Electroencephalogram-based cross-subject pain classification approach

Abstract

The electroencephalogram-based cross-subject pain classification method solves the problem of how to effectively improve the reliability of cross-subject pain classification and reduce the pressure of clinical data collection, and belongs to the field of electroencephalogram data evaluation. The present application comprises the following steps: identifying the resting EEG features highly correlated with the pain sensitivity of the target subject and the historical subjects in the historical data set, and screening out the historical subjects in the historical data set having similar pain response as the target subject to form a source subject set of the target subject; constructing pseudo-labels by using the source subject set, and optimizing the source subject set by considering cognitive and other experimental dynamic factors to obtain a source domain; performing adaptive transfer learning by using the source domain and the target domain, weighting and fusing the labels of the target domain obtained after learning all the source domains to obtain the predicted classification result of the target subject, and completing the classification.

Description

TECHNICAL FIELD

[0001] The present application relates to a cross-subject pain classification method guided by dynamic experimental factors such as pain sensitivity and cognition, and belongs to the field of electroencephalogram data evaluation. BACKGROUND

[0002] Existing research has confirmed that it is feasible to detect pain by using electroencephalogram (EEG). The detection of pain level is divided into two levels: one is intra-subject pain prediction, that is, training and testing are performed on the same subject; the other is cross-subject prediction, that is, a classifier is trained on one user and then tested on different subjects. Although intra-subject pain assessment is reliable, it requires the true pain label of a new subject, and obtaining a large amount of pain data of new users to train a single model is both time-consuming and faces the challenge of EEG collection. In addition, this method is not suitable for individuals who cannot reliably express pain perception. Therefore, the demand for cross-subject prediction in clinical application is more urgent. However, due to the inherent differences in pain perception and neural response between individuals, the performance of cross-subject pain prediction is significantly lower than that of intra-subject performance. Therefore, it is urgent to solve the problem of improving the reliability of cross-subject pain assessment.

[0003] Existing research on cross-subject pain assessment is relatively limited, and most methods rely on machine learning. However, the premise of machine learning is that the EEG data for training and testing share the same feature space and follow the same probability distribution. This assumption is often challenged by individual differences in practice, limiting the effectiveness of many models. Individual differences are reflected in many aspects, including neurophysiological characteristics, perception and expression of pain, etc., so cross-subject pain assessment usually faces the problem of insufficient reliability.

[0004] Therefore, in order to effectively improve the reliability of cross-subject pain classification and alleviate the pressure of clinical data collection, it is urgent to develop a cross-subject pain classification method in the context of pain. SUMMARY

[0005] In order to effectively improve the reliability of cross-subject pain classification and alleviate the pressure of clinical data collection, the present application provides a cross-subject pain classification method based on electroencephalogram.

[0006] The cross-subject pain classification method based on electroencephalogram provided by the present application comprises:

[0007] The target subject is a cross-subject, the historical subjects are screened based on the resting electroencephalogram signals in the historical data set, the screened historical subjects are used as source subjects of the target subject, and a source subject set is formed;

[0008] According to the source subject set and the target subject's electroencephalogram of pain, the pain characteristics of the source subject set and the target subject are obtained;

[0009] A first label prediction model is trained using the pain characteristics of the source subject set and the true labels, and the pseudo labels of the target subject are predicted using the first label prediction model, and the pain characteristics of the target subject and the pseudo labels are used as the target domain;

[0010] A second label prediction model is trained using the pain characteristics of the target subject and the pseudo labels, and the pain labels of the source subject set are predicted using the second label prediction model, and the classification accuracy is calculated according to the obtained prediction labels and the true labels, and the classification accuracy is ranked from high to low, and the pain characteristics and true labels of the top N S source subjects are selected as the source domain, and N S is a set value;

[0011] Adaptive transfer learning is performed using the source domain and the target domain to obtain the prediction results of the target subject after learning each source domain, fusion is performed to obtain the final prediction result, and classification is completed.

[0012] As preferred, the method for screening the historical subjects as the source subjects of the target subject based on the resting electroencephalogram signals in the historical data set comprises:

[0013] Obtain the resting electroencephalogram signals X* mapped by the pain brain area of the target subject and the historical subjects;

[0014] Calculate the relative power spectral density characteristics of each channel in X*, identify P statistically significant pain sensitivity characteristics v p , 1≤p≤P, and form a pain sensitivity characteristic vector V={v1,…,v P};

[0015] Calculate the similarity U i,j between the pain sensitivity characteristic vector V i of the target subject and the pain sensitivity characteristic vector V j of the historical subject j;

[0016] Cluster according to all the similarities U i,j calculated, select the cluster with the largest average similarity after clustering, and the historical subjects in the cluster wq constitute the source subject set of the target subject 1≤q≤N w , N W is the number of historical subjects in the source subject set W i .

[0017] Preferably, the method for obtaining the resting EEG signal X* mapped by the pain brain area of ​​the target subject and the historical subject comprises:

[0018] The resting EEG signals of the target subject and the source subject are X∈R N×T , the lead field matrix G∈R is obtained by linear configuration single-layer boundary element method N×D , and then solve the EEG inverse problem through dynamic statistical parameter mapping to obtain the brain source signal S∈R D×T :

[0019] X=GS

[0020] Where D represents the number of brain source dipoles, T is the number of time samples, and N is the number of sensors;

[0021] Retain the brain source signal corresponding to the painful brain area in the brain source signal S, set the other non-painful brain source signals to zero, and obtain S*, and obtain the resting EEG signal X mapped by the painful brain area * =GS * .

[0022] Preferably, the method for obtaining the pain characteristics of the source subject set and the target subject according to the pain EEGs of the source subject set and the target subject comprises:

[0023] Extracting pain-evoked EEG signals from pain EEG calculate Relative power spectral density in the Cz channel

[0024]

[0025] in, The subscript band is delta, theta, alpha, beta or gamma, indicating The power spectral density in the five frequency bands is: delta band 0.5-4 Hz, theta band 4-8 Hz, alpha band 8-12 Hz, beta band 12-30 Hz, and gamma band 31-49 Hz;

[0026] Absolute power spectral density

[0027] Relative power spectral density Normalization was performed to obtain pain features.

[0028] Preferably, both the No. 1 label prediction model and the No. 2 label prediction model are implemented using the SVM model.

[0029] Preferably, the method for adaptive transfer learning using the source domain and the target domain includes:

[0030] Minimizing the distribution difference between source domains and target domain by using balanced distribution adaptive learning method and target domain iteratively updating the labels of target subjects, with the pseudo-labels of target subjects as the initial values of the labels in the iterative updating process;

[0031] fusing the labels of the target domain obtained after learning all source domains as the predicted classification results of the target subjects.

[0032] As a preferred, the labels of the target domain obtained after learning all source domains are weighted fused, wherein the weight is the classification accuracy of the corresponding source domain.

[0033] The present application improves the reliability of cross-subject pain classification to address the challenges faced by current pain assessment. By remapping the resting electroencephalogram signals of pain-related brain regions and considering cognitive and other dynamic experimental factors that affect pain response, suitable source domains are effectively selected, the risk of negative transfer is reduced, and the high computational burden brought by traversing the entire historical subject database for transfer learning is alleviated. In order to reduce the dependence on labeled electroencephalogram signals of new subjects, the present application constructs pseudo-labels, which significantly reduces the pressure of data acquisition, making the method more feasible in clinical application. This innovation not only improves the efficiency of pain assessment for new subjects in a clinical environment, but also provides a more reliable tool for future pain management. At the same time, the present application designs high representation features driven by pain-induced responses, which can more accurately reflect the essence of pain signals and effectively alleviate data drift. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a schematic diagram of the principle of the method of the present application;

[0035] Figure 2 is the cross-subject pain classification accuracy effect, where (A) is low pain vs. high pain, B is low pain vs. medium pain vs. high pain, green circles represent average accuracy, 51 bars represent 51 subjects, and the black highlighted bar corresponds to subject No. 1, which is ordered clockwise to 51. The concentric circles from the center outward represent 25%, 50%, 75%, and 100%. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0037] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0038] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0039] The present embodiment provides an electroencephalogram-based cross-subject pain classification method, including:

[0040] Step 1: Identify resting EEG features that are highly correlated with pain sensitivity in the target subject and all historical subjects in the historical dataset, and select historical subjects in the historical dataset that have similar pain responses to the target subject to form the source subject set of the target subject:

[0041] In order to reduce the time cost of traversing the entire historical subject database in subsequent transfer learning, preliminary subject screening is first performed based on resting EEG signals.

[0042] Suppose there is a resting EEG signal X∈R N×T , the lead field matrix G∈R is obtained by linear configuration single-layer boundary element method N×D , and then solve the EEG inverse problem through dynamic statistical parameter mapping to obtain the brain source signal S∈R D×T .

[0043] X=GS (1)

[0044] where D represents the number of brain source dipoles, T is the number of time samples, and N is the number of sensors.

[0045] The brain source signals corresponding to the painful brain area in S are retained, and the remaining non-painful brain source signals are set to zero to obtain S*. Subsequently, (2) is used to obtain the resting EEG signal X* mapped by the painful brain area:

[0046] X * =GS * (2)

[0047] Calculate the relative power spectral density (PSD) feature of each channel in X* and combine it with the participant's pain sensitivity label (low sensitivity, high sensitivity) to identify P statistically significant features v p , 1≤p≤P, and form the pain sensitivity feature vector V={v1,…,v P These features effectively characterized the subjects’ pain sensitivity and were subsequently used to measure the similarity of resting EEG patterns across subjects.

[0048] Calculate the target subject's pain sensitivity feature vector V according to (3) iand the pain sensitivity feature vector V of historical subject j j The similarity between i,j .

[0049]

[0050] where ||V i || and ||V j || represent V i and V j The modulus of .

[0051] According to the calculated similarities U i,j Perform clustering and select the cluster with the largest average similarity after clustering. The historical subjects in the cluster constitute the source subject set of the target subject. Specifically, all U i,j The value will be clustered by K-means, and the number of clusters is set to 3. If U i,j The average value of the value is greater than that of the other two clusters, then the historical subject w in cluster h1 q , 1≤q≤N w , which are the initially selected source subjects and constitute the source subject set where N W It's W i The number of historical subjects in , which are expected to have similar pain performance as the target subject i.

[0052] Step 2: Construct pseudo-labels and optimize source subject selection while taking into account experimental dynamics such as cognition;

[0053] First, use the source subject set W i The pain EEG of the subject in is used to construct a pseudo label for the target subject i to reduce the dependence on the labeled pain EEG data of the new subject. Specifically, the NZT method is used to extract the pain-induced EEG signal to enhance the pain representation ability of the original signal. The extracted signal is recorded as calculate Relative power spectral density in the Cz channel

[0054]

[0055]

[0056] in, The subscript band is delta, theta, alpha, beta or gamma, indicating The power spectral density in the five frequency bands is: delta band 0.5-4 Hz, theta band 4-8 Hz, alpha band 8-12 Hz, beta band 12-30 Hz, and gamma band 31-49 Hz; Represents the absolute power spectral density.

[0057] Then, for the target subject and source subject set W i All subjects in Normalize to obtain pain features. For source subject j, its source domain can be expressed as in and Represents the corresponding pain features and true labels. For the target subject i, the target domain can be expressed as in represents the corresponding pain feature. n and m represent the number of samples in the source domain and target domain.

[0058] Using the source subject set W i The pain features and true labels of the target subject i are used to train a label prediction model for label 1. This label prediction model can be implemented using a machine learning model, specifically a support vector machine (SVM) classification model. This model is then used to predict the pain pseudo-label for the target subject i.

[0059] After determining the pain pseudo-label of the target subject i, this embodiment further optimizes the pain pseudo-label from the source subject set W by taking into account the influence of cognition and other dynamic experimental factors on pain response. i If the data of two subjects show similar marginal and conditional distributions, then they can accurately classify each other without relying on transfer learning. This means that using domain A as the training set and domain B as the test set (or vice versa) should produce satisfactory classification results. Based on this criterion, this embodiment uses the pseudo labels of the target subject i from the source subject set W i Identify the appropriate source domain and its corresponding weights for subsequent transfer learning.

[0060] Specifically, we use the pain features and pseudo labels of the target subject i to train a 2nd label prediction model, which adopts a new SVM model, and then test the source subject set W i The source subjects in the source subject set W i The source subjects in the dataset have known labels, so the classification accuracy of each source subject can be calculated. The classification accuracy is ranked from high to low, and the top N are selected. S The subjects, this NS The samples of the source subjects are the source domains used for subsequent transfer learning, and the classification accuracy is the weight of each source.

[0061] Step 3: Combine pseudo-labeling, Structural Risk Minimization (SRM), and balanced distribution adaptation to reduce the distribution difference between the source and target domains and train learning:

[0062] The goal of this step is to minimize the source domain With the target domain The distribution difference between To this end, this implementation introduces balanced distribution adaptation for learning. The source domain is the one selected in the two steps above. It should be noted that during the initial iteration of calculating the distribution difference, unlike the soft labeling method described in classic papers, this implementation uses pseudo labels instead, which are then continuously updated in subsequent iterations.

[0063] Step 4: Weighted fusion:

[0064] N S The target domain N is obtained after learning from the source domain S tags, for N S The labels are weightedly fused, where the weight is the classification accuracy of the corresponding source domain;

[0065] The greater the similarity between the two domains before transfer learning, the more reliable the prediction results obtained after transfer. Therefore, this embodiment performs weighted fusion on the labels predicted after learning each source domain to obtain the final prediction result, where the classification accuracy of each domain in step 2 is the N s The weight of the source domain e∈[1,2,3].

[0066]

[0067] in and They represent the number of predicted labels of 1 (low pain), 2 (medium pain), and 3 (high pain), respectively.

[0068] In order to verify the effectiveness of the method proposed in this embodiment, this embodiment uses EEG signals collected under pain stimulation to verify its performance, and compares it with 3 methods with the same purpose and 5 mainstream transfer learning technologies, a total of 8 methods. Figure 2 The accuracy of the nine methods for pain classification across subjects was presented.

[0069] Table 1 except Figure 2 In addition to the accuracy shown, other metrics used to evaluate pain classification performance across subjects

[0070]

[0071]

[0072] The method proposed in this embodiment achieved the highest cross-subject pain classification accuracy, reaching 90.39% accuracy in the binary classification (low pain vs. high pain) and 75.39% accuracy in the three-category classification (low pain, moderate pain, and high pain). In addition to the accuracy of pain intensity classification, Table 1 also provides four other evaluation metrics (precision, recall, F1 score, and Kappa value), all of which show that the method of this embodiment achieved the highest cross-subject pain classification level. These results clearly confirm that the present invention has significant advantages in improving the reliability of cross-subject pain intensity classification.

[0073] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be employed in conjunction with other described embodiments.

Claims

1. A cross-subject pain classification method based on electroencephalogram, characterized in that: The method comprises: Taking the cross-subjects as the target subjects, the historical subjects are screened based on the resting EEG signals in the historical dataset, and the screened historical subjects are used as the source subjects of the target subjects to form the source subject set; Obtaining pain features of the source subject set and the target subject according to the pain electroencephalograms of the source subject set and the target subject; The pain features and true labels of the source subject set are used to train a label prediction model No. 1, and the label prediction model No. 1 is used to predict the pseudo labels of the target subject, and the pain features and pseudo labels of the target subject are used as the target domain; The pain characteristics and pseudo labels of the target subjects are used to train the No. 2 label prediction model, and the No. 2 label prediction model is used to predict the pain labels of the source subject set. The classification accuracy is calculated based on the obtained predicted labels and true labels, and the classification accuracy is ranked from high to low, and the top ranked ones are selected. The pain features and true labels of the source subjects are used as the source domain, is the set value; Adaptive transfer learning is performed using the source and target domains to obtain the prediction results of the target subject after learning each source domain. The results are then fused to obtain the final prediction results and complete the classification. The method of screening historical subjects based on resting EEG signals in a historical data set and using the screened historical subjects as source subjects of target subjects includes: Obtain resting EEG signals mapped by pain brain areas of target subjects and historical subjects ; calculate The relative power spectral density features of each channel in the image were used to identify P statistically significant pain sensitivity features. , , and constitute the pain sensitivity feature vector ; Calculate the pain sensitivity feature vector of the target subject and the pain sensitivity feature vector of historical subject j The similarity between ; Based on all the similarities calculated Perform clustering and select the cluster with the largest average similarity after clustering. The set of source subjects that make up the target subject , , is the source subject set the number of historical subjects; Obtain resting EEG signals mapped by pain brain areas of target subjects and historical subjects The methods include: The resting EEG signals of the target subject and the historical subject are , the lead field matrix is ​​obtained by linear configuration single-layer boundary element method , and then solve the EEG inverse problem through dynamic statistical parameter mapping to obtain the brain source signal : in, represents the number of brain source dipoles, is the number of time samples, is the number of sensors; Preserving brain-derived signals The brain source signal corresponding to the pain brain area is set to zero, and the other non-pain brain source signals are set to zero. , obtaining resting EEG signals mapped by pain brain areas .

2. The cross-subject pain classification method based on electroencephalogram according to claim 1, characterized in that: The method for obtaining pain features of the source subject set and the target subject according to the pain electroencephalograms of the source subject set and the target subject includes: Extracting pain-evoked EEG signals from pain EEG ,calculate Relative power spectral density in the Cz channel : in, The subscript band is delta, theta, alpha, beta or gamma, indicating The power spectral density in five frequency bands is: delta band 0.5-4 Hz, theta band 4-8 Hz, alpha band 8-12 Hz, beta band 12-30 Hz, and gamma band 31-49 Hz; Absolute power spectral density ; Relative power spectral density Normalization was performed to obtain pain features.

3. The cross-subject pain classification method based on electroencephalogram according to claim 1, characterized in that: Both the No. 1 label prediction model and the No. 2 label prediction model are implemented using the SVM model.

4. The electroencephalogram-based cross-subject pain classification method according to claim 1, characterized in that: Methods for adaptive transfer learning using source and target domains include: Minimize the source domain using a balanced distribution adaptive learning method and target domain The distribution difference of the target subject is iteratively updated, and the pseudo label of the target subject is used as the initial value of the label in the iterative update process; The labels of the target domain obtained after learning all source domains are fused as the predicted classification results of the target subject.

5. The cross-subject pain classification method based on electroencephalogram according to claim 4, characterized in that: The labels of the target domain obtained after learning all source domains are weighted and fused, where the weight is the classification accuracy of the corresponding source domain.

6. A computer-readable storage device storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the electroencephalogram-based cross-subject pain classification method according to any one of claims 1 to 5 are implemented.

7. An electroencephalogram-based cross-subject pain classification device, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, characterized in that: The processor executes the computer program to implement the steps of the electroencephalogram-based cross-subject pain classification method according to any one of claims 1 to 5.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the electroencephalogram-based cross-subject pain classification method according to any one of claims 1 to 5 are implemented.

Images