Simulated medical information detection method and system based on AI algorithm

Abstract

The present invention provides a simulated medical information detection method and system based on an AI algorithm, which relates to the technical field of medical information detection, including obtaining simulated medical information to be detected, performing multimodal feature extraction on the simulated medical information, and mapping the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient; calculating the abnormal score of each feature in the comprehensive health feature portrait, identifying abnormal medical information fragments, and modeling the patient's historical medical information at the same time, building an abnormal evolution graph model in combination with medical prior knowledge, and predicting the patient's abnormal risk probability in a preset future time domain; automatically generating an abnormal warning report based on the obtained abnormal medical information fragments and abnormal risk probability, and continuously optimizing the detection strategy using a reinforcement learning algorithm, and using the optimized detection strategy for subsequent simulated medical information detection.

Description

Technical Field

[0001] The present invention relates to medical information detection technology, and in particular to a simulated medical information detection method and system based on AI algorithm. Background Art

[0002] With the rapid development and popularization of medical information technology, the amount of medical data generated and accumulated by medical institutions has exploded. However, in the massive amount of medical information, there are often various anomalies and errors, such as missing, inconsistent, and outliers in medical data. These abnormal information will have a negative impact on medical decision-making and the diagnosis and treatment process, and even endanger the patient's life safety. Traditional medical information anomaly detection methods mainly rely on manual review and simple statistical rules, which have problems such as low efficiency, low accuracy, and difficulty in dealing with complex medical scenarios. It is difficult to meet the growing demand for medical information quality management.

[0003] In recent years, the rapid development of artificial intelligence technology has provided new ideas and methods for solving the problem of medical information anomaly detection. Anomaly detection algorithms based on machine learning and deep learning can efficiently and accurately identify various anomalies and errors in medical information by automatically learning the normal and abnormal patterns of medical data. However, the existing AI-based medical information anomaly detection methods still have some limitations, such as the lack of full utilization of medical field knowledge, and the difficulty in effectively capturing the inherent associations and semantic information of medical data; the generalization ability of anomaly detection models is insufficient, and it is difficult to adapt to the changing distribution of medical data and anomaly types; the interpretability and credibility of detection results need to be improved, and it is difficult to provide effective decision support for medical professionals; the lack of fine-grained positioning and analysis of medical information anomalies makes it difficult to trace the causes of anomalies and quickly locate problems.

[0004] In order to solve the above problems, it is urgent to research and develop a simulated medical information detection method and system based on AI algorithm, make full use of medical field knowledge and artificial intelligence technology, realize efficient, accurate and fine-grained detection of medical information anomalies, and provide medical institutions with comprehensive and reliable medical information quality management solutions. Summary of the invention

[0005] The embodiments of the present invention provide a simulated medical information detection method and system based on AI algorithm, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a simulated medical information detection method based on an AI algorithm, comprising:

[0007] Acquire simulated medical information to be tested, the simulated medical information including the patient's electronic medical record, doctor's order information, test report, medical imaging data, and physiological parameter information collected by a wearable device, extract multimodal features from the simulated medical information, and map the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient;

[0008] Calculate the abnormal score of each feature in the comprehensive health feature portrait, identify abnormal medical information fragments, and model the patient's historical medical information. Combined with medical prior knowledge, build an abnormal evolution graph model to predict the patient's abnormal risk probability in the preset future time domain;

[0009] An abnormal warning report is automatically generated based on the abnormal medical information fragments and abnormal risk probabilities. At the same time, the reinforcement learning algorithm is used to continuously optimize the detection strategy, and the optimized detection strategy is used for subsequent simulated medical information detection.

[0010] In an optional embodiment, performing multimodal feature extraction on the simulated medical information and mapping the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient includes:

[0011] The key medical concepts in the text information of the simulated medical information are extracted through natural language processing technology to obtain the initial text features; the pre-trained convolutional neural network is used to perform end-to-end feature learning on the medical imaging data in the simulated medical information to obtain the initial image features; the gated recurrent unit is used to extract the initial time series features from the physiological parameter information in the simulated medical information;

[0012] A mapping function is introduced to map the extracted initial text features, initial image features and initial time series features into a shared semantic space, and positive sample pairs and negative sample pairs are constructed, where the positive sample pairs are composed of multimodal features of the same patient, and the negative sample pairs are composed of multimodal features of different patients;

[0013] A contrastive learning loss function is constructed based on positive and negative sample pairs. The similarity of the positive sample pairs is compared with the similarity of all negative sample pairs. By maximizing the similarity of the positive sample pairs and minimizing the similarity of the negative sample pairs, semantic alignment is performed in the shared semantic space to obtain the aligned multimodal features. At the same time, an attention mechanism is introduced to perform weighted fusion of the aligned multimodal features to obtain the final fused features as a comprehensive health feature portrait of the patient.

[0014] In an optional embodiment, constructing a loss function for contrastive learning based on positive sample pairs and negative sample pairs includes:

[0015] ;

[0016] Where L represents the loss function of contrastive learning, N represents the number of training samples, sim(·) represents the function of calculating the similarity between features, and f i represents the characteristics of the i-th training sample, f j represents the feature of the jth training sample, f k represents the feature of the kth training sample, and τ represents the temperature parameter.

[0017] In an optional embodiment, calculating the abnormal score of each feature in the comprehensive health feature profile and identifying abnormal medical information fragments include:

[0018] Construct a multi-layer anomaly detection model, including the bottom layer, middle layer and top layer. The bottom layer is used to extract local information of medical information and obtain local anomaly features; the middle layer is used to gradually aggregate local anomaly features of the lower layer to obtain regional anomaly features; the top layer is used to extract global information of medical information and obtain overall anomaly features.

[0019] The comprehensive health feature portrait is divided into high-level features and low-level features, and a local anomaly detection module is set on the bottom layer of the anomaly detection model. The local area features are extracted by performing a local pooling operation on the low-level features in the comprehensive health feature portrait, and the anomaly measurement of each local area feature is calculated. It is determined whether the current local area feature is abnormal according to the threshold of the anomaly measurement, and the abnormal area is marked;

[0020] A global anomaly detection module is set on the top layer of the anomaly detection model to perform a global pooling operation on high-level features in the comprehensive health feature portrait to obtain global features of the comprehensive health feature portrait, calculate an anomaly measure of the global features, determine whether the comprehensive health feature portrait has an anomaly according to a threshold of the anomaly measure, and give an anomaly score;

[0021] Design a composite loss function, including global anomaly detection loss and local anomaly detection loss, and obtain a trained anomaly detection model by jointly optimizing the global anomaly detection loss and the local anomaly detection loss;

[0022] The comprehensive health feature portrait is input into the trained anomaly detection model, and the global anomaly detection module is used to determine whether there is an abnormality in the comprehensive health feature portrait and give an anomaly score. At the same time, according to the abnormal information propagation mechanism of the multi-level graph convolutional network, the abnormal information detected by the global anomaly detection module is propagated from top to bottom to the local anomaly detection module to determine whether each local area has an abnormality, and mark the abnormal area to obtain abnormal medical information fragments.

[0023] In an optional embodiment, the patient's historical medical information is modeled, and an abnormal evolution graph model is constructed in combination with medical prior knowledge to predict the abnormal risk probability of the patient in a preset future time domain, including:

[0024] Obtain the patient's historical medical information and preprocess it, convert the preprocessed historical medical information into a time series feature vector, and build an abnormal evolution graph model based on the obtained time series feature vector, where the node represents the abnormal state, the edge represents the evolution relationship between the abnormal states, and the edge weight is the state transition probability;

[0025] Design a graph neural network architecture, encode the abnormal state into node features as the input of the graph neural network, aggregate the neighborhood information of the node through graph convolution operations, update the feature representation of the node, and use the constructed abnormal evolution graph model to train the graph neural network through supervised learning;

[0026] The patient’s current medical information is obtained and input into the trained graph neural network, and the abnormal risk probability of the patient in the preset future time domain is predicted through the abnormal evolution graph model.

[0027] In an optional embodiment, the calculation formula for predicting the abnormal risk probability of a patient in a preset future time domain by using the abnormal evolution graph model is as follows:

[0028] ;

[0029] Among them, P represents the abnormal risk probability, T represents the time domain length, M represents the number of abnormal states, ω a represents the weight of abnormal state a, y t,a represents the label of abnormal state a at time t, β represents the duration balance factor, and d t,a represents the duration of abnormal state a at time t, γ represents the correlation weight between abnormal states, ω ab Indicates the degree of correlation between abnormal state a and abnormal state b, y a The label of abnormal state a, y b The label representing the abnormal state b.

[0030] In an optional embodiment, an abnormal warning report is automatically generated based on the abnormal medical information fragments and abnormal risk probabilities, and a reinforcement learning algorithm is used to continuously optimize the detection strategy, and the optimized detection strategy is used for subsequent simulated medical information detection, including:

[0031] Constructing a transition dynamics model through a deep neural network, which takes the current state and action as input and outputs the predicted distribution parameters of the next state;

[0032] Based on the obtained transition dynamics model, a tree-based search algorithm is used to perform strategy search, where the tree represents the search process, the root node represents the current state, the child nodes represent the intermediate states, and the edges represent the actions;

[0033] In the selection phase, the transfer dynamics model is used to predict the state of the child node in the preset future time domain, and the optimal action sequence is selected in combination with the balance strategy of exploration and utilization; in the expansion phase, the transfer dynamics model is used to generate new states and update the statistical information of the tree;

[0034] The search results of the tree-based search algorithm are used as expert experience, and the policy gradient algorithm is used to update the detection strategy. By minimizing the difference between the policy network and the expert experience, the network parameters are updated, and iterations are repeated until the preset termination conditions are met to obtain the updated detection strategy. The optimized detection strategy is used for subsequent simulated medical information detection.

[0035] A second aspect of an embodiment of the present invention provides a simulated medical information detection system based on an AI algorithm, comprising:

[0036] The first unit is used to obtain simulated medical information to be tested, wherein the simulated medical information includes the patient's electronic medical record, doctor's order information, test report, medical imaging data, and physiological parameter information collected by the wearable device, extract multimodal features from the simulated medical information, and map the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient;

[0037] The second unit is used to calculate the abnormal score of each feature in the comprehensive health feature portrait, identify abnormal medical information fragments, model the patient's historical medical information, and build an abnormal evolution graph model based on medical prior knowledge to predict the patient's abnormal risk probability in the preset future time domain;

[0038] The third unit is used to automatically generate an abnormal warning report based on the abnormal medical information fragments and abnormal risk probabilities, and at the same time use the reinforcement learning algorithm to continuously optimize the detection strategy, and use the optimized detection strategy for subsequent simulated medical information detection.

[0039] According to a third aspect of an embodiment of the present invention, there is provided an electronic device, including:

[0040] processor;

[0041] a memory for storing processor-executable instructions;

[0042] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0043] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0044] Beneficial effects of the present invention:

[0045] (1) The present invention can construct a comprehensive health feature portrait of the patient by extracting multimodal features from the patient's electronic medical records, medical advice, examination reports, imaging data, and physiological parameter data, and mapping these heterogeneous features into the same semantic space. This multi-source heterogeneous data fusion helps provide a more comprehensive and holistic description of the patient's health status. Applying the trained anomaly detection model to the patient's health feature portrait can automatically identify abnormal medical information fragments and detect potential health anomalies. At the same time,

[0046] (2) The present invention can predict the abnormal risk probability of patients in the future by modeling the patient's historical data and combining it with medical prior knowledge, thereby issuing a risk warning. This is of great significance for early detection of diseases and timely intervention measures. The system can provide doctors with intelligent auxiliary diagnosis support, discover abnormal symptoms that may be missed by humans, and conduct comprehensive analysis from the perspective of multimodal data to assist doctors in making more accurate diagnostic judgments. By combining artificial intelligence technology with medical care data, the intelligence level of disease abnormality detection, risk warning and diagnostic assistance can be significantly improved, which will provide escort for precision medicine. . BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a simulated medical information detection method based on an AI algorithm according to an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of the structure of a simulated medical information detection system based on an AI algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0051] Figure 1 FIG. 1 is a flow chart of a simulated medical information detection method based on an AI algorithm according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0052] S101. Obtain simulated medical information to be tested, wherein the simulated medical information includes the patient's electronic medical record, doctor's order information, test report, medical imaging data, and physiological parameter information collected by wearable devices, perform multimodal feature extraction on the simulated medical information, and map the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient.

[0053] In an optional embodiment, performing multimodal feature extraction on the simulated medical information and mapping the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient includes:

[0054] The key medical concepts in the text information of the simulated medical information are extracted through natural language processing technology to obtain the initial text features; the pre-trained convolutional neural network is used to perform end-to-end feature learning on the medical imaging data in the simulated medical information to obtain the initial image features; the gated recurrent unit is used to extract the initial time series features from the physiological parameter information in the simulated medical information;

[0055] A mapping function is introduced to map the extracted initial text features, initial image features and initial time series features into a shared semantic space, and positive sample pairs and negative sample pairs are constructed, where the positive sample pairs are composed of multimodal features of the same patient, and the negative sample pairs are composed of multimodal features of different patients;

[0056] A contrastive learning loss function is constructed based on positive and negative sample pairs. The similarity of the positive sample pairs is compared with the similarity of all negative sample pairs. By maximizing the similarity of the positive sample pairs and minimizing the similarity of the negative sample pairs, semantic alignment is performed in the shared semantic space to obtain the aligned multimodal features. At the same time, an attention mechanism is introduced to perform weighted fusion of the aligned multimodal features to obtain the final fused features as a comprehensive health feature portrait of the patient.

[0057] For example, for electronic medical record text data, we first use a bidirectional long short-term memory network (BiLSTM) for feature extraction. First, the text is segmented and embedded, and then the word embedding sequence is encoded using BiLSTM. Through bidirectional information transfer, the contextual semantic information in the text is captured to obtain the initial text features.

[0058] For medical image data, convolutional neural networks (CNNs) are used for feature extraction. Specifically, pre-trained CNN models (such as ResNet, DenseNet, etc.) are used to perform end-to-end feature learning on medical images, and high-level abstract feature representations are extracted as initial image features through the stacking of convolutional layers and pooling layers.

[0059] For time series physiological signal data, we use gated recurrent units (GRU) for feature extraction. The time series signal is divided into time windows of fixed length, and then the signal sequence in each time window is encoded using GRU. The dynamic change pattern of the time series signal is adaptively captured through the gating mechanism to obtain the initial time series features.

[0060] For the initial features extracted from each modality, a multi-layer perceptron (MLP) is introduced as a mapping function to map the features of different modalities to a shared semantic space. MLP consists of multiple fully connected layers and nonlinear activation functions (such as ReLU). Through layer-by-layer nonlinear transformation, the original features are converted into representations with higher semantic consistency. At the same time, in order to maximize the mutual information between the features of different modalities, a loss function based on contrastive learning is designed. Specifically, for each training sample, positive sample pairs and negative sample pairs are constructed from the features of different modalities. Among them, the positive sample pairs are composed of different modal features of the same patient, such as the image features and text features of the same patient. The negative sample pairs are composed of different modal features of different patients, such as the image features and text features of different patients. Optionally, InfoNCE Loss can be used as the loss function for contrastive learning. This loss function encourages the maximization of the similarity of positive sample pairs by comparing the similarity of positive sample pairs with the similarity of all negative sample pairs, while suppressing the similarity of negative sample pairs.

[0061] Through contrastive learning, features of different modalities are semantically aligned in a shared semantic space. In order to further fuse these features, attention mechanisms and pooling operations are used. Specifically, the attention mechanism is first used to calculate the importance weights between features of different modalities. By learning the contribution of each modal feature to the final representation, their weights are adaptively adjusted to highlight more important and informative features. Then, the weighted features are pooled, such as maximum pooling or average pooling, to aggregate features of different modalities into a compact cross-modal feature representation. This representation fuses complementary information from different modalities and has stronger discriminative and generalization capabilities.

[0062] The fused cross-modal feature representation is used for downstream tasks such as anomaly detection and risk prediction. We input cross-modal features into task-specific models, such as support vector machines (SVM), logistic regression (LR), or neural networks, and optimize the performance of the entire system through end-to-end training. During the training process, we use cross-entropy loss functions or squared loss functions to measure the difference between the model's prediction results and the true label according to the type of task (classification or regression). Through the back-propagation algorithm, we adjust the parameters of the entire system to minimize the prediction error and improve the performance of downstream tasks.

[0063] Optionally, regularization methods such as adversarial training and data enhancement are introduced during the training process. By generating adversarial samples or applying random perturbations, the model's robustness to noise and interference is enhanced. At the same time, data enhancement techniques such as rotation, translation, and cropping are used to expand the diversity of training data and improve the generalization ability of the model.

[0064] Through multiple rounds of iterative unsupervised comparative learning and supervised fine-tuning, the performance of the model is continuously optimized so that it can better adapt to actual application scenarios and achieve better performance in different tasks.

[0065] In an optional embodiment, constructing a loss function for contrastive learning based on positive sample pairs and negative sample pairs includes:

[0066] ;

[0067] Where L represents the loss function of contrastive learning, N represents the number of training samples, sim(·) represents the function of calculating the similarity between features, and f i represents the characteristics of the i-th training sample, f j represents the feature of the jth training sample, f k represents the feature of the kth training sample, and τ represents the temperature parameter.

[0068] In this embodiment, by utilizing the contrastive learning framework, the features under different modalities are mapped to a shared semantic space, maximizing the similarity of features of different modalities of the same patient and minimizing the similarity of features of different patients. This method can achieve the fusion and alignment of different modal data at the semantic level, reveal the intrinsic relationship between the data, and thus construct a highly compact comprehensive health feature portrait. By weighted fusion of the aligned multimodal features through the attention mechanism, the advantageous features of each modal data can be fully utilized, so that the fused comprehensive feature portrait contains richer and more accurate patient health information. And without the need for manually annotated supervisory information, the feature representation can be automatically learned from a large amount of unlabeled multimodal medical data through the contrastive learning framework, reducing the cost and workload of manual annotation. Since the feature fusion is performed in the semantic space, it is independent of the specific task, and the obtained comprehensive health feature portrait has stronger versatility and generalization ability, which can be applied to a variety of downstream tasks, such as disease prediction, intelligent diagnosis, etc.

[0069] S102. Calculate the abnormal score of each feature in the comprehensive health feature portrait, identify abnormal medical information fragments, and model the patient's historical medical information. Combined with medical prior knowledge, build an abnormal evolution graph model to predict the patient's abnormal risk probability in a preset future time domain.

[0070] For example, by building a medical domain ontology knowledge base, professional knowledge in the medical field can be structured, including concepts such as diseases, symptoms, examinations, treatments, and their relationships. Using the ontology knowledge base, a semantic basis can be provided for heterogeneous medical data, achieving standardized and normalized representation of data, and promoting the interconnection and semantic integration of medical data.

[0071] Secondly, the use of graph convolutional neural networks to model and detect anomalies in the patient's comprehensive health feature portrait can make full use of the graph structure information of medical data and the interactive relationship between features. By introducing the attention mechanism, the importance of different features can be adaptively focused on, improving the accuracy and interpretability of anomaly detection.

[0072] Thirdly, by calculating the semantic anomaly score of each feature in the comprehensive health feature profile, abnormal medical information fragments can be identified, providing accurate positioning of abnormal information for clinical decision-making. This helps medical professionals quickly detect abnormal conditions of patients and take timely intervention measures.

[0073] Finally, by combining medical prior knowledge to build an abnormal evolution graph model and modeling the patient's historical medical information, we can reveal the development law and evolution pattern of the disease. By predicting the patient's future abnormal risk probability, we can achieve early warning and risk assessment, and provide patients with personalized health management and prevention strategies.

[0074] In an optional embodiment, calculating the abnormal score of each feature in the comprehensive health feature profile and identifying abnormal medical information fragments include:

[0075] Construct a multi-layer anomaly detection model, including the bottom layer, middle layer and top layer. The bottom layer is used to extract local information of medical information and obtain local anomaly features; the middle layer is used to gradually aggregate local anomaly features of the lower layer to obtain regional anomaly features; the top layer is used to extract global information of medical information and obtain overall anomaly features.

[0076] The comprehensive health feature portrait is divided into high-level features and low-level features, and a local anomaly detection module is set on the bottom layer of the anomaly detection model. The local area features are extracted by performing a local pooling operation on the low-level features in the comprehensive health feature portrait, and the anomaly measurement of each local area feature is calculated. It is determined whether the current local area feature is abnormal according to the threshold of the anomaly measurement, and the abnormal area is marked;

[0077] A global anomaly detection module is set on the top layer of the anomaly detection model to perform a global pooling operation on high-level features in the comprehensive health feature portrait to obtain global features of the comprehensive health feature portrait, calculate an anomaly measure of the global features, determine whether the comprehensive health feature portrait has an anomaly according to a threshold of the anomaly measure, and give an anomaly score;

[0078] Design a composite loss function, including global anomaly detection loss and local anomaly detection loss, and obtain a trained anomaly detection model by jointly optimizing the global anomaly detection loss and the local anomaly detection loss;

[0079] The comprehensive health feature portrait is input into the trained anomaly detection model, and the global anomaly detection module is used to determine whether there is an abnormality in the comprehensive health feature portrait and give an anomaly score. At the same time, according to the abnormal information propagation mechanism of the multi-level graph convolutional network, the abnormal information detected by the global anomaly detection module is propagated from top to bottom to the local anomaly detection module to determine whether each local area has an abnormality, and mark the abnormal area to obtain abnormal medical information fragments.

[0080] For example, first, a multi-layer graph convolutional neural network is designed to learn the hierarchical representation of medical data. Including the bottom layer, the middle layer and the top layer, each layer of the graph convolutional neural network corresponds to feature extraction and abnormal pattern capture of different granularities, among which the bottom layer is used to extract local information of medical information and obtain local abnormal features; the middle layer is used to gradually aggregate the local abnormal features of the lower layer to obtain regional abnormal features; the top layer is used to extract the global information of medical information and obtain the overall abnormal features.

[0081] Corresponding anomaly detection modules are set at different levels of the graph convolutional neural network to capture abnormal patterns at different granularities. Specifically, a local anomaly detection module is set at the bottom layer of the anomaly detection model, and local area features are extracted by performing local pooling operations on low-level features in the comprehensive health feature portrait, and the abnormality metric of each local area feature is calculated. According to the threshold of the abnormality metric, it is determined whether the current local area feature is abnormal, and the abnormal area is marked. Optionally, for each local area feature vector, the degree of deviation from the normal mode is calculated as an indicator of the local anomaly metric. For example, a special normal sample reconstruction model (such as an autoencoder, a generative adversarial network, etc.) is trained, and the local area feature vector is input into the reconstruction model to obtain the reconstructed local feature vector, and the distance between the original feature vector and the reconstructed vector (such as the Euclidean distance) is calculated as the local anomaly metric. The larger the distance, the more difficult it is to reconstruct using the normal mode, and the more likely it is to be abnormal.

[0082] A global anomaly detection module is set at the top layer of the anomaly detection model to perform global pooling operations on high-level features in the comprehensive health feature portrait to obtain the global features of the comprehensive health feature portrait, calculate the anomaly measurement of the global features, determine whether the comprehensive health feature portrait is abnormal based on the threshold of the anomaly measurement, and give an anomaly score. Specifically, the top global feature vector is input into the fully connected layer to obtain a single anomaly score, and the score value can be normalized to the interval [0, 1] as a global anomaly measurement. The closer the score value is to 1, the more abnormal the overall input is.

[0083] Use pre-trained models (such as VGG, ResNet, etc.) to extract features at different levels of comprehensive health feature portraits, with shallow features as low-level inputs and deep features as high-level inputs. Use small convolution kernels such as 3×3 to perform convolution operations on low-level features to extract local feature maps, calculate the reconstruction error between each local feature map and the normal feature map as the abnormal score, and use the top N% of the scores as abnormal areas based on the preset threshold.

[0084] Use global average pooling to obtain a single global feature vector, input the global vector into the fully connected layer to obtain a single anomaly score, and judge whether the entire input is abnormal based on the pre-set threshold.

[0085] Design a composite loss function, including local anomaly loss and global anomaly loss, where the local anomaly loss uses BCE loss or Focal Loss to supervise the anomaly classification of each local area, and the global anomaly loss uses BCE loss to supervise the anomaly classification of the entire input. The weighted sum of the local anomaly loss and the global anomaly loss is used as the final composite loss function.

[0086] The forward propagation obtains the global anomaly score and the score of each local area. The global score is used to determine whether the whole is abnormal. The global anomaly is passed to the bottom layer as prior knowledge, the local area anomaly score is adjusted, and the abnormal area is determined according to the adjusted local score.

[0087] Optionally, in order to further improve the accuracy and positioning capability of anomaly detection, an anomaly information propagation mechanism is designed to pass high-level global anomaly information to low-level layers to guide the detection and positioning of local anomalies. These include: top-down anomaly information propagation, in the global anomaly detection module, if an overall abnormal state is detected, the anomaly information is propagated from top to bottom to the lower levels. Through the back-propagation mechanism, the high-level anomaly information is passed to the bottom layer layer by layer to adjust the feature representation and anomaly measurement of the low level. Bottom-up anomaly information propagation, that is, in the local anomaly detection module, if a local abnormal area or node is detected, the anomaly information is propagated from bottom to top to the high level. Through the forward propagation mechanism, the bottom-level anomaly information is passed to the top layer layer by layer to affect the global anomaly judgment at the high level. Through the two-way propagation of anomaly information, we establish a connection between global anomalies and local anomalies, so that the anomaly detection model can capture abnormal patterns more comprehensively and accurately.

[0088] In this embodiment, by performing anomaly detection from both the global and local levels, the global module focuses on the overall anomaly score, and the local module focuses on discovering specific abnormal areas. The two complement each other and can discover anomalies more comprehensively. Through a top-down abnormal information transmission mechanism, the abnormal information detected globally is gradually transmitted to the bottom layer to guide the bottom layer to discover local regional anomalies. Modules at different levels work together to improve the accuracy of anomaly detection. There is no need for large-scale abnormal sample annotation data. Instead, normal patterns are learned from normal data through comparative learning, and abnormal situations that deviate from normal patterns are identified, which belongs to the category of weakly supervised anomaly detection. In addition to giving an overall anomaly score, specific abnormal medical information fragments can also be accurately located, providing more detailed reference for doctors' diagnosis and subsequent processing. A composite loss function of global anomaly loss and local anomaly loss is designed, and end-to-end training of the model is achieved by jointly optimizing the two losses, thereby improving the robustness of the model. Refined detection of medical data is achieved, which can provide important auxiliary support for disease diagnosis and medical decision-making.

[0089] In an optional embodiment, the patient's historical medical information is modeled, and an abnormal evolution graph model is constructed in combination with medical prior knowledge to predict the abnormal risk probability of the patient in a preset future time domain, including:

[0090] Obtain the patient's historical medical information and preprocess it, convert the preprocessed historical medical information into a time series feature vector, and build an abnormal evolution graph model based on the obtained time series feature vector, where the node represents the abnormal state, the edge represents the evolution relationship between the abnormal states, and the edge weight is the state transition probability;

[0091] Design a graph neural network architecture, encode the abnormal state into node features as the input of the graph neural network, aggregate the neighborhood information of the node through graph convolution operations, update the feature representation of the node, and use the constructed abnormal evolution graph model to train the graph neural network through supervised learning;

[0092] The patient’s current medical information is obtained and input into the trained graph neural network, and the abnormal risk probability of the patient in the preset future time domain is predicted through the abnormal evolution graph model.

[0093] Exemplarily, firstly, the patient's historical medical information is modeled in time series to construct an abnormal evolution graph model. Specifically, the patient's historical medical information is cleaned, normalized and feature extracted, and converted into a time series feature vector. Based on medical prior knowledge and expert experience, a set of abnormal states are defined, each of which corresponds to a specific clinical manifestation or abnormal physiological indicator. At the same time, medical prior knowledge and statistical methods are used to estimate the transition probability between abnormal states. The transition probability reflects the possibility of the patient switching between different abnormal states. With abnormal states as nodes and state transition probabilities as edge weights, an abnormal evolution graph model is constructed. The abnormal evolution graph model depicts the dynamic evolution process of the patient's abnormal state.

[0094] In order to improve the accuracy and interpretability of the abnormal evolution graph model, medical prior knowledge is introduced to constrain and optimize the model. Among them, the abnormal state constraint includes imposing constraints on the transition between abnormal states based on medical field knowledge. For example, the transition between certain abnormal states may be impossible or require specific clinical conditions to be met. The transition probability adjustment uses the experience and judgment of medical experts to adjust and optimize the transition probability between abnormal states. Expert knowledge can help us more accurately characterize the evolution law of abnormal states. The abnormal state weight setting assigns weights to each abnormal state according to the clinical severity and risk level of different abnormal states. The weight reflects the degree of impact of the abnormal state on the patient's health.

[0095] Design graph neural network architectures suitable for abnormal evolution graphs, such as graph convolutional neural networks (GCNs), graph attention networks (GATs), etc. Encode the clinical characteristics and temporal information of abnormal states into node feature vectors as input to the graph neural network. Through graph convolution operations, aggregate the neighborhood information of nodes and update the feature representation of nodes. Graph convolution operations can capture the interactions and influences between abnormal states. Introduce temporal modeling mechanisms in graph neural networks, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), to capture the temporal evolution patterns of abnormal states. Use patients' historical medical data to train graph neural network models through supervised learning. Among them, the optimization goal can be to predict the patient's future abnormal state or abnormal risk probability.

[0096] Given a patient’s current medical information, it is first converted into a feature vector, and then forward reasoning is performed through the abnormal evolution graph model to predict the probability distribution of the patient’s abnormal state in future time steps.

[0097] Optionally, the patient's abnormal risk score is calculated based on the predicted probability distribution of abnormal states. The risk score comprehensively considers the probability and severity of different abnormal states. A risk threshold is set, and when the patient's abnormal risk score exceeds the threshold, an early warning mechanism is triggered to remind medical personnel to intervene and treat.

[0098] In an optional embodiment, the calculation formula for predicting the abnormal risk probability of a patient in a preset future time domain by using the abnormal evolution graph model is as follows:

[0099] ;

[0100] Among them, P represents the abnormal risk probability, T represents the time domain length, M represents the number of abnormal states, ω a represents the weight of abnormal state a, y t,a represents the label of abnormal state a at time t, β represents the duration balance factor, and d t,a represents the duration of abnormal state a at time t, γ represents the correlation weight between abnormal states, ω ab Indicates the degree of correlation between abnormal state a and abnormal state b, y a The label of abnormal state a, y b The label representing the abnormal state b.

[0101] In this embodiment, by constructing an abnormal evolution graph model, the abnormal states and their evolutionary relationships in the patient's history are explicitly modeled, which can accurately describe and capture the temporal dynamic process of abnormal development. On this basis, the graph neural network is used to predict future abnormal risks, providing an effective solution for temporal anomaly detection. By combining the patient's actual medical history information with medical prior knowledge, the former provides a data-driven individual-specific representation for the model, while the latter introduces the universal inspiration of domain knowledge. The two complement each other and improve the accuracy of abnormal analysis. The abnormal evolution modeling based on the graph structure provides an intuitive explanation for abnormal analysis. The evolution path and risk source of the abnormal state are clear at a glance, providing a valuable reference for making subsequent diagnosis and treatment decisions.

[0102] S103. Automatically generate an abnormal warning report based on the abnormal medical information fragments and abnormal risk probabilities, and use the reinforcement learning algorithm to continuously optimize the detection strategy, and use the optimized detection strategy for subsequent simulated medical information detection.

[0103] Exemplarily, first, according to the output of the abnormal classification model, determine the type of abnormality that exists in the patient, such as medical accidents, medical disputes, medical fraud, etc. Natural language generation technology, such as template filling, sequence-to-sequence generation, etc., is used to automatically generate abnormal descriptions based on abnormal feature fragments and abnormal types. According to the abnormal risk probability and the severity of the abnormal type, the risk level of the abnormal situation is evaluated, such as low risk, medium risk, high risk, etc. And combined with medical accident prevention rules and clinical guideline knowledge, targeted disposal suggestions are given. Disposal suggestions may include further examinations, treatment plan adjustments, risk communication, etc. The abnormal type, abnormal description, risk level and disposal suggestions are integrated into an abnormal warning report and displayed to medical personnel in a visual way, so that they can understand and deal with abnormal situations in a timely manner.

[0104] In order to continuously improve the performance of anomaly detection and classification, case libraries and reinforcement learning algorithms are introduced. Specifically, the detected anomaly cases and the corresponding anomaly feature fragments, anomaly risk probability, anomaly type, disposal suggestions and other information are stored in the case library. The case library is regularly updated and maintained, new anomaly cases are added, and existing cases are reviewed and cleaned. Reinforcement learning algorithms suitable for anomaly detection and classification are designed, such as deep Q network (DQN), policy gradient, etc. The reinforcement learning algorithm continuously adjusts the detection strategy and optimizes the long-term return by interacting with the environment. The case library is used as the environment of the reinforcement learning algorithm, the feature fragments and risk probability of the anomaly cases are used as the state, the anomaly type and disposal suggestions are used as the action, and the performance indicators of anomaly detection and classification are used as the reward. By continuously interacting with the case library environment, the reinforcement learning algorithm is used to optimize the anomaly detection and classification strategy. In the optimization process, through the balance of exploration and utilization, while improving the detection performance, it continuously adapts to new anomaly patterns. Finally, the optimized detection strategy is applied to the actual anomaly detection and classification process to improve the system's adaptive ability and real-time feedback performance.

[0105] In an optional embodiment, an abnormal warning report is automatically generated based on the abnormal medical information fragments and abnormal risk probabilities, and a reinforcement learning algorithm is used to continuously optimize the detection strategy, and the optimized detection strategy is used for subsequent simulated medical information detection, including:

[0106] Constructing a transition dynamics model through a deep neural network, which takes the current state and action as input and outputs the predicted distribution parameters of the next state;

[0107] Based on the obtained transition dynamics model, a tree-based search algorithm is used to perform strategy search, where the tree represents the search process, the root node represents the current state, the child nodes represent the intermediate states, and the edges represent the actions;

[0108] In the selection phase, the transfer dynamics model is used to predict the state of the child node in the preset future time domain, and the optimal action sequence is selected in combination with the balance strategy of exploration and utilization; in the expansion phase, the transfer dynamics model is used to generate new states and update the statistical information of the tree;

[0109] The search results of the tree-based search algorithm are used as expert experience, and the policy gradient algorithm is used to update the detection strategy. By minimizing the difference between the policy network and the expert experience, the network parameters are updated, and iterations are repeated until the preset termination conditions are met to obtain the updated detection strategy. The optimized detection strategy is used for subsequent simulated medical information detection.

[0110] Exemplarily, a transfer dynamics model is first constructed. A deep neural network can be used as the main structure of the transfer dynamics model, and the appropriate number of network layers and neurons are designed to balance the model's expressiveness and computational efficiency. The current state and action are used as inputs to the deep neural network, and the predicted distribution parameters of the next state, such as the mean and variance, are calculated through forward propagation. Long short-term memory units or gated recurrent units are introduced into the deep neural network to capture the temporal dependency of state transfer and improve the model's modeling ability for complex medical information. A suitable loss function, such as a negative log-likelihood function or a mean square error function, is designed to measure the difference between the state distribution predicted by the model and the actual state, and the model parameters are optimized through a back-propagation algorithm.

[0111] The Monte Carlo tree search algorithm is selected as the tree-based search algorithm, and a suitable tree structure and node representation method are designed to adapt to the characteristics of the medical information detection task.

[0112] In the selection phase, starting from the current node, the transition dynamics model is used to predict the state distribution in the preset future time domain, and the optimal action sequence is selected by combining the exploration and exploitation balance strategy, such as the upper confidence interval algorithm or the soft maximum algorithm. In the expansion phase, new state nodes are generated by the transition dynamics model according to the selected action sequence and added to the search tree, while updating the tree statistics, such as the number of node visits and the average reward value. In the evaluation phase, the transition dynamics model is used to predict the state values ​​of the child nodes in the search tree, and the value estimate of the tree is updated by the back-propagation algorithm to guide the subsequent node selection and expansion.

[0113] The search results of the tree-based search algorithm are used as expert experience to construct a training data set, where each data sample includes a state, an action, and a corresponding reward value. The structure of the policy network, such as a deep neural network or a convolutional neural network, is designed to adapt to the characteristics of the medical information detection task and the network parameters are initialized.

[0114] Adopt policy gradient algorithms, such as REINFORCE algorithm or Actor-Critic algorithm, to optimize the policy network, and update the network parameters by maximizing the expected reward or minimizing the difference between the policy network and expert experience. During the training process, through experience replay or importance sampling and other technologies, improve sample utilization efficiency and training stability, and accelerate the convergence of the policy network.

[0115] The optimized strategy network is used as the decision model for anomaly detection. The medical information to be detected is input, and the action probability distribution is calculated through forward propagation. According to the action probability distribution, the action with the highest confidence is selected as the detection strategy, such as marking anomalies, triggering alarms or further analysis.

[0116] In practical applications, new medical information data is continuously collected and used for online learning and updating of the transfer dynamics model and policy network to adapt to the ever-changing medical environment. Optionally, the performance of the detection strategy is regularly evaluated, such as indicators such as accuracy, recall, and F1 value, and the model is fine-tuned and improved based on the evaluation results to maintain the effectiveness and robustness of the detection strategy.

[0117] In this embodiment, by combining the tree-based search algorithm and the policy gradient algorithm, the scheme can autonomously learn and optimize the anomaly detection strategy without a large amount of manually labeled data or complex domain rules, and has a certain degree of autonomous learning ability. The tree search-based method takes into account the multi-step decision-making process in the preset future time domain, and can better optimize long-term benefits rather than local greedy strategies. This long-term perspective helps to detect hidden abnormal evolution trajectories. By combining global exploration based on tree search and global optimization based on gradient update, the strategy optimization process can be effectively avoided from falling into the local optimum, and the robustness of the algorithm is improved. By optimizing the anomaly detection strategy, it is possible to continuously learn and adjust the detection decision, improve the accuracy and effectiveness of the detection, and at the same time have a certain degree of interpretability, which provides strong support for the realization of autonomous medical auxiliary diagnosis systems.

[0118] Figure 2 FIG. 1 is a schematic diagram of the structure of a simulated medical information detection system based on an AI algorithm according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0119] The first unit is used to obtain simulated medical information to be tested, wherein the simulated medical information includes the patient's electronic medical record, doctor's order information, test report, medical imaging data, and physiological parameter information collected by the wearable device, extract multimodal features from the simulated medical information, and map the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient;

[0120] The second unit is used to calculate the abnormal score of each feature in the comprehensive health feature portrait, identify abnormal medical information fragments, model the patient's historical medical information, and build an abnormal evolution graph model based on medical prior knowledge to predict the patient's abnormal risk probability in the preset future time domain;

[0121] The third unit is used to automatically generate an abnormal warning report based on the abnormal medical information fragments and abnormal risk probabilities, and at the same time use the reinforcement learning algorithm to continuously optimize the detection strategy, and use the optimized detection strategy for subsequent simulated medical information detection.

[0122] According to a third aspect of an embodiment of the present invention, there is provided an electronic device, including:

[0123] a processor; a memory for storing instructions executable by the processor;

[0124] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0125] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0126] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A simulated medical information detection method based on AI algorithm, characterized in that: include: Acquire simulated medical information to be tested, the simulated medical information including the patient's electronic medical record, doctor's order information, test report, medical imaging data, and physiological parameter information collected by a wearable device, extract multimodal features from the simulated medical information, and map the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient; Calculate the abnormal score of each feature in the comprehensive health feature portrait, identify abnormal medical information fragments, and model the patient's historical medical information. Combined with medical prior knowledge, build an abnormal evolution graph model to predict the patient's abnormal risk probability in the preset future time domain; Automatically generate abnormal warning reports based on the abnormal medical information fragments and abnormal risk probabilities, and use reinforcement learning algorithms to continuously optimize detection strategies, and use the optimized detection strategies for subsequent simulated medical information detection; Model the patient's historical medical information, combine it with medical prior knowledge to build an abnormal evolution graph model, and predict the patient's abnormal risk probability in the preset future time domain, including: Obtain the patient's historical medical information and preprocess it, convert the preprocessed historical medical information into a time series feature vector, and build an abnormal evolution graph model based on the obtained time series feature vector, where the node represents the abnormal state, the edge represents the evolution relationship between the abnormal states, and the edge weight is the state transition probability; Design a graph neural network architecture, encode the abnormal state into node features as the input of the graph neural network, aggregate the neighborhood information of the node through graph convolution operations, update the feature representation of the node, and use the constructed abnormal evolution graph model to train the graph neural network through supervised learning; Obtain the patient’s current medical information, input it into the trained graph neural network, and predict the patient’s abnormal risk probability in the preset future time domain through the abnormal evolution graph model; The calculation formula for predicting the abnormal risk probability of a patient in a preset future time domain through the abnormal evolution graph model is as follows: ； Among them, P represents the abnormal risk probability, T represents the time domain length, M represents the number of abnormal states, ω a represents the weight of abnormal state a, y t,a represents the label of abnormal state a at time t, β represents the duration balance factor, and d t,a represents the duration of abnormal state a at time t, γ represents the correlation weight between abnormal states, ω ab Indicates the degree of correlation between abnormal state a and abnormal state b, y a The label of abnormal state a, y b The label representing the abnormal state b; Automatically generate abnormal warning reports based on the abnormal medical information fragments and abnormal risk probabilities, and use reinforcement learning algorithms to continuously optimize detection strategies. The optimized detection strategies are then used for subsequent simulated medical information detection, including: Constructing a transition dynamics model through a deep neural network that takes the current state and action as input and outputs the predicted distribution parameters of the next state; Based on the obtained transition dynamics model, a tree-based search algorithm is used to perform strategy search, where the tree represents the search process, the root node represents the current state, the child nodes represent the intermediate states, and the edges represent the actions; In the selection phase, the transfer dynamics model is used to predict the state of the child node in the preset future time domain, and the optimal action sequence is selected in combination with the balance strategy of exploration and utilization; in the expansion phase, the transfer dynamics model is used to generate new states and update the statistical information of the tree; The search results of the tree-based search algorithm are used as expert experience, and the policy gradient algorithm is used to update the detection strategy. By minimizing the difference between the policy network and the expert experience, the network parameters are updated, and the iteration is repeated until the preset termination condition is met to obtain the updated detection strategy. The optimized detection strategy is then used for subsequent simulated medical information detection. Extracting multimodal features from the simulated medical information and mapping the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient includes: The key medical concepts in the text information of the simulated medical information are extracted through natural language processing technology to obtain the initial text features; the pre-trained convolutional neural network is used to perform end-to-end feature learning on the medical imaging data in the simulated medical information to obtain the initial image features; the gated recurrent unit is used to extract the initial time series features from the physiological parameter information in the simulated medical information; A mapping function is introduced to map the extracted initial text features, initial image features and initial time series features into a shared semantic space, and positive sample pairs and negative sample pairs are constructed, where the positive sample pairs are composed of multimodal features of the same patient, and the negative sample pairs are composed of multimodal features of different patients; A loss function for contrastive learning is constructed based on positive and negative sample pairs. The similarity of the positive sample pairs is compared with the similarity of all negative sample pairs. By maximizing the similarity of the positive sample pairs and minimizing the similarity of the negative sample pairs, semantic alignment is performed in the shared semantic space to obtain the aligned multimodal features. At the same time, an attention mechanism is introduced to perform weighted fusion on the aligned multimodal features to obtain the final fused features as a comprehensive health feature portrait of the patient. The loss functions for constructing contrastive learning based on positive and negative sample pairs include: ； Where L represents the loss function of contrastive learning, N represents the number of training samples, sim(·) represents the function of calculating the similarity between features, and f i represents the characteristics of the i-th training sample, f j represents the feature of the jth training sample, f k represents the characteristics of the kth training sample, and τ represents the temperature parameter; Calculate the anomaly score of each feature in the comprehensive health feature profile and identify abnormal medical information fragments including: Construct a multi-layer anomaly detection model, including the bottom layer, middle layer and top layer. The bottom layer is used to extract local information of medical information and obtain local anomaly features; the middle layer is used to gradually aggregate local anomaly features of the lower layer to obtain regional anomaly features; the top layer is used to extract global information of medical information and obtain overall anomaly features. The comprehensive health feature portrait is divided into high-level features and low-level features, and a local anomaly detection module is set on the bottom layer of the anomaly detection model. The local area features are extracted by performing a local pooling operation on the low-level features in the comprehensive health feature portrait, and the anomaly measurement of each local area feature is calculated. It is determined whether the current local area feature is abnormal according to the threshold of the anomaly measurement, and the abnormal area is marked; A global anomaly detection module is set on the top layer of the anomaly detection model to perform a global pooling operation on high-level features in the comprehensive health feature portrait to obtain global features of the comprehensive health feature portrait, calculate an anomaly measure of the global features, determine whether the comprehensive health feature portrait has an anomaly according to a threshold of the anomaly measure, and give an anomaly score; Design a composite loss function, including global anomaly detection loss and local anomaly detection loss, and obtain a trained anomaly detection model by jointly optimizing the global anomaly detection loss and the local anomaly detection loss; The comprehensive health feature portrait is input into the trained anomaly detection model, and the global anomaly detection module is used to determine whether there is an abnormality in the comprehensive health feature portrait and give an anomaly score. At the same time, according to the abnormal information propagation mechanism of the multi-level graph convolutional network, the abnormal information detected by the global anomaly detection module is propagated from top to bottom to the local anomaly detection module to determine whether each local area has an abnormality, and mark the abnormal area to obtain abnormal medical information fragments.

2. A simulated medical information detection system based on AI algorithm, used to implement the method described in claim 1, characterized in that: include: The first unit is used to obtain simulated medical information to be tested, wherein the simulated medical information includes the patient's electronic medical record, doctor's order information, test report, medical imaging data, and physiological parameter information collected by the wearable device, extract multimodal features from the simulated medical information, and map the extracted multimodal features to the same semantic space to construct a comprehensive health feature portrait of the patient; The second unit is used to calculate the abnormal score of each feature in the comprehensive health feature portrait, identify abnormal medical information fragments, model the patient's historical medical information, and build an abnormal evolution graph model based on medical prior knowledge to predict the patient's abnormal risk probability in the preset future time domain; The third unit is used to automatically generate an abnormal warning report based on the abnormal medical information fragments and abnormal risk probabilities, and at the same time use the reinforcement learning algorithm to continuously optimize the detection strategy, and use the optimized detection strategy for subsequent simulated medical information detection.

3. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method of claim 1.

4. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method of claim 1 is implemented.

Images