A multi-instance depression recognition method based on high-order micro-state features

Abstract

The application relates to the technical field of electroencephalogram signals, in particular to a multi-instance depression recognition method based on high-order microstate characteristics, which comprises the following steps: acquiring an electroencephalogram signal; extracting a microstate time sequence of the electroencephalogram by using an electroencephalogram microstate template; mining a frequent microstate transition mode by using the microstate time sequence, and constructing a global transition rule library; calculating the matching degree of the microstate time sequence of a sliding window and rules in the global rule library, and realizing rule activation recoding of the microstate time sequence; encoding a rule index sequence by using a multi-instance learning model to obtain a microstate high-order transition feature; and inputting the microstate high-order transition feature and a microstate time sequence feature after splicing into a classification network to output a depression judgment result. The application solves the problems of insufficient feature expression and unreliable diagnosis model of the existing method in early screening of depression.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) signal technology, and in particular to a method for identifying multiple instances of depression based on higher-order microstate features. Background Technology

[0002] With the development of neuroscience and EEG analysis technology, brain function assessment and auxiliary diagnosis of depression (MDD) have become an important direction for clinical research and application; EEG microstate analysis, due to its high temporal resolution, is widely used to characterize the spatiotemporal dynamic changes of the brain neural network in patients with depression.

[0003] The patent with publication number CN120227030A uses a wearable EEG device and a high-precision eye tracker to simultaneously collect brain electrical activity and gaze behavior data. After signal preprocessing and feature extraction, it uses a weighted fusion formula or graph neural network to achieve deep fusion of multimodal data. This method requires the addition of eye data from the eye tracker. The clock drift and trigger delay between EEG data and eye tracker data are difficult to completely eliminate. Furthermore, the asymmetry of alpha waves in EEG and the gaze patterns of eye movements vary significantly across different ages, genders, and cultural backgrounds, making it difficult to guarantee the model's cross-population generalization ability.

[0004] In addition, existing technologies rely on first-order statistical parameters, which has the following drawbacks: 1. High-order features are lost, and static statistics cannot capture the complex triggering logic and switching patterns between micro-states; 2. The lack of long-range dependency modeling makes it difficult for traditional first-order Markov chains to identify multi-state combination patterns across time scales. 3. Poor model interpretability: Although existing deep learning solutions have improved performance, their "black box" nature lacks clinical neurophysiological basis. Summary of the Invention

[0005] To address the shortcomings of existing methods, this invention solves the problems of insufficient feature expression and unreliable diagnostic models in the early screening of depression.

[0006] The technical solution adopted in this invention is: a multi-instance depression identification method based on high-order microstate features, comprising the following steps: Step 1: Acquire electroencephalogram (EEG) signals; In a preferred embodiment of the present invention, the electroencephalogram (EEG) signals are subjected to whole-brain average reference and ICA preprocessing.

[0007] In a preferred embodiment of the present invention, the electroencephalogram (EEG) signal is subjected to bandpass filtering and concave filtering.

[0008] Step 2: Extract the time sequence of brainwave microstates using the brainwave microstate template; In a preferred embodiment of the present invention, the microstate time sequence features include: average duration, frequency of occurrence, and time coverage.

[0009] Step 3: Utilize micro-state time series to mine frequent micro-state transition patterns and construct a global transition rule base; In a preferred embodiment of the present invention, a global transfer rule base is constructed using the RuleGrowth algorithm.

[0010] Step 4: Calculate the matching degree between the micro-state time sequence of the sliding window and the rules in the global rule base to realize the rule activation recoding of the micro-state time sequence; In a preferred embodiment of the present invention, step four specifically includes: Construct adjacent state transition sets for each sliding window subsequence. The set of adjacent state transitions with the regular sequence ; calculate and similarity ; Use the index number of the rule with the highest similarity score The micro-state time sequence within the window is activated and a regular index sequence is generated.

[0011] Step 5: Encode the rule index sequence using a multi-instance learning model to obtain higher-order transition features of microstates; In a preferred embodiment of the present invention, step five specifically includes: Define the rule index sequence as a package, and define the rule index sequence for a certain time period within the package as an instance; The packet is input into the LSTM encoder to capture the temporal dependencies within the instance, and the contribution weights of different time periods to the classification result are calculated through the attention aggregation mechanism to obtain the higher-order transition features of the microstate after weight aggregation.

[0012] Step 6: Concatenate the microstate high-order transition features and microstate temporal sequence features and input them into the classification network to output the discrimination result of depression; In a preferred embodiment of the present invention, the classification network includes: a first fully connected layer, a second fully connected layer, and a Softmax layer; wherein the number of neurons in the fully connected layers are 256 and 64, respectively, and the activation function is ReLU.

[0013] As a preferred embodiment of the present invention, a multi-instance depression recognition system based on higher-order microstate features includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a multi-instance depression recognition method based on higher-order microstate features.

[0014] As a preferred embodiment of the present invention, a computer-readable medium storing computer program code implements a multi-instance depression identification method with high-order micro-state features when executed by a processor.

[0015] The beneficial effects of this invention are: 1. This invention constructs a rule-driven EEG microstate sequence analysis and recognition method. By introducing the RuleGrowth algorithm to mine higher-order microstate transition rules, it effectively makes up for the shortcomings of traditional static parameters in capturing the dynamic switching logic of brain states and significantly enhances the discriminative power of features. 2. By using sequence-aware Jaccard similarity for rule activation recoding, while preserving the local temporal topology of microstate sequences, we have achieved effective modeling of long-range dependencies, thereby improving the model's ability to characterize the complex evolution of EEG signals. 3. The multi-instance learning architecture, which combines attention mechanisms, not only achieves higher recognition accuracy under weak supervision, but also provides an intuitive neurophysiological explanation through weight distribution, providing more robust and clinically valuable quantitative evidence for the early auxiliary diagnosis of depression. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-instance depression identification method based on high-order microstate features of the present invention; Figure 2 This is a flowchart of the electroencephalogram (EEG) microstate analysis for depression according to an embodiment of the present invention. Figure 3 This is a framework diagram for high-order feature extraction and feature fusion of EEG microstates in depression, according to an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments. The drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0018] like Figure 1 , Figure 2 As shown, a multi-instance depression identification method based on high-order microstate features includes the following steps: Step 1: Obtain multi-channel EEG signals from the target subject through an EEG acquisition system, and preprocess the EEG signals to obtain clean EEG signals; A multi-channel EEG acquisition device conforming to an approximate distribution of 10-20 can be selected, with the number of leads ranging from 64 to 128; the sampling frequency can be set to 250Hz-1000Hz; dry or wet electrodes can be used, and basic contact optimization and channel detection need to be completed before acquisition, controlling the electrode impedance to ≤40KΩ; complete 5 minutes of signal acquisition in the closed-eye resting state experimental paradigm, and record the corresponding experimental labels simultaneously; The acquired EEG signals were preprocessed using the MATLAB-based EEGLAB toolbox, specifically including the following steps: First, the raw signals were selected by selecting leads and removing useless leads. Then, bandpass filtering was performed from 0.5 to 45 Hz, followed by concave filtering from 49 to 51 Hz. Next, the EEG signals were averaged for the whole brain and then physiological artifacts such as electrooculography and electromyography were removed through independent component analysis. Finally, continuous 3-minute data of good quality were selected from each subject for subsequent microstate analysis.

[0019] Step 2: Extract typical EEG microstate templates from the preprocessed EEG signals using a clustering algorithm, and convert the EEG signals into microstate time sequences; Microstate analysis is performed using the CarTool microstate analysis toolkit, specifically including the following steps: First, the global field power (GFP) of the EEG signal is calculated. GFP is represented by the standard deviation of the electrode potentials at the same time point. The specific calculation formula is as follows: ; in, express Global field power of EEG signals at any given time; This indicates the total number of EEG electrode channels; Indicates the first Each channel is in The potential signal at a given moment; express The average value of the potential signals of all channels at any given time.

[0020] Then, local peak points in the GFP sequence are selected, and the topographic map of the local peak points is used as input. The modified K-means clustering algorithm is used for iteration, with the number of clusters set to 3-8 and the number of iterations set to 500. Next, the global interpretable variance (GEV) corresponding to each number of clusters is calculated, and the number of clusters corresponding to the highest GEV is selected to finally determine the global microstate template. Finally, based on the obtained global microstate template, the spatial correlation coefficient between the preprocessed EEG topography of each sampling point and each typical microstate template is calculated. Each sampling point is marked as the symbol corresponding to the typical microstate template with the largest absolute value of the spatial correlation coefficient, and the initial symbol sequence is obtained. The initial symbol sequence is denoised using a smoothing constraint algorithm to remove extremely short microstate fragments with a duration of less than 30 milliseconds, and the final continuous microstate time sequence is obtained. like Figure 3 Step 3: Use the RuleGrowth algorithm to mine frequent micro-state transition patterns and build a global transition rule base; First, the EEG microstate symbol sequence of each subject was arranged according to a preset length. K (For example K =12) Divide into multiple continuous subsequences; according to the category to which the subjects belong, the subsequences of all subjects in the depression group are summarized to form the depression group sequence database, and the subsequences of all subjects in the healthy control group are summarized to form the healthy group sequence database.

[0021] Then, using sequence databases from the depression group and the healthy group as inputs, respectively, the RuleGrowth sequence rule mining algorithm was used to mine EEG microstate transition patterns. The RuleGrowth algorithm is a sequence rule mining method based on pattern growth, which discovers frequent sequence rules that satisfy constraints by recursively expanding the antecedent and consequent of rules.

[0022] Specifically, the frequency of occurrence of each microstate symbol and its transition pattern in the sequence database is statistically analyzed, and initial rules that meet the minimum support requirement are generated. The rules are then expanded by progressively adding new microstate symbols as antecedents or consequents. During each expansion process, the support and confidence of candidate rules are calculated. Candidate rules that do not meet the constraints are pruned, while those that do meet the constraints are recursively expanded until no new effective rules can be generated. This yields a complete set of frequent EEG microstate transition rules. Among these, for sequence rules… Its support is defined as simultaneously including the rule antecedent. and successor The proportion of the number of sequences to the total number of sequences in the database, i.e.:

[0023] in, This represents the number of sequences that simultaneously satisfy both the antecedent and consequent of the rule. This represents the total number of sequences in the sequence database; its confidence level is defined as the number of sequences in the antecedent of the rule. The rule consequent occurs under the given conditions The conditional probability of occurrence, i.e.:

[0024] in, This indicates that the rule antecedent is included. The number of sequences. In this embodiment, the minimum support is set to 0.5, the minimum confidence is set to 0.5, and only rules with both support and confidence not lower than the preset threshold are retained as valid EEG microstate transition rules.

[0025] Furthermore, the RuleGrowth algorithm was executed on the sequence databases of the depression group and the healthy group, respectively, to obtain the rule sets of frequent EEG microstate transitions for the depression group and the healthy group. Since the two databases come from different categories of subjects, the obtained rule sets can respectively characterize the characteristic transition patterns of EEG microstates in the depression group and the healthy group. Subsequently, the two sets were merged to form a global transition rule base. ;in, The total number of rules, each rule is denoted as:

[0026] in, This indicates the number of microstate nodes included in the rule. The first in the rule A brainwave microstate symbol, A , B , C and D These represent the four standard categories of EEG microstates obtained through EEG microstate analysis, i.e., microstates. A microstates B microstates C and microstates D Each rule describes a continuous transition relationship of a brainwave microstate, such as rule ... This indicates that brain electrical activity sequentially transitions through microstates A, C, and D in chronological order; to facilitate subsequent rule matching, retrieval, and feature extraction, for the first... Each rule is assigned a unique number. .

[0027] Step 4: Use sequence-aware Jaccard similarity to calculate the matching degree between the sequence and the rules in the global rule base under the sliding window, and realize the rule activation recoding of the micro-state sequence.

[0028] Traverse the microstate time series using a fixed-length sliding window of 500 milliseconds, and extract the local microstate subsequence within each window:

[0029] in, Indicates the first The sequence of local microstate sub-states corresponding to each sliding window; Indicates the first in the window A brainwave microstate symbol, ; This represents the total number of microstate symbols contained in the window subsequence; the transition rules will be used to facilitate subsequent rule matching calculations. Convert the microstate nodes into their corresponding rule sequence representations according to their order of appearance:

[0030] in, Represents the first in the rule sequence Each microstate symbol; ; Indicates the length of the regular sequence.

[0031] window subsequence With regular sequence To preserve the temporal topological structure of the microstate sequence, a sequence-aware Jaccard similarity metric is used to quantify the degree of matching between the two states. The specific calculation method is as follows: First, construct the adjacent state transition sets for the window subsequence and the rule sequence respectively: ; in, Represents the sliding window subsequence The corresponding set of adjacent microstate transitions; Representing a regular sequence The corresponding set of adjacent microstate transitions.

[0032] Based on the above transformation set, the formula for calculating the sequence-aware Jaccard similarity is: ; in, The order-aware Jaccard similarity between the window subsequence and the regular sequence is represented. Indicates the number of elements in the set; , These represent the intersection and union operations of sets, respectively.

[0033] The following judgments are made based on the calculated similarity scores: the rule with the highest similarity score is selected, and the current window subsequence is determined to have successfully matched the corresponding rule, using the index number of that rule. The micro-state sequence within the window is activated to complete the recoding; after traversing all sliding windows, a complete rule index sequence is generated.

[0034] To further illustrate the rule matching process, assume that the local micro-state subsequence extracted by a certain sliding window is as follows: ; The corresponding set of state transitions is: ; If the following three rules exist in the global transfer rule base:

[0035] Based on the rule sequence representation, the corresponding state transition sets can be constructed respectively:

[0036] The matching degree between the window subsequence and each rule is calculated using sequence-aware Jaccard similarity; where, the rule... The state transition set is completely identical to the state transition set of the window subsequence, therefore their similarity is 1; rule Since there are no identical state transition relationships between the window subsequence and the window subsequence, their similarity is 0; rule Includes only state transitions It is consistent with the window subsequence, therefore its similarity is lower than that of the rule. .

[0037] Due to the rules It has the highest similarity score, therefore the current window subsequence is determined to be similar to the rule. Successful match, and the rules were applied. The index number "1" is used to activate the window; if subsequent sliding windows match the rules... ,rule and rules After rule activation and recoding, the rule index sequence can be obtained:

[0038] Each element in the rule index sequence represents the transition rule number corresponding to the activation of the sliding window. By converting the original microstate symbol sequence into a rule index sequence, not only is the temporal structural information of the EEG microstate transition process preserved, but also a high-level semantic representation based on frequent transition patterns is achieved. This provides a more stable and discriminative feature expression for subsequent rule-level feature extraction, brain network state analysis, and auxiliary identification of depression.

[0039] Compared to traditional Jaccard similarity, sequence-aware Jaccard similarity incorporates temporal topological information into its calculation. Traditional Jaccard similarity treats a sequence as an unordered set of elements, focusing only on co-occurrence and failing to distinguish sequences with identical constituent elements but different arrangements. In contrast, sequence-aware Jaccard similarity constructs a set of adjacent state transitions to differentiate between ordered sequences. As a basic comparison unit, the temporal structure of the sequence is encoded into the similarity calculation process; therefore, for sequences with the same constituent elements but different orders, the sequence-aware Jaccard similarity can give different matching scores, more accurately reflecting the structural differences of micro-state sequences in the time dimension, thereby improving the accuracy and discriminativeness of rule matching.

[0040] Step 5: Input the recoded rule index sequence into the attention-weighted multi-instance learning model to extract dynamic organizational features at the individual level.

[0041] The multi-instance learning model employs a hierarchical modeling strategy, including the following steps: First, the subject's rule index sequence is defined as a bag, and the rule index sequence within a short time period within the bag is defined as an instance. A short time period is a sequence that is further divided into small segments, and each segment is an instance. The defined bags are fed into an LSTM encoder to capture the temporal dependencies within the instances, and the contribution weights of different time periods to the classification results are calculated through an attention aggregation mechanism. Then, the higher-order transition features of the aggregated microstates are calculated according to the weights.

[0042] Specifically, for a containing Individual Instance Features The package, the attention weights for each instance Calculated in the following way: ; in, and These are trainable weight parameters; For the attention hidden layer dimension, is the instance feature dimension; T is the transpose.

[0043] Based on the aforementioned attention weights, all instance features are weighted and summed to obtain the subject-level microstate high-order transition feature vector. : ; Through this attention aggregation mechanism, the model can adaptively highlight time-segment features that contribute more to the classification task, while suppressing the influence of noise or irrelevant segments, ultimately generating more discriminative dynamic organizational features.

[0044] Step 6: Combine the EEG microstate time parameters and microstate dynamic organization features to construct a fusion feature set and feed it into a fully connected layer and a Softmax classifier to achieve the detection of depression.

[0045] First, temporal parameter features are extracted from the microstate time sequence of EEG, including the average duration, frequency of occurrence, and temporal coverage of each microstate category. Then, the subject-level microstate high-order transition features, i.e. dynamic organization features, output by the multi-instance learning model are extracted. The two types of features are then concatenated to construct a fusion feature set containing temporal statistics and temporal organization patterns.

[0046] Subsequently, the fused feature set is input into a classification network consisting of multiple fully connected layers. The classification network structure can be set as follows: the first fully connected layer has 256 neurons, the second layer has 64 neurons, the activation function is ReLU, and a Dropout layer is added after each layer to prevent overfitting. Finally, the classification probability distribution of depression / healthy controls is output through the Softmax layer, and the category with the highest probability is selected as the discrimination result.

[0047] During the training phase, the cross-entropy loss function is used as the optimization objective, and the model parameters are iteratively updated through the backpropagation algorithm to minimize the classification error. At the same time, an L2 regularization term can be introduced to constrain the network weights and avoid model overfitting. During the testing phase, the fused features of the subjects to be judged are input into the trained model, and the class probability output by the Softmax layer realizes the automatic judgment of depression.

[0048] Experimental setup and analysis: In this experiment, the EEG microstate signals of each subject were divided into multiple sample segments, each lasting 15 seconds, serving as an independent input sample. To evaluate the generalization performance of the model, a 5-fold cross-validation strategy was employed. When dividing the training and test sets, it was strictly ensured that all sample segments corresponding to the same subject appeared only in the same subset of data within the same fold; that is, different segments from the same subject were not simultaneously assigned to both the training and test sets. This was done to avoid inflated model performance due to data leakage within the subjects, ensuring the objectivity and reliability of the model evaluation.

[0049] All models used in this experiment were built on the PyTorch deep learning framework. The LSTM layer was set to 6 layers with an output encoding dimension of 128. The optimizer was Adam, with an initial learning rate of 0.0001. During training, the batch size was set to 16, the dropout rate to 0.3, and the number of epochs to 100. The experiment used the publicly available dataset MODMA released by the UAIS laboratory of Lanzhou University, which includes 5-minute resting-state EEG data from 24 patients with severe depression and 29 healthy controls, as well as self-collected resting-state EEG data from 24 patients with depression and 22 healthy controls. Accuracy, F1 score, and area under the curve (AUC) were selected as evaluation metrics.

[0050] The method of this invention was compared with existing mainstream methods such as SVM, logistic regression, bag-of-words model, 3D-CNN, and GCN; the experimental results are shown in Table 1; the results show that the model of this invention achieved the best performance in all evaluation metrics on both datasets; specifically, in the MODMA dataset, the accuracy, F1 score, and AUC of the model of this invention reached 96.67%, 94.12%, and 97.23%, respectively, which are significantly better than the second-best performing baseline model (accuracy is about 5.3% higher than GCN, and AUC is about 5.78% higher than 3D-CNN); in the self-collected dataset, as much as Despite the increased difficulty in overall brainwave recognition, the model of this invention still demonstrates strong robustness, achieving an accuracy of 92.67%, an F1 score of 92.78%, and an AUC of 94.50%. All indicators are 3 to 4 percentage points higher than the second-best 3D-CNN model. Comprehensive comparison shows that, compared to the limitations of traditional machine learning methods in handling high-dimensional and complex EEG features, and the insufficient generalization ability of conventional deep learning models, this invention, through optimized network structure design, can more efficiently extract specific representations of EEG signals, significantly improving the accuracy of depression recognition and the model's cross-dataset generalization ability.

[0051] To demonstrate the indispensability of each component module of this invention and its generalization ability across datasets, ablation experiments were conducted in parallel on the MODMA dataset and a self-collected dataset. As shown in Table 2, the combined results of the two sets of experiments show that the complete model of this invention achieved optimal recognition performance under both different data distributions, proving that there is a significant synergistic effect between the various technical features. Comparison of the various model variants reveals that each module exhibits complementary advantages for datasets with different features: In MODMA, removing the attention mechanism caused a sharp drop of over 17% in the model's F1 score, highlighting its decisive role in filtering redundant EEG noise and focusing on key features of depression; while in the self-collected data, removing the LSTM module and the multi-instance learning module caused performance degradation (the F1 score dropped to 80.56% after removing LSTM, and the AUC dropped to 82.33% after removing MIL), which strongly confirms that for 15-second EEG samples, the extraction of long-distance temporal features and the global aggregation of unevenly distributed features are key to dealing with complex data distributions. Furthermore, regardless of the dataset, using only the original sequences yields extremely poor results, and removing the feature fusion module consistently leads to a steady decline in performance. Moreover, using traditional Jaccard similarity for encoding also significantly reduces classification performance. In summary, the composite network architecture constructed in this invention has no redundant design, and the modules work together to solve the technical pain points of strong heterogeneity and large individual differences in EEG signals. This endows the model with excellent recognition accuracy and cross-domain robustness, possessing outstanding substantive characteristics.

[0052] Table 1. Depression identification results of different models

[0053] Table 2 Results of ablation experiments using proprietary models

[0054] This invention constructs a rule-driven method for analyzing and recognizing EEG microstate sequences. By introducing the RuleGrowth algorithm to mine higher-order microstate transition rules, it effectively compensates for the deficiency of traditional static parameters in capturing the dynamic switching logic of brain states, significantly enhancing the discriminative power of features. Simultaneously, it utilizes sequence-aware Jaccard similarity for rule activation recoding, achieving effective modeling of long-range dependencies while preserving the local temporal topology of microstate sequences, thus improving the model's ability to characterize the complex evolution of EEG signals. Furthermore, by combining a multi-instance learning architecture with attention mechanisms, it not only achieves higher recognition accuracy under weak supervision but also provides an intuitive neurophysiological explanation through weight distribution, offering more robust and clinically valuable quantitative evidence for the early auxiliary diagnosis of depression.

[0055] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A multi-instance depression identification method based on high-order microstate features, characterized in that, Includes the following steps: Step 1: Acquire electroencephalogram (EEG) signals; Step 2: Extract the time sequence of brainwave microstates using the brainwave microstate template; Step 3: Utilize micro-state time series to mine frequent micro-state transition patterns and construct a global transition rule base; Step 4: Calculate the matching degree between the micro-state time sequence of the sliding window and the rules in the global rule base to realize the rule activation recoding of the micro-state time sequence; Step 5: Encode the rule index sequence using a multi-instance learning model to obtain higher-order transition features of microstates; Step 6: Concatenate the microstate high-order transition features and microstate temporal sequence features and input them into the classification network to output the discrimination result of depression.

2. The multi-instance depression identification method based on high-order microstate features according to claim 1, characterized in that, Step four specifically includes: Construct adjacent state transition sets for each sliding window subsequence. The set of adjacent state transitions with the regular sequence ; Indicates the first in the window A brainwave microstate symbol, ; This represents the total number of microstate symbols contained in the window subsequence; Represents the first in the rule sequence Each microstate symbol; Indicates the length of the regular sequence; calculate and similarity ; Use the index number of the rule with the highest similarity score The micro-state time sequence within the window is activated and a regular index sequence is generated.

3. The multi-instance depression identification method based on high-order microstate features according to claim 2, characterized in that, Step five specifically includes: Define the rule index sequence as a package, and define the rule index sequence for a certain time period within the package as an instance; The packet is input into the LSTM encoder to capture the temporal dependencies within the instance, and the contribution weights of different time periods to the classification result are calculated through the attention aggregation mechanism to obtain the higher-order transition features of the microstate after weight aggregation.

4. The multi-instance depression identification method based on high-order microstate features according to claim 1, characterized in that, The classification network consists of a first fully connected layer, a second fully connected layer, and a Softmax layer; the number of neurons in the fully connected layers are 256 and 64, respectively, and the activation function is ReLU.

5. The multi-instance depression identification method based on high-order microstate features according to claim 1, characterized in that, A global transfer rule base is constructed using the RuleGrowth algorithm.

6. The multi-instance depression identification method based on high-order microstate features according to claim 1, characterized in that, The characteristics of microstate time series include: average duration, frequency of occurrence, and time coverage.

7. The multi-instance depression identification method based on high-order microstate features according to claim 1, characterized in that, Whole-brain average reference and ICA preprocessing were performed on the EEG signals.

8. The multi-instance depression identification method based on high-order microstate features according to claim 1, characterized in that, Bandpass filtering and concave filtering are applied to the EEG signals.

9. A multi-instance depression recognition system based on high-order micro-state features, characterized in that, include: Memory is used to store instructions that can be executed by the processor; A processor for executing instructions to implement the multi-instance depression identification method based on high-order microstate features as described in any one of claims 1-8.

10. A computer-readable medium storing computer program code, characterized in that, The computer program code, when executed by a processor, implements the multi-instance depression identification method based on any one of claims 1-8, which features high-order microstate characteristics.

Images