System for predicting the efficacy of deep brain stimulation surgery, medium and device

Abstract

The application relates to a deep brain stimulation surgery effect prediction system, medium and equipment, and particularly relates to the technical field of medical treatment. The system comprises the following: a data acquisition module, which is used for collecting electroencephalogram data and multi-modal neuroimaging data of a target object before surgery; collecting effect evaluation data of the target object under a target deep brain stimulation surgery scheme; a data preprocessing module, which is used for performing index calculation on the electroencephalogram data and the multi-modal neuroimaging data to obtain a feature matrix, wherein the feature matrix comprises an electroencephalogram data index feature matrix and an imaging index feature matrix; a feature fusion module, which is used for inputting the feature matrix into a self-encoder module to perform feature fusion, and the self-encoder module is used for converting the feature matrix into a latent representation; and a model training module, which is used for constructing a deep brain stimulation surgery effect prediction model based on the latent representation and the effect evaluation data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical treatment, in particular to a deep brain stimulation surgery efficacy prediction system, medium and equipment. BACKGROUND

[0002] Disorders of consciousness is an important but often neglected area in neurology, including cases of impaired or absent consciousness caused by severe brain injury. The two main subtypes of disorders of consciousness are minimally conscious state, characterized by inconsistent but reproducible evidence of consciousness, and unresponsive wakefulness syndrome, also known as vegetative state, characterized by a lack of awareness of self or environment.

[0003] Deep brain stimulation (DBS) is an invasive neurosurgical technique for neuromodulation in patients with disorders of consciousness. DBS parameters are adjustable and can be fine-tuned to optimize the patient's ability to respond to self-awareness and the surrounding environment. DBS surgical implant site options include the center-mesencephalic-parafascicular complex (cm-pf), the centromedian (CT), and the globus pallidus.

[0004] However, current DBS in clinical treatment of patients with disorders of consciousness faces many challenges, including patient selection, determining the optimal timing of DBS treatment, and selecting appropriate target points and stimulation parameters. Therefore, how to predict the efficacy of deep brain stimulation surgery to help patients choose the appropriate deep brain stimulation scheme needs to be solved. SUMMARY

[0005] The present application provides a deep brain stimulation surgery efficacy prediction system, medium and equipment, and the technical scheme is as follows.

[0006] In one aspect, a deep brain stimulation surgery efficacy prediction system is provided, the system comprising:

[0007] A data acquisition module is configured to collect preoperative electroencephalogram (EEG) data and multi-modal neuroimaging data of a target object, and collect efficacy evaluation data of the target object under a target deep brain stimulation (DBS) surgery scheme.

[0008] A data preprocessing module is configured to perform index calculation on the EEG data and multi-modal neuroimaging data to obtain a feature matrix, the feature matrix being used to quantify the level of consciousness of the target object, and the feature matrix comprising an EEG data index feature matrix and an imaging index feature matrix.

[0009] A feature fusion module is configured to input the feature matrix into a self-encoder module for feature fusion, and the self-encoder module is configured to convert the feature matrix into a latent representation.

[0010] a model training module configured to construct a prediction model of the efficacy of deep brain stimulation surgery based on the latent representation and the efficacy evaluation data, the prediction model being configured to predict the efficacy of various deep brain stimulation surgery schemes for the patient, the prediction model comprising a cascade of an input layer, a Stem layer, a plurality of mobile reverse bottleneck convolution modules, and a compression-excitation module, the compression-excitation module comprising a cascade of a global average pooling layer, a fully connected layer, and an activation function;

[0011] The input layer is configured to receive the latent representation; the Stem layer is configured to perform first-step feature extraction on the latent representation to obtain a first extracted feature map; the mobile reverse bottleneck convolution module is configured to perform second-step feature extraction on the first extracted feature map to obtain a second extracted feature map; and the compression-excitation module is configured to output a prediction result based on the second extracted feature map. In the compression-excitation module, the global average pooling layer is configured to compress the second extracted feature map to obtain a channel descriptor, the fully connected layer is configured to perform channel weighting on the channel descriptor to obtain a weight-applied feature map, and the activation function is configured to convert the weight-applied feature map into a probability output to obtain the prediction result.

[0012] In a possible implementation, in the data preprocessing module, the index calculation on the electroencephalogram data and the multi-modal neuroimaging data to obtain the feature matrix comprises:

[0013] The electroencephalogram data is subjected to index calculation in the time domain, the frequency domain, and the time-frequency domain respectively to obtain an electroencephalogram data index feature matrix composed of time domain features, frequency domain features, and time-frequency domain features;

[0014] The multi-modal neuroimaging data is subjected to index calculation in structural magnetic resonance imaging, functional magnetic resonance imaging, and diffusion magnetic resonance imaging respectively to obtain an imaging index feature matrix composed of structural magnetic resonance features, functional magnetic resonance features, and diffusion magnetic resonance features.

[0015] In a possible implementation, the autoencoder module comprises an encoder input layer, an encoder hidden layer, a decoder input layer, and a decoder hidden layer.

[0016] The encoder input layer is configured to receive the feature matrix.

[0017] The encoder hidden layer is configured to convert the feature matrix into a latent representation.

[0018] The decoder input layer is configured to receive the latent representation.

[0019] The decoder hidden layer is configured to restore the latent representation to obtain a reconstruction output.

[0020] In a possible implementation, in the feature fusion module, a mean square error is used as a loss function:

[0021] L= ;

[0022] where N is the number of samples, is the i-th feature matrix, is the i-th reconstruction output.

[0023] In a possible implementation, the encoder hidden layer performs data processing through the following formula:

[0024] ;

[0025] where W1 is a weight matrix, b1 is a bias, f is a ReLU activation function, X is a feature matrix, and Z is a latent representation.

[0026] In a possible implementation, the decoder hidden layer performs data processing through the following formula:

[0027] ;

[0028] where W2 is a weight matrix, b2 is a bias, g is an activation function, Z is a latent representation, and X' is a reconstruction output.

[0029] In a possible implementation, the mobile reverse bottleneck convolution module comprises:

[0030] a first point-by-point convolution module, configured to perform point-by-point convolution on an initial first extracted feature map to obtain a first extracted feature map with increased channels;

[0031] a deep convolution module, configured to independently perform convolution on each channel of the first extracted feature map to obtain a first extracted feature map after feature extraction;

[0032] a second point-by-point convolution module, configured to perform point-by-point convolution on the first extracted feature map after feature extraction to obtain a first extracted feature map with restored channels;

[0033] a skip connection module, configured to connect the initial first extracted feature map and the first extracted feature map with restored channels to obtain the second extracted feature map.

[0034] In a possible implementation, the first point-by-point convolution module comprises a Swish activation function.

[0035] In another aspect, a computer readable medium is provided, and the computer readable medium has a computer program stored thereon, where the program, when executed by a processor, implements the prediction method corresponding to the system.

[0036] In another aspect, a computer device is provided, comprising: one or more processors; a storage device storing one or more programs, which when executed by the one or more processors, cause the one or more processors to implement the prediction method corresponding to the system described above.

[0037] The technical solutions provided in the present application can include the following beneficial effects:

[0038] A deep brain stimulation surgery effect prediction system is provided, EEG data and multi-modal neural image data are collected, the EEG data and the multi-modal neural image data are subjected to index calculation to obtain a feature matrix, the feature matrix is fused through a self-encoder module, and then input into a prediction model composed of a cascaded input layer, a Stem layer, a plurality of mobile reverse bottleneck convolution modules and a compression-excitation module, so as to realize preoperative prediction of the deep brain stimulation surgery effect, and provide important model support for clinical doctors to formulate clinical strategies. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0040] Figure 1 is a structural block diagram of a deep brain stimulation surgery effect prediction system according to an exemplary embodiment.

[0041] Figure 2 is a structural block diagram of a prediction model according to an exemplary embodiment.

[0042] Figure 3 is a schematic diagram of a deep brain stimulation surgery effect prediction process according to an exemplary embodiment.

[0043] Figure 4 is a schematic diagram of a computer device according to an exemplary embodiment. DETAILED DESCRIPTION

[0044] The technical solutions of the present application will be described in detail below with reference to the drawings. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0045] It should be understood that the "indication" mentioned in the embodiments of the present application can be direct indication, or indirect indication, or can be an indication of an associated relationship. For example, A indicates B, which can mean that B can be obtained by A directly; or A indirectly indicates B, for example, A indicates C, and B can be obtained by C; or A and B have an associated relationship.

[0046] In the description of the embodiments of the present application, the term "corresponding" can mean a direct or indirect corresponding relationship between the two, or an associated relationship between the two, or an indication and an indicated, a configuration and a configured relationship.

[0047] In the embodiments of the present application, "predefined" can be realized by pre-saving corresponding codes, tables or other means for indicating related information in devices (for example, including terminal devices and network devices), and the specific implementation manner is not limited in the present application.

[0048] In view of the problem that the neural mechanism of patients with consciousness disorders is unclear, the potential consciousness level is difficult to detect, and the treatment effect of DBS surgery is difficult to predict, in the embodiments of the present application, a deep brain stimulation surgery effect prediction system is provided, which can realize individualized prediction of the consciousness state of a patient and a surgical plan before surgery, provide important model support for clinicians to develop clinical strategies, and specific embodiments are described below.

[0049] Figure 1 FIG. 1 is a schematic diagram of a deep brain stimulation surgery effect prediction system according to an example embodiment. The deep brain stimulation surgery effect prediction system can include:

[0050] (1) a data acquisition module 10.

[0051] The data acquisition module 10 is configured to collect preoperative electroencephalogram (EEG) data and multi-modal neuroimaging data of a target object, and collect efficacy evaluation data of the target object under a target deep brain stimulation surgery scheme.

[0052] The EEG data can be at least 10 minutes of resting state EEG data. For example, in a quiet environment, resting state EEG data is collected, and the target object is evaluated by easy-to-apply stimuli to evaluate EEG reactivity, each stimulus lasts for 15 seconds, including (i) passive eye opening; (ii) applying pressure to both earlobes (to reduce muscle artifacts); (iii) harmful stimulation to the fingertips and sternum; (iv) auditory stimulation, calling the patient's name; (v) applying tactile stimulation to the nostrils with a cotton swab. All stimuli are repeated at least once, with a pause of at least 60 seconds.

[0053] The multi-modal neural image data includes functional magnetic resonance image data, structural magnetic resonance image data, diffusion magnetic resonance imaging, and the like. The specific neural image type of the multi-modal neural image data is not limited in the application.

[0054] The efficacy evaluation data under the target deep brain stimulation surgery scheme can specifically divide the target object into one of an effective group or an ineffective group. The efficacy evaluation data under the target deep brain stimulation surgery scheme can be obtained by automatically evaluating the following types of data:

[0055] (1) Preoperative evaluation data: collect the patient's medical history, clinical manifestations and imaging data, and evaluate the indications.

[0056] (2) Surgical record data: detailed record of the operation process, stimulation parameter setting and complication situation.

[0057] (3) Postoperative follow-up data: regular evaluation of the patient's motor symptoms (such as the consciousness level score of patients with consciousness disorders), quality of life and complications.

[0058] The above data can be evaluated using standardized scales, such as the coma recovery scale revised edition (CRS-R) score or Rappaport disability score (RDR) scale, coma / near coma (C / NC) scale, and the like. The efficacy is quantitatively evaluated.

[0059] (2) The data preprocessing module 20.

[0060] The data preprocessing module 20 is configured to perform index calculation on the electroencephalogram data and the multi-modal neural image data to obtain a feature matrix, the feature matrix being used to quantify the consciousness state level of the target object, and the feature matrix including an electroencephalogram data index feature matrix and an imaging index feature matrix.

[0061] For example, the steps of preprocessing the electroencephalogram data for index calculation include: positioning the channel position; deleting useless channels; filtering; segmenting; baseline correction; re-referencing; reducing the sampling rate; interpolating bad leads; independent principal component analysis; rejecting bad segments, and the like.

[0062] For example, the multi-modal neural image data includes functional magnetic resonance image data. The steps of preprocessing the functional magnetic resonance image data for index calculation include: removing the first 5 time points of each subject data; time layer correction; head motion correction; discarding data with maximum translation greater than 2 mm or head rotation exceeding 2°; registration; normalization; resampling; mixed factor covariate regression; smoothing; removing linear drift trend; band-pass filtering.

[0063] Exemplarily, the multi-modal neuroimaging data includes diffusion magnetic resonance imaging data, and the steps of preprocessing the diffusion magnetic resonance imaging data for index calculation include: format conversion and extraction of b0 image; brain extraction BET; eddy correction; tensor calculation, etc.

[0064] Exemplarily, the multi-modal neuroimaging data includes T1 structural imaging data, and the steps of preprocessing the T1 structural imaging data for index calculation include: tissue segmentation, quality detection, total intracranial volume correction, resampling smoothing, etc.

[0065] In a possible implementation, in the data preprocessing module 20, the electroencephalogram data and the multi-modal neuroimaging data are subjected to index calculation to obtain a feature matrix, including: the electroencephalogram data is subjected to index calculation in time domain, frequency domain and time-frequency domain respectively to obtain an electroencephalogram data index feature matrix composed of time domain features, frequency domain features and time-frequency domain features; the multi-modal neuroimaging data is subjected to index calculation of structural magnetic resonance imaging, functional magnetic resonance imaging and diffusion magnetic resonance imaging respectively to obtain an imaging index feature matrix composed of structural magnetic resonance features, functional magnetic resonance features and diffusion magnetic resonance features.

[0066] In the implementation, the electroencephalogram data is subjected to index calculation in time domain, frequency domain and time-frequency domain respectively, and the multi-modal neuroimaging data is subjected to index calculation of structural magnetic resonance imaging, functional magnetic resonance imaging and diffusion magnetic resonance imaging respectively, so that the multi-aspect information of the data can be comprehensively utilized.

[0067] Exemplarily, the electroencephalogram data index can specifically include: visual assessment and Synek scale, ABCD spectral classification, spectral power of delta, theta, alpha, beta and Gamma, spectral summary, weighted symbolic mutual information, permutation entropy and Kolmogorov complexity, and the above indexes are calculated to obtain the frequency domain features F freq , the time domain features F time and the time-frequency features F time−freq , forming a feature vector F EEGi =[F freq , F time , F time−freq ] as the electroencephalogram data index feature matrix.

[0068] Exemplarily, the multi-modal neuroimaging data index can specifically include gray matter volume, cortical thickness, functional connectivity, low-frequency fluctuation amplitude, regional homogeneity, partial anisotropy, relative anisotropy, mean diffusivity, fiber bundle number matrix, fiber bundle length matrix, etc., and the above indexes are calculated to obtain the structural magnetic resonance features F sMRIi , the functional magnetic resonance features F fMRIi and the diffusion magnetic resonance features F DTIi , generating a feature vector [FDTIi , F sMRIi , F fMRIi ] as the imaging index feature matrix.

[0069] (3) a feature fusion module 30.

[0070] The feature fusion module 30 is configured to input the electroencephalogram data index feature matrix and the imaging index feature matrix into an autoencoder module for feature fusion, and the autoencoder module is configured to convert the feature matrix into a latent representation.

[0071] In a possible implementation, the autoencoder module comprises an encoder input layer, an encoder hidden layer, a decoder input layer, and a decoder hidden layer. The encoder input layer is configured to receive the feature matrix, the encoder hidden layer is configured to convert the feature matrix into the latent representation, the decoder input layer is configured to receive the latent representation, and the decoder hidden layer is configured to restore the latent representation to obtain a reconstructed output.

[0072] In this implementation, the autoencoder comprises two parts: an encoder and a decoder. The encoder part comprises an encoder input layer and an encoder hidden layer, and the decoder part comprises a decoder input layer and a decoder hidden layer. The encoder input layer receives the original feature matrix, the encoder hidden layer gradually reduces the feature dimension through a series of fully connected layers or convolutional layers, and finally generates a low-dimensional latent representation, the decoder input layer receives the latent representation, and the decoder hidden layer gradually restores the data to the same dimension as the encoder input layer, so as to calculate the reconstruction error during training. After training is completed, the encoder part is used to infer new data to obtain the low-dimensional latent representation, and the features of various modalities are integrated into a unified feature vector F EEGi , F fMRIi , F DTIi , F sMRIi ]. Further, the obtained low-dimensional latent representation can be visualized using t-distributed stochastic neighbor embedding (t-SNE) or uniform manifold approximation and projection (UMAP), so as to facilitate understanding of the data structure and clustering effect.

[0073] Further, in the feature fusion module, the mean square error is used as the loss function:

[0074] L= ;

[0075] wherein N is the number of samples, is the i-th feature matrix, is the i-th reconstructed output.

[0076] Further, the encoder hidden layer processes data through the following formula:

[0077] ;

[0078] wherein, W1 is a weight matrix, b1 is a bias, f is a ReLU activation function, X is a feature matrix, and Z is a latent representation.

[0079] Further, the decoder hidden layer processes data by the following formula:

[0080] ;

[0081] wherein, W2 is a weight matrix, b2 is a bias, g is an activation function, Z is a latent representation, and X' is a reconstruction output.

[0082] Further, in the training process, the autoencoder is trained, optimized and updated iteratively using the Adam optimization algorithm, including: calculating the gradient; updating the first moment and the second moment; then performing bias correction to complete the parameter update iteration; adjusting the weights and biases through the back propagation algorithm to minimize the loss function. The formula of the training process can be as follows:

[0083]

[0084] wherein, is the tth parameter value, is the previous parameter value of , is the bias correction of the first moment, is the bias correction of the second moment.

[0085] (4) The model training module 40.

[0086] The model training module 40 is configured to construct a prediction model of deep brain stimulation surgery efficacy based on the latent representation and the efficacy evaluation data, and the prediction model is configured to predict the efficacy of the patient for various deep brain stimulation surgery schemes.

[0087] wherein, the prediction model includes a cascade of: an input layer, a Stem layer, a plurality of mobile reverse bottleneck convolution modules, and a compression-excitation module. The compression-excitation module includes a cascade of: a global average pooling layer, a fully connected layer, and an activation function.

[0088] The input layer is configured to receive the latent representation; the Stem layer is configured to perform first-step feature extraction on the latent representation to obtain a first extracted feature map; the mobile inverted bottleneck convolution module is configured to perform second-step feature extraction on the first extracted feature map to obtain a second extracted feature map; and the compression-excitation module is configured to output a prediction result based on the second extracted feature map; in the compression-excitation module, the global average pooling layer is configured to compress the second extracted feature map to obtain a channel descriptor, the fully connected layer is configured to perform channel weighting on the channel descriptor to obtain a weight-applied feature map, and the activation function is configured to convert the weight-applied feature map into a probability output to obtain the prediction result.

[0089] For example, referring to Figure 2 , the input layer 401 is configured to obtain input features; the Stem layer 402 can include a 3x3 convolution operation, usually with a stride of 2, followed by a max pooling layer, which aims to extract preliminary features and reduce the spatial dimension to provide simplified data for subsequent layers; the mobile inverted bottleneck convolution module 403 includes a depth separable convolution operation: separating the convolution operation into a channel-wise convolution and a 1x1 point convolution to reduce the number of parameters and computational complexity, and using a Swish activation function and an inverted residual connection, which aims to efficiently extract features and improve the expression ability of the network; the compression-excitation module includes a global average pooling layer 404, a fully connected layer 405, and an activation function 406; the global average pooling can reduce the dimension of the feature map to provide a compact feature representation for the fully connected layer 405; the fully connected layer 405 connects the output of the global average pooling layer 404 to one or more fully connected layers for feature integration; and the activation function 406 uses a softmax activation function to convert the extracted features into class probabilities for final classification.

[0090] It can be understood that in the compression-excitation module, the global average pooling layer embodies the compression stage, which compresses the output feature map of the convolution layer into a feature vector through a global average pooling operation. Then, in the excitation stage, a channel weight vector is learned and generated by using a fully connected layer and a nonlinear activation function, which is applied to each channel of the original feature map to weight the features of different channels.

[0091] In one possible implementation, the mobile inverted bottleneck convolution module includes: a first point-wise convolution module configured to perform point-wise convolution on the initial first extracted feature map to obtain a first extracted feature map with increased channels; a depth convolution module configured to independently perform convolution on each channel of the first extracted feature map to obtain a first extracted feature map after feature extraction; a second point-wise convolution module configured to perform point-wise convolution on the first extracted feature map after feature extraction to obtain a first extracted feature map after channel restoration; and a skip connection module configured to connect the initial first extracted feature map with the first extracted feature map after channel restoration to obtain a second extracted feature map.

[0092] Exemplarily, in the mobile reverse bottleneck convolution module, the specific steps are as follows:

[0093] (1) input the initial first extraction feature map X ∈ R H×W×Cin , where H and W are the height and width of the feature map respectively, and C in is the number of input channels.

[0094] (2) increase the number of channels of the feature map by the first pointwise convolution module (1x1 convolution):

[0095] X′=Conv 1×1 (X)with C mid channels;

[0096] where X′ is the first extraction feature map after feature extraction, C mid is the number of intermediate channels, which is usually greater than C in .

[0097] where the expansion of the number of channels in the first pointwise convolution module can be performed after the Swish activation function, the purpose of which is to increase the feature dimension so that the subsequent convolution can learn more feature information. Swish is a nonlinear activation function used in neural network models, which is the product of the Sigmoid function and the input. The specific definition is as follows:

[0098]

[0099] where σ(x) is the Sigmoid function, σ(x)= .

[0100] (3) use the depthwise convolution (Depthwise Convolution) module to independently perform convolution on each channel to extract spatial features. Using the depthwise convolution module can significantly reduce the number of parameters and computational complexity:

[0101] X′′=DepthwiseConv(X′)with kernel size k×k;

[0102] where X′′ is the first extraction feature map after feature extraction, and X′ is the first extraction feature map after feature extraction.

[0103] (4) restore the number of channels to the original input number of channels by the second pointwise convolution:

[0104] Y=Conv 1×1 (X′′)to C out ;

[0105] Wherein, Y is the first extracted feature map after channel restoration, X" is the first extracted feature map after feature extraction, C out is the output channel number, equal to C in .

[0106] (5) Use the skip connection module to add the input feature map and the output feature map in the mobile reverse bottleneck convolution module to form a residual block, thereby ensuring the effective flow of information and helping to alleviate the gradient vanishing problem:

[0107] Z=Y+X if expand_ratio>1;

[0108] Wherein, Y is the first extracted feature map after channel restoration, X is the initial first extracted feature map.

[0109] It can be understood that the above mobile reverse bottleneck convolution module embodies an inverted residual structure, which specifically includes an expansion-compression convolution operation and a skip connection.

[0110] In the expansion-compression convolution operation, first, the channel number is expanded through 1x1 convolution, then the feature is extracted through depth separable convolution, and finally the channel number is compressed through 1x1 convolution. This "inverted" structure (expansion and then compression) improves the ability of feature learning, while also effectively reducing the amount of calculation.

[0111] In the skip connection, the input feature map and the output feature map are directly added through the residual connection to form an inverted residual block. The skip connection helps to alleviate the gradient vanishing problem, especially in deep networks, the input information is directly added to the output, helping the network to be easier to optimize in the training process.

[0112] In summary, the deep brain stimulation surgery efficacy prediction system provided by the embodiment, the system specifically collects electroencephalogram data and multi-modal neural image data, calculates the electroencephalogram data and multi-modal neural image data to obtain a feature matrix, and fuses the feature matrix through a self-encoder module, and then inputs a prediction model composed of a cascaded input layer, a stem layer, a plurality of mobile reverse bottleneck convolution modules, a compression-excitation module, a full connection layer and an activation function layer, thereby realizing preoperative prediction of deep brain stimulation surgery efficacy, and providing important model support for clinical doctors to develop clinical strategies.

[0113] Based on the above technical solution, Figure 3 a DBS surgery efficacy prediction model based on electroencephalogram and image features is shown, which can specifically include the following steps:

[0114] Step (1): simultaneously collect the preoperative electroencephalogram data and multi-modal neuroimaging data (functional magnetic resonance imaging, structural magnetic resonance imaging, diffusion magnetic resonance imaging, etc.) of patients with consciousness disorders; use the DBS surgery efficacy automatic evaluation system to evaluate the efficacy of the DBS surgery of the patient, and divide the patient into two groups: effective or ineffective.

[0115] Step (2): pre-process the electroencephalogram data and multi-modal imaging data respectively;

[0116] Step (3): calculate the indicators of the pre-processed electroencephalogram data and multi-modal imaging data respectively, extract each indicator to quantify the consciousness state level of the patient, use z-score to standardize the measurement of all features, and obtain multiple feature matrices;

[0117] Step (4): use the method of autoencoder to perform feature fusion and preliminary dimension reduction on the feature matrices of different dimensions;

[0118] Step (5): according to the evaluation of the DBS surgery efficacy of the patient with consciousness disorders, divide it into an effective group or an ineffective group. For the dimension-reduced neural signal and image features, use a convolutional neural network model with an EfficientNet architecture to train them. Divide the data set into a training set D train , a validation set D val and a test set D test , satisfying D=D train ∪D val ∪D test , and construct a DBS surgery efficacy prediction model based on electroencephalogram and image features.

[0119] Step (6): the performance of the prediction model can be evaluated by multiple metrics including accuracy, sensitivity, specificity, precision, recall, F1 score, AP and area under the curve.

[0120] It can be understood that this prediction model can be used to realize the individualized prediction of the preoperative clinical strategy and surgical plan of the patient, and provide important model support for the clinical doctors to formulate the clinical strategy.

[0121] Please refer to Figure 4 , which is a schematic diagram of a computer device according to an exemplary embodiment of the present application, the computer device comprising a memory and a processor, the memory being used to store a computer program, the computer program being executed by the processor to realize the prediction method corresponding to the prediction system of the deep brain stimulation surgery efficacy described above.

[0122] The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, or a combination thereof.

[0123] The memory is a non-transitory computer-readable storage medium, which can be used to store non-transitory software programs, non-transitory computer-executable programs and modules, such as program instructions / modules corresponding to the method in the embodiments of the present application. The processor executes various functions and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory, that is, implements the method in the above method embodiments.

[0124] The memory can include a program storage area and a data storage area. The program storage area can store an operating system and at least one application required by a function. The data storage area can store data created by the processor and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory remotely disposed relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0125] In an exemplary embodiment, a computer-readable storage medium for storing at least one computer program is also provided, and the at least one computer program is loaded and executed by the processor to implement all or part of the steps of the above method. For example, the computer-readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0126] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.

[0127] It is to be understood that the application is not limited to the precise construction herein disclosed and shown in the drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is limited only by the claims that follow.

Claims

1. A system for predicting the efficacy of deep brain stimulation surgery, comprising: The system comprises: a data acquisition module, configured to collect preoperative electroencephalogram data and multi-modal neuroimaging data of a target object, and collect efficacy evaluation data of the target object under a target deep brain stimulation surgery scheme; a data preprocessing module, configured to perform index calculation on the electroencephalogram data and the multi-modal neuroimaging data to obtain a feature matrix, the feature matrix being used to quantify a level of consciousness of the target object, and the feature matrix comprising an electroencephalogram data index feature matrix and an imaging index feature matrix; a feature fusion module, configured to input the feature matrix into a self-encoder module to perform feature fusion, and the self-encoder module being configured to convert the feature matrix into a latent representation; a model training module, configured to construct a deep brain stimulation surgery efficacy prediction model based on the latent representation and the efficacy evaluation data, the prediction model being configured to predict efficacy of patients for various deep brain stimulation surgery schemes, and the prediction model comprising a cascade of an input layer, a Stem layer, a plurality of mobile reverse bottleneck convolution modules, and a compression-excitation module, the compression-excitation module comprising a cascade of a global average pooling layer, a full connection layer, and an activation function; wherein the input layer is configured to receive the latent representation, the Stem layer is configured to perform first-step feature extraction on the latent representation to obtain a first extracted feature map, the mobile reverse bottleneck convolution module is configured to perform second-step feature extraction on the first extracted feature map to obtain a second extracted feature map, and the compression-excitation module is configured to output a prediction result based on the second extracted feature map; in the compression-excitation module, the global average pooling layer is configured to compress the second extracted feature map to obtain a channel descriptor, the full connection layer is configured to perform channel weighting on the channel descriptor to obtain a weight-applied feature map, and the activation function is configured to convert the weight-applied feature map into a probability output to obtain the prediction result; in the data preprocessing module, the index calculation on the electroencephalogram data and the multi-modal neuroimaging data to obtain the feature matrix comprises: performing index calculation on the electroencephalogram data in time domain, frequency domain, and time-frequency domain to obtain an electroencephalogram data index feature matrix composed of time domain features, frequency domain features, and time-frequency domain features; performing index calculation on the multi-modal neuroimaging data in structural magnetic resonance imaging, functional magnetic resonance imaging, and diffusion magnetic resonance imaging to obtain an imaging index feature matrix composed of structural magnetic resonance features, functional magnetic resonance features, and diffusion magnetic resonance features; the self-encoder module comprises an encoder input layer, an encoder hidden layer, a decoder input layer, and a decoder hidden layer; the encoder input layer is configured to receive the feature matrix; the encoder hidden layer is configured to convert the feature matrix into a latent representation; the decoder input layer is configured to receive the latent representation; the decoder hidden layer is configured to restore the latent representation to obtain a reconstruction output; the mobile reverse bottleneck convolution module comprises: a first point-by-point convolution module, configured to perform point-by-point convolution on an initial first extracted feature map to obtain a first extracted feature map with increased channels; a deep convolution module, configured to independently perform convolution on each channel of the first extracted feature map to obtain a first extracted feature map after feature extraction; a second point-by-point convolution module, configured to perform point-by-point convolution on the first extracted feature map after feature extraction to obtain a first extracted feature map after channel restoration; a skip connection module, configured to connect the initial first extracted feature map and the first extracted feature map after channel restoration to obtain the second extracted feature map.

2. The system of claim 1, wherein, In the feature fusion module, a mean square error is used as a loss function. L= ； where N is the number of samples, is the i-th feature matrix, is the i-th reconstructed output.

3. The system of claim 1, wherein, The encoder hidden layer performs data processing through the following formula: ； wherein W1 is a weight matrix, b1 is a bias, f is a ReLU activation function, X is a feature matrix, and Z is a latent representation.

4. The system of claim 1, wherein, The decoder hidden layer performs data processing through the following formula: ； wherein W2 is a weight matrix, b2 is a bias, g is an activation function, Z is a latent representation, and X' is a reconstruction output.

5. The system of claim 1, wherein, The first point-by-point convolution module comprises a Swish activation function.

6. A computer readable medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the following steps: Collecting electroencephalogram data and multi-modal neuroimaging data of a target object before surgery; collecting efficacy evaluation data of the target object under a target deep brain stimulation surgery scheme; Performing index calculation on the electroencephalogram data and multi-modal neuroimaging data to obtain a feature matrix, the feature matrix being used to quantify a level of consciousness of the target object, and the feature matrix comprising an electroencephalogram data index feature matrix and an imaging index feature matrix; The index calculation on the electroencephalogram data and multi-modal neuroimaging data to obtain the feature matrix comprises: Performing index calculation on the electroencephalogram data in time domain, frequency domain and time-frequency domain to obtain an electroencephalogram data index feature matrix composed of time domain features, frequency domain features and time-frequency domain features; Performing index calculation on the multi-modal neuroimaging data in structural magnetic resonance imaging, functional magnetic resonance imaging and diffusion magnetic resonance imaging to obtain an imaging index feature matrix composed of structural magnetic resonance features, functional magnetic resonance features and diffusion magnetic resonance features; Inputting the electroencephalogram data index feature matrix and the imaging index feature matrix into a self-encoder module for feature fusion, the self-encoder module being configured to convert the feature matrix into a latent representation; The self-encoder module comprises an encoder input layer, an encoder hidden layer, a decoder input layer and a decoder hidden layer; The encoder input layer is configured to receive the feature matrix; The encoder hidden layer is configured to convert the feature matrix into a latent representation; The decoder input layer is configured to receive the latent representation; The decoder hidden layer is configured to restore the latent representation to obtain a reconstruction output; constructing a deep brain stimulation surgery effect prediction model based on the latent representation and the therapeutic effect evaluation data, the prediction model being used to predict the therapeutic effect of a patient for various deep brain stimulation surgery schemes, the prediction model comprising a cascade of an input layer, a Stem layer, a plurality of mobile reverse bottleneck convolution modules, and a compression-excitation module, the compression-excitation module comprising a cascade of a global average pooling layer, a fully connected layer, and an activation function; wherein the input layer is configured to receive the latent representation, the Stem layer is configured to perform first-step feature extraction on the latent representation to obtain a first extracted feature map, the mobile reverse bottleneck convolution module is configured to perform second-step feature extraction on the first extracted feature map to obtain a second extracted feature map, and the compression-excitation module is configured to output a prediction result based on the second extracted feature map, in the compression-excitation module, the global average pooling layer is configured to compress the second extracted feature map to obtain a channel descriptor, the fully connected layer is configured to perform channel weighting on the channel descriptor to obtain a weight-applied feature map, and the activation function is configured to convert the weight-applied feature map into a probability output to obtain the prediction result; the mobile reverse bottleneck convolution module comprises: a first point-by-point convolution module configured to perform point-by-point convolution on an initial first extracted feature map to obtain a first extracted feature map with increased channels; a depth convolution module configured to independently perform convolution on each channel of the first extracted feature map to obtain a first extracted feature map after feature extraction; a second point-by-point convolution module configured to perform point-by-point convolution on the first extracted feature map after feature extraction to obtain a first extracted feature map after channel restoration; a skip connection module configured to connect the initial first extracted feature map and the first extracted feature map after channel restoration to obtain the second extracted feature map.

7. A computer device, comprising: comprise: one or more processors; a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the following steps: collecting preoperative electroencephalogram data and multi-modal neuroimaging data of a target object; collecting therapeutic effect evaluation data of the target object under a target deep brain stimulation surgery scheme; performing index calculation on the electroencephalogram data and the multi-modal neuroimaging data to obtain a feature matrix, the feature matrix being used to quantify the consciousness state level of the target object, the feature matrix comprising an electroencephalogram data index feature matrix and an imaging index feature matrix; the index calculation on the electroencephalogram data and the multi-modal neuroimaging data to obtain the feature matrix comprises: performing index calculation on the electroencephalogram data in the time domain, the frequency domain, and the time-frequency domain to obtain an electroencephalogram data index feature matrix composed of time domain features, frequency domain features, and time-frequency domain features; performing index calculation on the multi-modal neuroimaging data in structural magnetic resonance imaging, functional magnetic resonance imaging, and diffusion magnetic resonance imaging to obtain an imaging index feature matrix composed of structural magnetic resonance features, functional magnetic resonance features, and diffusion magnetic resonance features; The electroencephalogram data index feature matrix and the imaging index feature matrix are input into a self-encoder module for feature fusion, and the self-encoder module is used for converting the feature matrix into a latent representation; The self-encoder module comprises an encoder input layer, an encoder hidden layer, a decoder input layer and a decoder hidden layer; The encoder input layer is used for receiving the feature matrix; The encoder hidden layer is used for converting the feature matrix into a latent representation; The decoder input layer is used for receiving the latent representation; The decoder hidden layer is used for recovering the latent representation to obtain a reconstruction output; Based on the latent representation and the efficacy evaluation data, a deep brain stimulation surgery efficacy prediction model is constructed, the prediction model is used for predicting the efficacy of various deep brain stimulation surgery schemes for patients, and the prediction model comprises a cascade of an input layer, a Stem layer, a plurality of mobile reverse bottleneck convolution modules and a compression-excitation module, and the compression-excitation module comprises a cascade of a global average pooling layer, a full connection layer and an activation function; The input layer is used for receiving the latent representation; the Stem layer is used for performing first-step feature extraction on the latent representation to obtain a first extraction feature map; the mobile reverse bottleneck convolution module is used for performing second-step feature extraction on the first extraction feature map to obtain a second extraction feature map; and the compression-excitation module is used for outputting a prediction result based on the second extraction feature map; in the compression-excitation module, the global average pooling layer is used for compressing the second extraction feature map to obtain a channel descriptor, the full connection layer is used for performing channel weighting on the channel descriptor to obtain a weight-applied feature map, and the activation function is used for converting the weight-applied feature map into a probability output to obtain the prediction result; The mobile reverse bottleneck convolution module comprises: A first point-by-point convolution module is used for performing point-by-point convolution on the initial first extraction feature map to obtain a first extraction feature map with increased channels; A depth convolution module is used for independently performing convolution on each channel of the first extraction feature map to obtain a first extraction feature map after feature extraction; A second point-by-point convolution module is used for performing point-by-point convolution on the first extraction feature map after feature extraction to obtain a first extraction feature map after channel restoration; A skip connection module is used for connecting the initial first extraction feature map and the first extraction feature map after channel restoration to obtain the second extraction feature map.

Images