A method, system, and related device for determining the probability of depression.

Abstract

This invention discloses a method, system, and related apparatus for determining the probability of depression. The method includes: obtaining a dataset of N windows corresponding to the target EEG signal data of a patient to be tested through a preset filter; inputting the dataset into a trained prediction model to obtain corresponding prediction labels; and determining the probability of the patient suffering from depression based on the N prediction labels. The training set for the prediction model is the denoised data generated by training a preset conditional diffusion model based on EEG signal data, identification information, and labels indicating whether the patient suffers from depression, along with the initial EEG signal data of multiple subjects. By embedding identification information and label information as features into the diffusion process, individual patient characteristics and depression characteristics can be better incorporated, enabling the conditional diffusion model to generate more individualized EEG signals and improving the accuracy of the predicted probability values.

Description

Technical Field

[0001] This invention relates to the field of model prediction technology, and in particular to a method, system and related apparatus for determining the probability of depression. Background Technology

[0002] Existing diagnostic methods for depression include questionnaires, which are cumbersome, time-consuming, and prone to inaccuracies due to subjective user responses, ultimately affecting the accuracy of diagnosis. Alternatively, classification prediction models can be used, but clinical EEG data for depression is scarce, and these models often suffer from poor generalization and low accuracy.

[0003] Therefore, there is an urgent need for a predictive method to improve the accuracy of depression diagnosis. Summary of the Invention

[0004] The main objective of this invention is to provide a method, system, device, computer equipment, and storage medium for determining the probability of depression, which can solve the problem of low accuracy in probability prediction in the prior art.

[0005] To achieve the above objectives, the first aspect of the present invention provides a method for determining the probability of depression, the method comprising:

[0006] The brainwave signal data of the patient to be tested is acquired to obtain the target brainwave signal data. The target brainwave signal data is filtered using a preset filter to obtain a dataset with N windows after filtering.

[0007] N datasets are input into a preset prediction model to obtain N prediction labels. The prediction model is a model trained on a target training set. The target training set is a combination of a first training set and a second training set. The first training set includes the initial EEG signal data of multiple subjects. The second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data of multiple subjects, first information, and second information. The first information is the subject's identification information, and the second information is the label information indicating whether the subject has depression.

[0008] The number of target labels in the predicted labels is obtained, and the proportion of the number of target labels in the N predicted labels is determined as the probability that the patient to be tested has depression; wherein, the target label is a label that characterizes having depression.

[0009] In conjunction with the first aspect, in one possible implementation, denoised data is generated by using a conditional diffusion model based on initial EEG signal data, first information, and second information from multiple subjects. This includes: acquiring initial EEG signal data from multiple subjects; performing a global average reference on the initial EEG signal data to obtain reference EEG signal data; filtering the reference EEG signal data using a preset filter to obtain a filtered initial training set; inputting the initial training set into the conditional diffusion model; performing forward denoising on the initial training set to obtain forward-denoised data; inputting the first information and the second information into the conditional diffusion model to train the conditional diffusion model to predict the noise in the forward-denoised data, resulting in a trained conditional diffusion model; and inputting random noise, the subjects' first information, and second information into the trained conditional diffusion model to perform reverse denoising on the random noise, resulting in denoised generated data.

[0010] In conjunction with the first aspect, in one possible implementation, the aforementioned conditional diffusion model includes a first convolutional layer, a time-step embedding layer, and a feature embedding layer. The step of inputting the first information and the second information into the conditional diffusion model to train it to predict the noise in the forward-noised data, thereby obtaining the trained conditional diffusion model, includes: inputting the forward-noised data into the first convolutional layer to obtain first noisy data; processing the first noisy data using the SiLU activation function to obtain second noisy data; inputting the first information and the second information into the feature embedding layer to obtain corresponding first and second features; and inputting the time step into the time-step embedding layer to obtain a corresponding third feature; and training the conditional diffusion model to predict the noise in the forward-noised data based on the second noisy data, the first feature, the second feature, and the third feature, thereby obtaining the trained conditional diffusion model.

[0011] In conjunction with the first aspect, in one possible implementation, the conditional diffusion model further includes a linear layer and a dilated convolutional layer. The step of training the conditional diffusion model to predict noise in the forward-noisy data based on the second noisy data, the first feature, the second feature, and the third feature, to obtain the trained conditional diffusion model, includes: performing element-wise matrix addition on the first feature, the second feature, and the third feature to obtain a first matrix; inputting the first matrix into a linear layer for processing to obtain a corresponding second matrix; performing element-wise matrix addition on the second matrix and the second noisy data to obtain a third matrix; inputting the third matrix into a dilated convolutional layer to obtain a fourth matrix; performing convolution on the fourth matrix to obtain a feature matrix and the third noisy data; using the third noisy data as the second noisy data, and returning to perform element-wise matrix addition on the second matrix and the second noisy data to obtain the third matrix, until the loop repeats M times; training the conditional diffusion model to predict noise in the forward-noisy data based on the M feature matrices, to obtain the trained conditional diffusion model.

[0012] In conjunction with the first aspect, in one possible implementation, the aforementioned conditional diffusion model further includes a second convolutional layer and a third convolutional layer. The convolutional processing of the fourth matrix to obtain the feature matrix and the third noisy data includes: performing nonlinear transformations on the fourth matrix using the hyperbolic tangent function and the sigmoid activation function respectively to obtain a fifth matrix and a sixth matrix; performing element-wise matrix multiplication on the fifth matrix and the sixth matrix to obtain a seventh matrix; inputting the seventh matrix into the second convolutional layer to obtain the feature matrix; and inputting the seventh matrix into the third convolutional layer to obtain the third noisy data.

[0013] In conjunction with the first aspect, in one possible implementation, the aforementioned conditional diffusion model further includes a fourth convolutional layer and a fifth convolutional layer. The step of training the conditional diffusion model based on M feature matrices to predict noise in the forward-added noisy data, and obtaining the trained conditional diffusion model, includes: performing element-wise matrix addition on the obtained M feature matrices and a preset initial feature matrix to obtain a target feature matrix; inputting the target feature matrix into the fourth convolutional layer to obtain fourth-added noisy data; processing the fourth-added noisy data using the SiLU activation function to obtain fifth-added noisy data; inputting the fifth-added noisy data into the fifth convolutional layer to obtain predicted noise data; and training the conditional diffusion model based on the predicted noise data until convergence, thereby obtaining the trained conditional diffusion model.

[0014] To achieve the above objectives, a second aspect of the present invention provides a system for determining the probability of depression, the system comprising an electroencephalogram (EEG) signal acquisition device, a communication module, a host computer, and a server;

[0015] The host computer is connected to the EEG signal acquisition device and is used to send control signals characterizing the EEG signals of the patient to be tested to the EEG signal acquisition device; the EEG signal acquisition device is used to receive the control signals and acquire the EEG signal data of the patient to be tested.

[0016] The host computer and the EEG signal acquisition device are respectively connected to a communication module. The EEG signal acquisition device is also used to send the acquired target EEG signal data to the host computer through the communication module.

[0017] The host computer is also used to receive target EEG signal data and send the target EEG signal data to the server;

[0018] The server receives target EEG signal data and filters it using a preset filter to obtain a dataset with N windows. These N datasets are then input into a preset prediction model to obtain N prediction labels. The prediction model is trained on a target training set, which is a combination of a first training set and a second training set. The first training set includes initial EEG signal data from multiple subjects, and the second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data from multiple subjects, first information, and second information. The first information is the subject's identification number, and the second information is a label indicating whether the subject suffers from depression. The number of target labels in the prediction labels is obtained, and the proportion of the target labels among the N prediction labels is determined as the probability that the patient suffers from depression. The target labels are labels representing depression.

[0019] To achieve the above objectives, a third aspect of the present invention provides a device for determining the probability of depression, the device comprising:

[0020] Acquisition module: used to acquire EEG signal data of the patient to be tested, obtain target EEG signal data, filter the target EEG signal data using a preset filter, and obtain a dataset of N windows after filtering;

[0021] Label prediction module: used to input N datasets into a preset prediction model to obtain N predicted labels. The prediction model is a model trained on a target training set. The target training set is a set of a first training set and a second training set. The first training set includes the initial EEG signal data of multiple subjects. The second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data of multiple subjects, first information, and second information. The first information is the subject's identification information, and the second information is the label information indicating whether the subject has depression.

[0022] Probability prediction module: used to obtain the number of target labels in the predicted labels, and determine the proportion of the number of target labels in the N predicted labels as the probability that the patient to be tested has depression; wherein, the target label is a label that represents having depression.

[0023] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0024] The brainwave signal data of the patient to be tested is acquired to obtain the target brainwave signal data. The target brainwave signal data is filtered using a preset filter to obtain a dataset with N windows after filtering.

[0025] N datasets are input into a preset prediction model to obtain N prediction labels. The prediction model is a model trained on a target training set. The target training set is a combination of a first training set and a second training set. The first training set includes the initial EEG signal data of multiple subjects. The second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data of multiple subjects, first information, and second information. The first information is the subject's identification information, and the second information is the label information indicating whether the subject has depression.

[0026] The number of target labels in the predicted labels is obtained, and the proportion of the number of target labels in the N predicted labels is determined as the probability that the patient to be tested has depression; wherein, the target label is a label that characterizes having depression.

[0027] To achieve the above objectives, a fifth aspect of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0028] The brainwave signal data of the patient to be tested is acquired to obtain the target brainwave signal data. The target brainwave signal data is filtered using a preset filter to obtain a dataset with N windows after filtering.

[0029] N datasets are input into a preset prediction model to obtain N prediction labels. The prediction model is a model trained on a target training set. The target training set is a combination of a first training set and a second training set. The first training set includes the initial EEG signal data of multiple subjects. The second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data of multiple subjects, first information, and second information. The first information is the subject's identification information, and the second information is the label information indicating whether the subject has depression.

[0030] The number of target labels in the predicted labels is obtained, and the proportion of the number of target labels in the N predicted labels is determined as the probability that the patient to be tested has depression; wherein, the target label is a label that characterizes having depression.

[0031] The embodiments of the present invention have the following beneficial effects:

[0032] This invention provides a method for determining the probability of depression. By embedding the first and second information of multiple subjects as features into the diffusion process of a conditional diffusion model, it can better incorporate individual patient characteristics and depression characteristics. This allows the conditional diffusion model to generate more individualized EEG signals through denoising. Using the denoised EEG signals and the set of initial EEG signal data from multiple subjects as the training set of the prediction model can improve the accuracy of the prediction probability value. Furthermore, using the filtered dataset with N windows as the input of the prediction model can divide the overall target brain point signal data into more detailed data to obtain more accurate prediction labels, further improving the accuracy of the prediction probability value. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] in:

[0035] Figure 1 This is a flowchart illustrating a method for determining the probability of depression in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the training process of the conditional diffusion model in an embodiment of the present invention;

[0037] Figure 3 This is a flowchart illustrating a reverse denoising process in an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the prediction model in an embodiment of the present invention;

[0039] Figure 5 This is a structural block diagram of a device for determining the probability of depression in an embodiment of the present invention;

[0040] Figure 6 This is a structural block diagram of a system for determining the probability of depression according to an embodiment of the present invention;

[0041] Figure 7 This is a structural block diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] This invention provides a method for determining the probability of depression. This method is applicable to determining the probability that a patient has depression. Figure 1 , Figure 1 This is a flowchart illustrating a method for determining the probability of depression provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the specific steps of this method include the following:

[0044] Step S101: Obtain the EEG signal data of the patient to be tested, obtain the target EEG signal data, and filter the target EEG signal data using a preset filter to obtain a dataset with N windows after filtering.

[0045] Step S102: Input the N datasets into the preset prediction model to obtain the corresponding N prediction labels.

[0046] The prediction model is a model obtained by training on the target training set. The target training set is a combination of the first training set and the second training set. The first training set includes the initial EEG signal data of multiple subjects. The second training set includes the denoised generated data output by the conditional diffusion model based on the initial EEG signal data of multiple subjects, the first information, and the second information. The first information is the subject's identification information, and the second information is the label information indicating whether the subject has depression.

[0047] Step S103: Obtain the number of target labels in the predicted labels, and determine the proportion of the number of target labels in the N predicted labels as the probability that the patient to be tested has depression.

[0048] The target label is a label that characterizes someone who has depression.

[0049] In this embodiment, firstly, a preset conditional diffusion model is trained to obtain a conditional diffusion model for denoising the data. This conditional diffusion model specifically consists of a forward denoising process and a backward denoising process. In the forward denoising process, for each observed sample, a small amount of noise is continuously added to the sample until the sample is completely denoised as completely random noise, which follows a normal distribution. The backward denoising process randomly samples a noise sample from the normal distribution and continuously removes a small amount of noise from that sample until the noise sample becomes a generated sample resembling a real sample.

[0050] In this embodiment, a trained conditional diffusion model is obtained by training the conditional diffusion model, and denoised generation data is obtained based on the conditional diffusion model. (Refer to...) Figure 2 , Figure 2 This is a schematic diagram of the training process of the conditional diffusion model provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the steps for training the pre-defined conditional diffusion model are as follows:

[0051] Step S201: Obtain initial EEG signal data from multiple subjects, perform global averaging on the initial EEG signal data to obtain reference EEG signal data, and use a preset filter to filter the reference EEG signal data to obtain the filtered initial training set.

[0052] Step S202: Input the initial training set into the conditional diffusion model, add forward noise to the initial training set to obtain forward-noised data, and input the first information and the second information into the conditional diffusion model to train the conditional diffusion model to predict the noise of the forward-noised data, thereby obtaining the trained conditional diffusion model.

[0053] First, EEG signal data from multiple subjects were acquired. The subjects were patients whose depression status was known. Global average reference (CAR) was performed on the EEG signal data from multiple subjects to obtain reference EEG signal data. Global average reference changed the acquired signal from the original reference electrode-based to the reference signal based on the average signal.

[0054] Furthermore, the reference EEG signal data is filtered using a preset filter to obtain the initial training set. The filter can be a 50Hz band-stop FIR filter to remove high-energy 50Hz power frequency signals from the original signal. Then, a 1-50Hz bandpass FIR filter is used to retain signals containing EEG frequency components (Delta, Theta, Alpha, Beta, Gamma). Finally, these signals are segmented into 2-second non-overlapping windows. Data insufficient to form a 2-second window is discarded, resulting in N_trails of data points, which constitute the training set. For example, given a 3x2000 signal with a sampling rate of 250Hz, where 3 represents the channel and 2000 represents the number of data points, if each window is 3x500, then N_trails equals 4.

[0055] The initial training set is input into the conditional diffusion model, and forward noise is added to the initial training set to obtain forward-noised data. The conditional diffusion model is mainly used to amplify the EEG data samples to improve the accuracy and generalization of the subsequent classification prediction model. Since EEG signals have great heterogeneity in different populations, even under the same condition of having depression, different patients will show different EEG signals. Therefore, in order to better incorporate individual patient and depression characteristics, the subject's identification information and label information indicating whether they have depression are added to the diffusion process, so that the conditional diffusion model can generate more individualized EEG signals after denoising.

[0056] The conditional diffusion model includes a noise-adding module. The initial training set and random noise z0 sampled from a standard normal distribution are input into the conditional diffusion model. The noise-adding module performs forward noise-adding processing on the initial training set to obtain forward-noised data. Specifically, it... The original signal (E is the number of channels 3, L is the number of sampling points 500, initial training set) is denoised. The denoising module calculates the forward denoised data according to formula (2), which is based on the randomly generated time step t and the random noise z0 extracted from the standard normal distribution.

[0057] For the initial x0 and x t For example, the following relationship holds:

[0058]

[0059] Where α t =1-β t , Therefore, x is obtained. t The noise signal at time step t.

[0060] Where, β min β maxFor the hyperparameters of the predefined conditional diffusion model, in one possible implementation, βmin = 0.0001, βmax = 0.02, and t_step = 100.

[0061] In this embodiment, the first information and the second information are input into the conditional diffusion model to train the conditional diffusion model to predict the noise of the forward-noiseed data, thus obtaining the trained conditional diffusion model.

[0062] The specific steps for training the conditional diffusion model are shown in steps S301-S303:

[0063] Step S301: Input the forward-noise data into the first convolutional layer to obtain the first noise data, and process the first noise data using the SiLU activation function to obtain the second noise data.

[0064] Step S302: Input the first information and the second information into the feature embedding layer to obtain the corresponding first feature and the second feature, and input the time step into the time step embedding layer to obtain the corresponding third feature.

[0065] Step S303: Based on the second noisy data, the first feature, the second feature, and the third feature, train a conditional diffusion model to predict the noise in the forward noisy data, and obtain the trained conditional diffusion model.

[0066] The conditional diffusion model also includes a first convolutional layer, a time-step embedding layer, and a feature embedding layer. The forward-noised dataset is input into the first convolutional layer to obtain the first noisy data. The SiLU activation function is then applied to the first noisy data to obtain the second noisy data. The initial noisy dataset is convolved using a first convolutional layer of size (64, 3, 1), and then after passing through the SiLU activation function, a feature embedding layer of size (64, 1, 500) is obtained. This refers to the second noisy data. The second noisy data is represented in matrix form.

[0067] Simultaneously, the first and second information are input into the feature embedding layer to obtain the corresponding first and second features γ. * And by inputting the time step into the time step embedding layer, the corresponding third feature t is obtained. * Based on the second noisy data, the first feature, the second feature, and the third feature, a conditional diffusion model is trained to predict the noise in the forward noisy data, resulting in the trained conditional diffusion model. The first feature, the second feature, and the third feature are all represented by matrices.

[0068] For the subject ID and label information indicating whether they have depression in γ, feature embedding can be performed using the following formula:

[0069]

[0070] in Feature input representing subject identification information, The feature input represents the label information of the subjects, where m represents the number of subject IDs, n represents the number of subject labels, and k represents the number of feature dimensions in the embedding layer.

[0071] W represents a feature lookup table. For example, if the subject ID information is represented by a feature vector, and assuming the feature vector has a dimension of 1*k, then the feature vectors corresponding to the subject ID information are sequentially arranged into an m*k matrix, which yields the result. Based on the subject's number, it is possible to... Obtain the feature vector representing the subject's ID corresponding to the subject. Similarly, for the subject's label information, assuming a feature vector representing the label information has a dimension of 1*k, then according to the subject's ID, the feature vectors corresponding to the subject's label information are sequentially arranged into an n*k matrix, thus obtaining... Based on the subject number, it is possible to... The feature vectors that obtain the representational label information of the subjects are obtained.

[0072] The embedding layer at time step t is obtained by the following formula, where N is 512:

[0073] θ={X|x=2n,n∈N) (5)

[0074]

[0075] θ represents an even-numbered queue, t represents the current time step, ranging from 0 to 99, and Te represents a 2*N matrix.

[0076] The steps for training a conditional diffusion model based on the second noisy data, the first feature, the second feature, and the third feature to predict the noise in the forward noisy data, and obtaining the trained conditional diffusion model, are shown in steps S401-S402:

[0077] Step S401: Perform element-wise matrix addition on the first feature, the second feature, and the third feature to obtain a first matrix. Input the first matrix into a linear layer for processing to obtain a corresponding second matrix. Perform element-wise matrix addition on the second matrix and the second noisy data to obtain a third matrix. Input the third matrix into a dilated convolutional layer to obtain a fourth matrix.

[0078] Step S402: Perform convolution processing on the fourth matrix to obtain a feature matrix and third noise data; use the third noise data as the second noise data, and return to execute the step of performing matrix element-wise addition operation between the second matrix and the second noise data to obtain the third matrix, until the loop is repeated M times; train the conditional diffusion model based on the M feature matrices to predict the noise of the forward noise data, and obtain the trained conditional diffusion model.

[0079] The conditional diffusion model also includes a residual feature extraction module, which comprises linear layers and dilated convolutional layers, to extract the first feature and the second feature γ. * and the third feature t * Element-wise addition of matrices is performed to obtain a first matrix. The first matrix is ​​then input into a linear layer to obtain a corresponding second matrix. Element-wise addition of the second matrix with the second noisy data is performed to obtain a third matrix. The third matrix is ​​then input into a dilated convolutional layer to obtain a fourth matrix. In one possible implementation, the first matrix E is obtained from equation (7). The target matrix is ​​input into a linear layer to obtain a corresponding first matrix. Element-wise addition of the first matrix with the second noisy data is performed to obtain a third matrix. The third matrix is ​​then input into a dilated convolutional layer to obtain a fourth matrix.

[0080] E=T e +E s +E l (7)

[0081] The fourth matrix is ​​convolved to obtain the feature matrix and the third noisy data. The third noisy data is used as the second noisy data. The process is repeated to perform element-wise matrix addition on the second matrix and the second noisy data to obtain the third matrix. This process is repeated M times to obtain M feature matrices. Based on these M feature matrices, a conditional diffusion model is trained to predict the noise in the forward noisy data, resulting in the trained conditional diffusion model. The features corresponding to t and γ after the embedding layer are obtained according to formula (3-6). and Will t * γ * Together, they are used as input parameters and fed into the residual feature extraction module, where t... * and γ * The features are processed in a linear layer and finally combined with... Perform element-wise matrix addition to obtain the embedded third matrix, and then pass it through a dilated convolutional layer to obtain the fourth matrix (128, 1, 500).

[0082] The steps for performing convolution on the fourth matrix to obtain the feature matrix and the third noisy data are shown in steps S501-S502:

[0083] Step S501: After performing nonlinear transformations on the fourth matrix using the hyperbolic tangent function and the Sigmoid activation function respectively, the fifth and sixth matrices are obtained.

[0084] Step S502: Perform element-wise matrix multiplication on the fifth and sixth matrices to obtain the seventh matrix. Input the seventh matrix into the second convolutional layer to obtain the feature matrix. Input the seventh matrix into the third convolutional layer to obtain the third noisy data.

[0085] The residual feature extraction module also includes a second convolutional layer and a third convolutional layer. After performing nonlinear transformations on the fourth matrix using the hyperbolic tangent function and the Sigmoid activation function, the fifth and sixth matrices are obtained. The fifth and sixth matrices are then multiplied element-wise to obtain the seventh matrix. The seventh matrix is ​​then input into the second convolutional layer to obtain the feature matrix. The seventh matrix is ​​then input into the third convolutional layer to obtain the third noisy data.

[0086] The third noisy data is used as the second noisy data. The process is repeated until the second matrix is ​​obtained by performing element-wise matrix addition on the second noisy data. The conditional diffusion model is trained based on the M feature matrices to predict the noise of the forward noisy data, and the trained conditional diffusion model is obtained.

[0087] In this embodiment, the steps for training a conditional diffusion model based on M feature matrices to predict noise in forward-noiseed data and obtaining the trained conditional diffusion model are shown in steps S601-S602:

[0088] Step S601: Perform element-wise addition on the obtained M feature matrices and the preset initial feature matrix to obtain the target feature matrix. Input the target feature matrix into the fourth convolutional layer to obtain the fourth noisy data.

[0089] Step S602: The fourth noisy data is processed using the SiLU activation function to obtain the fifth noisy data. The fifth noisy data is then input into the fifth convolutional layer to obtain the predicted noise data. The conditional diffusion model is trained based on the predicted noise data until convergence, thus obtaining the trained conditional diffusion model.

[0090] The conditional diffusion model also includes a noise extraction module, which further includes a fourth convolutional layer and a fifth convolutional layer. The obtained M feature matrices are subjected to element-wise addition to obtain the target feature matrix. The target feature matrix is ​​then input into the fourth convolutional layer to obtain the fourth noisy data. The SiLU activation function is used to process the fourth noisy data to obtain the fifth noisy data. The fifth noisy data is then input into the fifth convolutional layer to obtain the predicted noise data.

[0091] The conditional diffusion model is trained until convergence based on the predicted noise data and random noise sampled from the standard normal distribution. Specifically, the predicted noise data and random noise are input into the MSE loss function for calculation. The MSE loss function is iterated until it converges, and the conditional diffusion model when the calculated value of the MSE loss function reaches a stable value is obtained. This completes the training of the conditional diffusion model and yields the trained target model.

[0092] After nonlinear transformations using the tanh function (hyperbolic tangent) and the sigmoid function, the network passes through two different convolutional layers, outputting two identical (64, 1, 500) matrices. One matrix is ​​used as the input to the next residual feature extraction module, and the other is added to the feature matrix. After calculation by the residual network extraction module with M layers (e.g., M = 32), M feature matrices are obtained. The M feature matrices and the preset initial feature matrix are then subjected to element-wise matrix addition to obtain a target feature matrix. This target feature matrix is ​​then fed into the noise extraction module to obtain noise data of the same size as the original input. The MSE loss is obtained according to equation (8), and the network is then trained. θ .

[0093]

[0094] Training model ∈ θ Essentially, at each time step t, based on the sample's input label γ and the currently predicted noise data... To predict the random noise ∈ added in the current step, such that equation (8) is minimized, where The mean squared error (MSE) loss function is used.

[0095] After the conditional diffusion model is trained, step S203 is executed: random noise, the first information and the second information of the subject are input into the trained conditional diffusion model, and the random noise is reversed to obtain the denoised generated data.

[0096] Random noise extracted from a standard normal distribution, along with the subject's first and second information, is input into a trained conditional diffusion model to perform reverse denoising on the random noise, resulting in denoised generated data.

[0097] Reference Figure 3 , Figure 3 This is a flowchart illustrating a reverse denoising process provided in an embodiment of the present invention, as shown below. Figure 3 As shown, in the reverse denoising process, the subject's ID, label information, time step t, and Gaussian signal x are used as the basis for the denoising process. t (Random noise), sampled according to the following formula:

[0098]

[0099] In this embodiment, the time step t ranges from 99 to 0, totaling 100 time steps, to handle the random noise x. t The time step t and the label γ are input together into the network ∈ θ In the middle, and according to formula (8), the denoised signal x is obtained. t-1 This process is repeated until x0 is generated. By sampling each subject, this method obtains K times the number of original samples. These generated data are then merged with the original EEG signal data to obtain the final data-enhanced sample.

[0100] Specifically, the initial EEG signal data of multiple subjects and the denoised generated data output by the conditional diffusion model are merged to obtain a target training set. The prediction model obtained by training a preset model using the target training set is then used. For example, the preset model is a neural network model.

[0101] The EEG signal data of the patient to be tested is acquired to obtain the target EEG signal data. The target EEG signal data is then globally averaged and referenced to obtain the target reference EEG signal data. The target reference EEG signal data is then filtered using a preset filter to obtain a dataset with N windows.

[0102] The datasets of N windows are input into the trained prediction model, resulting in N predicted labels. The number of target labels in the predicted labels is obtained, and the proportion of the target label among the N predicted labels is used to determine the probability that the patient has depression. Here, the target label is a label representing depression. For example, if N is 4, and the number of target labels in the predicted labels is 3, then the probability that the patient has depression is...

[0103] In one possible implementation, a pre-defined prediction model calculates the corresponding predicted probability values ​​based on N target noisy data inputs. Predicted probability values ​​greater than a pre-defined probability threshold are output as target labels, while predicted probability values ​​less than or equal to the pre-defined probability threshold are output as non-target labels. Non-target labels indicate that the test patient does not have depression. In other words, when the predicted probability value corresponding to the target noisy data is greater than the pre-defined probability threshold, the output label is the target label; when the predicted probability value is less than or equal to the pre-defined probability threshold, the output label is the non-target label. In another possible implementation, N target noisy data are input into a pre-defined prediction model, resulting in N predicted probability values. The number of predicted probability values ​​greater than a pre-defined probability threshold is obtained, representing the target number. The proportion of the target number to the total number of predicted probability values ​​is determined as the probability that the test patient has depression.

[0104] Reference Figure 4 , Figure 4 This is a schematic diagram of the prediction model provided in an embodiment of the present invention, such as... Figure 4 As shown, in the prediction model, for an input dataset with N windows... The overall computation employs three convolutional layers and one fully connected layer. Following each convolutional layer, a four-step process—BatchNormalized normalization, ReLU activation, average pooling, and Dropout—forms the general processing module. For the first convolutional layer (temporal convolution module), a kernel of size (1, 60) is used to perform convolution operations on each channel signal, extracting features from each temporal signal. Zeros are padded at both ends to ensure the final extracted feature size remains consistent with the original. This extracted feature is then fed into the general processing module for output. The second convolutional layer (temporal convolution module) then... The channel domain convolution module uses a (3,1) channel domain convolution kernel to merge the channels of the three features to obtain a one-dimensional feature vector, which is then sent to the general processing module for output. The third convolutional layer uses a (1,10) and (1,1) convolution kernel to extract the one-dimensional feature vector in sequence, which is then output through the general processing module. Finally, a fully connected layer is used to convert the size of the feature vector to 2. The final prediction model will output the probability values ​​corresponding to N_trails windows, count the number Q of N_trails probabilities greater than the probability threshold, and finally calculate the probability of having depression based on (Q / N_trails). The above is the reasoning process of the model. The training set for the prediction model consisted of original clinically collected EEG signals and EEG data generated by the subsequent conditional diffusion model. The training process used cross-entropy as the loss function and Adam optimizer, with a learning rate of 0.0001, a regularization parameter of 0.01, a batch size of 15, a dropout of 0.2, and a maximum number of epochs of 100. The training process used the PyTorch framework, an NVIDIA A40 graphics card, and Ubuntu 20.04 LST as the operating system.

[0105] Based on the above method, embedding the first and second information as features into the diffusion process can better incorporate individual patient characteristics and depression characteristics, enabling the conditional diffusion model to generate more individualized EEG signals through denoising. In addition, using the filtered dataset with N windows as input to the prediction model can divide the overall target brain point signal data into more detailed data to obtain more accurate prediction labels and improve the accuracy of prediction probability values.

[0106] To better implement the above method, embodiments of the present invention provide a device for determining the probability of depression, referring to... Figure 5 , Figure 5 This is a structural block diagram of a device for determining the probability of depression provided in an embodiment of the present invention, as shown below. Figure 5 As shown, the device 50 specifically includes:

[0107] Acquisition module 501: used to acquire EEG signal data of the patient to be tested, obtain target EEG signal data, filter the target EEG signal data using a preset filter, and obtain a dataset of N windows after filtering.

[0108] Label prediction module 502: used to input N datasets into a preset prediction model to obtain N predicted labels, wherein the prediction model is a model obtained by training on a target training set, the target training set is a set of a first training set and a second training set, the first training set includes the initial EEG signal data of multiple subjects, the second training set includes the denoised generated data output by the conditional diffusion model based on the initial EEG signal data of multiple subjects, the first information and the second information, the first information is the subject's identification information, and the second information is the label information indicating whether the subject has depression;

[0109] Probability prediction module 503: used to obtain the number of target labels in the predicted labels, and determine the proportion of the number of target labels in the N predicted labels as the probability that the patient to be tested has depression; wherein, the target label is a label that represents having depression.

[0110] In one possible design, the acquisition module 501 is specifically used to: acquire initial EEG signal data from multiple subjects, perform global averaging on the initial EEG signal data to obtain reference EEG signal data, and filter the reference EEG signal data using a preset filter to obtain a filtered initial training set.

[0111] In one possible design, the device 50 further includes a conditional diffusion model data generation module 504, which is specifically used to: input the initial training set into the conditional diffusion model, perform forward noise addition on the initial training set to obtain forward-noised data, and input the first information and the second information into the conditional diffusion model to train the conditional diffusion model to predict the noise of the forward-noised data, thereby obtaining a trained conditional diffusion model; and input random noise, the first information and the second information of the subject into the trained conditional diffusion model, and perform reverse denoising on the random noise to obtain denoised generated data.

[0112] In one possible design, the label prediction module 502 is specifically used to: train a preset model using the target training set to obtain the trained prediction model.

[0113] In one possible design, the conditional diffusion model includes a first convolutional layer, a time-step embedding layer, and a feature embedding layer. The conditional diffusion model data generation module 504 is specifically used to: input the forward-noised data into the first convolutional layer to obtain first noisy data; process the first noisy data using the SiLU activation function to obtain second noisy data; input the first information and the second information into the feature embedding layer to obtain corresponding first features and second features; input the time step into the time-step embedding layer to obtain corresponding third features; and train the conditional diffusion model based on the second noisy data, the first feature, the second feature, and the third feature to predict the noise of the forward-noisy data, thereby obtaining the trained conditional diffusion model.

[0114] In one possible design, the conditional diffusion model further includes a linear layer and a dilated convolutional layer. Specifically, the conditional diffusion model data generation module 504 is used to: input the first feature, second feature, and third feature into the linear layer for processing to obtain a corresponding first matrix; perform element-wise matrix addition on the first matrix and the second noisy data to obtain a third matrix; input the third matrix into the dilated convolutional layer to obtain a fourth matrix; perform convolution processing on the fourth matrix to obtain a feature matrix and third noisy data; use the third noisy data as the second noisy data, and return to execute the step of performing element-wise matrix addition on the first matrix and the second noisy data to obtain the third matrix, repeating this process M times; and train the conditional diffusion model based on the M feature matrices to predict the noise in the forward noisy data, thus obtaining the trained conditional diffusion model.

[0115] In one possible design, the conditional diffusion model further includes a second convolutional layer and a third convolutional layer. The conditional diffusion model data generation module 504 is specifically used to: perform nonlinear transformations on the fourth matrix using the hyperbolic tangent function and the sigmoid activation function respectively to obtain the fifth matrix and the sixth matrix; perform element-wise matrix multiplication on the fifth matrix and the sixth matrix to obtain the seventh matrix; input the seventh matrix into the second convolutional layer to obtain the feature matrix; and input the seventh matrix into the third convolutional layer to obtain the third noisy data.

[0116] In one possible design, the conditional diffusion model further includes a fourth convolutional layer and a fifth convolutional layer. The conditional diffusion model data generation module 504 is specifically used to: perform element-wise matrix addition on the obtained M feature matrices and the preset initial feature matrix to obtain the target feature matrix; input the target feature matrix into the fourth convolutional layer to obtain the fourth noisy data; use the SiLU activation function to process the fourth noisy data to obtain the fifth noisy data; input the fifth noisy data into the fifth convolutional layer to obtain the predicted noise data; and train the conditional diffusion model based on the predicted noise data until convergence to obtain the trained conditional diffusion model.

[0117] Based on the aforementioned device, a more accurate prediction of the probability of depression can be made.

[0118] This invention provides a system for determining the probability of depression, referring to... Figure 6 , Figure 6 A structural block diagram of a depression probability determination system provided in an embodiment of the present invention is shown below. Figure 6As shown, the system 60 includes an EEG signal acquisition device 601, a communication module 602, a host computer 603, and a server 604. The EEG signal acquisition device is connected to the host computer, and the host computer and the EEG signal acquisition device are respectively connected to the communication module. The host computer sends control signals representing the acquisition of EEG signals from the patient to the EEG signal acquisition device. The EEG signal acquisition device receives the control signals, acquires the EEG signal data from the patient, and sends the acquired target EEG signal data to the host computer via the communication module. The communication module may include Bluetooth, WIFI, USB, etc. The host computer also receives the target EEG signal data and sends it to the server. The server receives the target EEG signal data and filters it using a preset filter to obtain N filtered EEG signal data. The system generates a dataset with a set of N data points. N datasets are input into a pre-defined prediction model to obtain N predicted labels. The prediction model is trained on a target training set, which is a combination of a first training set and a second training set. The first training set includes initial EEG signal data from multiple subjects, and the second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data, first information, and second information from multiple subjects. The first information is the subject's identification number, and the second information is a label indicating whether the subject has depression. The system obtains the number of target labels in the predicted labels and determines the proportion of the target labels among the N predicted labels as the probability that the patient has depression. The target label is a label representing depression. The server also outputs the calculated probability of the patient having depression to a host computer for display.

[0119] The EEG signal acquisition device can be controlled by a microcontroller and a 24-bit high-precision ADC acquisition circuit. The sampling chip can use an 8-channel synchronous sampling ADS1299 chip, which can simultaneously acquire microvolt-level EEG signals from three leads: Fp1, Fz, and Fp2. Their positions are determined according to the 10-20 international standard lead system. The sampling rate is 250Hz. The acquired EEG signal can be transmitted to the host computer software for processing via USB, Bluetooth, or WIFI. When the acquisition ends, the software will persist the recorded EEG signal. The host computer software transmits and returns persistently stored EEG signal data to and from the server. Based on the TCP / IP protocol for byte-stream transmission, it uses the Socket application programming interface to achieve data signal transmission and return from the software to the server. In this case, the software acts as the client, transmitting and returning data to and from the server. Both the client and server use little-endian encoding and decoding. The specific steps are as follows: First, the server establishes a fixed IP address and port number. The client connects using the server's IP address and port number. Then, the client transmits the persistent EEG data file... The system reads and calculates the total number of bytes N of EEG data to be transmitted, and transmits this number of bytes N to the server via TCP / IP protocol. Once the server receives the number of bytes N, it begins listening for and receiving the next N bytes of data. At this point, the client will begin sending N bytes of EEG signal data. After receiving all the data, the server will use an online data preprocessing module and an online depression diagnosis module based on convolutional neural networks to analyze the data and finally obtain a predicted probability value (the probability value ranges from 0 to 1, with higher values ​​indicating a greater likelihood of depression). This predicted probability value is then sent back to the client and displayed on the client, providing the final diagnosis result. The online data preprocessing module is used to filter the target EEG signal data using a preset filter to obtain a dataset with N windows. These N datasets are then input into a preset prediction model to obtain N prediction labels. The prediction model is trained on a target training set, which is a combination of a first training set and a second training set. The first training set includes initial EEG signal data from multiple subjects, and the second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data from multiple subjects, first information, and second information. The first information is the subject's identification number, and the second information is a label indicating whether the subject suffers from depression. The number of target labels in the prediction labels is obtained, and the proportion of the target labels among the N prediction labels is determined as the probability that the patient suffers from depression. The target labels are labels representing depression.

[0120] Figure 6An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 6 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program causes the processor to perform all the steps of the above-described method. The internal memory may also store a computer program, which, when executed by the processor, causes the processor to perform all the steps of the above-described method. Those skilled in the art will understand that... Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0121] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the aforementioned method.

[0122] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the steps of the aforementioned method.

[0123] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0124] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for determining the probability of depression, characterized in that, The method includes: The brainwave signal data of the patient to be tested is acquired to obtain the target brainwave signal data. The target brainwave signal data is filtered using a preset filter to obtain a dataset with N windows after filtering. N datasets are input into a preset prediction model to obtain N prediction labels. The prediction model is a model trained on a target training set. The target training set is a combination of a first training set and a second training set. The first training set includes the initial EEG signal data of multiple subjects. The second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data of multiple subjects, first information, and second information. The first information is the subject's identification information, and the second information is the label information indicating whether the subject has depression. The number of target labels in the predicted labels is obtained, and the proportion of the number of target labels in the N predicted labels is determined as the probability that the patient to be tested has depression; wherein, the target label is a label that characterizes having depression.

2. The method according to claim 1, characterized in that, Data was generated by denoising and outputting initial EEG signal data, first information, and second information from multiple subjects using a conditional diffusion model, including: Initial EEG signal data from multiple subjects were acquired, and the initial EEG signal data were globally averaged to obtain reference EEG signal data. The reference EEG signal data was then filtered using a preset filter to obtain the filtered initial training set. The initial training set is input into the conditional diffusion model, and forward noise is added to the initial training set to obtain forward-noise data. The first information and the second information are then input into the conditional diffusion model to train the conditional diffusion model to predict the noise of the forward-noise data, thereby obtaining the trained conditional diffusion model. The random noise, the subject's first information, and the second information are then input into the trained conditional diffusion model to perform reverse denoising on the random noise, resulting in denoised generated data.

3. The method according to claim 2, characterized in that, The conditional diffusion model includes a first convolutional layer, a time-step embedding layer, and a feature embedding layer. The first information and the second information are input into the conditional diffusion model to train it to predict noise in the forward-added noisy data, resulting in a trained conditional diffusion model. The forward-added noisy data is input into the first convolutional layer to obtain the first noisy data. The first noisy data is then processed using the SiLU activation function to obtain the second noisy data. The first information and the second information are input into the feature embedding layer to obtain the corresponding first feature and the second feature, and the time step is input into the time step embedding layer to obtain the corresponding third feature; Based on the second noisy data, the first feature, the second feature, and the third feature, a conditional diffusion model is trained to predict the noise in the forward noisy data, thus obtaining the trained conditional diffusion model.

4. The method according to claim 3, characterized in that, The conditional diffusion model further includes linear layers and dilated convolutional layers. The process of training the conditional diffusion model based on the second noisy data, the first feature, the second feature, and the third feature to predict the noise in the forward noisy data, resulting in the trained conditional diffusion model, includes: The first feature, the second feature, and the third feature are subjected to element-wise matrix addition to obtain a first matrix. The first matrix is ​​then input into a linear layer for processing to obtain a corresponding second matrix. The second matrix is ​​then subjected to element-wise matrix addition with the second noisy data to obtain a third matrix. The third matrix is ​​then input into a dilated convolutional layer to obtain a fourth matrix. The fourth matrix is ​​convolved to obtain a feature matrix and the third noisy data; the third noisy data is used as the second noisy data, and the step of performing element-wise matrix addition on the second matrix and the second noisy data to obtain the third matrix is ​​repeated until M times; a conditional diffusion model is trained based on the M feature matrices to predict the noise of the forward noisy data, and the trained conditional diffusion model is obtained.

5. The method according to claim 4, characterized in that, The conditional diffusion model further includes a second convolutional layer and a third convolutional layer. The convolutional processing of the fourth matrix to obtain the feature matrix and the third noisy data includes: After performing nonlinear transformations on the fourth matrix using the hyperbolic tangent function and the sigmoid activation function respectively, the fifth and sixth matrices are obtained accordingly. The fifth and sixth matrices are multiplied element-wise to obtain the seventh matrix. The seventh matrix is ​​then input into the second convolutional layer to obtain the feature matrix. Finally, the seventh matrix is ​​input into the third convolutional layer to obtain the third noisy data.

6. The method according to claim 5, characterized in that, The conditional diffusion model further includes a fourth convolutional layer and a fifth convolutional layer. The conditional diffusion model, trained based on M feature matrices, predicts noise in the forward-noised data, resulting in the trained conditional diffusion model, including: The obtained M feature matrices and the preset initial feature matrix are subjected to element-wise matrix addition to obtain the target feature matrix. The target feature matrix is ​​then input into the fourth convolutional layer to obtain the fourth noisy data. The fourth noisy data is processed using the SiLU activation function to obtain the fifth noisy data. The fifth noisy data is then input into the fifth convolutional layer to obtain the predicted noise data. Based on the predicted noise data, the conditional diffusion model is trained until convergence to obtain the trained conditional diffusion model.

7. A system for determining the probability of depression, characterized in that, The system includes an EEG signal acquisition device, a communication module, a host computer, and a server; The host computer is connected to the EEG signal acquisition device and is used to send control signals that characterize the EEG signals of the patient to be tested to the EEG signal acquisition device. The EEG signal acquisition device is used to receive control signals and acquire EEG signal data of the patient to be tested; The host computer and the EEG signal acquisition device are respectively connected to a communication module. The EEG signal acquisition device is also used to send the acquired target EEG signal data to the host computer through the communication module. The host computer is also used to receive target EEG signal data and send the target EEG signal data to the server; The server receives target EEG signal data and filters it using a preset filter to obtain a dataset with N windows. These N datasets are then input into a preset prediction model to obtain N prediction labels. The prediction model is trained on a target training set, which is a combination of a first training set and a second training set. The first training set includes initial EEG signal data from multiple subjects, and the second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data from multiple subjects, first information, and second information. The first information is the subject's identification number, and the second information is a label indicating whether the subject suffers from depression. The number of target labels in the prediction labels is obtained, and the proportion of the target labels among the N prediction labels is determined as the probability that the patient suffers from depression. The target labels are labels representing depression.

8. A device for determining the probability of depression, characterized in that, The device includes: Acquisition module: used to acquire EEG signal data of the patient to be tested, obtain target EEG signal data, filter the target EEG signal data using a preset filter, and obtain a dataset of N windows after filtering; Label prediction module: used to input N datasets into a preset prediction model to obtain N predicted labels. The prediction model is a model trained on a target training set. The target training set is a set of a first training set and a second training set. The first training set includes the initial EEG signal data of multiple subjects. The second training set includes denoised data generated by a conditional diffusion model based on the initial EEG signal data of multiple subjects, first information, and second information. The first information is the subject's identification information, and the second information is the label information indicating whether the subject has depression. Probability prediction module: used to obtain the number of target labels in the predicted labels, and determine the proportion of the number of target labels in the N predicted labels as the probability that the patient to be tested has depression; wherein, the target label is a label that represents having depression.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, the processor performs the steps of the method as described in any one of claims 1 to 6.

10. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 6.

Images