CN-121980161-A - Brain function connection network prediction method for transcranial magnetic stimulation

CN121980161ACN 121980161 ACN121980161 ACN 121980161ACN-121980161-A

Abstract

The invention discloses a brain function connection network prediction method oriented to transcranial magnetic stimulation. The method comprises the steps of obtaining a functional near infrared spectrum signal, carrying out data preprocessing, calculating the concentration of hemoglobin and the concentration of oxygenated hemoglobin in a cerebral cortex from the preprocessed signal, generating a space-time characteristic network corresponding to a brain function connection network by using the concentration of hemoglobin and the concentration of oxygenated hemoglobin, training a pre-built neural network model by using the space-time characteristic network to obtain a time sequence brain function connection network prediction model based on a gate control time sequence graph convolution network, and estimating the dynamic connectivity between brain function connection network nodes of the next time slice. Compared with the traditional brain function connection network prediction method, the brain function connection network prediction method for transcranial magnetic stimulation has the advantages of higher accuracy and individuation, can identify and recommend treatment targets more comprehensively and efficiently, and provides more accurate data support for diagnosis and treatment of mental diseases.

Inventors

DU GUANGLONG

Assignees

佛山市智研未来科技服务有限公司
华南理工大学

Dates

Publication Date: 20260505
Application Date: 20251208

Claims (10)

1. The brain function connection network prediction method for transcranial magnetic stimulation is characterized by comprising the following steps of: s1, acquiring a functional near infrared spectrum (fNIRS) signal, and performing data preprocessing on the functional near infrared spectrum (fNIRS) signal; s2, according to the corrected lambert-beer law, the hemoglobin concentration and the oxygenated hemoglobin concentration in the cerebral cortex are calculated from the preprocessed functional near infrared spectrum (fNIRS) signals; s3, using the hemoglobin concentration and the oxygenated hemoglobin concentration calculated in the step S2 to generate a space-time characteristic network corresponding to the brain function connection network; and S4, training the pre-constructed neural network model to obtain a time sequence brain function connection network prediction model based on the gate control time sequence graph convolution network by utilizing the time-space characteristic network generated in the step S3, and estimating the dynamic connectivity between brain function connection network nodes of the next time slice.
2. The method for predicting brain function connection network for transcranial magnetic stimulation according to claim 1, wherein the step S1 comprises the steps of: Removing motion artifacts in the functional near infrared spectrum (fNIRS) signal; in order to preserve the signals of the desired frequency band, a bandpass filter is used to filter and smooth the functional near infrared spectrum (fNIRS) signals after removal of motion artifacts.
3. The method for predicting brain function connection network for transcranial magnetic stimulation according to claim 1, wherein in step S2, the concentration of hemoglobin, which includes hemoglobin and oxyhemoglobin, is deduced from the preprocessed functional near infrared spectrum (fNIRS) signal by a mathematical calculation method, comprising the steps of: calculating absorbance change by comparing incident light intensity and transmitted light intensity to obtain absorbance change value; Concentration change is deduced by calculating concentration change values of oxyhemoglobin HbO 2 and reduced hemoglobin Hb using the change value of absorbance, optical path length, and absorption coefficient of hemoglobin.
4. The method for predicting brain function connection network for transcranial magnetic stimulation according to claim 3, wherein in step S2, the relative amounts of the hemodynamic parameters, i.e. the concentrations of hemoglobin and oxyhemoglobin, are accurately calculated, and different wavelengths are required to be used 1 And 2 Different results are obtained The solution is performed as follows: To correspond to the wavelength of near infrared light Differential path factor under conditions Is a constant of (a); And The solution matrix equation is obtained for the hemoglobin concentration and the oxygenated hemoglobin concentration respectively: At a certain fixed wavelength The light intensity density variation in two moments The specific values of (2) are obtained by modified lambert-beer law: Taking two different wavelengths 1 And 2 Calculated as above And And substituting the values into a matrix equation to realize the calculation of the concentration of the hemoglobin and the concentration of the oxygenated hemoglobin.
5. The method for predicting brain function connection network for transcranial magnetic stimulation according to claim 1, wherein in step S3, the hemoglobin concentration and the oxygenated hemoglobin concentration calculated in step S2 are segmented into a plurality of time slices by using a sliding window analysis method; the temporal characteristics of the brain function connection network are then extracted for each time slice The oxygenated hemoglobin concentration curve for each sampling channel within was calculated as six statistics, average, maximum, minimum, variance, skewness (Skewness) and Kurtosis (Kurtosis).
6. The method of claim 5, wherein to increase data utilization, 50% overlap between each time slice is the same for the last 50% data point of the previous time slice and the first 50% data point of the next time slice.
7. The method for predicting brain function connection network for transcranial magnetic stimulation according to claim 1, wherein in step S4, the time sequence brain function connection network prediction model based on the gating time sequence graph convolution network comprises a space graph convolution network layer, a gating time convolution network layer, a bidirectional long-short-term neural network layer and a self-attention mechanism layer; The input of the time sequence brain function connection network prediction model based on the gating time sequence convolution network is a preprocessed functional near infrared spectrum (fNIRS) signal sequence and a space-time feature network generated in the step S3, firstly, two space-time convolution network layers are introduced to process input data, spatial features related to brain areas are excavated, output data of the space-time convolution network layers are input into a space-time embedding module to perform feature enhancement operation, then the output features of the space-time embedding module are sequentially transmitted to the gating time convolution network layer and the bidirectional long-short-term neural network layer to further capture short-term and long-term time correlation related to the brain areas, finally, the spatial features connected to the brain function connection network and the space-time features related to the brain areas are fused by means of a self-attention mechanism layer, and finally, the predicted value of the time sequence brain function connection network of the next time slice is obtained.
8. The brain function connection network prediction method for transcranial magnetic stimulation according to claim 7, wherein the spatial map convolution network layer adopts GCN spectral domain map convolution; The two space diagram convolution network layers are sequentially connected, data input is a fNIRS sequence and a characteristic matrix formed by a space-time characteristic network generated in the step S3, output is the characteristic matrix after spatial information extraction, input and output are three-dimensional matrices [ N, C and T ], wherein N represents the number of channels, C is a component of each channel, C is taken as 2, the concentrations of hemoglobin and oxyhemoglobin are respectively represented, and T represents a time point; The data output after the network layer is rolled by the two space diagrams is recorded as From the viewpoint of enhancing network feature expression capability, the graph structure data output by the network layer is rolled according to the space graph To construct a graph spatiotemporal sequence ; The time-space embedding module utilizes the time embedding and the space embedding to excavate the time and the space dependency relationship of the functional near infrared spectrum (fNIRS) signals, thereby the time sequence brain function connection network prediction model based on the gate control time sequence graph convolution network can fully consider the time and the space information of the brain function connection network contained in the last step output; In the space-time embedding module, input data for a given space-time embedding module, i.e. input graph structure data Not only will be Splitting into time-embedded tensors along the time dimension Splitting into spatially embedded tensors along the spatial dimension is also required Then, the convolution of 1 x1 is adopted to operate so that two embedded tensors contain necessary space-time information, and finally, the two embedded tensors are added to the graph structure data through a broadcasting mechanism In which the graph space-time sequence is finally obtained The input and output of the space-time embedding module are three-dimensional matrixes [ N, C and T ], wherein N represents the number of channels, C is a component of each channel, C is 2, the concentrations of hemoglobin and oxyhemoglobin are respectively represented, and T represents a time point; The method comprises the steps of acquiring the time characteristics of brain function connecting edge weights by a gating time convolution layer, wherein the time characteristics of the brain function connecting edge weights are acquired through a time convolution network TCN, the time convolution network TCN consists of two hole causal convolution networks and a full connection network, the hole causal convolution network (Dilated Causal Convolution) processes long-range sequences in a non-recursion mode, a bidirectional long-short-period neural network layer is realized by a bidirectional long-short-period neural network (Bi-LSTM), data input of the gating time convolution layer is a space-time characteristic matrix processed by a space-time embedding module, the space-time characteristic matrix is a three-dimensional matrix [ N, C, T ], and data output of the bidirectional long-short-period neural network layer is a two-dimensional matrix [ N, T ], wherein N represents the number of channels, and T represents a time point; The self-attention mechanism layer adopts the existing attention mechanism architecture to process the output characteristics of the bidirectional long-short period neural network layer, the relative important part in the targeted learning characteristics is adopted, the processed characteristic expression is output, the data input and output of the self-attention mechanism layer are two-dimensional matrixes [ N, T ], wherein N represents the number of channels, and T represents the time point; In a time sequence brain function connection network prediction model based on a gate control time sequence graph convolution network, output data of a self-attention mechanism layer is processed through a full connection layer, the output of the self-attention mechanism layer is taken as input, namely, an input format is a two-dimensional matrix [ N, T ], wherein N represents the number of channels, T represents a time point, and an output format is a two-dimensional matrix [ N, N ] which means that adjacent matrixes of a predicted brain function connection network are represented, namely, the predicted time sequence brain function connection network.
9. The method for predicting brain function connection network for transcranial magnetic stimulation according to claim 1, wherein two loss functions, mean square error (Mean Squared Error, MSE) and L2 regularization, are used in training the training of the pre-constructed neural network model.
10. The transcranial magnetic stimulation-oriented brain function connection network prediction method according to claim 9, wherein the mean square error (Mean Squared Error, MSE) is expressed as follows: Wherein, the Representing the actual observed brain function connection network The weight of the strip edge is calculated, First representing brain function connection network The predicted value of the weight of the stripe edge, Is the number of samples to be processed, The average value of the sum of squares of the differences between the true values of all the brain network edge weights and the model predicted values is calculated to evaluate the overall error of the whole network prediction, and the smaller the error is, the better the prediction effect of the model is.

Description

Brain function connection network prediction method for transcranial magnetic stimulation Technical field: the invention relates to the field of medical rehabilitation, in particular to a brain function connection network prediction method for transcranial magnetic stimulation. The background technology is as follows: The brain is the most important, complex organ of humans, and a wide variety of complex cognitive functions are dependent on brain realization. The exploration of the functional mechanism of the brain is helpful for human beings to deepen understanding of the neural mechanism principle of the brain, can promote the rapid development of interdisciplinary science, and has important promotion effect on the high-quality development of human production practice. With the proposal of high-level national strategy such as health China 2030 and China brain science program, the research around the non-invasive brain function detection technology and the application thereof is gradually rising. Such techniques include electroencephalography (electroencephalogram, EEG), functional near infrared spectroscopy (functional near-infrared spectroscopy, fNIRS), functional magnetic resonance imaging (functional magnetic resonance imaging, fMRI), and positron emission tomography (Positron Emission Tomography, PET), among others, fNIRS thus being of interest to researchers. On one hand, the kit is widely used for clinical detection and rehabilitation as a research tool of cognitive neuroscience. The technology is also used for exploring the neural mechanism of cognitive activities and researching the corresponding relation between the brain area structure and the function, and can be further combined with a neural feedback technology for application. On the other hand fNIRS is also a common monitoring means for rehabilitation of clinical mental diseases and organic lesions of the brain. A great deal of researches show that fNIRS can be applied to the mental diseases such as autism, epilepsy, schizophrenia, attention deficit hyperactivity disorder, bipolar disorder, anxiety and the like, and can also be applied to the continuous monitoring of brain organic lesions such as cerebral apoplexy, parkinson's disease, alzheimer's disease and the like. Brain networks are an important existence in the human brain, with obvious network characteristics. It is a typically complex network that naturally forms when the mutual transfer and co-processing of information occurs in different brain regions of the brain. Brain functions are used for the study of problems such as classification, prediction and brain-computer interfaces. Studies from the vast sea of cigarettes have shown that brain function is achieved by the transfer of substances and information between the entire brain's multiple functional networks. Wherein each brain function connection network corresponds to a specific brain region three-dimensional spatial distribution characteristic and a specific time-dependent activity pattern respectively. Most of the local brain functions in the human brain are connected to networks, such as auditory centers, visual centers, etc., which are responsible for a specific functional task. There are other networks of local brain function connections responsible for a number of tasks, such as the first somatosensory zone for receiving a variety of sensations of pain, temperature, touch, etc. Research on personalized space-time characteristics of the brain function connection network is carried out, and the method is favorable for promoting further exploration of brain mechanisms. The brain network is constantly changing according to the external environment. The transient cognitive state of the subject can be analyzed by extracting fNIRS signals of a time slice, and the method is widely used in the fields of brain-computer interfaces and the like. The brain function connection network is very closely related to the cognitive state and behavior of the subject. Obtaining the current fNIRS signals of the subject can predict the subsequent cognitive states and can be regarded as a link prediction problem in a complex network to be solved. Link prediction refers to a mathematical expectation that a connection exists between brain function connection network nodes of the next time slice by using known network information (nodes, edge weights, feature vectors, etc.). The solution of the problem is helpful for developing new mass productivity, provides reference for researching the change evolution law of the brain function connection network, and can provide potential biomarkers for diagnosing partial diseases. One of the hot spots in the brain network field is the brain function connection network prediction research. The problem is constantly researched, the algorithm is continuously optimized and iterated, the accuracy of predicting the brain function connection network is improved, and the development of relevant medical fields such as