Graph convolution addiction desire prediction method and system based on heart-brain coupling

Abstract

The invention provides a cardio-cerebral coupling-based graph convolution addiction craving prediction method and system. The method comprises the following steps of: obtaining an electrocardiosignal and an electroencephalogram signal generated when a subject watches a video containing a craving clue; performing fragment interception on the electroencephalogram signal near the R wave peak value by using the electrocardiosignal to obtain an electroencephalogram fragment sequence; and inputting the electroencephalogram fragment sequence into the trained desire prediction model so as to predict the desire state. According to the method, multi-mode data of electrocardio and electroencephalogram are fused, an electroencephalogram fragment sequence is intercepted in combination with heartbeat evoked potential, and a deep learning desire state prediction model is input to predict the current desire state of a subject. According to the method, subjective deviation of a traditional self-reporting method can be avoided, more objective and timely intervention measures are provided for treatment, the treatment effect is effectively improved, invalid treatment is reduced, and more scientific support is provided for rehabilitation of patients abstained from drugs.

Description

Technical Field

[0001] This application belongs to the field of addiction prediction technology, specifically relating to a graph convolutional method and system for predicting addiction cravings based on mind-brain coupling. Background Technology

[0002] As drug abuse becomes increasingly serious, drug addiction is causing more and more public health problems. Cravings among addicts are considered a key factor in relapse, and accurate and timely monitoring of cravings is crucial for effective individualized treatment. However, existing treatment methods (such as medication and psychological intervention) have low response rates and exhibit significant individual variability, making precision treatment difficult.

[0003] Electroencephalography (EEG), as a non-invasive, convenient, and high-temporal-resolution physiological signal detection technique, has been widely used in neuroscience research, especially in the study of addictive behaviors. Studies have shown that EEG signals can reflect the brain's neural responses to drug-related stimuli, making real-time monitoring of craving states possible.

[0004] Heartbeat evoked potentials (HEPs) are changes in electrical activity in the cerebral cortex triggered by heartbeats. They are typically extracted by simultaneously recording electrocardiogram (ECG) and electroencephalogram (EEG) signals and calculating heartbeat-related potentials (HEPs). Studies have shown that HEPs reflect bidirectional communication between the heart and brain, involving multiple physiological processes such as autonomic nervous system regulation, interoceptive processing, and emotional and cognitive functions. These factors all play important roles in addictive behavior and craving regulation.

[0005] In recent years, research on the neural mechanisms of addictive cravings has deepened. Zhao et al. (2021) found (Zhao, Diet al. “Neurophysiological correlate of incubation of craving in individuals with methamphetamine use disorder.” Molecular psychiatry vol. 26,11 (2021):6198-6208. doi:10.1038 / s41380-021-01252-5) that in patients with methamphetamine use disorder, the period of withdrawal lasting 1–3 months is the stage of the strongest craving, accompanied by a decrease in theta and alpha band power and an increase in beta band power, especially an enhanced beta band functional connectivity between the medial prefrontal cortex and the orbital gyrus, suggesting that EEG functional characteristics can be used to monitor the state of craving. Tian et al. (2024) proposed a craving state prediction method based on high-density electroencephalography (Tian, ​​Weiwen et al. "Anelectroencephalographic signature predicts craving form ethamphetamine." Cell reports. Medicine vol. 5, 1(2024): 101347. doi: 10.1016 / j.xcrm. 2023.101347). They divided 32 brain network regions using 128-channel EEG, calculated the functional connectivity network and extracted fictitious coherence features, and combined them with a machine learning model to predict the individual's craving level, providing data support for individualized intervention.

[0006] However, existing technologies have the following technical shortcomings: 1. Excessive use of leads leads to complex operation: Existing research typically uses high-density EEG, such as 128-lead EEG. While this can capture more brain activity, the excessive number of channels requires complex source localization and brain region segmentation during data processing. This makes data processing very cumbersome and places high demands on equipment, making it difficult to achieve portable and real-time applications.

[0007] 2. Static data collection, lacking dynamic characteristics: Traditional EEG collection is usually static data collection, with many studies conducted in a resting state. However, the craving state of drug addicts is usually triggered by external stimuli, and static collection cannot fully reflect changes in cravings and lacks sufficient dynamic response.

[0008] 3. Excessive data duration and low analysis efficiency: Current technologies typically involve long data durations, requiring extended EEG monitoring and complex post-processing analysis. However, processing long-term data is both time-consuming and labor-intensive, and it cannot achieve the immediate capture of desired changes within a short period.

[0009] 4. Machine learning training is time-consuming: Existing methods typically use traditional machine learning models, such as support vector machines and random forests. These models rely on hand-designed features, such as frequency domain analysis and functional connectivity features. Feature extraction and selection are complex, requiring not only extensive expertise but also increasing the time cost of data preprocessing. Summary of the Invention

[0010] To address the aforementioned problems in the existing technology, this application provides a graph convolutional method and system for predicting addiction cravings based on mind-brain coupling. The technical problem to be solved by this application is achieved through the following technical solution: A graph convolutional method for predicting addictive cravings based on mind-brain coupling includes: S100, acquire electrocardiogram and electroencephalogram signals generated by the subject while watching a video containing craving cues; S200, using the electrocardiogram signal to extract segments of the electroencephalogram signal near the R-wave peak, to obtain an electroencephalogram segment sequence; S300, the EEG segment sequence is input into the trained craving prediction model so that the craving prediction model extracts node features of different brain regions from the EEG segment sequence; the node features of different brain regions are aggregated to obtain local features; a global adjacency matrix is ​​constructed using the local features; the global adjacency matrix and the local features are aggregated to obtain a global feature representation, and the current craving state of the subject is predicted based on the global feature representation.

[0011] This application provides a graph convolutional addiction craving prediction system based on brain-heart coupling, comprising: The acquisition module is configured to acquire electrocardiogram (ECG) and electroencephalogram (EEG) signals generated by subjects while watching videos containing craving cues; The processing module is configured to use the electrocardiogram (ECG) signal to extract a segment sequence of EEG signals near the R-wave peak. The prediction module is configured to input the EEG segment sequence into a trained craving prediction model, so that the craving prediction model extracts node features of different brain regions from the EEG segment sequence; aggregates the node features of different brain regions to obtain local features; constructs a global adjacency matrix using the local features; aggregates the global adjacency matrix and the local features to obtain a global feature representation; and predicts the subject's current craving state based on the global feature representation.

[0012] Beneficial effects: 1. This application innovatively uses task-based data to collect electroencephalogram (EEG) signals while drug addicts are watching drug-related videos, and combines this with physiological stimulation to trigger a craving state. This method allows for more precise capture of brain activity in drug addicts during actual craving states, thereby improving the accuracy and real-time nature of predictions.

[0013] 2. This application uses simplified 6-lead data, which greatly reduces the complexity of signal processing and computation, and lowers the difficulty and cost of technical implementation. Furthermore, the use of 6-lead data makes the system more conveniently applicable for real-time prediction and intervention.

[0014] 3. This application uses only 0.8 seconds of EEG data for prediction. After each stimulation, 0.8 seconds of EEG data corresponding to the R-peak of the ECG signal are extracted for rapid analysis. This method enables rapid and efficient real-time prediction, making it particularly suitable for closed-loop neuromodulation system platforms, allowing for real-time adjustment of treatment parameters.

[0015] 4. This application collects data from only six key leads, significantly reducing hardware complexity and data dimensionality while ensuring predictive performance, thereby reducing system power consumption and improving model operating efficiency. Furthermore, the use of a deep learning model and end-to-end training method makes data processing simpler and more efficient.

[0016] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a graph convolutional method for predicting addiction cravings based on mind-brain coupling provided in this application. Figure 2 This is a schematic diagram illustrating the process of predicting the thirst state using a thirst prediction model provided in this application; Figure 3 This is a schematic diagram illustrating the construction of the HEP dataset provided in this application; Figure 4 This is a schematic diagram illustrating the process of video cues inducing cravings and the visual analog scale used to assess the degree of cravings provided in this application; Figure 5 This is a schematic diagram of the electroencephalogram (EEG) signal segmentation provided in this application; Figure 6 This is a schematic diagram illustrating the processing of collected data to obtain a training set and an independent test set, as provided in this application. Figure 7 This is a schematic diagram of the 10-fold cross-validation results of the model provided in this application; Figure 8 This is a schematic diagram of the results of the independent test set provided in this application. Detailed Implementation

[0018] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0019] like Figure 1 As shown, this application provides a graph convolutional method for predicting addictive cravings based on mind-brain coupling, including: S100, acquire electrocardiogram and electroencephalogram signals generated by the subject while watching a video containing craving cues; This step can collect EEG signals from 6 leads and ECG signals from 1 lead from the subject. This application effectively evokes cravings by having drug addicts watch drug-related videos, thereby improving the authenticity and predictive accuracy of the data; by using only 6 leads of EEG data, the amount of data is reduced, while ensuring the integrity of neural information and improving the portability and real-time monitoring capabilities of the device.

[0020] S200, using the electrocardiogram signal to extract segments of the electroencephalogram signal near the R-wave peak, to obtain an electroencephalogram segment sequence; This step involves extracting EEG segments and assembling them into 0.8-second EEG sequence sequences. This application analyzes each 0.8-second EEG signal segment separately to avoid data redundancy and enhance analysis accuracy and real-time response capabilities.

[0021] S300, the EEG segment sequence is input into the trained craving prediction model so that the craving prediction model extracts node features of different brain regions from the EEG segment sequence; the node features of different brain regions are aggregated to obtain local features; a global adjacency matrix is ​​constructed using the relationship between the local features; the global adjacency matrix and the local features are aggregated to obtain a global feature representation, and the current craving state of the subject is predicted based on the global feature representation.

[0022] refer to Figure 2 The thirst prediction model of this application includes a sequentially connected input layer, a temporal learning module, a graph learning module, a fully connected layer, and an output layer. The temporal learning module includes a sequentially connected temporal convolutional layer and a kernel-level attention fusion module; the graph learning module includes a graph convolutional neural network, a local learning module, and a global learning module; the outputs of the local learning module and the global learning module are both connected to the input of the graph convolutional neural network.

[0023] The core objective of the temporal learning module is to learn dynamic temporal features from EEG signals and capture patterns of signal change over time. The core objective of the graph learning module is to learn complex relationships between brain regions from the spatial features of EEG signals and model these spatial dependencies using graph neural networks. The local learning module specifically learns the inter-electrode dependencies within each local region by constructing local graph structures. The global learning module introduces a trainable adjacency matrix, enabling the model to learn global dependencies between brain regions, thereby improving the understanding of coordinated activities between brain regions. The collaborative work of the temporal and graph learning modules allows the model to learn EEG signals simultaneously from both temporal and spatial perspectives, thus enhancing the overall analytical capability of EEG activity.

[0024] This application predicts the subject's craving state as the subject's craving score corresponding to craving cues, which can be no craving, moderate craving, very craving, etc. The craving state can be set according to a craving score of 0-10.

[0025] This application utilizes a deep learning-based thirst prediction model to directly process time-series data, namely EEG segment sequences, and automatically extracts key features. While improving prediction accuracy and generalization ability, it supports rapid training and real-time inference, thereby providing efficient and scientific decision support for individualized intervention and effectively overcoming the limitations of traditional methods in terms of efficiency and application scenarios.

[0026] In one specific embodiment of this application, the trained desire prediction model is obtained by training a preset desire prediction model, and the training process includes: a. Collect electrocardiogram (ECG) signals, electroencephalogram (EEG) signals, and craving scores from multiple subjects at different sites while watching videos containing craving cues. refer to Figure 3 This application involves recruiting eligible drug addicts, requiring them to sit in a comfortable chair in a quiet environment, watch drug-related videos, and complete a craving assessment (using the Visual Analogue Scale). The craving score (0-10) is then used as a label, and electroencephalogram (EEG) data from six leads and electrocardiogram (ECG) data from one lead are collected during the viewing of the drug videos to construct the HEP dataset. Drug addicts meeting the inclusion and exclusion criteria are recruited, experimental precautions are clearly explained to each participant, and all participants sign written consent forms.

[0027] refer to Figure 4 Subjects watched a total of 10 minutes of video content containing craving cues, which consisted of alternating 10 30-second neutral videos and 10 30-second drug-related videos. Each video included a 3-second gaze, a 30-second video, and a 3-second craving score (VAS, 0–10).

[0028] Simultaneous acquisition of EEG and ECG signals was conducted in a quiet, soundproof environment. Subjects were instructed to sit upright in a comfortable chair and remain as quiet as possible, avoiding physical movement and verbal communication to minimize artifact interference. EEG signals were acquired using a Borricon EEG cap at a sampling frequency of 1 kHz. Before the experiment, conductive gel was applied to reduce electrode impedance, ensuring it remained below 10 kΩ to obtain high-quality EEG data.

[0029] b. Preprocess the electrocardiogram (ECG) signal and electroencephalogram (EEG) signal according to the corresponding method to obtain preprocessed ECG signal and preprocessed EEG signal; c. Using the preprocessed electrocardiogram signal, the preprocessed electroencephalogram signal is extracted to obtain electroencephalogram signal segments, and the electroencephalogram signal segments corresponding to the same site are combined into an HEP ​​dataset. d. Using the HEP dataset, and through a training strategy of ten-fold cross-validation combined with internal three-fold cross-validation, the preset thirst prediction model is iteratively trained to obtain the trained thirst prediction model.

[0030] In one specific embodiment of this application, b includes: b1, the EEG signal is sequentially subjected to electrode localization, filtering, linear drift removal, rereference, bad conduction detection, interpolation bad conduction, and independent component analysis to obtain a preprocessed EEG signal; The preprocessing workflow was implemented in the MATLAB environment using EEGLAB and its plugins. The workflow includes: electrode localization, filtering, linear drift removal, rereference, bad conduction detection, interpolation of bad conductions, and independent component analysis. Specifically, after electrode localization, the EEG signal is first subjected to bandpass filtering (0.5-100Hz) and power frequency notch filtering (48-52Hz). Then, the EEGLAB plugin Cleanlin is used to remove linear drift and eliminate slow fluctuations caused by external interference or physiological factors, followed by whole-brain averaging and rereference. Next, the EEG signal acquired during video viewing is segmented to obtain 10 segments of 30-second drug signals for each person. Bad conductions and segments are interpolated and repaired, downsampled to 500Hz, and artifacts are removed using ICA. Finally, the signal is filtered to the β band (13–30Hz) to extract features related to craving prediction. This completes the preprocessing of the EEG signal.

[0031] b2, bandpass filtering is performed on the electrocardiogram signal to obtain a preprocessed electrocardiogram signal.

[0032] The preprocessing of the ECG mainly involves a bandpass filter of 0.5-45Hz, followed by the Pan-Tompkin algorithm to identify the R-peak of the ECG signal.

[0033] In one specific embodiment of this invention, c includes: c1, with the R-wave peak value of the preprocessed ECG signal as the reference point; c2, from the first time period before the reference point to the second time period after the reference point, extract the EEG signal segment from the preprocessed EEG signal; Based on the extraction of EEG signals corresponding to the R-peak of the ECG signal. In this process, such as... Figure 5 As shown, using the R-wave peak of the electrocardiogram (ECG) signal as a reference point, this application first identifies the EEG signals associated with the ECG R-wave peak, and then extracts EEG signal segments 200ms before and 600ms after each R-wave peak, with a segment duration of 0.8 seconds. For each participant, the signal duration of watching the drug video was 30 seconds, with a total of 10 trials. In each 30-second video segment, according to the above requirements, this application can extract 22 0.8-second EEG signal segments. The extracted data consists of 10 trials per participant, with each trial containing 22 0.8-second EEG signal segments.

[0034] c3. Combine the EEG signal fragments of subjects at the same site into a HEP dataset to obtain the HEP dataset for each site.

[0035] This application can collect data from two sites: one site with 149 participants and another site with 38 participants, totaling 187 subjects. Each subject completed 10 trials, each lasting 30 seconds, yielding only 22 HEP fragments. Averaging these fragments resulted in 1870 sets of HEP data, as shown below. Figure 6 As shown, in the HEP dataset split, data from site one can be used for model training, while data from site two can be used as an independent test set, solely for evaluating the model's generalization performance.

[0036] In one specific embodiment of this application, d includes: d1. Divide the HEP dataset from the same site into 10 subsets, use 9 subsets as the training set and the remaining 1 subset as the test set; d2, the training set is divided into 3 subsets, 2 of which are used as training subsets and the remaining 1 is used as validation set; d3. The preset thirst prediction model is trained 9 times using 9 training sets. In each iteration, 2 training subsets are used for 2 internal iterations. The performance of the thirst prediction model after 2 internal iterations is verified using a validation set. The thirst prediction model after 9 iterations is then tested using the test set. The thirst prediction model that passes the test is determined as the trained thirst prediction model.

[0037] This application employs a training strategy combining ten iterations of ten-fold cross-validation with internal three-fold cross-validation. In the ten-fold cross-validation, the dataset is randomly divided into 10 subsets. Each iteration uses 9 subsets as the training set and the remaining subset as the test set, repeating this process 10 times to reduce evaluation bias and improve the model's generalization ability. In each training iteration, the training set is further subjected to three-fold cross-validation, dividing it into 3 subsets. Two of these subsets are used for training, and the remaining subset serves as the validation set for adjusting hyperparameters. Finally, the overall performance metric of the model is obtained by summing the average results of the ten iterations of ten-fold cross-validation, thus ensuring the model's stability and reliability across different data subsets. Figure 7 .

[0038] To evaluate the model's generalization ability, this application uses a separate test set for final performance evaluation. This test set is completely independent of the training and validation sets and is designed to simulate the model's performance in real-world environments. After initial optimization using 10-fold and 3-fold cross-validation, this application applies the model to this separate test set for testing, as follows: Figure 8 The results show that the model has good generalization ability.

[0039] In one specific embodiment of this application, The input layer is used to input EEG signal segments from the HEP dataset; This application directly uses 6-lead EEG signals, with an input shape of (batch_size, 6, 400), where 6 represents the number of EEG data leads and 400 represents the number of sampling points.

[0040] The temporal convolutional layer is the same as performing multi-scale convolution on the EEG signal segments to obtain multi-scale features; This application's temporal convolutional layers employ multi-scale 1D convolutional kernels to process EEG data in parallel, capturing features at different temporal resolutions. (Kernel length...) Depends on sampling rate and scaling factor :

[0041] The scaling factors for the three different sizes of convolution kernels are α=0.5, α=0.25, and α=0.125.

[0042] The kernel-level attention fusion module is used to adjust the weights of the temporal convolutional layer at different time scales and obtain channel-level features of the entire brain region through pooling operations. The kernel-level attention fusion module in this application is used to fuse features from different convolutional kernels to enhance the model's learning ability. Through the kernel-level attention mechanism, the network can adaptively adjust the weights of each convolutional kernel, thereby better integrating information from various time scales.

[0043] The local learning module is used to divide the entire brain region into multiple local sub-maps and map the channel-level features to the local sub-maps to obtain the local features of the entire brain region. The local learning module divides the entire brain region into multiple local sub-graphs and uses the channel-level features as node features in each local sub-graph. Within each local sub-graph, the node features are aggregated to obtain local features. The local features include node features of each brain region. The goal of the local learning module is to extract local EEG features from different brain regions. First, based on brain region functional division, EEG channels are constructed into multiple local subgraphs. Then, the channel-level features extracted by the temporal learning module are used as the initial attributes of each node in the subgraph. Within each local subgraph, node features are aggregated to obtain the overall representation of that brain region, i.e., the local feature matrix. (where R is the number of local graphs and T is the feature dimension).

[0044] The global learning module is used to calculate the global adjacency matrix using the local features and the local subgraphs of each brain region; wherein, the global learning module constructs the connection relationship between brain regions and calculates the dot product similarity between the node features of different brain regions; constructs a dynamic adjacency matrix and introduces a trainable attention mask M to calculate the global adjacency matrix. The global learning module aims to model the interaction relationships between different brain regions (local maps) to extract global-level feature representations. First, it calculates the dot product similarity between feature vectors of the local maps and constructs an instance-level dynamic adjacency matrix. Furthermore, a trainable attention mask M is introduced for weighted adjustment, resulting in a more discriminative global adjacency matrix: , where I is the identity matrix.

[0045] The graph convolutional neural network is used to aggregate the local features and the global adjacency matrix to obtain a global feature representation; The Graph Convolutional Neural Network (GCN) aggregates the local features and the global adjacency matrix to obtain a global feature representation. In this way, the model can not only learn the local features of a single brain region, but also use GCN to model the interactions between different brain regions, improving predictive ability. Among these, express The normalized global adjacency matrix; Represents the feature projection weight matrix; This represents a trainable bias vector.

[0046] The fully connected layer and the output layer are used to make predictions using the global feature representation to obtain the subject's current desire state.

[0047] This application provides a graph convolutional addiction craving prediction system based on brain-heart coupling, comprising: The acquisition module is configured to acquire electrocardiogram (ECG) and electroencephalogram (EEG) signals generated by subjects while watching videos containing craving cues; The processing module is configured to use the electrocardiogram (ECG) signal to extract a segment sequence of EEG signals near the R-wave peak. The prediction module is configured to input the EEG segment sequence into a trained craving prediction model, so that the craving prediction model extracts node features of different brain regions from the EEG segment sequence; aggregates the node features of different brain regions to obtain local features; constructs a global adjacency matrix using the local features; aggregates the global adjacency matrix and the local features to obtain a global feature representation; and predicts the subject's current craving state based on the global feature representation.

[0048] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0049] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A method for predicting addiction craving based on graph convolution of heart-brain coupling, characterized in that, The method comprises the following steps: S100, acquiring electrocardiogram signals and electroencephalogram signals generated by a subject when watching a video containing a craving clue; S200, using the electrocardiogram signals to segment the electroencephalogram signals near the R-wave peak to obtain an electroencephalogram segment sequence; S300, inputting the electroencephalogram segment sequence into a trained craving prediction model to enable the craving prediction model to extract node features of different brain regions from the electroencephalogram segment sequence, aggregate the node features of different brain regions to obtain local features, construct a global adjacency matrix using the local features, aggregate global information of the global adjacency matrix and the local features to obtain a global feature representation, and predict the current craving state of the subject based on the global feature representation.

2. The heart-brain coupling based graph convolutional addiction craving prediction method according to claim 1, characterized in that, The trained craving prediction model is obtained by training a preset craving prediction model, and the training process comprises the following steps: a, collecting electrocardiogram signals and electroencephalogram signals generated by multiple subjects at different sites when watching a video containing a craving clue, and collecting craving scores of the subjects; b, preprocessing the electrocardiogram signals and the electroencephalogram signals in a corresponding manner to obtain preprocessed electrocardiogram signals and preprocessed electroencephalogram signals; c, using the preprocessed electrocardiogram signals to segment the preprocessed electroencephalogram signals to obtain electroencephalogram signal segments, and grouping the electroencephalogram signal segments corresponding to the subjects at the same site to obtain an HEP data set; d, using the HEP data set and a training strategy of ten ten-fold cross-validation combined with internal three-fold cross-validation to iteratively train the preset craving prediction model to obtain the trained craving prediction model.

3. The heart-brain coupling based graph convolutional addiction craving prediction method according to claim 2, wherein b The method comprises the following steps: b1, sequentially performing electrode positioning, filtering, linear drift removal, re-referencing, bad lead detection, interpolation of bad leads, and independent component analysis on the electroencephalogram signals to obtain preprocessed electroencephalogram signals; b2, performing band-pass filtering on the electrocardiogram signals to obtain preprocessed electrocardiogram signals.

4. The heart-brain coupling based graph convolutional addiction craving prediction method according to claim 2, characterized in that, c comprises the following steps: c1, taking the R-wave peak of the preprocessed electrocardiogram signals as a reference point; c2, segmenting the preprocessed electroencephalogram signals to obtain electroencephalogram signal segments from a predetermined first time period before the reference point to a predetermined second time period after the reference point; c3, grouping the electroencephalogram signal segments corresponding to the subjects at the same site to obtain an HEP data set corresponding to each site.

5. The heart-brain coupling based graph convolutional addiction craving prediction method according to claim 2, characterized in that, d comprises the following steps: d1, dividing the HEP data set at the same site into 10 subsets, and using 9 subsets as a training set and the remaining 1 subset as a test set; d2, dividing the training set into 3 subsets, and using 2 subsets as a training subset and the remaining 1 subset as a validation set; d3, using the 9 training sets to iteratively train the preset craving prediction model 9 times, using the 2 training subsets to perform 2 internal iterations in each iteration, using the validation set to verify the performance of the craving prediction model in the 2 internal iterations, and using the test set to test the craving prediction model iterated 9 times, and determining the craving prediction model that passes the test performance as the trained craving prediction model.

6. The heart-brain coupling based graph convolutional addiction craving prediction method according to claim 1, characterized in that, The craving prediction model comprises an input layer, a time series learning module, a graph learning module, a fully connected layer, and an output layer connected in sequence.

7. The heart-brain coupling based graph convolutional addiction craving prediction method according to claim 6, characterized in that, The time sequence learning module comprises a time convolution layer and a kernel-level attention fusion module connected in sequence; the graph learning module comprises a graph convolutional neural network, a local learning module and a global learning module; the outputs of the local learning module and the global learning module are connected to the input of the graph convolutional neural network.

8. The heart-brain coupling based graph convolutional addiction craving prediction method according to claim 7, characterized in that, The input layer is configured to input an electroencephalogram (EEG) signal segment in a HEP dataset; The time convolution layer is configured to perform multi-scale convolution on the EEG signal segment to obtain multi-scale features; The kernel-level attention fusion module is configured to adjust the weights of the time convolution layer at different time scales and obtain channel-level features of the whole brain region through a pooling operation; The local learning module is configured to divide the whole brain region into a plurality of local sub-graphs and correspond the channel-level features with the local sub-graphs to obtain local features of the whole brain region; The global learning module is configured to calculate a global adjacency matrix by using the local features and the local sub-graphs of each brain region; The graph convolutional neural network is configured to aggregate global information of the local features and the global adjacency matrix to obtain a global feature representation; The fully connected layer and the output layer are configured to predict the current craving state of the subject by using the global feature representation.

9. The heart-brain coupling-based graph convolutional addiction craving prediction method according to claim 8, wherein the local learning module is specifically configured to: divide the whole brain region into a plurality of local sub-graphs, and take the channel-level features as node features in each local sub-graph, and aggregate the node features in each local sub-graph to obtain local features; wherein the local features comprise node features of each brain region; the global learning module is specifically configured to: construct the connection relationship between the brain regions, and calculate the dot product similarity between the local features of different brain regions; construct a dynamic adjacency matrix, and introduce a trainable attention mask M to calculate a global adjacency matrix. The method comprises:

10. A heart-brain coupling-based graph convolutional addiction craving prediction system, characterized in that, an acquisition module configured to acquire electrocardiogram (ECG) signals and EEG signals generated by a subject when watching a video containing a craving cue; a processing module configured to use the ECG signals to perform segment interception on the EEG signals near R-wave peaks to obtain an EEG segment sequence; a prediction module configured to input the EEG segment sequence into a trained craving prediction model, so that the craving prediction model extracts node features of different brain regions from the EEG segment sequence; aggregates the node features of different brain regions to obtain local features; constructs a global adjacency matrix by using the local features; aggregates global information of the global adjacency matrix and the local features to obtain a global feature representation, and predicts the current craving state of the subject based on the global feature representation. ​

Images