A system and method for predicting the effect of DBS surgery on Parkinson's disease patients based on brain functional network

Abstract

The present invention relates to a system and method for predicting the effect of DBS surgery on Parkinson's patients based on brain function network, and belongs to the field of computer application technology. The present invention establishes two new dynamic brain function networks, integrates the whole brain function network based on graph theory with the machine learning algorithm, and can better capture the functional connection information in the brain of Parkinson's patients and improve the accuracy of DBS surgery effect prediction. By analyzing the error evaluation index between the predicted recovery effect and the actual recovery effect of the system, the prediction effect is greatly improved.

Description

Technical Field

[0001] The present invention relates to a system and method for predicting the postoperative effect of DBS (Deep Brain Stimulation, DBS for short) on Parkinson's patients based on brain function network. A computer is used to fuse fMRI image information of Parkinson's patients with indicators based on graph theory to construct a brain function network, so as to predict the postoperative effect of DBS on Parkinson's patients. The invention belongs to the field of computer application technology. Background Art

[0002] At present, with the trend of population aging, the number of middle-aged and elderly people suffering from Parkinson's disease is increasing year by year. As one of the degenerative diseases of the nervous system, Parkinson's disease seriously threatens the quality of life of the elderly.

[0003] In daily life, common motor symptoms of Parkinson's disease include bradykinesia, freezing of gait and static tremor of the limbs, as well as common symptoms such as cognitive dysfunction, sleep disorders, depression or decreased sense of smell, which bring serious burdens to patients and their caregivers. Surgical treatment, as one of the effective means to relieve Parkinson's disease, plays an important role in improving the quality of life of patients.

[0004] Parkinson's disease is a typical neurodegenerative disease, and its development process is accompanied by the loss of neurons and myelin sheaths in patients. Deep Brain Stimulation (DBS) inserts electrodes into specific areas of the brain and causes the electrodes to emit electrical signals to stimulate surrounding tissues in order to improve the transmission pattern of human neural signals. Because of its reversibility and adjustability, it is considered an ideal treatment for Parkinson's disease that is not well controlled by drugs. However, this type of surgery requires the placement of electrodes and external control equipment in the human brain, which has a high economic cost and is an invasive surgery. It faces risks such as bleeding and infection, which brings inevitable economic and physical pressure to patients.

[0005] Therefore, it is necessary to develop an auxiliary evaluation system for the postoperative effect of DBS for Parkinson's disease to effectively predict the recovery effect of Parkinson's disease patients after DBS surgery, which can serve as one of the references for patients to undergo the surgery.

[0006] At present, the technical means used by the existing Parkinson's disease DBS postoperative recovery effect prediction system are mostly based on statistical analysis of clinical indicators, without combining the characteristics of Parkinson's disease as a neurodegenerative disease, lacking effective biomarkers, and having poor prediction effects. Although some fMRI-based studies have shown that the connection from the stimulated area to other brain regions has a predictive effect on the patient's DBS postoperative recovery effect, it does not take into account the coordination of the whole brain functional network as a whole, resulting in poor prediction results of the system. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the existing system for predicting the recovery effect of DBS surgery on Parkinson's patients, and to propose a system and method for predicting the effect of DBS surgery on Parkinson's patients based on brain functional networks. By establishing two new dynamic brain functional networks and integrating indicators based on graph theory, the system can better capture the functional connection information in the brain of Parkinson's patients and improve the accuracy of DBS surgery effect prediction.

[0008] The innovation of this invention is that it proposes a new prediction system, which integrates the whole brain functional network based on graph theory with the machine learning algorithm to achieve the prediction of the recovery effect of DBS after Parkinson's disease based on the brain functional network before surgery. By analyzing the error evaluation index between the predicted recovery effect and the actual recovery effect of the system, the prediction effect is greatly improved.

[0009] The two novel dynamic brain function networks designed by the present invention have the following advantages:

[0010] A sliding window strategy is adopted to construct a dynamic brain functional network with low time granularity. Taking dynamic factors into consideration, the functional matrix constructed through different windows captures the dynamic changes of the brain over a period of time.

[0011] The random dynamic brain functional network uses windows of different sizes, has the advantage of multi-scale, and can more comprehensively extract dynamic features in different small time periods.

[0012] Based on the analysis method of graph theory, the brain is divided into ROIs according to the brain map, and the Pearson correlation coefficient of the shortest path length time series or the global efficiency time series between any two ROIs is calculated as the brain connection strength of the two ROIs in the final brain network. The correlation between the information transmission capabilities of each ROI node is introduced.

[0013] The technical solution adopted by the present invention is as follows:

[0014] A system for predicting the postoperative effect of DBS on Parkinson's disease patients based on brain function network includes a data preprocessing module, a dynamic brain function network construction module, a feature screening module, a prediction model training module and a prediction module.

[0015] The data preprocessing module is responsible for dividing the brain region ROI of the preprocessed patient MRI image and generating the BOLD timing signal (txt format) of the ROI node. The input of this module is the patient's fMRI image.

[0016] The dynamic brain functional network construction module is responsible for constructing fixed dynamic brain functional networks and random dynamic brain functional networks based on BOLD timing signals. The input of this module is the BOLD timing signal generated by the data preprocessing module.

[0017] The feature screening module is responsible for using the random forest algorithm to perform feature screening on the input brain functional network. The input of this module is the brain functional network generated by the dynamic brain functional network construction module.

[0018] The prediction model training module is responsible for the regression model training. The input of this module is the characteristics of the brain functional network obtained by the feature screening module, as well as the preoperative UDPRS-III and postoperative UDPRS-III scores of the corresponding patients.

[0019] The prediction module is responsible for predicting the postoperative effect of DBS using a regression model. The input of this module is the fMRI image of the patient to be predicted.

[0020] The connection relationship between the above modules is:

[0021] The output end of the data preprocessing module is connected to the input end of the dynamic brain functional network construction module;

[0022] The output end of the dynamic brain function network construction module is connected to the input end of the feature screening module;

[0023] The output end of the feature screening module is connected to the input end of the prediction model training module and the prediction module respectively;

[0024] The output end of the prediction model training module is connected to the input end of the prediction module.

[0025] The working process of this system is as follows:

[0026] First, the fMRI image data of multiple patients are input into the data preprocessing module. After data preprocessing, the data preprocessing module obtains the patient's BOLD time series and sends it to the brain function network construction module.

[0027] Then, the brain functional network construction module constructs a fixed dynamic brain functional network and a random dynamic brain functional network according to the patient's BOLD time series, and inputs them into the feature screening module.

[0028] Afterwards, the feature screening module uses the random forest algorithm to perform feature screening on the patient's fixed dynamic brain functional network and random dynamic brain functional network to obtain the patient's important brain network features.

[0029] Then, the prediction model training module performs regression model training according to the patient's brain network characteristics and the patient's corresponding preoperative UDPRS-III and postoperative UDPRS-III scores, and inputs the trained regression model into the prediction module for subsequent prediction.

[0030] When predicting the postoperative effect of a new patient, the patient's fMRI image data is input into the data preprocessing module. Then, data preprocessing, brain functional network construction, and feature screening are performed in sequence. The trained regression model is used in the prediction module to predict the patient's postoperative recovery.

[0031] Furthermore, based on the above system, the present invention proposes a method for predicting the effect of DBS surgery on Parkinson's patients based on brain function network, comprising the following steps:

[0032] Step 1: Perform image preprocessing on the original MRI image data (e.g., image time correction, image motion correction, image space normalization, image space smoothing, image linear trend elimination, image noise reduction, brain map selection, etc.). Divide the brain region ROI on the preprocessed MRI image and generate the BOLD signal of the ROI node, which can be achieved using GRETNA software, etc.

[0033] Step 2: Construct a dynamic brain functional network, including constructing a fixed dynamic brain functional network and a random dynamic brain functional network.

[0034] Specifically, the steps include:

[0035] Step 2.1: Construct a fixed dynamic brain functional network. This includes the following steps:

[0036] Step 2.1.1: Select a sliding window of length l, and divide the BOLD time series signal of each ROI into n overlapping subsequences according to a fixed step length of s. Each subsequence is a window.

[0037] Step 2.1.2: Calculate the shortest path length of each ROI in each window. Take each window as a weighted network G with K nodes. For two nodes i and j, the path with the smallest weight accumulation among all their paths is called the shortest path between the two nodes, denoted as For a node i in a weighted network G with a total of K nodes, its shortest path length is defined as follows:

[0038]

[0039] Step 2.1.3: Calculate the Pearson correlation coefficient r of the shortest path length sequence between any two ROI X and ROI Y in all ROIs to obtain the final fixed dynamic brain functional network:

[0040]

[0041] Among them, x a represents the mean of vector X, y a Represents the mean of vector Y.

[0042] Step 2.2: Construct a random dynamic brain functional network. This includes the following steps:

[0043] Step 2.2.1: Select a sliding window with a length ranging from l_min to l_max, and divide the BOLD time series signal into n overlapping subsequences with no time series between different windows and time series within the window according to the step length ranging from s_min to s_max. Each subsequence is a window.

[0044] Step 2.2.2: Calculate the node global efficiency of each ROI in each window separately:

[0045]

[0046] Among them, K represents the total number of nodes, G represents the network, represents the global efficiency of node i, represents the weighted shortest distance between nodes i and j.

[0047] Step 2.2.3: Calculate the Pearson correlation coefficient of the global efficiency sequences of any two ROIs among all ROIs to obtain the final random dynamic brain network.

[0048] Step 3: Use the random forest algorithm to perform feature screening on the constructed brain network.

[0049] The importance of each brain connection feature is calculated using the random forest algorithm, and the importance is mapped to the left and right brain regions of the connection and then normalized. The above operation is repeated for each brain connection feature to obtain the importance weight of each brain region under the network.

[0050] Step 4: Training DBS postoperative prediction model.

[0051] The common linear regression model and nonlinear regression model were used to predict the postoperative effect of DBS, and the prediction effects were compared. Specifically, the following steps were included:

[0052] Step 4.1: Perform inner cross validation.

[0053] During the training process of each machine learning model, the model is optimized using a grid search method, that is, all the candidate parameters that are set are continuously looped through, and each parameter combination obtained is compared with its corresponding predicted value. When the predicted value is closest to the true value, the model at this time is the optimal model.

[0054] Step 4.2: Perform outer cross-validation. Split the source data and perform k-fold cross-validation to ensure that k-1 pieces of data are used for training and 1 piece of data is used for validation each time. Finally, take the average of the k validation results to get the final prediction effect.

[0055] Step 4.3: Train the least squares linear regression, ridge regression, lasso regression, support vector machine regression, and extreme gradient regression models respectively.

[0056] In the inner loop of nested cross-validation, the grid search algorithm was used to determine the optimal parameters of each model. Cross-validation was performed in the outer loop to obtain the predicted value of the postoperative UDPRS-III improvement rate of each subject, and the Pearson correlation coefficient, mean absolute error (MAE), and root mean square error (RMSE) with the true value were calculated as prediction effect evaluation indicators.

[0057]

[0058] Among them, UPDRS-III ratio Indicates the improvement rate of UDPRS-III, UPDRS-III preoperation Indicates the patient's preoperative UDPRS-III value, UPDRS-III postoperation It indicates the UDPRS-III value of the patient after surgery.

[0059] The optimal machine learning prediction model is obtained, and the optimal parameter machine learning regression model is stored for use in the prediction module.

[0060] Step 5: Use the prediction model to predict the patient's postoperative effect.

[0061] Input the patient's fMRI image data, then follow the methods described in steps 1 to 3 to perform data preprocessing, build a brain functional network, and screen features. Finally, use the prediction model trained in step 4 to predict the patient's postoperative recovery.

[0062] Beneficial Effects

[0063] Compared with the prior art, the method of the present invention has the following advantages:

[0064] (1) The system of the present invention adopts two brain functional networks, both of which take dynamic factors into consideration and have low time granularity. They are constructed using a sliding window strategy and capture the dynamic changes of the brain over a period of time through functional matrices constructed through different windows.

[0065] (2) The random dynamic brain functional network adopted by the system of the present invention has the advantage of multi-scale and can more comprehensively extract dynamic features within different small time periods.

[0066] (3) The system of the present invention adopts the analysis method of graph theory for both fixed dynamic brain function network and random dynamic brain function network, combining three kinds of information: dynamic brain activity, shortest path length or global efficiency of ROI nodes, and correlation between ROIs. The brain network formed has stable and effective brain connection characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a schematic diagram of the composition of the DBS postoperative effect prediction system for Parkinson's patients based on brain functional network proposed by the present invention.

[0068] Figure 2 This is a flow chart of the fixed dynamic brain network construction method proposed by the present invention for predicting the recovery effect after Parkinson's disease DBS surgery.

[0069] Figure 3 This is a flow chart of the random dynamic brain network construction method proposed by the present invention for predicting the recovery effect after Parkinson's DBS surgery.

[0070] Figure 4 Schematic diagram of the process for constructing a fixed dynamic brain functional network.

[0071] Figure 5 Schematic diagram of the process of constructing a random dynamic brain functional network.

[0072] Figure 6 Schematic diagram of nested cross validation.

[0073] Figure 7 This is a fitting diagram of the predicted value and true value of the postoperative UDPRS-III improvement rate based on the static brain functional network.

[0074] Figure 8 This is a fitting diagram of the predicted value and true value of the postoperative UDPRS-III improvement rate based on the root mean square dynamic brain functional network.

[0075] Fig. 9 This is a fitting diagram of the predicted and true values ​​of the postoperative UDPRS-III improvement rate based on the dynamic brain functional network with a tapered sliding window.

[0076] Fig.10 This is a fitting diagram of the predicted and true values ​​of the postoperative UDPRS-III improvement rate based on a fixed dynamic brain functional network.

[0077] Fig.11 This is a fitting diagram of the predicted and true values ​​of the postoperative UDPRS-III improvement rate based on the random dynamic brain functional network. DETAILED DESCRIPTION

[0078] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0079] A prediction system for DBS postoperative effect in Parkinson's patients based on brain functional network, such as Figure 1 As shown, it includes a data preprocessing module, a dynamic brain functional network construction module, a feature screening module, a prediction model training module and a prediction module.

[0080] First, the fMRI image data of multiple patients are input into the data preprocessing module. After data preprocessing, the data preprocessing module obtains the patient's BOLD time series and sends it to the brain function network construction module.

[0081] Then, the brain functional network construction module constructs a fixed dynamic brain functional network and a random dynamic brain functional network according to the patient's BOLD time series, and inputs them into the feature screening module.

[0082] Afterwards, the feature screening module uses the random forest algorithm to perform feature screening on the patient's fixed dynamic brain functional network and random dynamic brain functional network to obtain the patient's important brain network features.

[0083] Then, the prediction model training module performs regression model training according to the patient's brain network characteristics and the patient's corresponding preoperative UDPRS-III and postoperative UDPRS-III scores, and inputs the trained regression model into the prediction module for subsequent prediction.

[0084] When predicting the postoperative effect of a new patient, the patient's fMRI image data is input into the data preprocessing module. Then, data preprocessing, brain functional network construction, and feature screening are performed in sequence. The trained regression model is used in the prediction module to predict the patient's postoperative recovery.

[0085] A method for predicting the effect of DBS surgery on Parkinson's disease patients based on brain functional network. Figure 2 As shown, the following steps are included:

[0086] Step 1: Data preprocessing.

[0087] Step 1.1: Since MRI images are obtained by layer scanning, this method selects the middle layer of the brain as the standard and ensures the uniformity of the scanning time of each layer within a scanning cycle through the time correction step in GRETNA.

[0088] Step 1.2: This method evaluates the subject's head movement, takes the first image of each subject's image sequence as a reference, and aligns the remaining images to the reference to adjust the image misalignment at different times caused by the subject's head movement.

[0089] Step 1.3: This method takes into account standardized accuracy and time cost, uses a unified segmentation algorithm, and uses the subject's T1 structural image to align all subject functional images by aligning the subject's functional image to a standard brain template.

[0090] Step 1.4: In this method, the fMRI image of the subject is passed through a 4mm×4mm×4mm half-width (i.e., the wavelength width of the horizontal axis corresponding to half of the vertical axis peak) Gaussian filter, which significantly improves the signal-to-noise ratio of the image.

[0091] Step 1.5: Due to physiological changes in the scanned person or instrument instability, rsfMRI data may show systematic signal increases or decreases. Linear trend elimination can eliminate the impact of the offset generated by the instrument during scanning on subsequent data processing. This method uses a linear fitting strategy in the linear trend elimination step to reduce the error caused by the instrument itself when acquiring data.

[0092] Step 1.6: Because spontaneous brain function activities are most active in the low frequency band, this experiment uses a box filter of 0.01Hz to 0.1Hz to eliminate other frequency noises generated by the machine or the subject's heartbeat.

[0093] Step 1.7: BNA was proposed by the Brain Network Group Research Center of the Chinese Academy of Sciences. The atlas divides the human brain into 246 regions (36 subcortical nucleus subregions and 210 cortical regions), and the 246 brain regions can be divided into 24 gyri or 7 lobes. The data of this atlas comes from the brain images of 40 healthy adults, and its drawing integrates multimodal MRI data to make the partition more refined. Because the experimental data set is the resting brain function images of 54 people, and it is hoped to construct a brain function network based on the fine brain region division, this method selects the BNA template to construct the brain function network with 246 brain regions as ROI.

[0094] Step 2: Construct brain functional network.

[0095] In this embodiment, a total of 5 brain function networks were constructed, among which the latter two (fixed dynamic and random dynamic brain networks) are new brain function networks proposed by this method for predicting the recovery effect of Parkinson's disease after DBS surgery. The first three use the current common network construction method to compare the prediction effect with this method.

[0096] Step 2.1: Build a static functional network

[0097] Step 2.2.1: Calculate the Pearson correlation coefficient between any two time-series BOLD signals in the 246 ROIs to obtain the static brain functional network of each subject.

[0098] Step 2.2: Constructing a dynamic brain functional network based on RMS

[0099] Step 2.2.1: Select a sliding window of length 50, and divide the BOLD time series signal into 191 overlapping time series subsequences according to a fixed step length of 1. Each subsequence is a window.

[0100] Step 2.2.2: Calculate the root mean square value RMS between the i-th ROI and the j-th ROI through formula (5), where N is the number of windows. It represents the connection strength between the i-th ROI and the j-th ROI in the n-th sliding window, forming a dynamic brain functional network based on the root mean square.

[0101]

[0102] Among them, dFC ij Represents the position of the i-th row and j-th column in the RMS-based dynamic brain functional network.

[0103] Step 2.3: Construct a dynamic brain functional network based on a tapered sliding window.

[0104] Step 2.3.1: Select a tapered sliding window with a length of 50, where the tapered sliding window is a tapered cosine window, and the parameter correlation coefficient r of the window is 0.5. According to the fixed step length of 1, the time series BOLD signal is divided into 191 overlapping time series subsequences, and each subsequence is a window.

[0105] Step 2.3.2: Calculate the Pearson correlation coefficient between any two ROIs in the 246 ROIs in each conical window to form a 246 × 246 correlation matrix. Perform Fisher transformation on the matrix according to formula (6) to obtain the transformed brain connection matrix of each window.

[0106]

[0107] Wherein, z represents the result obtained by Fisher transform.

[0108] Step 2.3.3: Calculate the variance of the connection between the i-th ROI and the j-th ROI in 191 windows respectively, and obtain the final dynamic brain functional network based on the conical sliding window.

[0109] Step 2.4: Construct a fixed dynamic brain functional network.

[0110] Step 2.4.1: Select a sliding window of length 50, and divide the BOLD time series signal of each ROI into 191 overlapping subsequences according to a fixed step length of 1. Each subsequence is a window.

[0111] Step 2.4.2: Calculate the shortest path length of each small window of the 246 ROIs using formula (1).

[0112] Step 2.4.3: Calculate the Pearson correlation coefficient r of the shortest path length sequence between any two ROIs among the 246 ROIs using formula (2) to obtain the final fixed dynamic brain functional network.

[0113] The construction process is as follows Figure 3 shown.

[0114] Step 2.5: Constructing random dynamic brain functional networks

[0115] Step 2.5.1: Select a sliding window with a length ranging from 25 to 75, and divide the BOLD time series signal into 191 overlapping subsequences with no time series between different windows and time series within the window according to the step length ranging from 1 to 90. Each subsequence is a window.

[0116] Step 2.5.2: Calculate the node global efficiency of 246 ROIs in each window by formula

[0117] Step 2.5.3: Calculate the Pearson correlation coefficient of the global efficiency series of any two ROIs among the 246 ROIs to obtain the final random dynamic brain network.

[0118] The process of constructing random dynamic brain functional networks is as follows Figure 4 shown.

[0119] Step 3: Use the random forest algorithm to perform feature screening on the constructed brain network.

[0120] Step 3.1: Calculate the importance of each brain connection feature based on the random forest algorithm.

[0121] Step 3.2: Map the importance to the left and right brain regions of the connection and then normalize them.

[0122] Step 3.3: Repeat the above steps for each brain connection feature to obtain the importance weight of each brain region under the network.

[0123] Step 4: DBS postoperative prediction model training.

[0124] Common linear regression model and nonlinear regression model were used to predict the postoperative effect of DBS and the prediction effects were compared.

[0125] Step 4.1: Inner cross-validation. In the training process of each machine learning model, the model is optimized by grid search, that is, all candidate parameters are continuously looped through, and each parameter combination obtained by comparison and its corresponding predicted value is obtained. When the predicted value is closest to the true value, the model at this time is the optimal model. Taking the Ridge model as an example, its parameters include: alpha, solver, etc. Assuming that there are 5 possibilities for alpha and 3 possibilities for solve, if only two parameters are optimized, their situations are exhaustive.

[0126] Step 4.2: Outer cross validation. This experiment splits the data of 54 people and uses ten-fold cross validation to ensure that nine copies of the data are used for training and one copy of the data is used for validation each time. Finally, the ten validation results are averaged to obtain the final prediction effect. The schematic diagram of nested cross validation is shown in the figure below. Figure 5 shown.

[0127] Step 4.3: Use least squares linear regression, ridge regression, lasso regression, support vector machine regression, and extreme gradient regression for training. Use the grid search algorithm in the inner loop of nested cross-validation to obtain the optimal parameters of each model, and then perform ten-fold cross-validation in the outer loop. According to formula (4), the predicted value of the postoperative UDPRS-III improvement rate of each subject is obtained, and the Pearson correlation coefficient, mean absolute error (MAE), and root mean square error (RMSE) with the true value are calculated as the prediction effect evaluation indicators. The optimal machine learning prediction model is obtained, and the optimal parameter machine learning regression model is stored.

[0128] Step 5: Use the prediction model to predict the patient's postoperative effect.

[0129] Input the patient's fMRI image data, and then follow the methods described in steps 1 to 3 to perform data preprocessing, build a brain functional network, and screen features. Finally, use the trained prediction model to predict the patient's postoperative recovery.

[0130] When the brain connectivity features are derived from the static brain functional network, the Pearson correlation coefficient, MAE, and RMSE of each regression model are shown in Table 1. The fitting of the predicted value and the actual value of the postoperative UDPRS-III improvement rate of each regression model is shown in Table 1. Figure 6 shown.

[0131] Table 1 Prediction results of machine learning model based on static brain functional network

[0132]

[0133] When the brain connectivity features are derived from the dynamic brain functional network based on the root mean square, the Pearson correlation coefficient, mean absolute error and root mean square error of each regression model are shown in Table 2. The fitting of the predicted value and the actual value of the UDPRS-III improvement rate of each model is shown in Figure 7 shown.

[0134] Table 2 Prediction results of machine learning model based on RMS dynamic brain functional network

[0135]

[0136] When the brain connectivity features are derived from the dynamic brain functional network based on the tapered sliding window, the Pearson correlation coefficient, MAE, and RMSE of each regression model are shown in Table 3. The fitting of the predicted value and the actual value of the postoperative UDPRS-III improvement rate of each model is shown in Figure 8 shown.

[0137] Table 3 Prediction results of machine learning model of dynamic brain functional network based on conical sliding window

[0138]

[0139] When the brain connectivity features are derived from a fixed dynamic brain functional network, the Pearson correlation coefficient, MAE, and RMSE of each regression model are shown in Table 4. The predicted value and actual value of the postoperative UDPRS-III improvement rate of each model are shown in Table 4. Fig. 9 shown.

[0140] Table 4 Prediction results of machine learning model based on fixed dynamic brain functional network

[0141]

[0142] When the brain connectivity features are derived from random dynamic brain functional networks, the Pearson correlation coefficients, MAE, and RMSE of each regression model are shown in Table 5. The predicted values ​​and actual values ​​of the postoperative UDPRS-III improvement rate of each model are shown in Table 5. Fig.10 shown.

[0143] Table 5 Prediction results of machine learning model based on random dynamic brain functional network

[0144]

[0145] According to the database used in this experiment, compared with the two dynamic networks proposed by this method, the prediction effects of the static functional network, the dynamic brain functional network based on root mean square and the dynamic brain functional network based on conical sliding window are relatively poor.

[0146] The dynamic brain functional network based on root mean square and the brain dynamic network based on conical sliding window introduced dynamic characteristics of brain connections, but because they did not conduct deep information mining, the created dynamic brain networks were not suitable for the goal of predicting the recovery effect of DBS surgery in Parkinson's patients in this paper, resulting in poor final prediction results.

[0147] In the process of constructing fixed dynamic brain functional networks and random dynamic brain functional networks, the advantages of the first two comparative dynamic brain networks were borrowed, and the sliding window strategy was also adopted to extract the dynamic changes of brain connections. It introduced the node shortest path length index and global efficiency index in graph theory, starting from the node, and calculated the shortest path length or global efficiency of 246 ROIs, thereby obtaining the information transmission capacity of each ROI under a certain window. Then, the Pearson correlation coefficient of the shortest path length time series or global efficiency time series between any two ROIs was calculated as the brain connection strength of the final brain network, and the correlation between the information transmission capacity of each ROI node was introduced. Therefore, the prediction accuracy of these two brain functional networks is higher than that of other methods.

Claims

1. A system for predicting the effect of DBS surgery on Parkinson's disease patients based on brain function network, characterized in that: It includes data preprocessing module, dynamic brain function network construction module, feature screening module, prediction model training module and prediction module; Among them, the data preprocessing module is responsible for dividing the brain region ROI of the preprocessed patient MRI image and generating the BOLD timing signal of the ROI node. The input of this module is the patient's fMRI image; The dynamic brain function network construction module is responsible for constructing fixed dynamic brain function networks and random dynamic brain function networks based on BOLD timing signals; the input of this module is the BOLD timing signal generated by the data preprocessing module; The feature screening module is responsible for using the random forest algorithm to perform feature screening on the input brain functional network; the input of this module is the brain functional network generated by the dynamic brain functional network construction module; The prediction model training module is responsible for the regression model training; the input of this module is the characteristics of the brain functional network obtained by the feature screening module, as well as the preoperative UDPRS-III and postoperative UDPRS-III scores of the corresponding patients; The prediction module is responsible for predicting the postoperative effect of DBS on patients using a regression model. The input of this module is the fMRI image of the patient to be predicted. The connection relationship between the above modules is: The output end of the data preprocessing module is connected to the input end of the dynamic brain functional network construction module; The output end of the dynamic brain function network construction module is connected to the input end of the feature screening module; The output end of the feature screening module is connected to the input end of the prediction model training module and the prediction module respectively; The output end of the prediction model training module is connected to the input end of the prediction module; The system works as follows: First, the fMRI image data of multiple patients are input into the data preprocessing module; after the data preprocessing module performs data preprocessing, the BOLD time series of the patients is obtained and sent to the brain function network construction module; Then, the brain function network construction module constructs a fixed dynamic brain function network and a random dynamic brain function network according to the patient's BOLD time series, and inputs them into the feature screening module; The fixed dynamic brain function network is constructed in the following manner, including the following steps: Select a sliding window of length l, and divide the BOLD time series signal of each ROI into n overlapping subsequences according to a fixed step length of s. Each subsequence is a window. Calculate the shortest path length of each ROI in each window respectively; regard each window as a weighted network G with K nodes. For two nodes i and j, among all their paths, the path with the smallest accumulated weight is called the shortest path between the two nodes, denoted as For a node i in a weighted network G with a total of K nodes, its shortest path length is defined as follows: The Pearson correlation coefficient r of the shortest path length sequence between any two ROIX and ROIY in all ROIs was calculated to obtain the final fixed dynamic brain functional network: Among them, x a represents the mean of vector X, y a represents the mean of vector Y; The random dynamic brain functional network is constructed in the following way, including the following steps: Select a sliding window with a length range of l_min to l_max, and divide the BOLD time series signal into n overlapping subsequences with no time series between different windows and time series within the window according to the step length range of s_min to s_max. Each subsequence is a window. Calculate the global efficiency of the nodes for each ROI in each window separately: Among them, K represents the total number of nodes, G represents the network, represents the global efficiency of node i, represents the weighted shortest distance between nodes i and j; The Pearson correlation coefficient of the global efficiency series of any two ROIs among all ROIs was calculated to obtain the final random dynamic brain network; Afterwards, the feature screening module uses the random forest algorithm to perform feature screening on the patient's fixed dynamic brain functional network and random dynamic brain functional network to obtain the patient's brain network characteristics; Then, the prediction model training module performs regression model training according to the patient's brain network characteristics and the patient's corresponding preoperative UDPRS-III and postoperative UDPRS-III scores, and inputs the trained regression model into the prediction module for subsequent prediction; When predicting the postoperative effect of a new patient, the patient's fMRI image data is input into the data preprocessing module. Then, data preprocessing, brain functional network construction, and feature screening are performed in sequence. The trained regression model is used in the prediction module to predict the patient's postoperative recovery.

2. A method for predicting the effect of DBS surgery on Parkinson's disease patients based on brain functional network, characterized in that: The following steps are involved: Step 1: Perform image preprocessing on the original MRI image data; divide the brain region ROI on the preprocessed MRI image and generate the BOLD signal of the ROI node; Step 2: construct a dynamic brain function network, including constructing a fixed dynamic brain function network and a random dynamic brain function network, including the following steps: Step 2.1: Construct a fixed dynamic brain functional network: Step 2.1.1: Select a sliding window of length l, and divide the BOLD time series signal of each ROI into n overlapping subsequences according to a fixed step length of s. Each subsequence is a window. Step 2.1.2: Calculate the shortest path length of each ROI in each window respectively; treat each window as a weighted network G with K nodes. For two nodes i and j, among all their paths, the path with the smallest accumulated weight is called the shortest path between the two nodes, denoted as For a node i in a weighted network G with a total of K nodes, its shortest path length is defined as follows: Step 2.1.3: Calculate the Pearson correlation coefficient r of the shortest path length sequence between any two ROI X and ROI Y in all ROIs to obtain the final fixed dynamic brain functional network: Among them, x a represents the mean of vector X, y a represents the mean of vector Y; Step 2.2: Construct a random dynamic brain functional network, including the following steps: Step 2.2.1: Select a sliding window with a length range of l_min to l_max, and divide the BOLD time series signal into n overlapping subsequences with no time series between different windows and time series within the window according to the step length range of s_min to s_max. Each subsequence is a window. Step 2.2.2: Calculate the node global efficiency of each ROI in each window separately: Among them, K represents the total number of nodes, G represents the network, represents the global efficiency of node i, represents the weighted shortest distance between nodes i and j; Step 2.2.3: Calculate the Pearson correlation coefficient of the global efficiency series of any two ROIs among all ROIs to obtain the final random dynamic brain network; Step 3: Use the random forest algorithm to perform feature screening on the constructed brain network; The importance of each brain connection feature is calculated using the random forest algorithm, and the importance is mapped to the left and right brain regions of the connection and then normalized; the above operation is repeated for each brain connection feature to obtain the importance weight of each brain region under the network; Step 4: Training DBS postoperative prediction model; Step 5: Use the prediction model to predict the patient's postoperative effect.

3. The method for predicting the effect of DBS surgery on Parkinson's disease patients based on brain function network according to claim 2, characterized in that: Step 4 includes the following steps: Step 4.1: Perform inner cross validation; In the training process of each machine learning model, the model is optimized by grid search, that is, all candidate parameters are continuously looped through, and each parameter combination obtained is compared with its corresponding predicted value. When the predicted value is closest to the true value, the model at this time is the optimal model. Step 4.2: Perform outer cross-validation; split the source data, and use k-1 pieces of data for training and 1 piece of data for validation each time through k-fold cross-validation; finally, take the average of the k-fold validation results to get the final prediction effect; Step 4.3: Train the least squares linear regression, ridge regression, lasso regression, support vector machine regression and extreme gradient regression models respectively; In the nested cross-validation inner loop, the grid search algorithm was used to determine the optimal parameters of each model. In the outer loop, cross-validation was performed to obtain the predicted value of the postoperative UDPRS-III improvement rate of each subject, and the Pearson correlation coefficient, mean absolute error, and root mean square error with the true value were calculated as the prediction effect evaluation indicators: Among them, UPDRS-III ratio Indicates the improvement rate of UDPRS-III, UPDRS-III preoperation Indicates the patient's preoperative UDPRS-III value, UPDRS-III postoperation It represents the UDPRS-III value of the patient after surgery; Obtain the optimal machine learning prediction model from it.

4. The method for predicting the effect of DBS surgery on Parkinson's disease patients based on brain function network according to claim 2, characterized in that: Step 5 is implemented as follows: Input the patient's fMRI image data, and then perform data preprocessing, brain functional network construction and feature screening in sequence according to the methods described in steps 1 to 3; Finally, the prediction model trained in step 4 is used to predict the patient's postoperative recovery.

Images