Drug addict sleep quality classification method based on dynamic function connection and 3DGraphCAN model
By combining dynamic effective connection and 3D graph convolution neural network model, we can handle the dynamic changes in the brain of drug users, and solve the problem of poor sleep quality classification effect in the existing technology, achieving higher accuracy and sensitivity, suitable for drug addicts and other research on complex brain networks.
Patent Information
- Application Number
- CN202411106360.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-13
- Publication Date
- 2024-11-22
AI Technical Summary
When the prior art deals with dynamic changes in the brain of drug users, the sleep quality classification effect is poor, the sensitivity is low, and static analysis cannot fully capture the complexity and time-varying characteristics of brain functional networks.
Dynamic effective connection (DEC) is used to combine with 3D graph convolutional neural network (3DGraphCAN) model, and dynamic functional connection matrix is constructed through sliding time window technology, combined with deep learning technology, using 3D convolutional layer, graph convolutional layer and channel attention Mechanism, extracting spatiotemporal and spatial characteristics and complex interactions of brain intervals, and enhancing the model's sensitivity to key brain region connections.
It significantly improves the classification accuracy and sensitivity of sleep quality for drug addicts, can effectively distinguish the sleep quality of drug addicts and healthy people, and is suitable for the processing of dynamic characteristics of complex brain networks.
Smart Images

Figure CN119006420A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the application fields of neuroimaging, machine learning and neural network models, and in particular to a method for classifying the sleep quality of drug addicts based on dynamic functional connectivity analysis and a 3D graph convolutional neural network (3DGraphCAN) model. The method combines functional magnetic resonance imaging (fMRI) data analysis and deep learning technology, and is suitable for sleep research and clinical diagnosis in the fields of psychiatry, addiction medicine and neuroscience. Background Art
[0002] Drug addiction not only seriously affects the physical health of individuals, but also leads to significant mental and psychological problems, especially in terms of sleep quality. Drug addicts often show significant sleep disorders, such as insomnia, intermittent sleep, and non-restorative sleep. Existing sleep research mainly relies on effective connection analysis. Although this method reveals the connection relationship between brain regions to a certain extent, it ignores the dynamic changes of the brain in different states. Especially for a group with highly dynamic brain changes such as drug addicts, static analysis cannot fully capture the complexity and time-varying characteristics of their brain functional networks.
[0003] In recent years, Dynamic Effective Connectivity (DEC) has gradually become an important tool for studying the spatiotemporal dynamic characteristics of the brain. DEC can provide more detailed brain network dynamic information by using a sliding time window method to record the changes in brain functional connectivity in real time. However, most of the current research on DEC focuses on the analysis of resting-state brain connectivity patterns, and lacks in-depth research on the special sleep problems of drug addicts. At the same time, traditional machine learning models often show low classification performance when processing DEC data due to its complex spatiotemporal structure.
[0004] In order to solve the above problems, this paper proposes a sleep quality classification method that combines a dynamic functional connectivity matrix and a 3D graph convolutional neural network (3DGraphCAN) model, aiming to improve the classification accuracy and sensitivity of sleep disorders in drug addicts. By combining DEC with a deep learning model, the characteristics of dynamic changes in the brain of drug addicts can be more effectively captured, thereby providing a scientific basis for clinical treatment and rehabilitation. Summary of the invention
[0005] The main purpose of the present invention is to provide a method for classifying the sleep quality of drug addicts based on the Dynamic Effective Connectivity (DEC) and 3D Graph Convolutional Neural Network (3DGraphCAN) model. This method improves the accuracy and sensitivity of classifying the sleep quality of drug addicts by dynamically monitoring the spatiotemporal changes of brain functional connections in drug addicts and combining deep learning technology. It aims to solve the problem of poor classification effect and low sensitivity in the existing technology when processing the dynamic characteristics of complex brain networks.
[0006] The technical solution adopted by the present invention is:
[0007] Step 1: Obtain resting-state functional magnetic resonance imaging (rs-fMRI) data from drug addicts (including those who are using methadone and those who are not) and healthy controls. Each subject is required to stay awake but as relaxed as possible during the scan, and their Pittsburgh Sleep Quality Index (PSQI) score is recorded to assess sleep quality.
[0008] Step 2, using standard fMRI data preprocessing steps, including:
[0009] · De-artifacting: Eliminate artifacts caused by equipment or environmental factors.
[0010] Time correction: Perform time correction on time series data.
[0011] ●Head motion correction: Correct the signal drift caused by the subject's head movement.
[0012] Spatial normalization: Unify all data into the same spatial coordinate system.
[0013] Smoothing: A 6 mm FWHM Gaussian kernel was used to smooth the data.
[0014] Finally, the AAL template was used to extract time series signals from 90 brain regions to provide standardized data for subsequent analysis.
[0015] Step 3, using sliding time window technology to segment fMRI data into multiple overlapping time segments, calculate the Granger causal relationship between brain regions in each time window, and thus construct a dynamic effective connectivity matrix (DECM). DECM can capture the dynamic changes of causal relationships between brain regions over time, which is one of the core features of the present invention.
[0016] Step 4: One-way analysis of variance (ANOVA) was performed on all constructed DECMs to determine the statistical significance of the connections between brain regions. A weight matrix (EC_weighted_Matrix) was constructed based on the results of ANOVA, which was used to enhance the connections of brain regions that were important in the classification of sleep quality in drug addicts, ensuring that the model could better identify key brain network features.
[0017] Step 5: The 3DGraphCAN model proposed in the present invention combines the advantages of convolutional neural network (CNN) and graph convolutional neural network (GCN). The model includes:
[0018] 3D convolutional layers: used to extract spatial and temporal features in fMRI data.
[0019] Graph convolutional layer: used to process the graph structure in DECM data and capture the complex dynamic interactions between brain regions.
[0020] Channel attention mechanism: further improve the model's sensitivity to key brain area information and optimize classification results.
[0021] The model is trained on the training set, and parameters are tuned on the validation set to ensure the accuracy and generalization ability of the model.
[0022] Step 6: Use the trained 3DGraphCAN model to classify the test set. The performance of the model was evaluated by cross-validation, and the main indicators included classification accuracy, precision, recall, and F1-Measure value. The experimental results showed that the model can effectively distinguish the sleep quality of drug addicts and healthy people, especially showing high sensitivity and accuracy when processing complex brain networks.
[0023] Compared with the prior art, the present invention has the following beneficial effects:
[0024] 1. Improve classification accuracy: By combining dynamic functional connectivity analysis and 3DGraphCAN model, the present invention significantly improves the classification accuracy of sleep quality of drug addicts, especially in processing complex spatiotemporal dynamic characteristics of brain networks;
[0025] 2. Enhance model sensitivity: Use weight matrices to enhance the connections between important brain regions, so that the model can more sensitively capture brain region connections related to sleep disorders and improve classification results;
[0026] 3. Potential for wide application: This method is not only suitable for the research and clinical diagnosis of sleep quality in drug addicts, but can also be extended to other fields involving dynamic analysis of complex brain networks, such as the research on depression, anxiety and other mental disorders. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 The present invention relates to an overall idea and a data processing scheme;
[0028] Figure 2 The present invention relates to a 3DGraphCAN model construction scheme;
[0029] Figure 3 The invention relates to the construction of a weight matrix result based on variance analysis and multiple hypothesis verification. DETAILED DESCRIPTION
[0030] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0031] The embodiments of the present invention and their implementation process are as follows:
[0032] Taking the data provided by Yunnan Drug Dependence Treatment Technology Innovation Center as an example, the implementation mode of the present invention is described in detail based on processing its real data.
[0033] The data was provided by Yunnan Drug Dependence Treatment Technology Innovation Center. All data were collected on GE MEDICAL SYSTEMs, model Discovery MR750w, using EPI sequence to obtain rs-fMRI images. The main parameters are as follows: thickness = 3mm; TR = 2000ms; TE = 30ms; FOV = 224mm2. Each subject was subjected to a resting state operation lasting 8 minutes (240TRs). After the rs-fMRI image data were collected, the PSQI scale test was performed. All healthy subjects had no sleep problems. The PSQI scores of methadone addicts were all less than or equal to 5, while the PSQI scores of drug addicts who were not treated with methadone were all greater than or equal to 6.
[0034] In addition, the data included 30 healthy subjects and 96 drug addicts. All subjects were asked to stay relaxed but not fall asleep during the scanning process. Due to the missing of some data, the valid data included 29 healthy subjects and 90 drug addicts, of which 35 drug addicts used methadone to treat their sleep condition, and the other 55 subjects did not use methadone to treat their sleep. Table 1 shows the demographic and clinical characteristics of all subjects in this experiment.
[0035] Table 1 Summary of demographic information and other sample characteristics
[0036]
[0037] In order to apply the sleep quality classification method of drug addicts based on dynamic functional connectivity and 3DGraphCAN model, the following steps were formulated:
[0038] Step 1: Data preprocessing was performed using GRETNA software, including the following steps:
[0039] Artifact removal: Eliminate artifacts caused by magnetic field inhomogeneity or equipment problems during fMRI scanning.
[0040] Time correction: Time series data is time-corrected to compensate for the differences in acquisition time of different slices.
[0041] Head motion correction: Correct the signal shift caused by the subject's head movement through rigid registration.
[0042] Spatial normalization: All data are converted to a standardized spatial coordinate system, usually using the MNI (Montreal Neurological Institute) coordinate system.
[0043] Smoothing: The data were spatially smoothed using a 6 mm FWHM (full width at half maximum) Gaussian kernel to improve the signal-to-noise ratio.
[0044] After preprocessing, the AAL (Automated Anatomical Labeling) template was used to extract the time series signals of 90 brain regions. These signals will be used for the subsequent construction of the dynamic functional connectivity matrix.
[0045] Step 2, the dynamic functional connectivity matrix (Dynamic Effective Connectivity Matrix, DECM) was constructed using the sliding time window method. The sliding window parameters were set, the window width was set to 50 time points (corresponding to 100 seconds of fMRI data), and the overlap between windows was 96%. The sliding window method was used to divide the time series data into multiple overlapping subsequences, sliding 2 time points each time. Then, in each time window, the Granger causal relationship between each pair of brain regions was calculated to determine the strength of the causal relationship between the two brain regions. According to the strength of the causal relationship, the DECM matrix in the corresponding time window was constructed. The elements of each matrix represent the strength of the causal influence of one brain region on another brain region. The generated DECM matrix will capture the dynamic changes of the brain functional network in different time windows, reflecting the differences in brain functional connectivity between drug addicts and the healthy control group.
[0046] Step 3: Perform one-way analysis of variance (ANOVA) on the connection strength between each pair of brain regions, calculate the P value, and determine the significant difference in connection between different groups (drug addicts and healthy people). Benjamini-Hochberg multiple hypothesis test correction is performed on the P value to determine the significance threshold. According to the corrected P value, brain region connections with larger significant differences are given higher weights to construct a weight matrix (EC_weighted_Matrix). This weight matrix is used for subsequent 3DGraphCAN model training to help the model better identify brain region connections that contribute significantly to classification.
[0047] Step 4, a 3DGraphCAN model combining convolutional neural network (CNN) and graph convolutional neural network (GCN) is proposed to process DECM data and perform classification. The model structure is as follows:
[0048] 3D convolution layer: Input the preprocessed DECM matrix and use the 3D convolution layer to extract spatial and temporal features. The convolution kernel size is 3x3x3.
[0049] Max pooling layer: Reduces the dimension of features through pooling operation and retains the most important feature information.
[0050] Graph convolution layer: The pooled features are input into the graph convolution layer to process graph structure data and capture the complex interactions between brain regions.
[0051] Channel attention mechanism: further optimizes the feature map and improves the classification performance of the model by increasing the weights of the connections between key brain regions.
[0052] The constructed model was applied to the training set, optimized using the cross entropy loss function, and the validation set was used to tune the model parameters to ensure the generalization ability of the model. The trained 3DGraphCAN model was classified using the test set to evaluate the performance of the model. The model effect was comprehensively evaluated using indicators such as classification accuracy, precision, recall, and F1-Measure value. Finally, all data were randomly divided into training set, validation set, and test set in a ratio of 6:2:2, and ten cross-validations were performed to ensure the stability and reliability of the classification results. The average classification accuracy of the model on the test set was 81%, the precision was 83%, the recall was 81%, and the F1-Measure value was 80%. The results showed that the 3DGraphCAN model can effectively distinguish the sleep quality of drug addicts from the healthy control group, especially when dealing with the dynamic characteristics of complex brain networks, showing high sensitivity and accuracy.
[0053] In summary, the present invention provides an efficient method for classifying sleep quality of drug addicts, which combines dynamic functional connectivity and 3DGraphCAN model to improve classification accuracy and sensitivity, and has broad application prospects.
Claims
1. A sleep quality classification method for drug addicts based on dynamic functional connectivity and 3DGraphCAN model, characterized in that: The method comprises the following steps: In step 1, resting-state functional magnetic resonance imaging (rs-fMRI) data were collected from drug addicts and healthy controls, and the Pittsburgh Sleep Quality Index (PSQI) scores of the subjects were recorded. In step 2, the collected rs-fMRI data were preprocessed, including artifact removal, time correction, head motion correction, spatial normalization, and smoothing, and the time series signals of 90 brain regions were extracted using the AAL template. Step 3: Use the sliding time window technique to divide the preprocessed fMRI data into multiple overlapping time segments, calculate the Granger causal relationship between brain regions in each time window, and thus construct a dynamic effective connectivity matrix (DECM). Step 4: Perform one-way analysis of variance (ANOVA) on all constructed DECMs to construct a weight matrix (EC_weighted_Matrix) for enhancing the classification effect, which gives higher weights to brain region connections with significant differences. Step 5, a 3DGraphCAN model combining convolutional neural network (CNN) and graph convolutional neural network (GCN) is proposed. This model extracts the spatial and temporal features of fMRI data through 3D convolutional layers, uses graph convolutional layers to process the graph structure in DECM, and further optimizes the classification effect through the channel attention mechanism. Step 6: Use the trained 3DGraphCAN model to classify the sleep quality of the fMRI data in the test set, and evaluate the model performance through indicators such as classification accuracy, precision, recall rate, and F1-Measure value.
2. According to claim 1, a method for classifying sleep quality of drug addicts based on dynamic functional connectivity and 3DGraphCAN model, characterized in that: The sliding time window technique comprises the following steps: Step 3.1: The window width is set to 50 time points, and the overlap between windows is 96%, that is, 2 time points are slid each time; Step 3.2: Within each time window, the causal relationship between each pair of brain regions is calculated by Granger causality analysis to construct a dynamic effective connectivity matrix (DECM).
3. The sleep quality classification method for drug addicts based on dynamic functional connectivity and 3DGraphCAN model according to claim 1, characterized in that: The method for constructing the weight matrix includes: Step 4.1: Perform one-way analysis of variance (ANOVA) on the strength of each pair of brain region connections in the DECM matrix and calculate the significant P value of each brain region connection; Step 4.2: Perform Benjamini-Hochberg multiple hypothesis test correction on the significant P value, determine the significance threshold, and construct a weight matrix based on the corrected P value.
4. The sleep quality classification method for drug addicts based on dynamic functional connectivity and 3DGraphCAN model according to claim 1, characterized in that: The 3DGraphCAN model includes: Multiple 3D convolutional layers to extract spatial and temporal features in rs-fMRI data. Multiple graph convolutional layers are used to process the graph structure in the DECM matrix and extract the complex interactions between brain regions. Channel attention mechanism, which is used to optimize the classification performance of the model and increase the sensitivity of the model to sleep disorders by increasing the weights of connections between key brain regions.
5. The sleep quality classification method for drug addicts based on dynamic functional connectivity and 3DGraphCAN model according to claim 1, characterized in that: The classification and evaluation steps include: Step 6.1: Use the cross-validation method to evaluate the model and calculate the classification accuracy, precision, recall and F1-Measure value of the model on the test set; Step 6.2: Based on these evaluation indicators, analyze the overall classification performance of the model and tune the model parameters according to the results of the validation set.
