Emotional state objective evaluation system based on electroencephalogram signals

Abstract

The invention discloses an emotional state objective evaluation system based on electroencephalogram signals, and relates to the technical field of mental health assistance for art therapy healing, and the system comprises a data collection module which is used for synchronously collecting multi-modal physiological and behavior data when a subject accepts an art therapy healing intervention process based on audio-visual content; and the input end of the emotional state evaluation module is connected to the data acquisition module, and the emotional state evaluation module is used for receiving and processing the multi-modal data. According to the emotional state objective evaluation system based on the electroencephalogram signals, multi-modal physiological behavior data and subjective feedback are fused, multiple independent verification and artificial arbitration mechanisms are introduced, the objectivity and reliability of emotional state evaluation are enhanced, and the system can accurately evaluate the emotional state according to the real-time evaluation result. According to the method, personalized healing intervention is dynamically matched and triggered from the pre-labeled content library, so that the conversion from static observation to dynamic response is realized, and the artistic healing process can adapt to the emotional change of the subject in real time.

Description

Technical Field

[0001] This invention relates to the field of mental health support technology for art therapy, specifically to an objective assessment system for emotional states based on electroencephalogram (EEG) signals. Background Technology

[0002] Today, with increasing attention being paid to mental health issues, art therapy, as a non-pharmacological intervention, is gradually being widely used. Among these, film therapy has become an important healing medium due to its ease of implementation and strong emotional arousal capabilities. However, traditional film therapy relies on the therapist's subjective observation and the patient's self-report, lacking objective and quantitative methods for assessing emotional states. This makes it difficult to accurately measure the therapeutic effect and to form a systematic "assessment-intervention-reassessment" treatment loop.

[0003] In the fields of affective computing and physiological signal analysis, electroencephalogram (EEG) signals are widely used in emotion recognition research due to their advantages such as high temporal resolution, strong objectivity, and resistance to subjective interference. Several EEG-based emotion classification methods already exist, such as the "Multi-label Classification Method and Device for Positive Emotions Based on Multi-channel EEG Data" described in patent publication number "CN113598774A". This method extracts EEG time-frequency features by segmenting panes and combines bidirectional long short-term memory networks, attention mechanisms, and graph convolutional neural networks to classify multi-label positive emotions. This technology improves the accuracy of emotion recognition to a certain extent, and is particularly suitable for multi-emotional scenes induced by movies.

[0004] However, the above methods still have the following limitations: Lack of adaptability to treatment scenarios: This technology focuses on emotion classification itself and has not been optimized for application scenarios of art therapy (such as film therapy), nor has it been deeply integrated with the treatment process and efficacy evaluation system; Lack of a closed-loop system: This solution is only an emotion recognition module and has not constructed a complete treatment closed loop of "assessment-intervention-reassessment," making it impossible to achieve dynamic adjustment of the treatment process and continuous tracking of efficacy; Lack of integration of multimodal treatment data: In real treatment scenarios, relying solely on EEG signals may ignore multi-dimensional information such as the patient's behavioral feedback and subjective feelings, affecting the comprehensiveness and accuracy of the assessment; Lack of personalized intervention mechanisms: Existing methods do not dynamically adjust the treatment content (such as film clips) according to real-time emotional states, making it difficult to achieve personalized and adaptive art therapy interventions.

[0005] Therefore, there is an urgent need for a system that can integrate EEG signal analysis, emotion recognition, and treatment process management to achieve objective assessment of patients' emotional states, quantification of efficacy, and optimization of the treatment closed loop, so as to improve the scientificity and effectiveness of art therapy. Summary of the Invention

[0006] The purpose of this invention is to provide an objective assessment system for emotional states based on electroencephalogram (EEG) signals, in order to solve the problems mentioned in the background art.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: an objective emotional state assessment system based on electroencephalogram (EEG) signals, comprising: The data acquisition module is used to simultaneously collect multimodal physiological and behavioral data during the process of subjects receiving art therapy intervention based on audiovisual content; The emotional state assessment module, whose input is connected to the data acquisition module, is used to receive and process the multimodal data and output the subject's real-time emotional state vector. The healing intervention control module, whose input is connected to the output of the emotional state assessment module, is used to dynamically select and output appropriate intervention content from a pre-established healing content library based on the emotional state vector. The closed-loop optimization module, whose input is connected to the output of the emotional state assessment module and the output of the healing intervention control module, is used to calculate efficacy assessment indicators based on the mapping relationship between emotional state vectors and intervention content in a single intervention and historical intervention cycles, and to update the subject's personal model and the strategy of the healing intervention control module. The emotional state assessment module, the healing intervention control module, and the closed-loop optimization module are sequentially connected to form a closed-loop workflow of "assessment-intervention-reassessment".

[0008] Furthermore, the data acquisition module includes an electroencephalogram (EEG) signal acquisition unit, a behavioral data acquisition unit, and a subjective feedback acquisition unit; The emotional state assessment module includes: The feature extraction and fusion submodule is used to extract time-frequency domain features from the EEG signals acquired by the EEG signal acquisition unit to obtain EEG features; and to perform weighted fusion of the EEG features, the behavioral features acquired by the behavioral data acquisition unit, and the subjective feedback features acquired by the subjective feedback acquisition unit to generate a multimodal fusion feature vector. The multi-label sentiment classification submodule, whose input is connected to the feature extraction and fusion submodule, is used to output a preliminary sentiment state label set containing multiple sentiment categories and their corresponding intensity probabilities based on the multimodal fusion feature vector.

[0009] Furthermore, the emotional state assessment module also includes: The verification and arbitration submodule has its input connected to the output of the multi-label sentiment classification submodule. The verification and arbitration submodule is configured to perform no fewer than two independent verification processes on the preliminary emotional state tag set; The verification process includes a combination of at least two of the following methods: confidence verification of emotion classification based on adversarial network reconstruction, temporal logic verification based on the alignment analysis of temporal changes in emotional state with key events in healing content, and verification by comparison with the subject's historical emotional baseline data. When the results of all verification processes are consistent with the preliminary emotional state label set, the verification and arbitration submodule outputs the preliminary emotional state label set as the final emotional state vector. When the result of the verification process is inconsistent with the preliminary emotional state label set, the verification and arbitration submodule triggers the manual auxiliary labeling interface and feeds back the manually confirmed labeling result as a training sample to the multi-label emotional classification submodule, driving the parameters of the multi-label emotional classification submodule to be adjusted online adaptively.

[0010] Furthermore, after the verification and arbitration submodule triggers the manual-assisted labeling interface, it packages the newly generated labeled samples, the corresponding multimodal fusion feature vectors, and the identification information of verification inconsistencies into an incremental training data package, and inputs it into the multi-label sentiment classification submodule. The multi-label sentiment classification submodule includes a lightweight fine-tuning routine. This routine is configured to iteratively update some network layer parameters of the classification model based on the incremental training data package during off-peak system working hours, while retaining the model's core knowledge, thereby achieving progressive performance optimization of the sentiment state assessment module during continuous use.

[0011] Furthermore, the healing intervention control module includes: The content-emotion mapping library contains several pre-stored healing content fragments, each of which is associated with one or more preset emotional arousal tags and intensity values; A dynamic matching engine is used to match the current emotional state vector output by the emotional state assessment module with the content-emotion mapping library; The matching strategy includes: when the current emotional state vector shows that a specific negative or low-arousal emotion is dominant, selecting content segments with positive emotional arousal tags and intensity values ​​that form a preset compensation relationship with the dominant emotion; when the current emotional state vector shows that a specific positive emotion is dominant, selecting content segments with emotional arousal tags that belong to the same category as the dominant emotion and intensity values ​​that form a continuation or enhancement relationship.

[0012] Furthermore, the healing intervention control module also includes: A personalized strategy library stores intervention strategy rules for different subjects or subject populations. These rules are generated and updated by the closed-loop optimization module based on historical efficacy data using a reinforcement learning algorithm. When performing matching, the dynamic matching engine prioritizes calling the strategy rules in the personalized strategy library that correspond to the current subject, and sorts and selects the candidate content fragments filtered from the content-emotion mapping library.

[0013] Furthermore, the closed-loop optimization module includes: The efficacy quantification unit is used to calculate efficacy indicators on one or more dimensions based on the time-series emotional state vector sequence output by the emotional state assessment module within a complete intervention cycle; the efficacy indicators include emotional diversity indicators, emotional stability indicators, and area under the curve of target emotional intensity change indicators. The model update unit, whose input is connected to the output of the efficacy quantification unit, is used to adjust the parameters in the subject's personal model according to the efficacy indicators and the corresponding intervention content sequence. The subject's personal model is used to characterize the subject's emotional response pattern to different healing content.

[0014] Furthermore, the model update unit of the closed-loop optimization module also outputs the evaluation data it generates regarding the effectiveness of the intervention strategy to the personalized strategy library of the healing intervention control module, as an environmental feedback signal for training the reinforcement learning algorithm, and is used to iteratively optimize the intervention strategy rules.

[0015] Furthermore, the multi-label sentiment classification submodule adopts a sentiment association modeling structure based on graph neural networks; This structure uses pre-trained word vectors to represent various emotion categories as graph nodes. It constructs a graph adjacency matrix based on the emotion co-occurrence probability matrix obtained from the data of the healing experiment. It learns the association between emotion categories through graph convolution operations and integrates the learned association information into the weights of the classifier, so that the emotion classification process can utilize the prior association between emotion semantics.

[0016] Furthermore, the audiovisual content on which the art therapy intervention is based is a film or film clip; the emotional state vector contains emotional categories selected from one or more of the set consisting of friendship, romance, tenderness, respect, pride, awe, gratitude, joy, and longing.

[0017] This invention provides an objective assessment system for emotional states based on electroencephalogram (EEG) signals. It has the following beneficial effects: This objective emotional state assessment system based on electroencephalogram (EEG) signals enhances the objectivity and reliability of emotional state assessment by integrating multimodal physiological and behavioral data with subjective feedback and introducing multiple independent verification and human arbitration mechanisms. Based on real-time assessment results, the system can dynamically match and trigger personalized therapeutic interventions from a pre-labeled content library, achieving a shift from static observation to dynamic response, enabling the art therapy process to adapt to the subject's emotional changes in real time.

[0018] This objective emotional state assessment system based on electroencephalogram (EEG) signals constructs a complete "assessment-intervention-reassessment" closed loop, driving continuous iterative optimization of the subject's personal model and intervention strategies through quantitative efficacy indicators. This closed-loop design makes the system not just a one-time assessment tool, but an adaptive healing platform with self-correction and evolutionary capabilities, providing a complete technical path for the quantitative measurement of art therapy effects and the continuous improvement of personalized intervention programs. Attached Figure Description

[0019] Figure 1 This is a data flow diagram showing the modules of an objective emotional state assessment system based on electroencephalogram (EEG) signals according to the present invention. Figure 2 This is a flowchart of the verification and arbitration mechanism of an objective emotional state assessment system based on electroencephalogram (EEG) signals according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 and Figure 2 This invention provides a technical solution: an objective assessment system for emotional states based on electroencephalogram (EEG) signals, comprising: The data acquisition module is used to simultaneously collect multimodal physiological and behavioral data during the process of subjects receiving art therapy intervention based on audiovisual content; The emotional state assessment module, whose input is connected to the data acquisition module, is used to receive and process multimodal data and output the subject's real-time emotional state vector. The healing intervention control module, whose input is connected to the output of the emotional state assessment module, is used to dynamically select and output appropriate intervention content from a pre-established healing content library based on the emotional state vector. The closed-loop optimization module, whose input is connected to the output of the emotional state assessment module and the output of the healing intervention control module, is used to calculate efficacy assessment indicators based on the mapping relationship between emotional state vectors and intervention content in a single intervention and historical intervention cycles, and to update the strategies of the subject's personal model and the healing intervention control module. Among them, the emotional state assessment module, the healing intervention control module, and the closed-loop optimization module are connected in sequence to form a closed-loop workflow of "assessment-intervention-reassessment".

[0022] It should be further explained that, firstly, while the subjects watched the film clips that were preset as a healing medium, the data acquisition module operated synchronously, continuously recording scalp EEG signals through a multi-channel EEG acquisition device, while simultaneously capturing facial expressions and body movements video streams using a camera, and collecting self-reported emotional keywords or intensity slider data from the subjects through an interactive interface, thus forming a multimodal data stream that includes physiological signals, behavioral videos, and subjective feedback.

[0023] Next, the emotional state assessment module receives this data stream. Its internal feature extraction and fusion submodule preprocesses the EEG signal to extract power spectral density features, performs feature point analysis on the behavioral video to extract the intensity of action units, and normalizes and weights these objective features with subjective feedback to form a unified multimodal feature vector. This vector is then input into a multi-label classification sub-model based on a bidirectional long short-term memory network and attention mechanism. This model combines the prior knowledge of the association between emotional categories learned by the graph convolutional network to output a preliminary emotional state label set containing various discrete emotions such as pleasure, gentleness, and awe, as well as their probability intensities.

[0024] Subsequently, the dynamic matching engine of the healing intervention control module is activated. It compares the received real-time emotional state vector with a pre-built content-emotion mapping library. The metadata of each movie clip in the library is labeled with the type and intensity of the emotions it may induce. According to the built-in strategy, if the current emotional state is mainly negative, the matching engine selects content labeled with strong positive emotions for compensatory intervention; otherwise, it selects content with similar emotional types for continuous reinforcement, thereby realizing the dynamic and personalized push of intervention content.

[0025] Finally, the closed-loop optimization module is activated after the entire intervention period ends. Its efficacy quantification unit analyzes the temporal evolution of the emotional state vector throughout the process and calculates quantitative indicators such as emotional diversity and stability. The model update unit then associates these indicators with the sequence of intervention content used, adjusts the subject's personal response model parameters using incremental learning, and uses the efficacy feedback to optimize the matching strategy of the intervention control module.

[0026] The four modules described above are sequentially linked, with the output of the preceding module serving as the input for the subsequent module. This allows the emotional assessment results to directly drive the selection of intervention content, while the post-intervention effects are fed back to optimize the assessment model and intervention strategies, thus forming a self-iterable and continuously adaptable "assessment-intervention-reassessment" treatment loop. This system enhances the robustness of state assessment through multimodal data fusion, achieves dynamic intervention through real-time content matching, and ensures the adaptability and sustainable optimization capabilities of the entire healing process through a closed-loop feedback mechanism.

[0027] The data acquisition module includes an EEG signal acquisition unit, a behavioral data acquisition unit, and a subjective feedback acquisition unit; The emotional state assessment module includes: The feature extraction and fusion submodule is used to extract time-frequency domain features from the EEG signals acquired by the EEG signal acquisition unit to obtain EEG features; and to perform weighted fusion of the EEG features, the behavioral features acquired by the behavioral data acquisition unit, and the subjective feedback features acquired by the subjective feedback acquisition unit to generate a multimodal fusion feature vector. The multi-label sentiment classification submodule, whose input is connected to the feature extraction and fusion submodule, is used to output a preliminary sentiment state label set containing multiple sentiment categories and their corresponding intensity probabilities based on the multimodal fusion feature vector.

[0028] It should be further explained that in the data acquisition module, the EEG signal acquisition unit uses a multi-lead EEG device to continuously acquire raw EEG signals at a preset sampling rate while the subject watches movie clips; the behavioral data acquisition unit uses a camera to capture the subject's facial video stream at a fixed frame rate and extracts action unit parameters such as the degree of mouth twitching and changes in eyebrow-eye distance through algorithms; the subjective feedback acquisition unit provides interactive controls on the side of the screen for the subject to select or slide the rating in real time to mark the emotional keywords they perceive and their intensity.

[0029] The feature extraction and fusion submodule in the emotional state assessment module first performs bandpass filtering and segmentation on the EEG signal, calculates the power spectral density characteristics of each segment in different frequency bands, and forms an EEG feature vector; at the same time, it standardizes the temporal action unit parameters extracted from the behavioral video to form a behavioral feature vector; the above EEG feature vector and behavioral feature vector, together with the encoded subjective feedback data, are input into a feature fusion layer with learnable weight parameters.

[0030] This fusion layer determines the contribution weights of various features in the final fusion vector through training, thereby generating a unified multimodal fusion feature vector. The multi-label sentiment classification submodule receives this fusion feature vector, and its model architecture consists of a feature encoding network and an image convolutional integral class head connected in sequence. The feature encoding network adopts a bidirectional long short-term memory network structure to capture the temporal dependencies of multimodal features and output high-level feature representations.

[0031] The graph convolutional classifier head pre-constructs a graph structure where nodes are word vectors for each sentiment category obtained using a pre-trained GloVe word embedding model, and edge weights are calculated based on the co-occurrence statistics of sentiment categories in the training dataset. This classifier head integrates the semantic relationships between sentiments into the classifier parameters through graph convolution operations, ultimately processing the high-level feature representation of the input and outputting a label set containing multiple sentiment categories and their corresponding probability values, thus completing the mapping from multimodal data to multi-label sentiment states.

[0032] The emotional state assessment module also includes: The verification and arbitration submodule has its input connected to the output of the multi-label sentiment classification submodule. The verification and arbitration submodule is configured to perform no fewer than two independent verification processes on the initial sentiment state tag set; The verification process includes a combination of at least two of the following methods: confidence verification of emotion classification based on adversarial network reconstruction, temporal logic verification based on the alignment analysis of temporal state changes with key events in the healing content, and verification by comparison with the subject's historical emotional baseline data. When the results of all verification processes are consistent with the preliminary sentiment state label set, the verification and arbitration submodule outputs the preliminary sentiment state label set as the final sentiment state vector. When the results of the verification process are inconsistent with the initial sentiment state label set, the verification and arbitration submodule triggers the manual auxiliary labeling interface and feeds back the manually confirmed labeling results as training samples to the multi-label sentiment classification submodule, driving the parameters of the multi-label sentiment classification submodule to make online adaptive adjustments.

[0033] It should be further explained that the specific implementation of the verification and arbitration submodule is as follows: After receiving the preliminary sentiment state label set output by the multi-label sentiment classification submodule, this submodule starts at least two independent verification routines in parallel. The first verification routine is based on an adversarial reconstruction network, which includes a sentiment label generator conditioned on multimodal fusion features and a discriminator; the generator attempts to reconstruct multimodal features from the preliminary label set, and obtains the confidence score of sentiment classification by calculating the mean squared error between the original features and the reconstructed features. If the score is lower than a preset threshold, the classification result is determined to be uncertain.

[0034] The confidence verification of sentiment classification reconstructed by adversarial networks in the verification and arbitration submodule is based on a Generative Adversarial Network (GAN). The generator adopts a stacked structure of three fully connected layers. The input is the feature encoding corresponding to the initial sentiment state label set, and the output is the reconstructed multimodal fusion feature. The discriminator also adopts three fully connected layers to distinguish between the original multimodal fusion feature and the feature reconstructed by the generator. The specific process of feature reconstruction is as follows: the intensity probability of each sentiment category in the initial sentiment state label set is converted into a fixed-dimensional encoding vector, which is input into the generator and outputs a reconstructed feature with the same dimension as the original multimodal fusion feature. The mean squared error between the original feature and the reconstructed feature is calculated as the reconstruction error. The confidence is negatively correlated with the reconstruction error. When the reconstruction error is below 0.05, the confidence is judged as qualified; when it is above this threshold, it is judged as unqualified. This threshold is determined by statistical analysis of 1000 sets of historical experimental data.

[0035] The second verification routine performs time logic verification. By analyzing the change curve of the probability of emotional intensity on the time axis, it detects whether there are abrupt peaks or jumps. It then aligns the timestamps of these change points with the timestamps of key plot events pre-marked in the currently playing movie clip and calculates their synchronous correlation. If the correlation coefficient is lower than the preset value, it is determined that the logical consistency between the emotional change and the content stimulus is insufficient.

[0036] In the temporal logic verification, the alignment analysis uses the Pearson correlation coefficient algorithm. The labeling standard for key events is the plot turning point in the film clip, including scene changes, emotional outbursts of characters, and the appearance of core lines. These are independently labeled by two or more experts with professional backgrounds in art therapy. The labeling results are determined to be key event timestamps after passing the consistency test (Kappa coefficient ≥ 0.7). The correlation threshold is set at 0.6. By calculating the Pearson correlation coefficient between the emotional state temporal change curve and the key event timestamp sequence, if the coefficient is lower than 0.6, it is determined that the logical consistency between the emotional change and the content stimulus is insufficient; if it is higher than or equal to the threshold, it is determined to be consistent.

[0037] If the results of both verification routines support the preliminary label set, the verification and arbitration submodule directly outputs it as the final emotional state vector. If either verification routine generates an inconsistent warning signal, the submodule will immediately trigger a graphical user interface that displays multimodal data fragments near the current time point, preliminary classification results, and verification anomaly prompts to the operator (e.g., therapist), requesting manual confirmation and labeling by the operator.

[0038] After the operator completes the labeling, the system automatically encapsulates the multimodal fusion feature vector corresponding to this event, the manually confirmed accurate sentiment label, and the verification anomaly type identifier into a single data packet and pushes it to a waiting queue in the multi-label sentiment classification submodule. The classification submodule contains an online fine-tuning process. When system computing resources are idle, this process retrieves the data packet from the queue and iteratively updates the parameters of specific network layers in the classification model with a small learning rate. This allows the model to gradually adapt to individual differences or correct systematic biases, achieving continuous self-correction of assessment accuracy. This mechanism transforms a one-time static classification into a dynamic, feedback-driven optimization process.

[0039] After the verification and arbitration submodule triggers the manual auxiliary labeling interface, it packages the newly labeled samples, the corresponding multimodal fusion feature vectors, and the identification information of verification inconsistencies into an incremental training data package and inputs it into the multi-label sentiment classification submodule. The multi-label sentiment classification submodule includes a lightweight fine-tuning routine. This routine is configured to iteratively update some network layer parameters of the classification model based on incremental training data packets during off-peak system working hours, while retaining the model's core knowledge, thereby achieving incremental performance optimization of the sentiment state assessment module during continuous use.

[0040] It should be further explained that after the verification and arbitration submodule triggers the manual auxiliary labeling interface and receives the operator's confirmation result, it performs the following operations: First, the submodule encapsulates the entire data instance that caused the verification inconsistency. This data instance includes the multimodal fusion feature vector sequence within a time window before and after the triggering time, the preliminary sentiment label set output by the multi-label sentiment classification submodule, the detailed judgment data and confidence scores output by all verification routines, and the final authoritative sentiment label confirmed or corrected by the operator.

[0041] During encapsulation, the specific type of inconsistency in this verification is recorded in the data packet header information, such as "low confidence in adversarial reconstruction" or "insufficient temporal logical relevance". Subsequently, this structured incremental training data packet is transmitted and stored in a first-in-first-out cache queue maintained by the multi-label sentiment classification submodule.

[0042] The lightweight fine-tuning routines pre-defined within the submodules are managed by a background scheduler that continuously monitors system load. When the scheduler detects that CPU utilization is below a set threshold and there are no tasks being evaluated in real time, it activates the fine-tuning routine. This routine retrieves a data packet from the head of the queue, first parsing its header information to identify the inconsistency type, and then selectively unfreezing specific network layer parameters in the classification model based on the type. In the lightweight fine-tuning routine, "partial network layers" specifically refers to the two fully connected layers at the end of the feature encoding network in the multi-label sentiment classification submodule. This selection is based on sensitivity analysis of the model structure. These two network layers are more adaptable to personalized data, and fine-tuning will not significantly affect the overall stability of the model. The learning rate ranges from 1e-5 to 1e-3. When the sample size of the incremental training data package is less than 50, a learning rate of 1e-5 to 5e-5 is selected; when the sample size is between 50 and 200, a learning rate of 5e-5 to 5e-4 is selected; and when the sample size exceeds 200, a learning rate of 5e-4 to 1e-3 is selected. The stopping condition for iterative updates is that the change in the loss function value is less than 1e-4 after three consecutive iterations, or the number of iterations reaches 50. The fine-tuning process is terminated when either condition is met, and the updated parameters are fixed as the current model parameters.

[0043] For example, to combat the reconstruction confidence problem, the focus might be on adjusting the parameters of the fully connected layer at the end of the feature encoding network; for the temporal logic problem, the focus might be on adjusting the parameters of the attention mechanism layer.

[0044] During parameter updates, a small learning rate is used, and training is performed only a limited number of times using data from the current incremental data packet. Meanwhile, the elastic weight integration algorithm is used to constrain other parameters in the model that have not been updated, in order to protect the previously learned core knowledge from being destroyed.

[0045] The core of the elastic weight integration algorithm is to apply dynamic decay constraints to the parameters of unupdated network layers. The constraint strength is determined by the contribution of each parameter in historical training, which is quantified by the parameter's impact on the model's classification accuracy. Parameters with better historical performance are subject to stronger constraints. In practice, a decay factor is introduced into the parameters of unupdated layers during parameter updates, keeping the parameter adjustment within 10% of its original value. This reduces parameter fluctuations, thereby preserving the model's learned core knowledge and preventing overall performance degradation due to local fine-tuning.

[0046] After the update is completed, the updated model parameters are seamlessly switched to the online service model, while the original parameters are saved as a backup. This process ensures that the sentiment assessment module can continuously optimize itself using human feedback without interrupting service, thereby gradually improving the accuracy of identifying complex or individualized sentiment patterns.

[0047] The healing intervention control module includes: The content-emotion mapping library contains several pre-stored healing content fragments, each of which is associated with one or more preset emotional arousal tags and intensity values; A dynamic matching engine is used to match the current sentiment state vector output by the sentiment state assessment module with the content-sentiment mapping library; The matching strategy includes: when the current emotional state vector shows that a specific negative or low-arousal emotion is dominant, selecting content segments with positive emotional arousal tags and intensity values ​​that form a preset compensatory relationship with the dominant emotion; when the current emotional state vector shows that a specific positive emotion is dominant, selecting content segments with emotional arousal tags that belong to the same category as the dominant emotion and intensity values ​​that form a continuation or enhancement relationship.

[0048] It should be further explained that the specific implementation of the healing intervention control module is as follows: The content-emotion mapping library is established through a pre-annotation process before the system is deployed. For each movie or video clip in the healing content library, multiple experts independently annotate the expected one or more emotion categories and corresponding intensity levels that may be induced by its audiovisual content. Finally, a structured emotion attribute description file is generated for each clip through aggregation and statistics. This file contains a list of emotion tags and the normalized intensity value of each tag.

[0049] The dynamic matching engine receives emotional state vectors from the emotional state assessment module in real time. These vectors describe the prevalence of various emotions in the form of a probability distribution. The engine first extracts the dominant emotional components from the emotional state vectors, which are defined as a number of emotional tags and their intensities that have a probability value exceeding a preset threshold and rank highly among all components.

[0050] Next, the engine queries the content-sentiment mapping library to filter out content fragments in the sentiment attribute description files that contain sentiment tags that match or are consistent with the current dominant sentiment tag, forming a candidate set. For each selected candidate content fragment, the engine calculates multiple relationship metrics between its labeled sentiment intensity vector and the current user's sentiment state vector. Specifically: when the dominant sentiment component is identified as negative or low-arousal, the engine prioritizes calculating "sentiment compensation," which is a complementary function value between the positive sentiment label intensity of the candidate fragment and the negative sentiment intensity of the current user; when the dominant sentiment component is positive, the engine prioritizes calculating "sentiment continuity," which is a function value of the intensity difference between the candidate fragment and the current user in the same or similar sentiment.

[0051] Ultimately, the engine determines the most suitable healing content segment from the candidate set based on preset rules (e.g., selecting the segment with the highest compensation or continuity), and sends the segment's identifier to the playback subsystem for loading and presentation, thereby achieving real-time, data-driven dynamic control of the healing intervention content.

[0052] The healing intervention control module also includes: The personalized strategy library stores intervention strategy rules for different subjects or subject populations. The rules are generated and updated by the closed-loop optimization module based on historical efficacy data through reinforcement learning algorithms. When performing matching, the dynamic matching engine prioritizes calling the strategy rules in the personalized strategy library that correspond to the current subject, and sorts and selects the candidate content fragments filtered from the content-sentiment mapping library.

[0053] It should be further explained that the specific implementation of the collaborative work between the personalized strategy library and the dynamic matching engine in the therapeutic intervention control module is as follows: Logically, the personalized strategy library consists of a set of queryable policy functions or decision tables, with each policy associated with a specific subject's unique identifier or a group label of subjects with similar response patterns. The generation and iterative updating of strategies are driven by a closed-loop optimization module, the core of which is a reinforcement learning agent running in the background.

[0054] The agent models a single intervention decision as a Markov decision process: where the state (S) is defined as the concatenation of the emotional state vector output by the emotional state assessment module at the current moment and the subject's recent historical efficacy trend; the action (A) is defined as the operation of selecting a specific content segment from the content-emotion mapping library; and the reward (R) is calculated based on the positive change in the subject's emotional state vector observed in a subsequent time window after performing the action (playing the segment), such as calculating the increase in the intensity of the target positive emotion or the decrease in the intensity of the negative emotion.

[0055] The reinforcement learning algorithm in the healing intervention control module adopts the PPO (Proximal Policy Optimization) algorithm, which can effectively control the update amplitude during policy updates and avoid training instability. The state space is defined as a 9-dimensional vector composed of the intensity probability values ​​of 9 emotion categories in the emotional state vector, which is concatenated with the mean of the efficacy indicators (emotional diversity, stability, and area under the curve) of the last 3 interventions to form a 12-dimensional extended state vector to comprehensively reflect the current emotional state and the historical intervention effect. The action space encodes each segment in the healing content library into a unique integer identifier according to the storage order of the content-emotion mapping library, with the encoding range from 1 to the total number of segments in the content library. The reward function's "positive change in emotion" is calculated by subtracting the average increase in the intensity probability of the 9 preset positive emotions (friendship, romance, gentleness, respect, pride, awe, gratitude, joy, and longing) after the current intervention from the average increase in the intensity probability of low-arousal negative emotions (such as depression and numbness). The increase is calculated by the difference in the intensity probability of the corresponding emotions before and after the intervention.

[0056] Reinforcement learning agents collect state-action-reward sequence data through continuous interaction with the system environment (i.e., the real healing process) and periodically update their internal value function or policy network parameters using this data. The stable policy obtained after training convergence is compiled or simplified into a series of executable rules and stored in a personalized policy library. When the dynamic matching engine performs content matching for a subject, it first checks if an effective policy for that subject exists in the personalized policy library. If it does, the engine invokes that policy, which directly recommends one or more content fragment identifiers based on the current state, or reorders and weights the candidate fragment list filtered by general rules. This ensures that the final content selection is not only based on immediate emotional matching but also incorporates optimal intervention experience learned from the subject's historical interaction data.

[0057] The closed-loop optimization module continuously collects newly generated intervention effect data and provides new training samples for the reinforcement learning agent, thereby enabling regular and automated updates of the strategies in the personalized strategy library, and giving the intervention control the ability to adapt and optimize over a long period of time.

[0058] The closed-loop optimization module includes: The efficacy quantification unit is used to calculate efficacy indicators on one or more dimensions based on the time-series emotional state vector sequence output by the emotional state assessment module within a complete intervention cycle. The efficacy indicators include emotional diversity indicators, emotional stability indicators, and the area under the curve of the target emotional intensity change. The model update unit, whose input is connected to the output of the efficacy quantification unit, is used to adjust the parameters in the subject's personal model according to the efficacy indicators and the corresponding intervention content sequence. The subject's personal model is used to characterize the subject's emotional response pattern to different healing content.

[0059] It should be further explained that the specific process of the efficacy quantification unit and the model update unit working together in the closed-loop optimization module is as follows: After a single complete intervention cycle, the efficacy quantification unit receives a sequence of emotional state vectors arranged in chronological order from the emotional state assessment module. This unit first extracts the probability intensity of several emotional categories set as treatment targets by the system from the vectors at each time point, forming multiple curves showing the change of target emotional intensity over time.

[0060] The efficacy quantification unit performs three core calculations: First, it calculates the emotional diversity index by counting the total number of different emotional categories that have occurred at all time points throughout the entire cycle and whose probability intensity exceeds a set threshold, and then normalizes this total number. In the closed-loop optimization module, the normalization of the emotional diversity index adopts the min-max normalization method. This maps the total number of valid emotional categories (those with a probability intensity ≥ 0.3 are included) throughout the entire intervention period to the 0-1 interval. The mapping formula references the industry-standard min-max normalization logic, facilitating horizontal comparison of index values. The threshold for "valid emotional category" is uniformly set to a probability intensity ≥ 0.3. The statistics must cover all time points within the intervention period; any emotional category whose probability intensity exceeds this threshold at any time point is considered a valid category. Finally, the total number of unique valid categories is calculated.

[0061] Second, calculate the emotional stability index, select one or more core target emotions, calculate the variance of their intensity curve in the time dimension or calculate the local volatility through a sliding window, and then aggregate these volatility measures. The variance of the emotional stability index is calculated using a 10-second time window. That is, an emotional intensity sequence is extracted every 10 seconds, and the variance of the core target emotion (determined according to the subject's personal model, usually 3-5 emotions with the most significant healing effect on the subject) intensity within the window is calculated. The local volatility is calculated by the ratio of the difference in variance between two adjacent windows to the variance of the previous window. For example, if the variance of the nth window is σn and the variance of the (n+1)th window is σn+1, then the local volatility is |σn+1-σn| / σn. The average of the local volatility of all adjacent windows throughout the entire intervention period is taken as the final emotional stability index.

[0062] Third, calculate the area under the target emotion intensity change curve. By numerically integrating the target emotion intensity curve on the time axis, the cumulative change in emotion intensity from the beginning to the end of the cycle is obtained.

[0063] The area under the target emotional intensity change curve was calculated using the trapezoidal integral method. The starting time point of the integration interval was the moment the subject began playing the healing content, and the ending time point was 30 seconds after the healing content ended (with no additional intervention during this period), ensuring coverage of the period of sustained emotional impact after the intervention. During integration, time was plotted on the horizontal axis and the target emotional intensity on the vertical axis, dividing the emotional intensity curve of the entire intervention period into several consecutive trapezoids. The upper and lower bases of each trapezoid represented the emotional intensity values ​​of two adjacent time points, and the height was the time interval between the two time points. The sum of the areas of all trapezoids was the final value of this indicator.

[0064] The model update unit receives the calculated efficacy indicators and the identifier sequence of all content fragments used by the healing intervention control module within the current cycle. This unit maintains a personal response model for the current subject, which can be a parameterized function whose input is the feature encoding of the content fragments and whose output is the predicted emotional response vector.

[0065] The model update unit encodes the actual content segment sequence played in the current period and inputs it into the personal response model to obtain the predicted emotional response sequence. It then compares this predicted sequence with the actual observed emotional state vector sequence and calculates the prediction error.

[0066] Then, the model update unit constructs a composite loss function, which not only includes the prediction error term mentioned above, but also introduces various indicators calculated by the efficacy quantification unit as part of the regularization constraint or optimization objective. For example, it requires that the update direction of the model parameters should make the predicted sentiment stability index move closer to a better value.

[0067] Through backpropagation or other optimization algorithms, the model update unit adjusts the internal parameters of the individual response model, thereby narrowing the gap between its predictions and the actual responses of the subjects, and embedding the goal of improving efficacy into the model's learning process. The updated individual response model will be used for personalized prediction and strategy optimization in subsequent intervention cycles.

[0068] The model update unit of the closed-loop optimization module also outputs the evaluation data it generates regarding the effectiveness of the intervention strategy to the personalized strategy library of the healing intervention control module, as an environmental feedback signal for training the reinforcement learning algorithm, and is used to iteratively optimize the intervention strategy rules.

[0069] It should be further explained that the collaborative data interaction and update mechanism between the model update unit of the closed-loop optimization module and the personalized strategy library of the healing intervention control module is implemented as follows: After the end of each intervention cycle and the model update unit completes the adjustment of the parameters of the subject's individual response model, the unit will simultaneously generate a structured intervention strategy effectiveness evaluation data package.

[0070] The core content of this data package includes: the identifier of each content segment actually selected and played by the healing intervention control module during this cycle, its contextual state information when it was selected (i.e., the emotional state vector at that time), the data on the changes in the subject's emotional state vector actually observed after playing the segment, and the estimated value of the immediate reward calculated based on these changes.

[0071] The model update unit formats and encapsulates this evaluation data packet as a new "experience trajectory" through an internally defined data interface that is fully compatible with the training environment of reinforcement learning agents in the personalized policy library. It then sends and adds the packet to the experience replay buffer in the personalized policy library.

[0072] The reinforcement learning agent training process, running within a personalized policy library, is periodically awakened, either offline or asynchronously. Upon awakening, the process randomly samples a batch of such experience trajectory data from the experience replay buffer to update the parameters of its policy network or value function. During training, the agent utilizes contextual state information and content fragment identifiers (as actions) from the trajectory data to update its estimate of the expected long-term reward for taking a specific action in a given state.

[0073] In particular, the actual emotional changes and reward predictions provided by the trajectory data are directly used as environmental feedback signals to calculate temporal difference errors or policy gradients, thereby guiding the policy network parameters to update in a direction that generates more positive emotional responses. After one round of training iterations, the updated policy network parameters are solidified and compiled into new, executable decision rules or policy functions, which are then overridden or merged into the corresponding subject's policy records in the personalized policy library. This process ensures that the optimization of intervention strategies not only relies on long-term statistical efficacy indicators but also directly incorporates the immediate effect feedback generated by each specific intervention action, making the policy learning process more refined and the response more rapid, forming a two-way closed-loop optimization channel from individual efficacy assessment to group (or individual) intervention strategy iteration.

[0074] The multi-label sentiment classification submodule adopts a sentiment association modeling structure based on graph neural networks; This structure uses pre-trained word vectors to represent various emotion categories as graph nodes. It constructs a graph adjacency matrix based on the emotion co-occurrence probability matrix obtained from the data of the healing experiment. It learns the association between emotion categories through graph convolution operations and integrates the learned association information into the weights of the classifier, so that the emotion classification process can utilize the prior association between emotion semantics.

[0075] It should be further explained that the specific implementation of the sentiment association modeling structure based on graph neural network in the multi-label sentiment classification submodule is as follows: The structure first constructs a graph, where each node corresponds to a discrete sentiment category. The initial feature vector of the node is obtained by loading a pre-trained natural language word embedding model, such as using the GloVe model trained on a large corpus to generate a dense vector representation of fixed dimensions for each sentiment category word such as "friendship", "romance", and "awe".

[0076] The edge structure of the graph is defined by the emotion co-occurrence probability matrix, which is obtained by statistically analyzing the historical healing experiment dataset used in the training phase. The matrix calculates the frequency of each pair of emotion categories appearing in the same sample and converts them into conditional probabilities or smoothed joint probabilities, which serve as the quantitative basis for the connection strength between nodes.

[0077] The statistical sample size of the emotion co-occurrence probability matrix is ​​no less than 1000 complete art therapy experiment samples, with each sample containing a complete emotional state sequence of the subject within one intervention period. Co-occurrence is defined as occurring within the playback duration of the same therapeutic content segment; that is, if two emotion categories are both detected as valid emotions (intensity probability ≥ 0.3) during the playback of the same film segment, it is considered a co-occurrence. The probability is calculated using the joint probability P(A,B), which is the number of co-occurrence samples of emotion A and emotion B divided by the total number of samples. Each element in the matrix corresponds to the joint probability of a pair of emotion categories. Finally, the elements in each row of the matrix are normalized so that the sum of each row's elements is 1, facilitating weight calculation in subsequent graph convolution operations.

[0078] Next, the probability matrix is ​​thresholded and normalized to form the adjacency matrix required for graph convolution operations. During classification, the multimodal fusion feature vector is first subjected to a depth transformation through a feature encoding network to obtain a high-level abstract feature representation.

[0079] Meanwhile, the graph convolutional network takes the aforementioned sentiment category node vectors as input and the constructed adjacency matrix as structural constraints, and performs at least one layer of graph convolution operation. In each layer, each node aggregates the feature information of its neighboring nodes and updates it in combination with its own features, thereby outputting a new node representation that can reflect the semantic correlation of sentiment.

[0080] The multi-label sentiment classification submodule employs a graph neural network-based sentiment association modeling structure. The graph neural network consists of two layers, using the ReLU activation function to effectively mitigate the vanishing gradient problem and meet the learning requirements of sentiment features. The graph convolution propagation process involves updating the features of each sentiment category's corresponding node by aggregating the weighted sum of the features of its neighboring nodes and its own features. The weights are determined by the sentiment co-occurrence probability values ​​in the adjacency matrix, ensuring that closely related sentiment categories contribute more to each other's feature updates. Pre-trained word vectors are generated using a GloVe model with a 200-dimensional dimension. The training corpus covers literature in the field of sentiment computing, descriptive texts of art therapy cases, and a corpus of everyday sentiment expressions, ensuring that the word vectors accurately represent the semantic associations of nine preset sentiment categories.

[0081] Finally, the classifier interacts with the high-level abstract feature representations and the sentiment node representations updated by graph convolution. Specifically, this is achieved by performing a dot product operation between the feature representations and each sentiment node representation, followed by further processing, to generate probability scores for each sentiment category. This mechanism ensures that classification decisions not only rely on the input multimodal data itself but are also guided by the inherent symbiotic relationships or semantic similarities between sentiment categories statistically learned from the data, thus explicitly embedding prior knowledge about sentiment associations within the model.

[0082] The audiovisual content on which art therapy interventions are based is film or film clips; the emotional state vector contains emotional categories selected from one or more of the set consisting of friendship, romance, tenderness, respect, pride, awe, gratitude, joy, and longing.

[0083] It should be further explained that the specific implementation of art therapy intervention content and emotional categories is as follows: The content library managed by the system's therapy intervention control module consists of selected and edited film clips. Each clip has a defined start and end time, and the duration is adapted to the duration requirements of a single intervention.

[0084] These film clips undergo preprocessing before being stored in the database. In addition to storing their audiovisual data files, a structured metadata file is generated and associated with them, which records the basic information of the film from which the clips originated, a content summary, and pre-annotated emotional arousal attributes.

[0085] The system's content management and playback subsystem can quickly locate and smoothly play corresponding movie clips from the content library based on the clip identifiers sent by the healing intervention control module. Regarding emotion categories, in the multi-label emotion classification submodule of the emotion state assessment module, the system explicitly limits the emotion recognition task to a closed set consisting of nine discrete emotion words, including: friendship, romance, tenderness, respect, pride, awe, gratitude, joy, and longing.

[0086] These nine emotion categories are uniquely encoded in the system's internal data processing, for example, by being assigned fixed index numbers from 1 to 9. During the model's training and inference, the number of neurons in the output layer of the emotion classification submodule strictly corresponds to these nine emotions, and the activation value of each neuron represents the probability that the input data belongs to the corresponding emotion category.

[0087] In the pre-annotation stage of the content-sentiment mapping library, experts only use these nine sentiment words to annotate movie clips; in the reinforcement learning model of the personalized strategy library, the calculation of actions and rewards is also based on the state changes of these nine sentiments.

[0088] In visualization reports for therapists or subjects, the trend charts of emotional state changes and the calculation of quantitative indicators of therapeutic efficacy are all based on these nine emotions as analytical dimensions. This specific definition of the emotion set provides a consistent and operational emotional semantic framework for the entire system, ensuring the consistency, interpretability, and computability of the emotional dimensions throughout the entire process from signal recognition and content matching to effect evaluation. This allows the system to focus on identifying and regulating specific positive emotional spectrums closely related to art therapy, especially those evoked by film media.

[0089] This system enhances the objectivity and reliability of emotional state assessment by integrating multimodal physiological and behavioral data with subjective feedback and introducing multiple independent verification and human arbitration mechanisms. Based on real-time assessment results, the system can dynamically match and trigger personalized therapeutic interventions from a pre-labeled content library, achieving a shift from static observation to dynamic response, enabling the art therapy process to adapt to the subject's emotional changes in real time.

[0090] Furthermore, the system constructs a complete "assessment-intervention-reassessment" closed loop, driving continuous iterative optimization of the subject's personal model and intervention strategies through quantitative efficacy indicators. This closed-loop design makes the system not just a one-time assessment tool, but an adaptive healing platform with self-correction and evolutionary capabilities, providing a complete technical implementation path for the quantitative measurement of art therapy effects and the continuous improvement of personalized intervention programs.

[0091] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An objective assessment system for emotional states based on electroencephalogram (EEG) signals, characterized in that, include: The data acquisition module is used to simultaneously collect multimodal physiological and behavioral data during the process of subjects receiving art therapy intervention based on audiovisual content; The emotional state assessment module, whose input is connected to the data acquisition module, is used to receive and process the multimodal data and output the subject's real-time emotional state vector. The healing intervention control module, whose input is connected to the output of the emotional state assessment module, is used to dynamically select and output appropriate intervention content from a pre-established healing content library based on the emotional state vector. The closed-loop optimization module, whose input is connected to the output of the emotional state assessment module and the output of the healing intervention control module, is used to calculate efficacy assessment indicators based on the mapping relationship between emotional state vectors and intervention content in a single intervention and historical intervention cycles, and to update the subject's personal model and the strategy of the healing intervention control module. The emotional state assessment module, the healing intervention control module, and the closed-loop optimization module are sequentially connected to form a closed-loop workflow of "assessment-intervention-reassessment".

2. The objective assessment system for emotional state based on electroencephalogram (EEG) signals according to claim 1, characterized in that: The data acquisition module includes an electroencephalogram (EEG) signal acquisition unit, a behavioral data acquisition unit, and a subjective feedback acquisition unit. The emotional state assessment module includes: The feature extraction and fusion submodule is used to extract time-frequency domain features from the EEG signals acquired by the EEG signal acquisition unit to obtain EEG features; and to perform weighted fusion of the EEG features, the behavioral features acquired by the behavioral data acquisition unit, and the subjective feedback features acquired by the subjective feedback acquisition unit to generate a multimodal fusion feature vector. The multi-label sentiment classification submodule, whose input is connected to the feature extraction and fusion submodule, is used to output a preliminary sentiment state label set containing multiple sentiment categories and their corresponding intensity probabilities based on the multimodal fusion feature vector.

3. The objective assessment system for emotional state based on electroencephalogram (EEG) signals according to claim 2, characterized in that: The emotional state assessment module also includes: The verification and arbitration submodule has its input connected to the output of the multi-label sentiment classification submodule. The verification and arbitration submodule is configured to perform no fewer than two independent verification processes on the preliminary emotional state tag set; The verification process includes a combination of at least two of the following methods: confidence verification of emotion classification based on adversarial network reconstruction, temporal logic verification based on the alignment analysis of temporal changes in emotional state with key events in healing content, and verification by comparison with the subject's historical emotional baseline data. When the results of all verification processes are consistent with the preliminary emotional state label set, the verification and arbitration submodule outputs the preliminary emotional state label set as the final emotional state vector. When the result of the verification process is inconsistent with the preliminary emotional state label set, the verification and arbitration submodule triggers the manual auxiliary labeling interface and feeds back the manually confirmed labeling result as a training sample to the multi-label emotional classification submodule, driving the parameters of the multi-label emotional classification submodule to be adjusted online adaptively.

4. The objective assessment system for emotional state based on electroencephalogram (EEG) signals according to claim 3, characterized in that: After the verification and arbitration submodule triggers the manual-assisted labeling interface, it packages the newly generated labeled samples, the corresponding multimodal fusion feature vectors, and the identification information of verification inconsistencies into an incremental training data package, and inputs it into the multi-label sentiment classification submodule. The multi-label sentiment classification submodule includes a lightweight fine-tuning routine. This routine is configured to iteratively update some network layer parameters of the classification model based on the incremental training data package during off-peak system working hours, while retaining the model's core knowledge, thereby achieving progressive performance optimization of the sentiment state assessment module during continuous use.

5. The objective assessment system for emotional state based on electroencephalogram (EEG) signals according to claim 4, characterized in that: The therapeutic intervention control module includes: The content-emotion mapping library contains several pre-stored healing content fragments, each of which is associated with one or more preset emotional arousal tags and intensity values; A dynamic matching engine is used to match the current emotional state vector output by the emotional state assessment module with the content-emotion mapping library; The matching strategy includes: when the current emotional state vector shows that a specific negative or low-arousal emotion is dominant, selecting content segments with positive emotional arousal tags and intensity values ​​that form a preset compensation relationship with the dominant emotion; when the current emotional state vector shows that a specific positive emotion is dominant, selecting content segments with emotional arousal tags that belong to the same category as the dominant emotion and intensity values ​​that form a continuation or enhancement relationship.

6. The objective assessment system for emotional state based on electroencephalogram (EEG) signals according to claim 5, characterized in that: The healing intervention control module also includes: A personalized strategy library stores intervention strategy rules for different subjects or subject populations. These rules are generated and updated by the closed-loop optimization module based on historical efficacy data using a reinforcement learning algorithm. When performing matching, the dynamic matching engine prioritizes calling the strategy rules in the personalized strategy library that correspond to the current subject, and sorts and selects the candidate content fragments filtered from the content-emotion mapping library.

7. The objective assessment system for emotional state based on electroencephalogram (EEG) signals according to claim 6, characterized in that: The closed-loop optimization module includes: The efficacy quantification unit is used to calculate efficacy indicators on one or more dimensions based on the time-series emotional state vector sequence output by the emotional state assessment module within a complete intervention cycle; the efficacy indicators include emotional diversity indicators, emotional stability indicators, and area under the curve of target emotional intensity change indicators. The model update unit, whose input is connected to the output of the efficacy quantification unit, is used to adjust the parameters in the subject's personal model according to the efficacy indicators and the corresponding intervention content sequence. The subject's personal model is used to characterize the subject's emotional response pattern to different healing content.

8. The objective assessment system for emotional state based on electroencephalogram (EEG) signals according to claim 7, characterized in that: The model update unit of the closed-loop optimization module also outputs the evaluation data it generates regarding the effectiveness of the intervention strategy to the personalized strategy library of the healing intervention control module, as an environmental feedback signal for training the reinforcement learning algorithm, and is used to iteratively optimize the intervention strategy rules.

9. The objective assessment system for emotional state based on electroencephalogram (EEG) signals according to claim 8, characterized in that: The multi-label sentiment classification submodule adopts a sentiment association modeling structure based on graph neural networks; This structure uses pre-trained word vectors to represent various emotion categories as graph nodes. It constructs a graph adjacency matrix based on the emotion co-occurrence probability matrix obtained from the data of the healing experiment. It learns the association between emotion categories through graph convolution operations and integrates the learned association information into the weights of the classifier, so that the emotion classification process can utilize the prior association between emotion semantics.

10. The objective assessment system for emotional state based on electroencephalogram (EEG) signals according to claim 9, characterized in that: The audiovisual content on which the art therapy intervention is based is a movie or movie clip; the emotional state vector contains emotional categories selected from one or more of the set consisting of friendship, romance, tenderness, respect, pride, awe, gratitude, joy, and longing.

Images