both 3, and the upsampling layer corresponding to the pooling layer is the nearest neighbor upsampling with a step size of 3.

[0082] In one example, the first time domain module is a second single-branch convolutional autoencoder. Figure 4 As shown in Figure 1, it includes convolutional layers and deconvolutional layers. The number of channels of the second single-branch convolutional autoencoder changes to [1, 2, 1] in sequence.

[0083] according to Figure 3 The first frequency domain module provided, based on the first sample data set, the specific implementation of iterative training of the first frequency domain module may include: for each first sample data included in the first sample data set, obtaining the first sample frequency domain features in each first sample data, and obtaining multiple first sample frequency domain features. Iterative training is performed on the first K-branch convolutional autoencoder based on multiple first sample frequency domain features. The hidden layer features of each branch in the first K-branch convolutional autoencoder after iterative training are extracted respectively to obtain the hidden layer features of K branches, and for the hidden layer features of each branch in the K branches, the first single-branch convolutional autoencoder corresponding to each branch is iteratively trained based on the hidden layer features of each branch. The weights of the first K-branch convolutional autoencoder after iterative training and the weights of each first single-branch convolutional autoencoder after iterative training are synchronized as the weights of the second K-branch convolutional autoencoder. Based on multiple first sample frequency domain features, the synchronized second K-branch convolutional autoencoder is iteratively trained.

[0084] That is, the electronic device selects the first sample frequency domain features of each first sample data from the first sample data set, so as to perform a training operation using the selected multiple first sample frequency domain features. Specifically, it can be divided into three processes (1)-(3):

[0085] (1) Using multiple first sample frequency domain features as input features of the first K-branch convolutional autoencoder, using the predicted value and the true value to calculate the minimum mean square error loss as the loss function, iteratively training the first K-branch convolutional autoencoder, and the number of iterations can be a preset number of iterations. In this way, the first K-branch convolutional autoencoder has a preliminary feature extraction capability, and feature extraction is performed through the first K-branch convolutional autoencoder. Among them, the electronic device can calculate the minimum mean square error loss through formula (8):

[0086]

[0087] Among them, L1 represents the minimum mean square error loss, Z i represents the true value of sample i, Z' i Represents the predicted value.

[0088] (2) extracting the hidden layer features of each branch in the first K-branch convolutional autoencoder, using the hidden layer features of each branch as the input features of each first single-branch autoencoder, and iteratively training each first single-branch autoencoder, where the number of iterations can be a preset number of iterations. For example, please refer to Figure 3 , extract the features extracted by the convolution layer in the bra11 branch, and based on the features, iterate the training of AE1 for a preset number of iterations; extract the features extracted by the convolution layer in the bra12 branch, and based on the features, iterate the training of AE2 for a preset number of iterations; extract the features extracted by the convolution layer in the bra13 branch, and based on the features, iterate the training of AE3 for a preset number of iterations.

[0089] Since each first single-branch autoencoder includes a pooling layer, after iterative training of each first single-branch autoencoder, the features extracted by the convolutional layers of each branch in each first single-branch autoencoder are low-dimensional and can fully represent the original features.

[0090] (3) Extract the weights of each layer in the first K-branch convolutional autoencoder after iterative training, and extract the weights of each layer in each first single-branch convolutional autoencoder after iterative training. Since the model structure of the second K-branch convolutional autoencoder is the same as the model structure after the first K-branch convolutional autoencoder and multiple first single-branch convolutional autoencoders are combined, for example Figure 3As shown, the weights of each layer in the first K-branch convolutional self-encoder after iterative training can be synchronized to each layer of the corresponding branch in the second K-branch convolutional self-encoder, and the weights of each layer in each first single-branch convolutional self-encoder after iterative training can be synchronized to each layer of the corresponding branch in the second K-branch convolutional self-encoder. For example, the weights of each layer in the bra11 branch are synchronized to the corresponding layer of the bra21 branch of the second K-branch convolutional self-encoder, and the weights of each layer in AE1 are synchronized to the corresponding layer of the bra21 branch of the second K-branch convolutional self-encoder; the weights of each layer in the bra12 branch are synchronized to the corresponding layer of the bra22 branch of the second K-branch convolutional self-encoder, and the weights of each layer in AE2 are synchronized to the corresponding layer of the bra22 branch of the second K-branch convolutional self-encoder; the weights of each layer in the bra13 branch are synchronized to the corresponding layer of the bra23 branch of the second K-branch convolutional self-encoder, and the weights of each layer in AE3 are synchronized to the corresponding layer of the bra23 branch of the second K-branch convolutional self-encoder. Then, based on the frequency domain features of the multiple first samples, the second K-branch convolutional autoencoder obtained after synchronization is iteratively trained to fine-tune the parameters of the second K-branch convolutional autoencoder, and the number of iterations can be a preset number of iterations. At this point, iterative training of the first frequency domain module is achieved.

[0091] The preset number of iterations can be set according to actual needs, for example, the preset number of iterations can be 30 times.

[0092] For example, taking the first sample frequency domain feature including the phase and correlation information of 45 frequency points as an example, please refer to Figure 3, based on each first sample frequency domain feature in the multiple first sample frequency domain features, a 2*45 matrix is ​​generated. The generated matrix is ​​input into the first K branch convolutional autoencoder for iterative training, for example, 30 iterations. Afterwards, the hidden layer features of each branch of the first K branch convolutional autoencoder are extracted, that is, the hidden layer features of the bra11 branch, the hidden layer features of the bra12 branch, and the hidden layer features of the bra13 branch are extracted. For example, the hidden layer features of each branch extracted are all 8*45 matrices. The hidden layer features of the bra11 branch are used as the input features of AE1 for iterative training of AE1 30 times, the hidden layer features of the bra12 branch are used as the input features of AE2 for iterative training of AE2 30 times, and the hidden layer features of the bra13 branch are used as the input features of AE3 for iterative training of AE3 30 times. Then, the weights of each layer of the bra11 branch after iterative training and the weights of each layer in AE1 are extracted, and synchronized as the weights of each layer of the bra21 branch in the second K-branch convolutional autoencoder; the weights of each layer of the bra12 branch after iterative training and the weights of each layer in AE2 are extracted, and synchronized as the weights of each layer of the bra22 branch in the second K-branch convolutional autoencoder; the weights of each layer of the bra13 branch after iterative training and the weights of each layer in AE3 are extracted, and synchronized as the weights of each layer of the bra23 branch in the second K-branch convolutional autoencoder. After the synchronization process, the synchronized second K-branch convolutional autoencoder is iteratively trained again based on the generated 2*45 matrix, for example, iterated 30 times. In this way, the iterative training of the first frequency domain module is completed.

[0093] It should be noted that, in each of the above-mentioned convolutional autoencoders, the number of convolutional layers included in each branch may be one or more, which is not limited in the embodiments of the present application.

[0094] It is worth mentioning that since each convolutional autoencoder in the first frequency domain module is multi-branched, multi-scale local feature extraction can be performed on the frequency domain features of the first sample, which retains the characteristics of spatial local information and extracts rich local information under different receptive fields, which is beneficial to the learning of subsequent models.

[0095] according to Figure 4 The first time domain module provided, based on the first sample data set, the specific implementation of iterative training of the first time domain module may include: for each first sample data included in the first sample data set, obtaining the first sample time domain features in each first sample data, and obtaining multiple first sample time domain features. The multiple first sample time domain features are used as input features of a single-branch convolutional autoencoder, and the single-branch convolutional autoencoder is iteratively trained.

[0096] like Figure 2As shown, the electronic device selects the first sample time domain features of each first sample data from the first sample data set, and uses the selected multiple first sample time domain features as input features of the first time domain module, and iteratively trains the first time domain module, for example, iteratively preset the number of iterations. For example, taking the first sample time domain features including the step response of 12 time points as an example, please refer to Figure 4 , based on each of the first sample time domain features in the plurality of first sample time domain features, a 1*12 matrix is ​​generated. The generated matrix is ​​input into the second single-branch convolutional autoencoder for iterative training 30 times.

[0097] It is worth mentioning that by incorporating an appropriate amount of unlabeled sample signals, the universality of discriminating CA functions is improved, and the noise impact caused by fragmented data is avoided as much as possible.

[0098] It is not difficult to understand that since the training in the first stage is performed based on the first sample data set corresponding to the unlabeled sample signal, the training process in the first stage can be called unsupervised learning.

[0099] Step 102: construct a second frequency domain module according to the iteratively trained first frequency domain module, and construct a second time domain module according to the iteratively trained first time domain module.

[0100] The second frequency domain module has a feature extraction function and a classification function, and the second time domain module also has a feature extraction function and a classification function.

[0101] In the second stage of model training, the first frequency domain module and the first time domain module after iterative training in the first stage are used to reconstruct the module (including the second frequency domain module and the second time domain module) and train it, specifically including the contents of step 102 and step 103.

[0102] As an example of the present application, the specific implementation of constructing the second frequency domain module according to the iteratively trained first frequency domain module may include: removing the decoder part in the second K-branch convolutional autoencoder after iterative training. After removing the processed second K-branch convolutional autoencoder, the first pooling layer, the first fully connected layer and the first classifier are sequentially connected to obtain the second frequency domain module.

[0103] Please continue to refer to Figure 2 , the electronic device retains the encoder of the first frequency domain module, removes the decoder part, and connects a first pooling layer to further compress the extracted features. Then a first fully connected layer is connected as a bottleneck layer, and finally a first classifier is connected. For example, a Soft-Max layer can be used as the first classifier. In this way, a second frequency domain module is constructed. For example, the second frequency domain module is as follows Figure 5 As shown in Figure (a).

[0104] As an example of the present application, the specific implementation of constructing the second time domain module according to the iteratively trained first time domain module may include: removing the decoder part in the second single-branch convolutional autoencoder after iterative training. After removing the processed second single-branch convolutional autoencoder, the second pooling layer, the second fully connected layer and the second classifier are sequentially connected to obtain the second time domain module, and the second fully connected layer has the same number of neurons as the first fully connected layer. Exemplarily, the number of neurons in the first fully connected layer and the second fully connected layer can be 15.

[0105] Please continue to refer to Figure 2 , the electronic device retains the encoder of the first time domain module, removes the decoder part, and connects a second pooling layer to further compress the extracted features. Then a second fully connected layer is connected as a bottleneck layer, and finally a second classifier is connected. For example, a Soft-Max layer can be used as the second classifier. In this way, a second time domain module is constructed. For example, the second time domain module is as follows Figure 5 As shown in Figure (b).

[0106] Step 103: Based on the second sample data set, iteratively train the second frequency domain module and the second time domain module respectively.

[0107] As mentioned above, the second sample data in the second sample data set are second sample frequency domain features and second sample time domain features corresponding to the labeled sample signal.

[0108] Please refer to Figure 2 In the second stage of training, after the second frequency domain module and the second time domain module are constructed, the second frequency domain module and the second time domain module are iteratively trained respectively using the second sample data set, so that the second frequency domain module has the ability to classify based on the frequency domain features corresponding to the CA capability of any individual, and the second time domain module has the ability to classify based on the time domain features corresponding to the CA capability of any individual.

[0109] The specific implementation of iteratively training the second frequency domain module based on the second sample data set includes: for each second sample data included in the second sample data set, obtaining the second sample frequency domain features in each second sample data, obtaining multiple second sample frequency domain features. The multiple second sample frequency domain features are used as input features of the second frequency domain module, and the second frequency domain module is iteratively trained.

[0110] Please refer to Figure 5In (a) of FIG. 1 , the electronic device generates a 2*45 matrix based on each second sample frequency domain feature of the plurality of second sample frequency domain features. The generated matrix is ​​used as the input feature of the second frequency domain module, and the cross entropy loss is calculated using the predicted value and the true label value as the loss function, and the second frequency domain module is iteratively trained, for example, 60 times. The electronic device can calculate the cross entropy loss by formula (9):

[0111]

[0112] Among them, y i represents the true label value of sample i. For example, the true label value of individuals with severe symptoms (positive) is 1, and the true label value of healthy (negative) individuals is 0; p i It represents the probability of predicting sample i as positive.

[0113] The specific implementation of iteratively training the second time domain module based on the second sample time domain set includes: for each second sample data included in the second sample data set, obtaining the second sample time domain features in each second sample data, obtaining multiple second sample time domain features. The multiple second sample time domain features are used as input features of the second time domain module, and the second time domain module is iteratively trained.

[0114] For example, please refer to Figure 5 In (b) of FIG. 1 , a 1*12 matrix is ​​generated based on each second sample time domain feature of the plurality of second sample time domain features. The generated matrix is ​​used as the input feature of the second time domain module, and is input into the second time domain module. The cross entropy loss is calculated using the predicted value and the true label value as the loss function, and the second time domain module is iteratively trained, for example, 60 times.

[0115] Step 104: construct a multi-scale fusion perception model according to the iteratively trained second frequency domain module and the iteratively trained second time domain module.

[0116] Among them, the multi-scale fusion perception module has feature extraction and classification functions.

[0117] Since the second frequency domain module after iterative training is determined based on multiple second sample frequency domain features, classification can only be performed based on frequency domain features. Similarly, since the second time domain module after iterative training is determined based on multiple second sample time domain features, classification can only be performed based on time domain features. In an embodiment of the present application, the ultimate goal is to perform classification based on frequency domain features and time domain features to determine whether the CA function to be monitored is impaired. Therefore, in the third stage of training, a multi-scale fusion perception model is constructed based on the second frequency domain module after iterative training and the second time domain module after iterative training, that is, the second frequency domain module after iterative training is fused with the time domain module after iterative training, and the multi-scale fusion perception module obtained after fusion is fine-tuned in parameters, so as to obtain the final target multi-scale fusion perception model. The third stage specifically includes the contents of step 104 and step 105.

[0118] For an example, see Figure 2 , the specific implementation of step 104 may include: removing the first classifier in the second frequency domain module after iterative training, and removing the second classifier in the second time domain module after iterative training. Connecting the intermediate hidden layer in the removed second frequency domain module along the channel, stacking the first fully connected layer of the connected second frequency domain module and the second fully connected layer of the removed second time domain module, so as to fuse the removed second frequency domain module with the removed second time domain module. Connecting the third classifier to the fused module to obtain a multi-scale fusion perception model.

[0119] That is, the electronic device constructs a multi-scale fusion perception model based on the second frequency domain module and the second time domain module obtained in the second stage. As shown in Figure (6), during the construction process, the middle hidden layer in the second frequency domain module is fused. The fusion method is to connect along the channel, for example Figure 5 As shown in Figure (a), three 2*15 features are connected to become 6*15 features. In addition, before accessing the fully connected layer, the features are flattened into a single dimension, for example, the flattened 6*15 features are 90 single-dimensional features, and pooled through the pooling layer (90 / 3=30). In the time domain module, the size of the single-scale convolution kernel is fixed, the size and step size of the pooling layer are both 2, and the nearest neighbor upsampling step size is also 2. It is flattened and pooled once before accessing the fully connected layer. When fusing the second frequency domain module and the second time domain module, the weights of the fully connected layers are stacked. For example, the weight of the first fully connected layer is 30*15, the weight of the second fully connected layer is 12*15, and the weight of the stacked fully connected layer is 30*15+12*15, that is, 42*15. Since classification discrimination is still required, the third classifier is finally connected to output two-dimensional prediction values, representing the probabilities of each category (including positive and negative).

[0120] It is not difficult to understand that since the multi-scale fusion perception model is obtained based on the second frequency domain module after iterative training and the second time domain module after iterative training, the multi-scale fusion perception model includes a frequency domain branch and a time domain branch, wherein the frequency domain branch is used to extract frequency domain features, and the time domain branch is used to extract time domain features.

[0121] Step 105: Based on the second sample data set, iteratively train the multi-scale fusion perception model to obtain a target multi-scale fusion perception model.

[0122] Please continue to refer to Figure 2 After the module fusion is performed, the electronic device fine-tunes the multi-scale fusion perception model based on the second sample data set. In the implementation, a 2*45 matrix is ​​generated based on each second sample frequency domain feature in each second sample data, and a 1*12 matrix is ​​generated based on each second sample time domain feature in each second sample data. All the generated matrices are used as input features of the multi-scale fusion perception model, and the multi-scale fusion perception model is iteratively trained, wherein the 2*45 matrix is ​​processed by the frequency domain branch in the multi-scale fusion perception model, and the 1*12 matrix is ​​processed by the time domain branch of the multi-scale fusion perception model.

[0123] It should be noted that step 103 to step 105 is an exemplary implementation method of determining a target multi-scale fusion perception model based on the second sample data set, the second frequency domain module, and the second time domain module.

[0124] It is worth mentioning that the embodiment of the present application, by fusing frequency domain features and time domain features at different scales and supplemented by two fine-tunings, can form information complementarity between frequency domain modules and time domain modules, and between multiple scales within a module, making the prediction result more accurate.

[0125] Please refer to Figure 2As an example of the present application, before the iteratively trained multi-scale fusion perception model is determined as the target multi-scale fusion perception model, a model testing operation can usually be performed to evaluate the performance of the iteratively trained multi-scale fusion perception model. In implementation, the electronic device obtains a test data set, and uses the test data set as the input feature of the iteratively trained multi-scale fusion perception model. The output value is the predicted probability value of whether CA is damaged, and the performance of the iteratively trained multi-scale fusion perception model can be evaluated according to the accuracy of the predicted probability value. If it is determined after evaluation that the performance meets the requirements, the iteratively trained multi-scale fusion perception model is determined as the target multi-scale fusion perception model. Otherwise, if it is determined after evaluation that the performance does not meet the requirements, the iteratively trained multi-scale fusion perception model can continue to be trained until it is determined after evaluation that the performance meets the requirements, so as to obtain a multi-scale fusion perception model with good predictive performance.

[0126] In one embodiment, regarding the acquisition of the test data set, the electronic device may determine the sample data set and select part of the second sample data from it as the test data set. Since the second sample data set is relatively small, a k-fold cross-validation method may be used to determine the training data set and the test data set from the second sample data set. Exemplarily, n=5 may be taken, that is, 4 / 5 of the second sample data is used for training and 1 / 5 of the second sample data is used for testing in each experiment, and the electronic device may iterate the experiment 5 times without repeating the test set.

[0127] It is not difficult to understand that since the training process of the second stage and the third stage is based on the second sample data set corresponding to the labeled sample signal, the training of the second stage and the third stage can be called supervised learning.

[0128] As an example of this application, the model design can be performed using the Pytorch open source framework. The model parameters can be randomly initialized, and the tanh function can be selected as the activation function.

[0129] In each of the above iterative training processes, the gradient is calculated and back-propagated to update the model parameters, and the weights of the trained model are saved for each training. The optimizer used for gradient back-propagation can use the Adam optimizer and the SGD (Stochastic Gradient Descent) optimizer for combined learning. In one example, the Adam optimizer is used in the first preset number of iterations (such as 30 times) of each training, and the learning rate is set to 0.01, and the SGD optimizer is used in the subsequent preset number of iterations (such as 30 times), and the learning rate is set to 0.001. As an example, in the first stage of model training, only the Adam optimizer can be used.

[0130] As an example of the present application, in the first stage, 32 first sample data can be selected as input features for each iterative training, that is, the small batch of unsupervised learning can be set to 32; in the second and third stages, 10 second sample data can be selected as input features for each iterative training, that is, the small batch of the supervised learning part can be set to 10.

[0131] As an example of the present application, each of the above-mentioned convolutional layers and deconvolutional layers has a batch normalization operation, and the momentum can be set according to actual needs. The momentum of the unsupervised learning part can be set to 0.1, and the momentum of the supervised learning part can be set to 0.5.

[0132] It is worth mentioning that the embodiment of the present application has been preliminarily verified, proving the superiority of the current algorithm, and prediction through the trained multi-scale fusion perception model only requires millisecond-level calculations. The target multi-scale fusion perception model (abbreviated as: MEFP) proposed in the embodiment of the present application is compared with the traditional method (Convention) and the semi-supervised stacked ensemble learning method (abbreviated as: SSEL) SSEL in terms of accuracy, ROC (receiver operating characteristic curve) curve and AUC (AUthentication Center) value. The accuracy is shown in Table 1 below, and the ROC and AUC are shown in Table 2 below. Figure 7 shown.

[0133] Table 1

[0134] Convention SSEL MSFP Accuracy 79.92％ 85.67％ 96.34％

[0135] In an embodiment of the present application, based on the first sample data set, the first frequency domain module and the first time domain module are iteratively trained respectively. The first sample data in the first sample data set includes the first sample frequency domain features and the first sample time domain features corresponding to the unlabeled sample signal. The first frequency domain module and the first time domain module are independent modules, and both have feature extraction functions. The second frequency domain module is constructed according to the first frequency domain module after iterative training, and the second time domain module is constructed according to the first time domain module after iterative training. Based on the second sample data set, the second frequency domain module, and the second time domain module, the target multi-scale fusion perception model is determined. The second sample data in the second sample data set is the second sample frequency domain features and the second sample time domain features corresponding to the labeled sample signal. The obtained target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and time domain features corresponding to the CA function of any individual, that is, the CA function of the target individual can be monitored by the target multi-scale fusion perception model. Compared with the logistic regression model, the target multi-scale fusion perception model is more robust and can improve the accuracy of the detection results.

[0136] In addition, the number of floating-point operations in the embodiment of the present application is extremely low, and the predicted operation time is in the millisecond level, which can meet the real-time requirements.

[0137] Please refer to Figure 8 , Figure 8 The flowchart of a method for monitoring a CA function according to an exemplary embodiment is shown. The method can be detected by the target multi-scale fusion perception model and executed by an electronic device. The method may include the following steps:

[0138] Step 801: Based on the arterial blood pressure signal and cerebral blood flow signal of the target individual, determine a first transfer function analysis TFA variable for representing the current state of the cerebral blood flow autoregulation CA function of the target individual, wherein the first TFA variable includes frequency domain features and time domain features.

[0139] Its specific implementation can be found in the previous article.

[0140] Step 802: Based on the frequency domain features and the time domain features, determine whether the CA function of the target individual is damaged through the target multi-scale fusion perception model. The target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and the time domain features corresponding to the CA function of any individual.

[0141] In implementation, a corresponding matrix can be generated based on frequency domain features and time domain features, and the generated matrix can be input into the target multi-scale fusion perception model for prediction, and the probability values ​​of each type can be output. Therefore, it can be determined whether the CA function of the target individual is impaired based on the output probability values.

[0142] In an embodiment of the present application, since the target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain characteristics and time domain characteristics corresponding to the CA function of any individual, the CA function of the target individual can be monitored through the target multi-scale fusion perception model. Compared with the logistic regression model, the target multi-scale fusion perception model is more robust and can improve the accuracy of the detection results.

[0143] Fig. 9 This is a schematic diagram of a device for monitoring the automatic regulation function of cerebral blood flow according to an exemplary embodiment. The device can be implemented as the above electronic device by software, hardware or a combination of both. The device may include:

[0144] A first determination module 910 is used to determine a first transfer function analysis TFA variable for representing a current state of a cerebral blood flow autoregulation (CA) function of the target individual based on an arterial blood pressure signal and a cerebral blood flow signal of the target individual, wherein the first TFA variable includes a frequency domain feature and a time domain feature;

[0145] The second determination module 920 is used to determine whether the CA function of the target individual is damaged based on the frequency domain features and the time domain features through a target multi-scale fusion perception model. The target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and time domain features corresponding to the CA function of any individual.

[0146] In an embodiment of the present application, since the target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain characteristics and time domain characteristics corresponding to the CA function of any individual, the CA function of the target individual can be monitored through the target multi-scale fusion perception model. Compared with the logistic regression model, the target multi-scale fusion perception model is more robust and can improve the accuracy of the detection results.

[0147] Fig.10 1 is a schematic diagram of a structure of a training module device according to an exemplary embodiment. The device can be implemented as the above electronic device by software, hardware or a combination of both. The device may include:

[0148] A first training module 1010 is used to iteratively train a first frequency domain module and a first time domain module based on a first sample data set, wherein the first sample data in the first sample data set includes a first sample frequency domain feature and a first sample time domain feature corresponding to an unlabeled sample signal, and the first frequency domain module and the first time domain module are independent modules and both have a feature extraction function;

[0149] A construction module 1020 is used to construct a second frequency domain module according to the iteratively trained first frequency domain module, and to construct a second time domain module according to the iteratively trained first time domain module, wherein both the second frequency domain module and the second time domain module have a feature extraction function and a classification function;

[0150] A second training module 1030 is used to determine a target multi-scale fusion perception model based on a second sample data set, the second frequency domain module, and the second time domain module, wherein the second sample data in the second sample data set is a second sample frequency domain feature and a second sample time domain feature corresponding to the labeled sample signal;

[0151] Among them, the target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain characteristics and time domain characteristics corresponding to the CA function of any individual.

[0152] As an example of the present application, the first frequency domain module includes mutually independent first K-branch convolutional autoencoders, K first single-branch convolutional autoencoders, and second K-branch convolutional autoencoders, the model structure of the second K-branch convolutional autoencoder is the same as the model structure after the first K-branch convolutional autoencoder and the multiple first single-branch convolutional autoencoders are combined, the K first single-branch convolutional autoencoders correspond one-to-one to the K branches in the first K-branch convolutional autoencoder, and K is an integer greater than 1;

[0153] The first training module 1010 is used to:

[0154] For each first sample data included in the first sample data set, obtaining a first sample frequency domain feature in each first sample data to obtain a plurality of first sample frequency domain features;

[0155] Based on the multiple first sample frequency domain features, iteratively training the first K-branch convolutional autoencoder;

[0156] The hidden layer features of each branch in the first K-branch convolutional autoencoder after iterative training are extracted respectively to obtain the hidden layer features of K branches;

[0157] For the hidden layer features of each branch in the K branches, based on the hidden layer features of each branch, iteratively train the first single-branch convolutional autoencoder corresponding to each branch;

[0158] The weight of the first K-branch convolutional autoencoder after iterative training and the weight of each first single-branch convolutional autoencoder after iterative training are synchronized as the weight of the second K-branch convolutional autoencoder;

[0159] Based on the multiple first sample frequency domain features, the synchronized second K-branch convolutional autoencoder is iteratively trained.

[0160] As an example of the present application, the building block 1020 is used to:

[0161] Remove the decoder part in the second K branch convolutional autoencoder after iterative training;

[0162] After removing the processed second K-branch convolutional autoencoder, the first pooling layer, the first fully connected layer and the first classifier are sequentially connected to obtain the second frequency domain module.

[0163] As an example of the present application, the first time domain module is a second single-branch convolutional autoencoder, and the first training module 1010 is used to:

[0164] For each first sample data included in the first sample data set, obtaining a first sample time domain feature in each first sample data to obtain a plurality of first sample time domain features;

[0165] The multiple first sample time domain features are used as input features of the second single-branch convolutional autoencoder, and the second single-branch convolutional autoencoder is iteratively trained.

[0166] As an example of the present application, the building module 1020 is used to:

[0167] Remove the decoder part in the second single-branch convolutional autoencoder after iterative training;

[0168] After removing the second single-branch convolutional autoencoder after the processing, the second pooling layer, the second fully connected layer and the second classifier are sequentially connected to obtain the second time domain module, and the second fully connected layer has the same number of neurons as the first fully connected layer.

[0169] As an example of the present application, the third training module 1030 is used to:

[0170] Based on the second sample data set, iteratively training the second frequency domain module and the second time domain module respectively;

[0171] Constructing a multi-scale fusion perception model according to the iteratively trained second frequency domain module and the iteratively trained second time domain module;

[0172] Based on the second sample data set, the multi-scale fusion perception model is iteratively trained to obtain the target multi-scale fusion perception model.

[0173] As an example of the present application, the third training module 1030 is used to:

[0174] For each second sample data included in the second sample data set, obtaining a second sample frequency domain feature in each second sample data to obtain a plurality of second sample frequency domain features;

[0175] The multiple second sample frequency domain features are used as input features of the second frequency domain module, and the second frequency domain module is iteratively trained.

[0176] As an example of the present application, the third training module 1030 is used to:

[0177] For each second sample data included in the second sample data set, obtaining a second sample time domain feature in each second sample data to obtain a plurality of second sample time domain features;

[0178] The multiple second sample time domain features are used as input features of the second time domain module, and the second time domain module is iteratively trained.

[0179] As an example of the present application, the third training module 1030 is used to:

[0180] Removing the first classifier in the second frequency domain module after iterative training, and removing the second classifier in the second time domain module after iterative training;

[0181] Connect the middle hidden layer in the second frequency domain module after the removal process along the channel;

[0182] stacking the first fully connected layer of the connected second frequency domain module and the second fully connected layer of the removed second time domain module to fuse the removed second frequency domain module with the removed second time domain module;

[0183] A third classifier is connected to the module obtained after fusion to obtain the multi-scale fusion perception model.

[0184] In an embodiment of the present application, based on the first sample data set, the first frequency domain module and the first time domain module are iteratively trained respectively. The first sample data in the first sample data set includes the first sample frequency domain features and the first sample time domain features corresponding to the unlabeled sample signal. The first frequency domain module and the first time domain module are independent modules, and both have feature extraction functions. The second frequency domain module is constructed according to the first frequency domain module after iterative training, and the second time domain module is constructed according to the first time domain module after iterative training. Based on the second sample data set, the second frequency domain module, and the second time domain module, the target multi-scale fusion perception model is determined. The second sample data in the second sample data set is the second sample frequency domain features and the second sample time domain features corresponding to the labeled sample signal. The obtained target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and time domain features corresponding to the CA function of any individual, that is, the CA function of the target individual can be monitored by the target multi-scale fusion perception model. Compared with the logistic regression model, the target multi-scale fusion perception model is more robust and can improve the accuracy of the detection results.

[0185] Fig.11 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Fig.11 As shown, the electronic device 11 of this embodiment includes: at least one processor 1100 ( Fig.10 Only one is shown), a memory 1101, and a computer program 1102 stored in the memory 1101 and executable on the at least one processor 1100, wherein the processor 1100 implements the steps of any of the above-mentioned method embodiments when executing the computer program 1102.

[0186] The electronic device 11 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The electronic device may include, but is not limited to, a processor 1100 and a memory 1101. Those skilled in the art will appreciate that Fig.11 It is only an example of the electronic device 11 and does not constitute a limitation on the electronic device 11. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, it may also include input and output devices, network access devices, etc.

[0187] The processor 1100 may be a CPU (Central Processing Unit), or other general-purpose processors, DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0188] In some embodiments, the memory 1101 may be an internal storage unit of the electronic device 11, such as a hard disk or memory of the electronic device 11. In other embodiments, the memory 1101 may also be an external storage device of the electronic device 11, such as a plug-in hard disk, SMC (Smart Media Card), SD (Secure Digital) card, Flash Card, etc. equipped on the electronic device 11. Further, the memory 1101 may also include both an internal storage unit and an external storage device of the electronic device 11. The memory 1101 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program. The memory 1101 may also be used to temporarily store data that has been output or is to be output.

[0189] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0190] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0191] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for monitoring cerebral blood flow autoregulation function, characterized in that: The method comprises: Based on the arterial blood pressure signal and the cerebral blood flow signal of the target individual, determining a first transfer function analysis TFA variable for representing the current state of the cerebral blood flow autoregulation (CA) function of the target individual, wherein the first transfer function analysis TFA variable includes frequency domain features and time domain features; Based on the frequency domain features and the time domain features, determining whether the CA function of the target individual is damaged through a target multi-scale fusion perception model, wherein the target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and the time domain features corresponding to the CA function of any individual; The target multi-scale fusion perception model is determined based on a second sample data set, a second frequency domain module, and a second time domain module. The second sample data in the second sample data set is a second sample frequency domain feature and a second sample time domain feature corresponding to a labeled sample signal. The labeled sample signal includes a labeled ABP signal and a labeled CBFV signal. The second frequency domain module and the second time domain module both have a feature extraction function and a classification function. The construction process of the second frequency domain module includes: iteratively training the first frequency domain module based on the first sample data set, the first sample data in the first sample data set includes the first sample frequency domain features and the first sample time domain features corresponding to the unlabeled sample signal, and the unlabeled sample signal includes the unlabeled arterial blood pressure ABP signal and the unlabeled cerebral blood flow velocity CBFV signal; the first frequency domain module includes a mutually independent first K-branch convolutional autoencoder, K first single-branch convolutional autoencoders, and a second K-branch convolutional autoencoder, the model structure of the second K-branch convolutional autoencoder is the same as the model structure after the first K-branch convolutional autoencoder and the K first single-branch convolutional autoencoders are combined, the K first single-branch convolutional autoencoders correspond to the K branches in the first K-branch convolutional autoencoder one by one, and K is an integer greater than 1; remove the decoder part of the second K-branch convolutional autoencoder after iterative training, and connect the first pooling layer, the first fully connected layer and the first classifier in sequence after the second K-branch convolutional autoencoder after removal; The construction process of the second time domain module includes: iteratively training the first time domain module based on the first sample data set, the first time domain module is a second single-branch convolutional autoencoder, the first frequency domain module and the first time domain module are independent modules, and both have feature extraction functions; removing the decoder part of the second single-branch convolutional autoencoder after iterative training, and connecting the second pooling layer, the second fully connected layer and the second classifier in sequence after the removal of the second single-branch convolutional autoencoder, and the second fully connected layer has the same number of neurons as the first fully connected layer.

2. A method for training a model, characterized in that: The method comprises: Based on the first sample data set, the first frequency domain module and the first time domain module are iteratively trained respectively, the first sample data in the first sample data set includes the first sample frequency domain features and the first sample time domain features corresponding to the unlabeled sample signal, the first frequency domain module and the first time domain module are independent modules, and both have feature extraction functions, the unlabeled sample signal includes an unlabeled arterial blood pressure ABP signal and an unlabeled cerebral blood flow velocity CBFV signal; the first frequency domain module includes a mutually independent first K-branch convolutional autoencoder, K first single-branch convolutional autoencoders, and a second K-branch convolutional autoencoder, the model structure of the second K-branch convolutional autoencoder is the same as the model structure after the first K-branch convolutional autoencoder and the K first single-branch convolutional autoencoders are combined, the K first single-branch convolutional autoencoders correspond one-to-one to the K branches in the first K-branch convolutional autoencoder, and K is an integer greater than 1; the first time domain module is a second single-branch convolutional autoencoder; The decoder part of the second K-branch convolutional autoencoder after iterative training is removed, and the first pooling layer, the first fully connected layer and the first classifier are sequentially connected after the second K-branch convolutional autoencoder after the removal process to obtain a second frequency domain module, and the decoder part of the second single-branch convolutional autoencoder after iterative training is removed, and the second pooling layer, the second fully connected layer and the second classifier are sequentially connected after the second single-branch convolutional autoencoder after the removal process to obtain a second time domain module, the second fully connected layer has the same number of neurons as the first fully connected layer, and the second frequency domain module and the second time domain module both have feature extraction function and classification function; Determine a target multi-scale fusion perception model based on a second sample data set, the second frequency domain module, and the second time domain module, wherein the second sample data in the second sample data set is a second sample frequency domain feature and a second sample time domain feature corresponding to a labeled sample signal, and the labeled sample signal includes a labeled ABP signal and a labeled CBFV signal; Among them, the target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain characteristics and time domain characteristics corresponding to the CA function of any individual.

3. The method according to claim 2, characterized in that Based on the first sample data set, iteratively training the first frequency domain module includes: For each first sample data included in the first sample data set, obtaining a first sample frequency domain feature in each first sample data to obtain a plurality of first sample frequency domain features; Based on the multiple first sample frequency domain features, iteratively training the first K-branch convolutional autoencoder; The hidden layer features of each branch in the first K-branch convolutional autoencoder after iterative training are extracted respectively to obtain the hidden layer features of K branches; For the hidden layer features of each branch in the K branches, based on the hidden layer features of each branch, iteratively train the first single-branch convolutional autoencoder corresponding to each branch; The weight of the first K-branch convolutional autoencoder after iterative training and the weight of each first single-branch convolutional autoencoder after iterative training are synchronized as the weight of the second K-branch convolutional autoencoder; Based on the multiple first sample frequency domain features, the synchronized second K-branch convolutional autoencoder is iteratively trained.

4. The method according to claim 2, characterized in that Iteratively training the first time domain module based on the first sample data set includes: For each first sample data included in the first sample data set, obtaining a first sample time domain feature in each first sample data to obtain a plurality of first sample time domain features; The multiple first sample time domain features are used as input features of the second single-branch convolutional autoencoder, and the second single-branch convolutional autoencoder is iteratively trained.

5. The method according to claim 2, characterized in that The determining the target multi-scale fusion perception model based on the second sample data set, the second frequency domain module, and the second time domain module includes: Based on the second sample data set, iteratively training the second frequency domain module and the second time domain module respectively; Constructing a multi-scale fusion perception model according to the iteratively trained second frequency domain module and the iteratively trained second time domain module; Based on the second sample data set, the multi-scale fusion perception model is iteratively trained to obtain the target multi-scale fusion perception model.

6. The method according to claim 5, characterized in that Based on the second sample data set, iteratively training the second frequency domain module includes: For each second sample data included in the second sample data set, obtaining a second sample frequency domain feature in each second sample data to obtain a plurality of second sample frequency domain features; The multiple second sample frequency domain features are used as input features of the second frequency domain module, and the second frequency domain module is iteratively trained.

7. The method according to claim 5, characterized in that Based on the second sample data set, iteratively training the second time domain module includes: For each second sample data included in the second sample data set, obtaining a second sample time domain feature in each second sample data to obtain a plurality of second sample time domain features; The multiple second sample time domain features are used as input features of the second time domain module, and the second time domain module is iteratively trained.

8. The method according to any one of claims 5 to 7, characterized in that The multi-scale fusion perception model is constructed according to the iteratively trained second frequency domain module and the iteratively trained second time domain module, including: Removing the first classifier in the second frequency domain module after iterative training, and removing the second classifier in the second time domain module after iterative training; Connect the middle hidden layer in the second frequency domain module after the removal process along the channel; stacking the first fully connected layer of the connected second frequency domain module and the second fully connected layer of the removed second time domain module to fuse the removed second frequency domain module with the removed second time domain module; A third classifier is connected to the module obtained after fusion to obtain the multi-scale fusion perception model.

9. A device for monitoring the autoregulation function of cerebral blood flow, characterized in that: The device comprises: A first determination module is used to determine a first transfer function analysis TFA variable for representing a current state of a cerebral blood flow autoregulation (CA) function of the target individual based on an arterial blood pressure signal and a cerebral blood flow signal of the target individual, wherein the first transfer function analysis TFA variable includes a frequency domain feature and a time domain feature; a second determination module, configured to determine whether the CA function of the target individual is damaged based on the frequency domain features and the time domain features through a target multi-scale fusion perception model, wherein the target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain features and the time domain features corresponding to the CA function of any individual; The target multi-scale fusion perception model is determined based on a second sample data set, a second frequency domain module, and a second time domain module. The second sample data in the second sample data set is a second sample frequency domain feature and a second sample time domain feature corresponding to a labeled sample signal. The labeled sample signal includes a labeled ABP signal and a labeled CBFV signal. The second frequency domain module and the second time domain module both have a feature extraction function and a classification function. The construction process of the second frequency domain module includes: iteratively training the first frequency domain module based on the first sample data set, the first sample data in the first sample data set includes the first sample frequency domain features and the first sample time domain features corresponding to the unlabeled sample signal, and the unlabeled sample signal includes the unlabeled arterial blood pressure ABP signal and the unlabeled cerebral blood flow velocity CBFV signal; the first frequency domain module includes a mutually independent first K-branch convolutional autoencoder, K first single-branch convolutional autoencoders, and a second K-branch convolutional autoencoder, the model structure of the second K-branch convolutional autoencoder is the same as the model structure after the first K-branch convolutional autoencoder and the K first single-branch convolutional autoencoders are combined, the K first single-branch convolutional autoencoders correspond to the K branches in the first K-branch convolutional autoencoder one by one, and K is an integer greater than 1; remove the decoder part of the second K-branch convolutional autoencoder after iterative training, and connect the first pooling layer, the first fully connected layer and the first classifier in sequence after the second K-branch convolutional autoencoder after removal; The construction process of the second time domain module includes: iteratively training the first time domain module based on the first sample data set, the first time domain module is a second single-branch convolutional autoencoder, the first frequency domain module and the first time domain module are independent modules, and both have feature extraction functions; removing the decoder part of the second single-branch convolutional autoencoder after iterative training, and connecting the second pooling layer, the second fully connected layer and the second classifier in sequence after the removal of the second single-branch convolutional autoencoder, and the second fully connected layer has the same number of neurons as the first fully connected layer.

10. A device for training a model, characterized in that: The device comprises: A first training module is used to iteratively train a first frequency domain module and a first time domain module based on a first sample data set, wherein the first sample data in the first sample data set includes a first sample frequency domain feature and a first sample time domain feature corresponding to an unlabeled sample signal, the first frequency domain module and the first time domain module are independent modules, and both have a feature extraction function, the unlabeled sample signal includes an unlabeled arterial blood pressure ABP signal and an unlabeled cerebral blood flow velocity CBFV signal; the first frequency domain module includes a first K-branch convolutional autoencoder, K first single-branch convolutional autoencoders, and a second K-branch convolutional autoencoder that are independent of each other, the model structure of the second K-branch convolutional autoencoder is the same as the model structure after the first K-branch convolutional autoencoder and the K first single-branch convolutional autoencoders are combined, the K first single-branch convolutional autoencoders correspond one-to-one to the K branches in the first K-branch convolutional autoencoder, and K is an integer greater than 1; the first time domain module is a second single-branch convolutional autoencoder; A construction module is used to remove the decoder part of the second K-branch convolutional autoencoder after iterative training, and sequentially connect the first pooling layer, the first fully connected layer and the first classifier after the second K-branch convolutional autoencoder after the removal process to obtain a second frequency domain module, and remove the decoder part of the second single-branch convolutional autoencoder after iterative training, and sequentially connect the second pooling layer, the second fully connected layer and the second classifier after the second single-branch convolutional autoencoder after the removal process to obtain a second time domain module, wherein the second fully connected layer has the same number of neurons as the first fully connected layer, and both the second frequency domain module and the second time domain module have feature extraction and classification functions; A second training module is used to determine a target multi-scale fusion perception model based on a second sample data set, the second frequency domain module, and the second time domain module, wherein the second sample data in the second sample data set is a second sample frequency domain feature and a second sample time domain feature corresponding to a labeled sample signal, and the labeled sample signal includes a labeled ABP signal and a labeled CBFV signal; Among them, the target multi-scale fusion perception model can determine whether the CA function of any individual is damaged based on the frequency domain characteristics and time domain characteristics corresponding to the CA function of any individual.

Images