Intraoperative intervention assessment method and apparatus based on generative adversarial network

Abstract

The application relates to an intraoperative intervention measure evaluation method and device based on a generative adversarial network. The method first encodes an intraoperative electroencephalogram signal into a neural disturbance field by introducing a space-time encoder, effectively capturing the space-time propagation mode of a nociceptive stimulus in the brain; then, a generative adversarial network is used to predict a postoperative immune state vector trajectory, capturing the mapping relationship between the neural imprint and the immune imprint left by the nociceptive stimulus in different space-time modes during the operation; then, two different postoperative immune state vector trajectories with and without intervention measures are predicted, the effects of different intervention measures can be simulated and quantitatively compared, so as to assist doctors in optimizing the postoperative recovery and long-term prognosis of patients. The application deeply excavates the space-time characteristics of the intraoperative neural signal, dynamically predicts the postoperative immune state by combining the generative adversarial network, and introduces an intervention operator to simulate the intervention effect, thereby providing an innovative way of prospectively and quantitatively evaluating the effect of intraoperative intervention measures.

Description

Technical Field

[0001] This application relates to the field of postoperative intervention assessment technology, and in particular to a method and apparatus for intraoperative intervention assessment based on generative adversarial networks. Background Technology

[0002] Perioperative trauma, anesthetic stimulation, and tissue traction can trigger immune response disorders through the central nervous system and neuroimmune axis, often manifesting as hyperinflammatory postoperatively and immune cell imbalance. This directly increases the risk of postoperative infection, delays organ function recovery, and affects long-term prognosis. Accurately assessing the intensity of stimulation during surgery, predicting the evolution of postoperative immune status, and objectively evaluating the clinical effects of interventions such as analgesia and sedation have become key requirements for perioperative precision medicine.

[0003] Current clinical intraoperative monitoring mainly relies on routine physiological indicators such as hemodynamics, electrocardiogram, and electroencephalogram. Immune status assessment depends on laboratory tests at single or few time points before and after surgery. It mainly uses retrospective statistics or traditional machine learning methods to establish simple correlations between intraoperative stimulation parameters and postoperative immune indicators, thereby conducting perioperative risk assessment and evaluating intervention effects.

[0004] These methods can only achieve static correlation analysis of postoperative immune indicators, but cannot capture the spatiotemporal changes of intraoperative nerve signals, nor can they construct an accurate predictive model of the dynamic trajectory of postoperative immune response from intraoperative state to postoperative state. Furthermore, they cannot simulate and quantify the effects of different intervention measures, making it difficult to provide forward-looking and individualized support for clinical decision-making during surgery. Summary of the Invention

[0005] A first aspect of this application provides a method for evaluating intraoperative interventions based on generative adversarial networks, the method comprising: Acquire intraoperative EEG signals of the target, preoperative immune status vector of the target, and intervention measures; The target intraoperative EEG signal is encoded into a first neural perturbation field based on a spatiotemporal encoder; wherein, the neural perturbation field is a latent variable sequence, and the latent variable sequence is the spatiotemporal propagation pattern of the captured noxious stimulus in the brain; Based on the first neural perturbation field and the target preoperative immune state vector, a generative adversarial network is used to predict the trajectory of the first postoperative immune state vector; wherein, the trajectory of the first postoperative immune state vector consists of postoperative immune state vectors at multiple time steps. Based on the intervention parameters of the aforementioned intervention measures, the first neural perturbation field is adjusted to obtain the second neural perturbation field; The second neural perturbation field and the target preoperative immune state vector are input into the generative adversarial network to obtain the predicted trajectory of the second postoperative immune state vector. Based on the trajectory of the first postoperative immune state vector and the trajectory of the second postoperative immune state vector, the intervention effect of the intervention measures is determined.

[0006] The intraoperative intervention evaluation method based on generative adversarial networks provided in this embodiment has at least the following beneficial effects: First, by introducing a spatiotemporal encoder, the intraoperative EEG signals are encoded into a first neural perturbation field, effectively capturing the spatiotemporal propagation pattern of noxious stimuli in the brain. Then, a generative adversarial network (GAN) is used to predict the postoperative immune status vector trajectory, achieving precise modeling of the dynamic evolution of the postoperative immune status. This captures the mapping relationship between the specific neural imprints left by noxious stimuli in different spatiotemporal patterns during surgery in the central nervous system and the immunological imprints left at the epigenetic level of immune cells. Finally, by predicting two different postoperative immune status vector trajectories (with and without intervention), the effects of different interventions can be simulated and quantitatively compared to assist physicians in optimizing postoperative recovery and long-term prognosis. This embodiment, by deeply mining the spatiotemporal characteristics of intraoperative neural signals, combining them with a GAN to dynamically predict postoperative immune status, and introducing intervention quantum simulation to assess intervention effects, provides a prospective and quantitative approach to evaluating the effectiveness of intraoperative interventions.

[0007] In a second aspect, this application provides an intraoperative intervention assessment device based on generative adversarial networks, the device comprising the following: The data acquisition module is used to acquire the target intraoperative EEG signals, the target preoperative immune status vector, and intervention measures; A neural feature extraction module is used to encode the target intraoperative EEG signal into a first neural perturbation field based on a spatiotemporal encoder; wherein, the neural perturbation field is a latent variable sequence, and the latent variable sequence is the spatiotemporal propagation pattern of the captured noxious stimulus in the brain; An immune feature extraction module is used to predict the trajectory of a first postoperative immune state vector using a generative adversarial network based on the first neural perturbation field and the target preoperative immune state vector; wherein the trajectory of the first postoperative immune state vector consists of postoperative immune state vectors at multiple time steps. The intervention module is used to adjust the first neural perturbation field based on the intervention quantifier of the intervention measures to obtain the second neural perturbation field; The effect prediction module is used to input the second neural perturbation field and the target preoperative immune state vector into the generative adversarial network to obtain the predicted second postoperative immune state vector trajectory, and to determine the intervention effect of the intervention measures based on the first postoperative immune state vector trajectory and the second postoperative immune state vector trajectory.

[0008] A third aspect of this application provides an electronic device including at least one controller and a memory for communicatively connecting to the controller; the memory stores instructions executable by the at least one controller to cause the at least one controller to perform an intraoperative intervention assessment method based on a generative adversarial network as described above.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform an intraoperative intervention assessment method based on a generative adversarial network as described above.

[0010] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of an intraoperative intervention evaluation method based on generative adversarial networks provided in an embodiment of this application; Figure 2 This is a schematic diagram of an intraoperative intervention assessment device based on generative adversarial networks provided in an embodiment of this application; Figure 3 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0014] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0015] like Figure 1 One embodiment of the application provides a method for evaluating intraoperative interventions based on generative adversarial networks, the method comprising: Step S110: Obtain the target intraoperative EEG signal, the target preoperative immune status vector, and the intervention measures.

[0016] Step S120: The intraoperative EEG signal of the target is encoded into a first neural perturbation field based on the spatiotemporal encoder; wherein, the neural perturbation field is a latent variable sequence, and the latent variable sequence is the spatiotemporal propagation pattern of the captured noxious stimulus in the brain.

[0017] Step S130: Based on the first neural perturbation field and the target preoperative immune state vector, a generative adversarial network is used to predict the trajectory of the first postoperative immune state vector; wherein, the trajectory of the first postoperative immune state vector is composed of postoperative immune state vectors at multiple time steps.

[0018] Step S140: Based on the intervention quantifier, adjust the first neural perturbation field to obtain the second neural perturbation field.

[0019] Step S150: Input the second neural perturbation field and the target preoperative immune state vector into the generative adversarial network to obtain the predicted trajectory of the second postoperative immune state vector, and determine the intervention effect of the intervention measures based on the trajectory of the first postoperative immune state vector and the trajectory of the second postoperative immune state vector.

[0020] Due to the noxious stimuli occurring in different spatiotemporal patterns during surgery, specific neural imprints are left in the central nervous system, and immunoimprints are left at the epigenetic level of immune cells (such as macrophages and T cells) through the hypothalamus-pituitary-adrenal axis and the sympathetic nervous system. Therefore, in this embodiment, the mapping relationship between neural imprints and immunoimprints is extracted based on deep learning technology, and then the intervention effect is evaluated based on this mapping relationship to assist doctors in optimizing intervention strategies to improve patient prognosis. It should be noted that this embodiment does not directly apply to surgical treatment or disease diagnosis.

[0021] For ease of understanding, the following explains some key terms in this embodiment: Intraoperative EEG signals refer to the brain electrical activity data of the target patient collected in real time during the operation. Intraoperative EEG signals reflect the neurophysiological state of the brain under the influence of surgical stimulation, anesthetic drugs, etc., and are key information for assessing the intensity of noxious stimuli and neural responses.

[0022] The target preoperative immune status vector refers to a set of indicators related to the patient's immune system collected before surgery. These indicators may include immune cell counts, inflammatory factor levels, etc., and are represented in vector form to characterize the patient's basic immune status before surgery.

[0023] Interventions refer to medical measures that may be taken during surgery, such as the use of analgesics, adjustment of sedation dosage, and fluid management. These measures aim to regulate the patient's physiological response to optimize surgical outcomes and postoperative recovery.

[0024] A spatiotemporal encoder is a neural network structure that transforms input data into a more compact and informative representation, namely latent variables. In this embodiment, the spatiotemporal encoder transforms complex electroencephalogram (EEG) signals into a neural perturbation field. A sequence of latent variables refers to a series of hidden, unobservable variables that are correlated with observable data through mathematical models. In this method, the sequence of latent variables is used to represent the neural perturbation field, thereby capturing complex spatiotemporal dynamic information. The neural perturbation field is a sequence of latent variables designed to capture the spatiotemporal propagation patterns of noxious stimuli in the brain. The perturbation field can abstractly represent the nervous system's response to stimuli, including the intensity, duration, and diffusion of the stimulus across different brain regions. A spatiotemporal propagation pattern refers to the evolution of a phenomenon in the temporal and spatial dimensions. For the neural perturbation field, it describes the dynamic process of how neural activity induced by noxious stimuli propagates between different brain regions and changes over time.

[0025] Noxious stimuli are stimuli that may cause tissue damage or pain, such as surgical incisions, tissue traction, and inflammatory reactions. These stimuli can trigger stress responses and immune responses in the body.

[0026] Generative Adversarial Networks (GANs) are deep learning models consisting of a generator and a discriminator. The generator learns the distribution of real data and generates new samples, while the discriminator is responsible for distinguishing real samples from those generated by the generator.

[0027] Intervention effect refers to the degree and nature of the influence of a specific intervention on the postoperative immune status vector trajectory of a patient. By comparing the postoperative immune status vector trajectories with and without intervention, the effectiveness of the intervention can be quantitatively evaluated.

[0028] In step S110, the target intraoperative EEG signal can be acquired in real time by placing electrodes on the scalp of the target patient, for example, using a standard 10-20 international lead system to record multichannel EEG data. The target preoperative immune status vector can be obtained by performing blood tests on the patient before surgery to obtain immunological indicators such as white blood cell count, C-reactive protein, and cytokine levels, and these indicators can be integrated into a numerical vector. Interventions can be pre-set or input by the clinician, for example, specifying the dosage of a certain analgesic or adjusting the depth of anesthesia.

[0029] In step S120, the neural perturbation field is defined as a latent variable sequence. This latent variable sequence aims to capture the spatiotemporal propagation pattern of noxious stimuli in the brain. For example, a spatiotemporal encoder can consist of a graph convolutional layer and a Transformer, receiving EEG signals as input and compressing them into a low-dimensional latent variable sequence through a series of nonlinear transformations. Each element of the latent variable sequence can represent the brain's response intensity to noxious stimuli at different time points or in different regions. As another implementation, the spatiotemporal encoder can also consist of a recurrent neural network and a Transformer, which can process the temporal series characteristics of EEG signals and extract spatial domain features based on the Transformer, outputting a latent variable sequence reflecting the dynamic changes in neural activity.

[0030] In step S130, based on the first neural perturbation field and the target preoperative immune state vector (as the initial vector) extracted in step S120, a generative adversarial network (GAN) is used to predict the trajectory of the first postoperative immune state vector. The trajectory of the postoperative immune state vector consists of postoperative immune state vectors at multiple time steps. Specifically, the GAN can receive the encoded first neural perturbation field and the target preoperative immune state vector as input. The network is trained to learn the mapping relationship from neural perturbation and preoperative immune state to postoperative immune dynamic changes. For example, the generator part in the GAN can be a sequence generation model, which generates a series of vectors representing the postoperative immune state at different time points based on the input. These vectors together constitute the trajectory of the first postoperative immune state vector, reflecting the evolution of the immune state without additional intervention.

[0031] In step S140, based on the intervention inductor, the first neural perturbation field is adjusted to obtain the second neural perturbation field. The intervention inductor can be a predefined mathematical function or a small neural network model that modifies the first neural perturbation field according to the type and intensity of the intervention. For example, if the intervention is to increase the dosage of analgesic drugs, the intervention inductor can be designed to reduce the value of specific pain-related components in the neural perturbation field. Alternatively, the intervention inductor can be constructed by learning the patterns of the intervention's impact on neural activity from historical data, thereby more accurately simulating the intervention effect. The adjusted neural perturbation field is the second neural perturbation field, which reflects the possible changes in the effect of noxious stimuli on the brain under the intervention.

[0032] In step S150, the second neural perturbation field and the target preoperative immune state vector are input into the generative adversarial network (GAN) to obtain the predicted trajectory of the second postoperative immune state vector. Subsequently, the intervention effect of the intervention is determined based on the first and second postoperative immune state vector trajectories. Specifically, the GAN is used again for prediction, but this time based on the adjusted second neural perturbation field. The resulting second postoperative immune state vector trajectory represents the evolution of the immune state after implementing a specific intervention.

[0033] To determine the effectiveness of the intervention, the two trajectories can be directly compared. For example, the Euclidean distance between the immune status vectors of the two trajectories at each time step can be calculated, or the overall similarity index of the two trajectories can be calculated. Through this comparison, the degree and direction of the change in the postoperative immune status trajectory caused by the intervention can be quantitatively assessed.

[0034] By introducing a spatiotemporal encoder to encode the target intraoperative EEG signal into a first neural perturbation field, the spatiotemporal propagation pattern of noxious stimuli in the brain is effectively captured. This extraction of deep dynamic features of neural signals overcomes the limitations of existing technologies in capturing spatiotemporal changes in intraoperative neural signals. Then, this embodiment utilizes a generative adversarial network to predict the postoperative immune status vector trajectory, achieving precise modeling of the dynamic evolution of the postoperative immune status. Furthermore, by predicting two different postoperative immune status vector trajectories (with and without intervention), the effects of different interventions can be simulated and quantitatively compared. This assists physicians in optimizing postoperative recovery and long-term prognosis. In summary, this embodiment, by deeply mining the spatiotemporal features of intraoperative neural signals, combining generative adversarial networks for dynamic prediction of postoperative immune status, and introducing intervention quantitative simulation of intervention effects, provides a forward-looking and quantitatively innovative approach to evaluating the effectiveness of intraoperative interventions.

[0035] In some embodiments of this application, the generative adversarial network includes a generator and a discriminator; wherein, the training process of the generative adversarial network includes: Step S210: Obtain preoperative immune status vector samples, real postoperative immune status vector trajectory samples, random noise, and the third neural perturbation field extracted by the spatiotemporal encoder based on intraoperative EEG signal samples.

[0036] Step S220: The generator pools the third neural perturbation field to obtain a global neural perturbation representation. Based on the global neural perturbation representation, random noise, and preoperative immune state vector samples, an initial hidden state is generated. Based on the initial hidden state, the trajectory of the third postoperative immune state vector is predicted.

[0037] Step S230: The discriminator extracts the trajectory code of the postoperative immune status vector trajectory to be discriminated, splices the trajectory code, global neural perturbation representation and preoperative immune status vector sample to generate a spliced ​​vector, and determines the probability that the postoperative immune status vector trajectory to be discriminated is the true postoperative immune status vector trajectory based on the spliced ​​vector; the postoperative immune status vector trajectory to be discriminated is the third postoperative immune status vector trajectory or the true postoperative immune status vector trajectory.

[0038] Step S240: Based on the loss function, train the generator and discriminator alternately until the generative adversarial network training is complete.

[0039] This embodiment mainly focuses on the training process of a generative adversarial network composed of a generator and a discriminator. It should also be noted that preoperative immune state vector samples, actual postoperative immune state vector trajectory samples, and intraoperative EEG signal samples refer to the data samples used for network training.

[0040] In this embodiment, the generator is responsible for predicting the trajectory of the postoperative immune state vector based on the input neural perturbation field, the preoperative immune state vector, and random noise. The generator can be implemented using various neural network architectures. For example, it can be a sequence generation model based on a recurrent neural network (RNN), such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU), which can process time series data and generate new sequences. Alternatively, it can be a decoder based on a Transformer architecture, which captures long-distance dependencies in the input sequence through a self-attention mechanism, thereby generating a high-quality trajectory.

[0041] The discriminator receives the trajectory of the postoperative immune status vector to be judged, along with relevant contextual information (such as global neural perturbation representation and preoperative immune status vector samples), and outputs the probability that the trajectory is the true trajectory. The discriminator can be composed of a convolutional neural network (CNN) or a multilayer perceptron (MLP). For example, it can be a one-dimensional convolutional network used to extract features from the trajectory and then input these features into a fully connected layer for classification; or it can be a network based on an attention mechanism that can focus on the part of the trajectory that has a greater impact on the discrimination result.

[0042] Preoperative immune state vector samples, real postoperative immune state vector trajectory samples, random noise, and neural perturbation field samples are the data inputs required for training the generative adversarial network. Among them, the preoperative immune state vector samples and real postoperative immune state vector trajectories are derived from actual clinical data to provide real-world immune state change patterns; random noise introduces diversity into the generator to prevent pattern collapse; and neural perturbation field samples are extracted by the spatiotemporal encoder from intraoperative EEG signal samples, representing the brain's response to noxious stimuli.

[0043] Pooling samples from the neural perturbation field to aggregate them into a global neural perturbation representation aims to compress complex spatiotemporal neural perturbation information into a concise, fixed-dimensional vector representation, facilitating its use as input to the generator and discriminator. Pooling operations can be implemented using max pooling, average pooling, or more complex attention mechanisms. For example, an attention layer can be used to weight and sum the pooled features based on the importance of different time steps and spatial locations within the neural perturbation field to obtain the global representation; alternatively, a fully connected layer can directly map the pooled features to the global representation.

[0044] The initial hidden state is generated by concatenating global neural perturbation representations, random noise, and preoperative immune state vector samples. This integrates all relevant input information, providing the generator with a comprehensive starting point so that it can generate the postoperative immune state vector trajectory based on this information. The concatenation operation typically connects these vectors in a dimensional way to form a longer vector. Then, the third postoperative immune state vector trajectory is predicted based on the initial hidden state. The generator, based on its learned data distribution, starts from this initial hidden state and gradually generates a series of postoperative immune state vectors at different time steps, thus forming a complete trajectory.

[0045] The discriminator extracts the trajectory encoding of the postoperative immune status vector trajectory to be discriminated, aiming to convert the variable-length trajectory sequence into a fixed-length vector representation so that the discriminator can process it. This can be achieved by a sequence encoder, such as a bidirectional LSTM network or a Transformer encoder, which can capture the temporal features in the trajectory and compress them into meaningful codes.

[0046] By concatenating trajectory encoding, global neural perturbation representation, and preoperative immune status vector samples to generate a concatenated vector, all the contextual information required by the discriminator is integrated together so that it can comprehensively judge the authenticity of the trajectory.

[0047] The probability that the trajectory of the postoperative immune state vector to be determined is the true postoperative immune state vector trajectory, based on the concatenated vector, is the final output of the discriminator. This is typically a value between 0 and 1, representing the likelihood that the discriminator considers the trajectory to be the true trajectory. This can be achieved using a fully connected layer with a sigmoid activation function. The postoperative immune state vector trajectory to be determined can be either a third postoperative immune state vector trajectory generated by the generator or a true postoperative immune state vector trajectory; the discriminator needs to distinguish between these two.

[0048] Based on the loss function, the generator and discriminator are trained alternately until the generative adversarial network is trained. By continuously optimizing the parameters of the generator and discriminator, the data generated by the generator becomes more and more realistic, and the discriminator becomes more and more capable of distinguishing between real and generated data, eventually reaching a dynamic equilibrium.

[0049] This embodiment ensures that the generative adversarial network (GAN) has extremely high accuracy and realism in predicting the postoperative immune status vector trajectory. This adversarially trained GAN can capture the complex nonlinear relationship between neural perturbations, preoperative immune status, and postoperative immune trajectory, thereby generating a predicted trajectory that is highly consistent with the actual physiological process. This makes the prediction of both the first and second postoperative immune status vector trajectories more reliable and precise when evaluating intraoperative interventions. Ultimately, the intervention effect determined based on these high-precision predicted trajectories will more accurately reflect the actual physiological response, providing clinicians with more instructive decision-making basis, thereby optimizing the patient's intraoperative management and postoperative recovery.

[0050] In some embodiments of this application, the trajectory encoding of the postoperative immune status vector trajectory to be determined in step S230 includes: Step S2310: Extract trajectory codes from the postoperative immune status vector trajectory to be determined based on a bidirectional LSTM network; Step S230, determining the probability that the trajectory of the postoperative immune status vector to be identified is the true postoperative immune status vector trajectory based on the spliced ​​vector, includes: Step S2320: Input the spliced ​​vector into the multilayer perceptron to obtain the probability that the output postoperative immune status vector trajectory to be judged is the true postoperative immune status vector trajectory.

[0051] Among them, bidirectional LSTM can capture the contextual information of any time step in the sequence, thereby gaining a more comprehensive understanding of the dynamic changes and long-distance dependencies of the sequence. In addition to bidirectional LSTM networks, Transformer encoders or GRU networks can also be used to extract trajectory codes. These networks are also good at processing sequence data and capturing its inherent temporal patterns.

[0052] Multilayer perceptrons can learn and recognize complex patterns in data. They are used to map spliced ​​vectors to a probability value to determine the likelihood that the trajectory of the postoperative immune status vector to be identified is the true postoperative immune status vector trajectory. In addition to multilayer perceptrons, classifiers such as logistic regression models or support vector machines (SVMs) can also be used to make probability judgments. However, multilayer perceptrons have a stronger nonlinear fitting ability and can improve the accuracy of probability prediction.

[0053] This embodiment employs a bidirectional LSTM network to extract trajectory codes from the postoperative immune state vector trajectory to be determined. This fully captures the complex temporal features and long-range dependencies of the trajectory, making the trajectory codes more representative and discriminative. Furthermore, this refined trajectory code is concatenated with global neural perturbation representations and preoperative immune state vector samples, and then input into a multilayer perceptron for probabilistic judgment. The multilayer perceptron, with its powerful nonlinear mapping capabilities, can perform in-depth analysis of the comprehensive information, significantly improving the discriminator's ability to distinguish between real and generated trajectories. This improved discriminative mechanism provides the generator with more accurate and effective training signals, helping the generative adversarial network to converge more stably and efficiently, ultimately generating postoperative immune state vector trajectories that more closely resemble the actual physiological process. This provides more reliable and refined prediction results for the evaluation of intraoperative interventions.

[0054] In some embodiments of this application, the loss function includes: a discriminator loss function and a generator loss function; The discriminator loss function includes the expected output of the real samples, the expected output of the generated samples, and a gradient penalty term; The generator loss function includes adversarial loss and reconstruction loss.

[0055] A loss function is a mathematical metric used to quantify the difference between a model’s predicted output and the actual target value.

[0056] The expected output of a real sample refers to the ideal output value of the discriminator when it receives a real data sample. Usually, this expected value is set to a high value, such as 1, to indicate that the sample is real.

[0057] The expected output of the generated sample refers to the ideal output value of the discriminator when it receives a fake data sample generated by the generator. This expected value is set to a low value, such as 0, to indicate that the sample is fake.

[0058] Gradient penalty is a regularization technique primarily used to stabilize the training process of generative adversarial networks, especially in Wasserstein GANs with Gradient Penalty (WGAN-GP). It forces the discriminator to satisfy the Lipschitz continuity condition by penalizing the gradient norm of the discriminator output relative to its input. This helps prevent gradient vanishing or gradient exploding problems, thereby improving the stability of training and the quality of generated samples.

[0059] Adversarial loss is a core component of the generator loss function, quantifying the generator's ability to deceive the discriminator. The generator learns to generate samples that are difficult to distinguish from real data by minimizing adversarial loss. Specifically, the generator attempts to make the discriminator output high probability values ​​for its generated samples, thus leading the discriminator to believe these samples are real. Adversarial loss can be calculated as the negative log-likelihood of the discriminator's output on the generated samples, or, in WGAN, directly using the discriminator's output values ​​on the generated samples. Reconstruction loss is an additional component of the generator loss function, used to measure the similarity between the generator's output and a reference input or target output. In some generation tasks, the generator not only needs to generate realistic samples but also needs to ensure that these samples are consistent with the input conditions or original data in specific aspects. Reconstruction loss guides the generator to learn this consistency by calculating the difference between the generated samples and the reference samples (e.g., using mean squared error, mean absolute error, or L1 / L2 loss), thereby generating samples that are both realistic and possess specific attributes.

[0060] In some embodiments of this application, step S120, which encodes the target intraoperative EEG signal into a first neural perturbation field based on a spatiotemporal encoder, includes: Step S310: Extract spatial features from the target intraoperative EEG signal based on graph convolutional network; Step S320: Based on the spatial features, the first neural perturbation field is encoded using Transformer.

[0061] Among them, graph convolutional networks are a type of neural network model that processes graph-structured data. They can effectively learn representations from node features and graph structures.

[0062] When processing EEG signals, electrode channels can be viewed as nodes in a graph, and the functional connections or spatial proximity relationships between electrode channels can be considered as edges. Graph convolutional networks update node representations by aggregating information from neighboring nodes, thereby capturing complex spatial dependencies and local structural information in EEG signals. Implementation methods can include, but are not limited to: using spectral domain graph convolutional networks, performing convolution operations in the spectral domain through graph Fourier transform; or using spatial domain graph convolutional networks, directly defining convolution operations in the spatial domain of the graph, for example, by aggregating features from neighboring nodes.

[0063] Transformer is a deep learning model based on self-attention mechanisms. Its core strength lies in its ability to capture long-distance dependencies between arbitrary positions in a sequence and its advantage in parallel computation. When processing spatial feature sequences extracted from EEG signals, Transformer can effectively model the dynamic evolution of these spatial features in the temporal dimension, thereby integrating and encoding spatiotemporal information into a neural perturbation field. Its implementation can include, but is not limited to: using a standard Transformer encoder containing multi-head self-attention layers and feedforward neural network layers to process the sequence input and output the encoded sequence; or using a temporal Transformer specifically designed for processing time-series data, capturing temporal dependencies through self-attention mechanisms.

[0064] This embodiment decomposes the encoding process of intraoperative EEG signals into two stages: spatial feature extraction and temporal encoding, thereby capturing the complex spatiotemporal dynamics of EEG signals more precisely. Specifically, firstly, the graph structure characteristics of EEG signals are processed using graph convolutional networks to extract representative spatial features from the connections and activities between electrode channels. This approach can effectively handle the high dimensionality and nonlinearity of EEG signals, ensuring accurate capture of spatial information. Subsequently, these extracted spatial features are input as sequences into the Transformer model. With its powerful self-attention mechanism, the Transformer can capture the long-distance dependencies and dynamic evolution patterns of spatial features in the temporal dimension, thereby integrating and encoding this spatiotemporal information into a first neural perturbation field. This phased and collaborative processing mechanism enables the encoder to more comprehensively and accurately represent the spatiotemporal propagation patterns of noxious stimuli in the brain, providing high-quality input for subsequent generative adversarial networks to predict the postoperative immune status vector trajectory, thus improving the accuracy and reliability of intervention assessment.

[0065] In some embodiments of this application, step S310, which involves extracting spatial features from the target intraoperative EEG signal using a graph convolutional network, includes: Step S3110: Divide the target intraoperative EEG signal into an outflow window signal sequence according to the time window. Step S3120: Extract the phase lock value and directed transfer function of each window signal between each electrode channel pair; Step S3130: Each window signal is used as a graph signal on the graph structure, the electrode channel is used as a graph node of the graph structure, and the weighted matrix of the phase lock value and the directed transfer function is used as the edge of the graph structure. Step S3140: Extract spatial features from the graph structure using a graph convolutional network.

[0066] First, the continuous intraoperative EEG signal is segmented into out-of-window signal sequences according to time windows. This aims to discretize the long-duration EEG signal into a series of processable short signal segments, allowing for independent evaluation and feature extraction within each time window. This helps capture the local dynamic characteristics of the EEG signal. For example, a fixed-length sliding window can be used for segmentation, or adaptive segmentation can be performed based on specific events of the signal. Next, the phase-lock value and directed transfer function (DJF) of each window signal are extracted between electrode channel pairs. The phase-lock value is an indicator of the degree of phase synchronization between two signals, reflecting the strength of synchronous oscillations between neuronal populations and characterizing functional connectivity. The DJF is a frequency domain analysis method based on Granger causality, used to assess the causal influence of one signal on another, thus revealing the direction of information transmission in neural activity. By extracting these indicators, a dynamic network structure reflecting brain functional connectivity and information flow can be constructed. In addition to phase-lock values ​​and DJF, indicators such as coherence and mutual information can be used to measure functional connectivity, or methods such as partial Granger causality can be used to assess directed information flow.

[0067] Furthermore, each window signal is treated as a graph signal on a graph structure, the electrode channels as graph nodes, and the weighted matrix of phase-locked values ​​and directed transfer functions as edges of the graph structure. This step aims to transform the complex relationships of EEG signals into graph-structured data for processing by graph convolutional networks. Each electrode channel represents a specific region or measurement point in the brain and is thus mapped as a node in the graph. The amplitude of each window signal or the energy of a specific frequency band can serve as a feature on the graph node, i.e., a graph signal. The phase-locked values ​​and directed transfer functions extracted in the previous step quantify the functional connectivity strength and directionality between different electrode channels. These values ​​can be used to construct the adjacency matrix of the graph, serving as edges, and these edges are weighted to reflect the strength of the connections. This graph structure can effectively represent the topological structure and dynamic characteristics of the brain's functional networks.

[0068] Finally, spatial features are extracted from the graph structure using a graph convolutional network (GCN). This step leverages the powerful capabilities of GCNs to learn and extract representative spatial features from the constructed graph structure. A GCN is a deep learning model specifically designed for graph-structured data. It updates the feature representations of nodes by aggregating information from their neighbors, thereby capturing local and global structural information in the graph. By applying a GCN to the constructed graph structure, spatial patterns related to neural perturbations can be effectively learned from the functional connections and information flow between electrode channels. Various variants of the GCN can be employed, such as spectral domain GCNs or spatial domain GCNs, to adapt to different graph structures and feature extraction requirements.

[0069] This embodiment segments the intraoperative EEG signal into time windows and extracts the phase-locking value and directed transfer function between each electrode channel pair to construct a dynamic graph structure. This allows the graph convolutional network to extract spatial features from more biologically meaningful functional connections and information flows. This method can capture the dynamic changes and information transmission direction of the brain's functional network more precisely, overcoming the limitations that may exist when directly inputting the raw EEG signal into the graph convolutional network. Therefore, the extracted spatial features can more accurately and comprehensively reflect the spatiotemporal propagation pattern of neural perturbations in the brain, significantly improving the accuracy and robustness of neural perturbation field encoding.

[0070] In some embodiments of this application, step S150, determining the intervention effect of the intervention based on the first postoperative immune status vector trajectory and the second postoperative immune status vector trajectory, includes: Step S510: Calculate the change index based on the first postoperative immune status vector trajectory and the second postoperative immune status vector trajectory. Step S520: Determine the intervention effect of the intervention measures based on the change indicators; The trajectory change includes at least one or a weighted combination of any two of the following: peak value change, area under the curve change, peak time change, and steady-state recovery time change.

[0071] The calculation of trajectory changes refers to quantifying the differences between two postoperative immune status vector trajectories in a specific dimension. This can be a measure of the overall morphological differences of the trajectories or a measure of the differences in key feature points. For example, the overall difference can be represented by calculating the sum of the Euclidean distances between the two trajectories at each time step, or by using the Dynamic Time Warping (DTW) algorithm to measure the nonlinear alignment differences that may exist between the two trajectories in the time dimension. Determining the intervention effect refers to judging and evaluating the effectiveness, strength, or direction of the intervention based on the calculated trajectory changes. This can be a simple threshold judgment, for example, if the trajectory change exceeds a certain preset threshold, the intervention is considered effective; or it can be a more complex decision model, such as inputting the trajectory change into a classifier to output the level of intervention effect.

[0072] Trajectory changes can include peak changes, which refer to the difference between the maximum values ​​of the two postoperative immune status vector trajectories. For example, the absolute difference or percentage change between the peak value of the first postoperative immune status vector trajectory and the peak value of the second postoperative immune status vector trajectory can be calculated.

[0073] Trajectory changes can also include changes in the area under the curve, which refers to the difference between the area enclosed by the two postoperative immune status vector trajectories and the time axis. For example, the two trajectories can be integrated separately, and then the difference between the two integral values ​​can be calculated.

[0074] Trajectory changes can also include peak time changes, which refer to the difference between the time points required for two postoperative immune status vector trajectories to reach their respective peak values. For example, the time points at which the peak values ​​of the two trajectories appear can be identified, and then the difference between these two time points can be calculated.

[0075] Trajectory changes can also include steady-state recovery time changes, which refer to the difference between the time required for two postoperative immune state vector trajectories to recover to or approach a certain steady-state level after being disturbed. For example, a recovery threshold can be set, and then the time required for each of the two trajectories to reach the threshold for the first time can be calculated, and the difference between the two time points can be calculated.

[0076] In addition, trajectory change can also be a weighted combination of any two or more of the above indicators. For example, by assigning different weights to different change indicators and then combining them linearly, a comprehensive trajectory change score can be formed.

[0077] This embodiment quantifies the specific differences between the first and second postoperative immune status vector trajectories, such as changes in peak value, area under the curve, time to peak, and steady-state recovery time. This quantitative assessment method makes the judgment of intervention effect clearer, provides clinicians with more operational feedback, optimizes intervention strategies, and improves the quality of postoperative recovery for patients. At the same time, by comprehensively considering multiple indicators, the impact of intervention measures on the dynamic process of immune status can be fully reflected, which helps to deepen the understanding of the intervention mechanism and provides data support for the development of personalized medical plans.

[0078] To facilitate understanding by those skilled in the art, one embodiment of this application provides a method for evaluating intraoperative interventions based on generative adversarial networks. This method includes: Step S910: Acquire data and preprocess the data (training phase); Get initial data; make For each surgical patient, the group is assembled. ,Record: (1) Intraoperative EEG signal samples: Among them, intraoperative EEG signals This represents the number of electrode channels (e.g., 256). This represents the total number of sampling time points (e.g., the entire surgical procedure, sampled at 250Hz). (2) Preoperative immune status vector sample: The preoperative immune status vector is obtained by high-dimensional immune profiling (e.g., the proportion of cell subpopulations and the expression intensity of surface markers obtained by mass flow cytometry). (3) True postoperative immune status vector sequence: ,in Postoperative day Postoperative immune status vector (including various immune cell counts, cytokine concentrations, etc.) at each time point (e.g., every hour). This refers to the duration of postoperative monitoring.

[0079] Data preprocessing; (1) Apply bandpass filtering (0.5–45 Hz) and independent component analysis.

[0080] Will The signal is divided into segments by sliding along a fixed-length window (e.g., 10 seconds) to obtain the window signal sequence. Adjacent windows overlap by 50%. Each window is considered a time step. (2) Spline interpolation was used to obtain a window signal sequence with the same temporal resolution as the EEG window (i.e., one immune status point every 10 seconds). As a monitoring signal; (3) Perform z-score normalization on each feature dimension.

[0081] Step S920, Spatiotemporal encoder training; The goal of a spacetime encoder is to: This is mapped to a low-dimensional sequence of latent variables (i.e., a neural perturbation field). Latent variable sequences are used to capture the spatiotemporal propagation patterns of noxious stimuli in the brain.

[0082] Step S9210, for each time window Calculate the phase lock value (represented by the PLV matrix) and the directed transfer function (represented by the DTF matrix) between each electrode channel pair. make for Electrode channels in The window signal. Here, the analytic signal is obtained through Hilbert transform, and the instantaneous phase is extracted. Then PLV matrix Defined as: ; in, This represents the number of time points within the window.

[0083] DTF matrix The causal flow is reflected by calculation using a multivariate autoregressive model.

[0084] Step S9220, will Each window signal in the graph structure is considered as defined in the graph structure. The above graph signals and graph structures The graph nodes are electrode channels, graph structure The edges are passed through the matrix and Weighted composition; then graph convolution is used to extract spatial features: ; in (Add self-loop) For degree matrix, For learnable weights, This is the ReLU activation function.

[0085] Step S9230, output the graph convolution. Inputting the data along the time axis into the Transformer model yields the neural perturbation field output by the model. First, global average pooling is performed on the graph node features at each time step to obtain a vector sequence. , ( (Number of output channels for graph convolution).

[0086] Then, add position encoding. , to obtain the input sequence .

[0087] Transformer encoder is composed of Composed of multi-head self-attention and feedforward networks: ; in, .

[0088] Finally, the output is The vector at each time step is regarded as the neural perturbation field at that moment. Therefore, the entire surgical procedure was encoded as a sequence of latent variables. The overall network structure can be represented as follows: ; in, These are encoder parameters.

[0089] Step S930: Generate adversarial network training; The purpose of training a generative adversarial network is to use neural perturbation fields. and preoperative immune status vector samples (As an initial vector) as a condition, generate the postoperative immune status vector trajectory. This makes the generated trajectory Vector trajectory of actual postoperative immune status Difficult to distinguish.

[0090] Step S9310: The generator uses a conditional temporal generation network, takes conditional information as input, and outputs an immune trajectory. First, the neural perturbation field Global neural perturbation representations are obtained by aggregating data through a temporal pooling layer (such as average pooling or attention pooling). ; Then , and random noise (Introducing randomness to simulate individual differences) The concatenation is used as the initial hidden state. ; Finally, GRU units are used to generate the immune status at each time step step by step: ; in, This is the initial state. A generator can be written as: ; in, These are the model parameters for the generator.

[0091] Step S9320, the discriminator distinguishes the true trajectory. and generated trajectory Simultaneously, the neural perturbation field and preoperative immune state vector samples are used for discrimination; a conditional temporal discriminator, such as a bidirectional LSTM-based encoder, is employed, and finally the discrimination probability is output through a fully connected layer.

[0092] First, the vector trajectory of the postoperative immune status to be determined is... Input a bidirectional LSTM encoder to obtain the hidden state at each time step, and take the concatenated vector at the last time step as the trajectory code. .

[0093] Then , , The concatenation process inputs a multilayer perceptron (MLP) and outputs a scalar. This indicates the probability that the input is the true trajectory. ; Step S9330, design of the loss function; WGAN-GP is used to improve training stability.

[0094] Discriminator loss: ; in, It is a linear interpolation of the real and generated trajectories. For the expectation of the real sample, To generate the expected value of the samples, It is a gradient penalty term. For weights.

[0095] Generator loss (adversarial loss + reconstruction loss): ; in, To reconstruct the loss (L1 or L2 loss), the generated trajectory should approximate the true trajectory. For balance coefficient, To combat the losses.

[0096] Step S9340: Alternately optimize the discriminator and generator until convergence.

[0097] For training the spatiotemporal encoder and the generative adversarial network, a joint training strategy can be adopted: First, a pre-trained spatiotemporal encoder (self-supervised or combined with downstream tasks), for example, can be trained using contrastive learning (such as SimCLR) to enable the spatiotemporal encoder to extract meaningful perturbation representations.

[0098] Secondly, a fixed spatiotemporal encoder is used to train the generator and discriminator.

[0099] Finally (optional): End-to-end fine-tuning allows for minor updates to generator parameters to optimize generation quality.

[0100] Step S940, Model Inference; First, for the target patient, obtain their target intraoperative EEG signal and target preoperative immune status vector; Next, the target intraoperative EEG signal is input into the spatiotemporal encoder to obtain the first neural perturbation field output. Next, the first neural perturbation field and the target preoperative immune state vector are input into the generator, and the generator outputs the trajectory of the first postoperative immune state vector.

[0101] Step S950: By modifying the input first neural perturbation field, a second postoperative immune status vector trajectory is generated to evaluate the effects of different intervention measures. Assuming the first neural perturbation field potential ; Define the budget ,right After modification, the second neural perturbation field is obtained. ,For example: Erasing contributions from specific brain regions: certain latent variable dimensions Activity corresponding to specific brain regions (such as the insula).

[0102] The intervention operations corresponding to the budget can set these dimensions to zero: ; Reduce stimulus intensity: Scale all perturbation field vectors proportionally. ,in, .

[0103] Time truncation: Only retain the first part A perturbation at each time step, followed by a reset to zero.

[0104] The revised and Input generator and fix noise (Or, by averaging multiple samples), the postoperative immune status vector is obtained: ; Composition of multiple postoperative immune status vectors; Second postoperative immune status vector trajectory One row corresponds to the immune status vector at a certain point after surgery (including the counts of various immune cells, cytokine concentrations, etc.). By plotting a curve and overlaying it with the original first postoperative immune status vector trajectory, the deviation of the immune evolution path caused by the intervention can be visually displayed (e.g., a decrease in inflammation peak, a shortened recovery period, etc.).

[0105] Step S960: Quantification of intervention effect indicators; By comparing the second postoperative immune status vector trajectory with the first postoperative immune status vector trajectory, a series of clinically relevant scalar indicators (such as peak change, area under the curve change, steady-state recovery time change, and steady-state recovery time change, or a weighted combination of multiple indicators) are calculated to objectively evaluate the intervention effect.

[0106] With multiple candidate interventions, corresponding second postoperative immune status vector trajectories are generated in batches. The candidate interventions are then ranked based on preset clinical goals (e.g., minimizing IL-6 peak, minimizing AUC, and fastest recovery to steady state). Each intervention includes an intervention identifier, a corresponding quantitative indicator value, and the degree of improvement relative to no intervention (first postoperative immune status vector trajectory). See the table below: Table 1

[0107] like Figure 2 One embodiment of this application provides an intraoperative intervention assessment device based on generative adversarial networks, the device comprising the following: The data acquisition module 1001 is used to acquire the target intraoperative EEG signal, the target preoperative immune status vector, and intervention measures; The neural feature extraction module 1002 is used to encode the target intraoperative EEG signal into a first neural perturbation field based on a spatiotemporal encoder; wherein, the neural perturbation field is a latent variable sequence, and the latent variable sequence is the spatiotemporal propagation pattern of the captured noxious stimulus in the brain; The immune feature extraction module 1003 is used to predict the trajectory of the first postoperative immune state vector using a generative adversarial network based on the first neural perturbation field and the target preoperative immune state vector; wherein, the trajectory of the postoperative immune state vector is composed of postoperative immune state vectors at multiple time steps. The intervention module 1004 is used to adjust the first neural perturbation field based on the intervention measures to obtain the second neural perturbation field; The effect prediction module 1005 is used to input the second neural perturbation field and the target preoperative immune state vector into the generative adversarial network to obtain the predicted trajectory of the second postoperative immune state vector, and to determine the intervention effect of the intervention measures based on the trajectory of the first postoperative immune state vector and the trajectory of the second postoperative immune state vector.

[0108] It should be noted that the intraoperative intervention assessment device based on generative adversarial networks provided in this embodiment is based on the same inventive concept as the intraoperative intervention assessment method based on generative adversarial networks described above. Therefore, the content of the intraoperative intervention assessment method based on generative adversarial networks described above is also applicable to the content of the intraoperative intervention assessment device based on generative adversarial networks in this embodiment, and will not be repeated here.

[0109] like Figure 3 One embodiment of this application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method for evaluating intraoperative interventions based on generative adversarial networks. The electronic device includes: At least one battery; At least one memory; At least one processor; At least one program; The program is stored in memory, and the processor executes at least one program to implement the above-described method for evaluating intraoperative interventions based on generative adversarial networks.

[0110] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0111] The electronic devices according to embodiments of this application will now be described in detail.

[0112] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure. The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to execute the intraoperative intervention assessment method based on generative adversarial networks of this disclosure.

[0113] The input / output interface 1800 is used to implement information input and output. The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900); The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0114] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the above-described method for evaluating intraoperative interventions based on generative adversarial networks.

[0115] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0116] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0117] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0118] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0119] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0120] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0121] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0122] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0123] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0126] The above is a detailed description of the preferred embodiments of this application. However, the embodiments of this application are not limited to the above-described implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the embodiments of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of the embodiments of this application.

Claims

1. A method for evaluating intraoperative interventions based on generative adversarial networks, characterized in that, The method includes: Acquire intraoperative EEG signals of the target, preoperative immune status vector of the target, and intervention measures; The target intraoperative EEG signal is encoded into a first neural perturbation field based on a spatiotemporal encoder; wherein, the neural perturbation field is a latent variable sequence, and the latent variable sequence is the spatiotemporal propagation pattern of the captured noxious stimulus in the brain; Based on the first neural perturbation field and the target preoperative immune state vector, a generative adversarial network is used to predict the trajectory of the first postoperative immune state vector; wherein, the trajectory of the first postoperative immune state vector consists of postoperative immune state vectors at multiple time steps. Based on the intervention parameters of the aforementioned intervention measures, the first neural perturbation field is adjusted to obtain the second neural perturbation field; The second neural perturbation field and the target preoperative immune state vector are input into the generative adversarial network to obtain the predicted trajectory of the second postoperative immune state vector. Based on the trajectory of the first postoperative immune state vector and the trajectory of the second postoperative immune state vector, the intervention effect of the intervention measures is determined.

2. The method for evaluating intraoperative interventions based on generative adversarial networks according to claim 1, characterized in that, The generative adversarial network includes a generator and a discriminator; wherein, the training process of the generative adversarial network includes: Acquire preoperative immune status vector samples, actual postoperative immune status vector trajectories, random noise, and the third neural perturbation field extracted by the spatiotemporal encoder based on intraoperative EEG signal samples; The generator is used to pool the third neural perturbation field and aggregate it to obtain a global neural perturbation representation. The global neural perturbation representation, the random noise, and the preoperative immune state vector sample are concatenated to generate an initial hidden state. The trajectory of the third postoperative immune state vector is predicted based on the initial hidden state. The discriminator extracts the trajectory encoding of the postoperative immune status vector trajectory to be determined, concatenates the trajectory encoding, the global neural perturbation representation, and the preoperative immune status vector sample to generate a concatenated vector, and determines the probability that the postoperative immune status vector trajectory to be determined is the true postoperative immune status vector trajectory based on the concatenated vector; the postoperative immune status vector trajectory to be determined is either the third postoperative immune status vector trajectory or the true postoperative immune status vector trajectory; Based on the loss function, the generator and the discriminator are trained alternately until the generative adversarial network is trained.

3. The method for evaluating intraoperative interventions based on generative adversarial networks according to claim 2, characterized in that, The trajectory encoding of the extracted postoperative immune status vector trajectory to be determined includes: Trajectory codes are extracted from the postoperative immune status vector trajectory to be determined based on a bidirectional LSTM network; The step of determining the probability that the trajectory of the postoperative immune status vector to be determined is the true postoperative immune status vector trajectory based on the spliced ​​vector includes: The spliced ​​vector is input into a multilayer perceptron to obtain the probability that the output trajectory of the postoperative immune state vector to be determined is the trajectory of the true postoperative immune state vector.

4. The method for evaluating intraoperative interventions based on generative adversarial networks according to claim 2, characterized in that, The loss function includes a discriminator loss function and a generator loss function; wherein, the discriminator loss function includes a gradient penalty term.

5. The method for evaluating intraoperative interventions based on generative adversarial networks according to claim 1, characterized in that, The process of encoding the target intraoperative EEG signal into a first neural perturbation field based on a spatiotemporal encoder includes: Spatial features were extracted from the intraoperative EEG signals of the target based on graph convolutional networks; The spatial features are encoded into a first neural perturbation field based on the Transformer.

6. The method for evaluating intraoperative interventions based on generative adversarial networks according to claim 5, characterized in that, The extraction of spatial features from the target intraoperative EEG signal using a graph convolutional network includes: The target intraoperative EEG signal is divided into window signal sequences according to a fixed time window; Extract the phase lock value and directed transfer function of each window signal in the window signal sequence between each electrode channel pair; Each window signal is used as a graph signal on the graph structure, the electrode channel is used as a graph node of the graph structure, the weighted matrix of the phase lock value and the directed transfer function is used as the edge of the graph structure, and spatial features are extracted from the graph structure according to the graph convolutional network.

7. The method for evaluating intraoperative interventions based on generative adversarial networks according to claim 1, characterized in that, The step of determining the intervention effect of the intervention measures based on the first postoperative immune status vector trajectory and the second postoperative immune status vector trajectory includes: Calculate the change index between the first postoperative immune status vector trajectory and the second postoperative immune status vector trajectory; The intervention effect of the intervention measures is determined based on the change indicators; wherein the change indicators include at least one or a weighted combination of any two of the following: peak value change, area under the curve change, time to peak value change, and steady-state recovery time change.

8. A device for evaluating intraoperative interventions based on generative adversarial networks, characterized in that, The device includes the following: The data acquisition module is used to acquire the target intraoperative EEG signals, the target preoperative immune status vector, and intervention measures; The neural feature extraction module is used to encode the target intraoperative EEG signal into a first neural perturbation field based on a spatiotemporal encoder; wherein, the neural perturbation field is a latent variable sequence, and the latent variable sequence is the spatiotemporal propagation pattern of the captured noxious stimulus in the brain; An immune feature extraction module is used to predict the trajectory of a first postoperative immune state vector using a generative adversarial network based on the first neural perturbation field and the target preoperative immune state vector; wherein the trajectory of the first postoperative immune state vector consists of postoperative immune state vectors at multiple time steps. The intervention module is used to adjust the first neural perturbation field based on the intervention quantifier of the intervention measures to obtain the second neural perturbation field; The effect prediction module is used to input the second neural perturbation field and the target preoperative immune state vector into the generative adversarial network to obtain the predicted second postoperative immune state vector trajectory, and to determine the intervention effect of the intervention measures based on the first postoperative immune state vector trajectory and the second postoperative immune state vector trajectory.

9. An electronic device, characterized in that, It includes at least one controller and a memory for communicatively connecting with the controller; the memory stores instructions that can be executed by the at least one controller to cause the at least one controller to perform a method for evaluating intraoperative interventions based on generative adversarial networks as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions for causing a computer to perform a method for evaluating intraoperative interventions based on generative adversarial networks as described in any one of claims 1 to 7.

Images