Endocrine health monitoring method and system based on artificial intelligence

Abstract

The invention discloses an endocrine health monitoring method and system based on artificial intelligence. The method comprises the following steps: S1, generating a preprocessed standard physiological signal data set; s2, constructing an initial physiological state map based on the preprocessed standard physiological signal data set to form an initial physiological state map; s3, expanding and refining the preliminary physiological state atlas, constructing a multilayer heterogeneous physiological state atlas, and forming a combined physiological state atlas comprehensively reflecting various physiological interaction mechanisms; s4, outputting graph embedding representation reflecting the states of the endocrine glands and the dynamic linkage relation of the endocrine glands; and S5, on the basis of the graph embedding representation, analyzing and judging the linkage state of each gland in the endocrine system by using a preset anomaly judgment strategy, detecting a node or edge relationship with an abnormal mode, and outputting an anomaly monitoring result serving as anomaly early warning information. The endocrine health monitoring precision can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of health monitoring, and particularly to an endocrine health monitoring method and system based on artificial intelligence. Background Art

[0002] With the continuous development of intelligent medicine and artificial intelligence algorithms, more and more health monitoring systems are applied to the field of clinical auxiliary diagnosis. Especially in the management of long-term courses of chronic diseases and endocrine and metabolic diseases, they have become important tools for improving the diagnosis and treatment efficiency and the quality of life of patients. As the core physiological system that regulates multiple functions of the body's metabolism, immunity, and reproduction, the changes in the health status of the endocrine system are often hidden and complex, and are often manifested through the complex interactions between multiple glands. Therefore, the comprehensive monitoring of the state of the endocrine system, especially the identification and abnormal detection of the linkage mechanism between glands, has become a key issue in medical research and the construction of intelligent diagnosis and treatment systems.

[0003] Currently, endocrine health monitoring mainly relies on periodic physical examinations, blood hormone level tests, or remote monitoring systems based on single physiological indicators, mostly based on linear analysis or single-dimensional index evaluation. It is difficult to capture the linkage patterns of multi-dimensional data between endocrine glands in the time dimension and interaction structure. For example, some monitoring platforms try to identify abnormal fluctuations by continuously recording changes in hormone concentrations, but often ignore the functional coupling mechanism between endocrine glands and lack a comprehensive evaluation of neuroendocrine signals and other key parameters of physiological rhythms. In addition, traditional methods still mainly use tabular sequences in data structure processing and lack modeling means at the structured atlas level, which limits the modeling and expression ability of complex physiological interaction relationships.

[0004] On the other hand, some studies have tried to use machine learning methods to detect abnormalities in endocrine data, but there are problems such as simple training data structure, insufficient sensitivity of the model to the synergistic effects between glands, and lack of high-order spatio-temporal feature extraction ability. More importantly, most existing methods regard the endocrine system as an independent collection of glands and fail to build a dynamic monitoring model that reflects its overall synergistic mechanism, resulting in unsatisfactory effects in identifying linkage abnormalities and predicting the risk of systemic imbalance.

[0005] In summary, the existing endocrine health monitoring technologies still have significant deficiencies in the data collection dimension, gland linkage modeling ability, and abnormal identification mechanism, and it is difficult to comprehensively and dynamically reflect the true physiological state of the endocrine system. These series of technical shortcomings urgently need to be broken through and solved in the new generation of intelligent monitoring methods. Summary of the Invention

[0006] An object of the present invention is to propose an endocrine health monitoring method and system based on artificial intelligence, and the present invention can improve the accuracy of endocrine health monitoring.

[0007] An endocrine health monitoring method based on artificial intelligence according to an embodiment of the present invention includes the following steps: S1. Synchronously collect a physiological signal dataset from multiple endocrine glands using a multi-channel real-time acquisition device, preprocess the collected physiological signal dataset, and generate a preprocessed standard physiological signal dataset; S2. Construct a preliminary physiological state map based on the preprocessed standard physiological signal dataset, correspond each endocrine gland to a node in the physiological state map, and construct edges between the nodes according to the physiological interaction relationships between the glands to form a preliminary physiological state map; S3. Expand and refine the preliminary physiological state map to construct a multi-layer heterogeneous physiological state map, where different layers represent different types of physiological interaction mechanisms, and perform information fusion on the nodes and edges in each layer to form a combined physiological state map that comprehensively reflects various physiological interaction mechanisms; S4. Use the combined physiological state map and the preprocessed standard physiological signal dataset as inputs, and transfer them to a spatio-temporal fusion graph convolutional neural network model, and the output is a graph embedding representation reflecting the states of each endocrine gland and their dynamic linkage relationships; S5. Based on the graph embedding representation, use a preset abnormal determination strategy to analyze and discriminate the linkage states of each gland in the endocrine system, detect nodes or edge relationships with abnormal patterns, and output an abnormal monitoring result, and the abnormal monitoring result is used as abnormal warning information.

[0008] Optionally, the S1 includes the following steps: S11. Collect raw physiological signal data from multiple endocrine glands, set a set G of target endocrine glands, and for each endocrine gland in the set G of target endocrine glands , set the raw physiological signal data , where represents the th sampling moment, represents the hormone concentration of the endocrine gland collected at time , represents the neuroendocrine signal value, represents the physiological rhythm parameter, is the number of samples within the total sampling duration; S12. Combine the original sampling data of all endocrine glands to form a complete raw physiological signal dataset, and the raw physiological signal dataset represents a unified data structure of multiple endocrine glands and multiple types of indicators in a continuous time series; S13. Preprocess the original physiological signal dataset, perform signal denoising using the weighted moving average method, normalize the denoised data, and use the min-max normalization method to normalize each type of signal to the interval , perform time alignment on the normalized data, and perform linear interpolation or sample interpolation on the normalized data of each endocrine gland along the time axis based on the unified sampling time axis, so that the data of each endocrine gland has corresponding values at all time points , obtain the standardized physiological signal vector after complete time alignment, and organize the standardized physiological signal vectors of all endocrine glands at each time point to form a standardized physiological signal dataset with a unified structure after preprocessing .

[0009] Optionally, the S2 includes the following steps: S21. Based on the standardized physiological signal dataset after preprocessing , construct a preliminary physiological state map including structural information and state-dependent information. The preliminary physiological state map is a structured graph, including a node set, an edge set, a node feature matrix, and a weighted adjacency matrix. The node set V is used to represent multiple target endocrine glands involved in monitoring, and the nodes correspond to the target endocrine glands one by one. The node feature matrix X is used to represent the standardized physiological state information of each target endocrine gland in the current time window. Each node feature vector consists of the normalized value of hormone concentration, the normalized value of neuroendocrine signal, and the normalized value of physiological rhythm parameters; S22. Introduce information dependence based on joint state changes to construct the structure of the structured graph, calculate the interaction intensity between endocrine glands using the dynamic mutual information of physiological states between nodes, and construct the edge set and the weighted adjacency matrix . Calculate the information dependence intensity between the physiological states of every two different target endocrine glands. The information dependence intensity uses the state mutual information and the state joint entropy as evaluation indicators to reflect the state linkage synergy of two target endocrine glands in the standardized physiological state feature space. After calculating the information dependence intensity among all target endocrine glands, set the mutual information threshold, and only retain the connection relationships between gland pairs whose information dependence intensity exceeds the set threshold. Finally, form an edge set reflecting high state synergy, and record this connection intensity as the edge weight in the weighted adjacency matrix ; S23. Construct a weighted adjacency matrix based on the edge weights. The edge weights recorded in the weighted adjacency matrix are used to represent the state linkage dependence between two target endocrine glands. The larger the edge weight value, the stronger the synchronization and mutual responsiveness of the state changes between the target endocrine glands in the current time window. Connections are not established for pairs of target endocrine glands whose values do not exceed the threshold, and the values at the corresponding positions in the adjacency matrix are set to zero; S24. When constructing the preliminary physiological state map, introduce the co-perturbation sensitivity matrix S to characterize the state perturbation response degree between target endocrine glands. The co-perturbation sensitivity indicates the degree of synchronous change in the standard physiological states of other target endocrine glands when the standard physiological state of one target endocrine gland is perturbed in the current time window. The higher the co-perturbation sensitivity of an endocrine gland pair, the stronger its linkage reaction, which is indicative of endocrine system abnormalities; S25. Comprehensive node set 、Node feature matrix 、Weighted adjacency matrix 、Co-perturbation sensitivity matrix to construct a preliminary physiological state map driven by the state dependence structure among multiple endocrine glands .

[0010] Optionally, S3 includes the following steps: S31. Classify the physiological interaction mechanisms between target endocrine glands based on the preliminary physiological state map and set the physiological interaction mechanism set . The physiological interaction mechanism set includes hormone regulation mechanism, neural regulation mechanism, and metabolic feedback mechanism. Each type of physiological interaction mechanism in the physiological interaction mechanism set represents a different type of gland linkage method; S32. For each type of physiological interaction mechanism in the physiological interaction mechanism set , extract the graph structure information related to this type of physiological interaction mechanism from the preliminary physiological state map and construct the corresponding layer. Each layer contains a node set, a node feature matrix, an edge set, and a weighted adjacency matrix. The node set and node feature matrix in the layer are consistent with the preliminary physiological state map. The edge set is the set of connection relationships of target endocrine glands related to this type of physiological interaction mechanism, and the edge weights recorded in the weighted adjacency matrix represent the state linkage strength between target endocrine glands under this type of physiological interaction mechanism; S33. Uniformly fuse the structural information in all layers, set a fusion weight vector, where each weight value in the fusion weight vector is used to represent the contribution degree of the corresponding layer during the fusion process. Perform weighted summation on the weighted adjacency matrices in the layers according to the fusion weights corresponding to each layer to generate a fused joint weighted adjacency matrix. The weight value at each position in the joint weighted adjacency matrix represents the joint linkage strength of the pair of target endocrine glands under the combined action of all physiological interaction mechanisms. S34. Combine the node sets, edge sets, node feature matrices, and weighted adjacency matrices of all layers with the node set, edge set, node feature matrix, and joint weighted adjacency matrix of the fused layer to form a multi-layer heterogeneous physiological state atlas set. Each layer respectively reflects the linkage relationship of the target endocrine glands under a specific physiological interaction mechanism, and the fused layer is used to reflect the system-level target endocrine gland linkage structure under the combined action of all physiological interaction mechanisms.

[0011] Optionally, the S4 includes the following steps: S41. Use the multi-layer heterogeneous physiological state atlas set and the preprocessed standard physiological signal dataset as input data, and input them into a spatio-temporal fusion graph convolutional neural network model. The input data includes the weighted adjacency matrix, node feature matrix, and layer identification information of each layer in the multi-layer heterogeneous physiological state atlas. S42. Divide the preprocessed standard physiological signal dataset into time windows, set the length of the continuous time window to , and divide each target endocrine gland into time series segments of length in the time dimension to form a time-node joint input tensor, which represents the multi-dimensional standard physiological state change characteristics of all target endocrine glands under each time window. S43. Construct a spatio-temporal fusion graph convolutional neural network model. The structure of the spatio-temporal fusion graph convolutional neural network model is composed of alternating stacks of a time convolutional module and a graph structure convolutional module. The time convolutional module performs one-dimensional convolutional operations on the standard physiological state feature sequences of each target endocrine gland within continuous time windows based on the time-node joint input tensor to extract the dynamic evolution characteristics of the target endocrine gland in the time dimension. S44. Perform graph convolutional operations on the output result of the time convolutional module and the structural information of each layer in the multi-layer heterogeneous physiological state atlas set. The graph structure convolutional module performs multi-layer graph convolutional calculations based on the weighted adjacency matrix and node feature matrix between target endocrine glands in the graph structure, combined with the layer identification information, to extract the linkage relationship characteristics between target endocrine glands under different physiological interaction mechanisms, and output the graph convolutional output representation of each gland under this layer. S45. Perform layer aggregation processing on the graph convolutional output representations of all layers, and perform weighted combination on the multi-graph convolutional output representations according to the layer fusion weights to form the graph embedding representation of each target endocrine gland under the comprehensive action of all physiological interaction mechanisms; S46. Output the graph embedding representation, which is the high-dimensional state representation of each target endocrine gland within the current time window, and is used to reflect the individual state characteristics of the gland and its dynamic linkage relationship under the multi-endocrine gland linkage mechanism.

[0012] Optionally, the S5 includes the following steps: S51. Construct an abnormal determination input feature set based on the graph embedding representation. The abnormal determination input feature set includes the graph embedding vectors of each target endocrine gland within the current time window. The graph embedding vectors comprehensively reflect the linkage relationship characteristics between the standard physiological state of the endocrine gland and other endocrine glands under various physiological interaction mechanisms; S52. Construct an abnormal determination strategy, which is constructed based on historical monitoring data and clinical annotation information, and through training, obtain the classification boundary conditions for determining whether the current graph embedding representation shows an abnormal state; S53. Perform abnormal determination on the graph embedding vectors of each target endocrine gland to determine whether there are abnormal node patterns and abnormal edge patterns; S54. Mark the target endocrine gland or the relationship between endocrine glands with abnormal node patterns and abnormal edge patterns, and output the abnormal monitoring result, which includes node abnormal information and edge abnormal information; S55. Output the abnormal monitoring result, and the abnormal monitoring result is used as abnormal warning information for clinical auxiliary decision-making.

[0013] Optionally, the abnormal node pattern is that when the distance distribution of the graph embedding vector in the high-dimensional embedding space significantly deviates from the clustering center of normal training samples, or the embedding eigenvalue exceeds the preset physiological reasonable threshold range, it is determined that the target endocrine gland has individual functional abnormalities within the current time window; the abnormal edge pattern is that when the linkage representation value between a certain target endocrine gland and its associated endocrine gland in the graph embedding representation significantly deviates from the historical cooperation pattern in the normal state, or there is a situation where the decline speed of the linkage intensity is higher than the threshold, or the rise speed is higher than the threshold, it is determined that there is linkage abnormality or feedback abnormality between endocrine gland pairs.

[0014] Optionally, the abnormal monitoring result includes node abnormal information and edge abnormal information, which respectively record the endocrine gland numbers determined to be abnormal, the abnormal types, and their abnormal degree index values within the current time window.

[0015] Optionally, the node anomaly information is classified as individual endocrine gland anomaly, and the judgment condition is that if the distribution value of the graph embedding vector in the standard physiological embedding space satisfies one of the following rules: the hormone concentration embedding component is lower than 0.2 or higher than 0.85, the change rate of the neuroendocrine signal embedding component is greater than the set dynamic fluctuation threshold, and the rhythmic parameter embedding feature presents an aperiodic feature pattern; The edge anomaly information is classified as abnormal linkage between endocrine glands, and the judgment condition is that if the graph embedding linkage component between two target endocrine glands continuously deviates outside the historical mean interval for three time windows, or the sensitivity mutation rate of this pair of endocrine glands in the co-perturbation sensitivity matrix exceeds the set threshold.

[0016] An artificial intelligence-based endocrine health monitoring system for implementing an artificial intelligence-based endocrine health monitoring method, including the following modules: A physiological signal acquisition module for collecting the hormone concentration, neuroendocrine signal, and physiological rhythm parameters of multiple target endocrine glands, and performing denoising, normalization, and time alignment processing to generate a standard physiological signal dataset; A graph construction module for constructing a preliminary physiological state graph based on the standard physiological signal dataset, and extracting the mutual information and joint entropy of the gland states to construct edge weights; A heterogeneous graph expansion module for refining the preliminary physiological state graph into a multi-layer heterogeneous physiological state graph; A graph convolutional neural network module for fusing the standard physiological signal dataset and the multi-layer heterogeneous physiological state graph, extracting the dynamic evolution features and linkage structure features of the target endocrine glands, and generating a graph embedding representation; An anomaly recognition module for identifying node or edge anomalies according to the graph embedding representation and a preset anomaly determination strategy, and outputting an anomaly monitoring result.

[0017] The beneficial effects of the present invention are: (1) In the present invention, multiple endocrine glands are regarded as nodes in the graph structure, and a triple representation model of a node state feature matrix + weighted adjacency matrix + co-perturbation sensitivity matrix is introduced to accurately model the physiological interaction relationship between glands. In the edge construction strategy, the mutual information and joint entropy of the states between nodes are used to measure the dependence strength between glands, and low-coordination connections are dynamically removed, effectively reducing the redundancy and noise sensitivity of the graph structure.

[0018] (2) In the present invention, the hormone regulation mechanism, neuroregulation mechanism, and metabolic feedback mechanism in the endocrine system are respectively modeled as independent layers. Before the input of the graph neural network, the node and edge structures under each mechanism are hierarchically modeled and weighted and fused. By constructing a joint weighted adjacency matrix, cross-mechanism linkage information comprehensive processing is realized in the graph neural network, improving the model's perception ability of systematic dynamic linkage changes.

[0019] (3) The present invention designs an alternating stacked structure of a temporal convolution module + a multi-layer graph convolution module, captures the evolution trend of each gland at different time periods through temporal convolution, fuses the linkage features between glands in the multi-layer structure through graph convolution, and finally outputs a high-dimensional graph embedding vector for anomaly determination. Description of the Drawings

[0020] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 It is a flowchart of a method for monitoring endocrine health based on artificial intelligence proposed by the present invention. Detailed Embodiments

[0021] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0022] Refer to Figure 1 , a method for monitoring endocrine health based on artificial intelligence, includes the following steps: S1. Synchronously collect a physiological signal dataset from multiple endocrine glands using a multi-channel real-time acquisition device, and preprocess the collected physiological signal dataset to generate a preprocessed standard physiological signal dataset; S2. Construct a preliminary physiological state map based on the preprocessed standard physiological signal dataset, correspond each endocrine gland to a node in the physiological state map, and construct edges between the nodes according to the physiological interaction relationships between the glands to form a preliminary physiological state map; S3. Expand and refine the preliminary physiological state map to construct a multi-layer heterogeneous physiological state map. Different layers respectively represent different types of physiological interaction mechanisms, and perform information fusion on the nodes and edges in each layer to form a combined physiological state map that comprehensively reflects various physiological interaction mechanisms; S4. Use the combined physiological state map and the preprocessed standard physiological signal dataset as inputs, and transfer them to a spatio-temporal fusion graph convolutional neural network model, and the output is a graph embedding representation reflecting the states of each endocrine gland and their dynamic linkage relationships; S5. Based on the graph embedding representation, use a preset anomaly determination strategy to analyze and discriminate the linkage states of each gland in the endocrine system, detect nodes or edge relationships with abnormal patterns, and output an anomaly monitoring result, and the anomaly monitoring result is used as an anomaly warning message.

[0023] In this embodiment, S1 includes the following steps: S11. Collect the original physiological signal data from multiple endocrine glands, set the collection target endocrine gland set G, and for each endocrine gland in the target endocrine gland set G , set the original physiological signal data , where represents the th sampling moment, represents at the moment the hormone concentration of the collected endocrine gland , represents the neuroendocrine signal value, represents the physiological rhythm parameter, is the number of samples within the total sampling duration; S12. Combine the original sampling data of all endocrine glands to form a complete original physiological signal data set, which represents a unified data structure of multiple endocrine glands and multiple types of indicators in a continuous time series; S13. Preprocess the original physiological signal data set, use the weighted moving average method for signal denoising, normalize the denoised data, use the min-max normalization method to normalize each type of signal to the interval , perform time alignment on the normalized data, linearly interpolate or sample interpolate the normalized data of each endocrine gland along the time axis based on the unified sampling time axis, so that the data of each endocrine gland has corresponding values at all time points , obtain the complete time-aligned standard physiological signal vector, and organize the standard physiological signal vectors of all endocrine glands at each time point to form a preprocessed standard physiological signal data set with a unified structure .

[0024] In this embodiment, S2 includes the following steps: S21. Based on the preprocessed standard physiological signal data set , construct a preliminary physiological state map containing structural information and state-dependent information. The preliminary physiological state map is a structured graph, including a node set, an edge set, a node feature matrix, and a weighted adjacency matrix. The node set V is used to represent multiple target endocrine glands involved in the monitoring, and the nodes correspond to the target endocrine glands one by one. The node feature matrix X is used to represent the standard physiological state information of each target endocrine gland in the current time window. Each node feature vector consists of the normalized hormone concentration value, the normalized neuroendocrine signal value, and the normalized physiological rhythm parameter value; S22. Introduce information dependence based on joint state changes to construct the structure of the structured graph, calculate the interaction strength between endocrine glands using the dynamic mutual information of the physiological states between nodes, and construct the edge set and weighted adjacency matrix , calculate the information dependence strength between the physiological states of every two different target endocrine glands. The information dependence strength uses state mutual information and state joint entropy as evaluation indicators, which is used to reflect the state linkage synergy of two target endocrine glands in the standard physiological state feature space. After calculating the information dependence strength between all target endocrine glands, set a mutual information threshold, and only retain the connection relationships between gland pairs whose information dependence strength exceeds the set threshold. Finally, form an edge set reflecting high state synergy, and record this connection strength as the edge weight in the weighted adjacency matrix ; S23. Construct a weighted adjacency matrix based on the edge weights. The edge weights recorded in the weighted adjacency matrix are used to represent the state linkage dependence between two target endocrine glands. The larger the edge weight value, the stronger the synchronization and mutual responsiveness of the state changes between the target endocrine glands in the current time window. No connection is constructed between target endocrine gland pairs that do not exceed the threshold, and the value of the corresponding position in the adjacency matrix is set to zero; S24. Introduce a co-perturbation sensitivity matrix S while constructing the preliminary physiological state map, which is used to characterize the state perturbation response degree between target endocrine glands. Co-perturbation sensitivity represents the degree of synchronous change caused to the standard physiological states of other target endocrine glands when the standard physiological state of one target endocrine gland is perturbed in the current time window. The higher the co-perturbation sensitivity of an endocrine gland pair, the stronger its linkage reaction, which has an indicative significance for endocrine system abnormalities; Co-perturbation sensitivity matrix , defined as follows:

[0025] where represents the degree of influence on the state when the physiological state of endocrine gland is perturbed. Based on the state change gradient estimation of the historical window, the node pairs with high sensitivity respond more strongly to linkage abnormalities, which is used to assist the abnormal path attention mechanism learning of the subsequent graph convolutional neural network.

[0026] S25. Synthesize the node set , node feature matrix , weighted adjacency matrix , co-perturbation sensitivity matrix , and construct a preliminary physiological state map driven by the state dependence structure among multiple endocrine glands .

[0027] The present invention regards multiple endocrine glands as nodes in a graph structure, introduces a triple representation model of a node state feature matrix + weighted adjacency matrix + co-perturbation sensitivity matrix to accurately model the physiological interaction relationships between the glands. In the edge construction strategy, the mutual information of states and the joint entropy of states between nodes are used to measure the dependence strength between the glands, and low-cooperativity connections are dynamically removed, effectively reducing the redundancy and noise sensitivity of the graph structure.

[0028] In this embodiment, S3 includes the following steps: S31. Based on the preliminary physiological state atlas classify the physiological interaction mechanisms between the target endocrine glands, and set a set of physiological interaction mechanisms , the set of physiological interaction mechanisms includes a hormone regulation mechanism, a neural regulation mechanism, and a metabolic feedback mechanism. Each type of physiological interaction mechanism in the set of physiological interaction mechanisms represents a different type of gland linkage mode; S32. For each type of physiological interaction mechanism in the set of physiological interaction mechanisms , extract the graph structure information related to this type of physiological interaction mechanism from the preliminary physiological state atlas, and construct a corresponding layer. Each layer contains a node set, a node feature matrix, an edge set, and a weighted adjacency matrix. The node set and the node feature matrix in the layer are consistent with the preliminary physiological state atlas. The edge set is a set of connection relationships of the target endocrine glands related to this type of physiological interaction mechanism. The edge weights recorded in the weighted adjacency matrix represent the state linkage strength between the target endocrine glands under this type of physiological interaction mechanism; S33. Uniformly fuse the structure information in all layers, set a fusion weight vector. Each weight value in the fusion weight vector is used to represent the contribution degree of the corresponding layer in the fusion process. Perform weighted summation on the weighted adjacency matrices in the layers according to the fusion weights corresponding to each layer to generate a fused joint weighted adjacency matrix. The weight value at each position in the joint weighted adjacency matrix represents the joint linkage strength of the pair of target endocrine glands under the combined action of all physiological interaction mechanisms; S34. Combine the node sets, edge sets, node feature matrices, and weighted adjacency matrices of all layers with the node set, edge set, node feature matrix, and joint weighted adjacency matrix of the fusion layer to form a multi-layer heterogeneous physiological state atlas set , each layer respectively reflects the linkage relationship of the target endocrine glands under a specific physiological interaction mechanism, and the fusion layer is used to reflect the system-level target endocrine gland linkage structure under the combined action of all physiological interaction mechanisms.

[0029] In the present invention, the hormone regulation mechanism, neural regulation mechanism, and metabolic feedback mechanism in the endocrine system are respectively modeled as independent layers. Before inputting into the graph neural network, hierarchical modeling and weighted fusion are performed on the node and edge structures under each mechanism. By constructing a joint weighted adjacency matrix, cross-mechanism linkage information comprehensive processing is realized in the graph neural network, improving the model's perception ability of systematic dynamic linkage changes.

[0030] In this embodiment, S4 includes the following steps: S41. Use the multi-layer heterogeneous physiological state map set and the preprocessed standard physiological signal data set as input data, and input them into the spatio-temporal fusion graph convolutional neural network model. The input data includes the weighted adjacency matrix, node feature matrix, and layer identification information of each layer in the multi-layer heterogeneous physiological state map; S42. Divide the preprocessed standard physiological signal data set in the time window, set the length of the continuous time window as , and divide each target endocrine gland into time series segments with a length of in the time dimension, forming a time-node joint input tensor, which represents the multi-dimensional standard physiological state change characteristics of all target endocrine glands under each time window; S43. Construct a spatio-temporal fusion graph convolutional neural network model. The structure of the spatio-temporal fusion graph convolutional neural network model is composed of alternating stacks of a time convolutional module and a graph structure convolutional module. The time convolutional module performs one-dimensional convolutional operations on the standard physiological state feature sequences of each target endocrine gland within continuous time windows based on the time-node joint input tensor, for extracting the dynamic evolution characteristics of the target endocrine gland in the time dimension; S44. Perform graph convolutional operations on the output result of the time convolutional module and the structure information of each layer in the multi-layer heterogeneous physiological state map set. The graph structure convolutional module performs multi-layer graph convolutional calculations based on the weighted adjacency matrix and node feature matrix between target endocrine glands in the graph structure, combined with the layer identification information, for extracting the linkage relationship characteristics between target endocrine glands under different physiological interaction mechanisms, and outputting the graph convolutional output representation of each gland under this layer; S45. Perform layer aggregation processing on the graph convolutional output representations of all layers, and perform weighted combination on the multi-graph convolutional output representations according to the layer fusion weights, forming the graph embedding representation of each target endocrine gland under the comprehensive action of all physiological interaction mechanisms; S46. Output the graph embedding representation, which is the high-dimensional state representation of each target endocrine gland within the current time window, used to reflect the individual state characteristics of the gland and its dynamic linkage relationship under the multi-endocrine gland linkage mechanism.

[0031] The present invention designs an alternating stacked structure of a temporal convolution module + a multi-layer graph convolution module, captures the evolution trend of each gland at different time periods through temporal convolution, fuses the linkage features between glands in the multi-layer structure through graph convolution, and finally outputs a high-dimensional graph embedding vector for anomaly determination.

[0032] In this embodiment, S5 includes the following steps: S51. Construct an input feature set for anomaly determination based on the graph embedding representation. The input feature set for anomaly determination includes the graph embedding vectors of each target endocrine gland within the current time window. The graph embedding vectors comprehensively reflect the linkage relationship features between the standard physiological state of the endocrine gland and other endocrine glands under various physiological interaction mechanisms. S52. Construct an anomaly determination strategy. The anomaly determination strategy is constructed based on historical monitoring data and clinical annotation information, and a classification boundary condition for determining whether the current graph embedding representation shows an abnormal state is obtained through training. S53. Perform anomaly determination on the graph embedding vectors of each target endocrine gland to determine whether there are abnormal node patterns and abnormal edge patterns. S54. Mark the target endocrine gland or the relationship between endocrine glands with abnormal node patterns and abnormal edge patterns, and output the anomaly monitoring result. The anomaly monitoring result includes node anomaly information and edge anomaly information. S55. Output the anomaly monitoring result, and the anomaly monitoring result is used as anomaly warning information for clinical auxiliary decision-making.

[0033] In this embodiment, the abnormal node pattern is that when the distance distribution of the graph embedding vector in the high-dimensional embedding space significantly deviates from the clustering center of normal training samples, or the embedding eigenvalue exceeds the preset physiological reasonable threshold range, it is determined that the target endocrine gland has an individual functional abnormality within the current time window. The abnormal edge pattern is that when the linkage representation value between a certain target endocrine gland and its associated endocrine gland in the graph embedding representation significantly deviates from the historical cooperation pattern in the normal state, or there is a situation where the decline speed of the linkage intensity is higher than the threshold, or the increase speed is higher than the threshold, it is determined that there is a linkage abnormality or feedback abnormality between the endocrine gland pairs.

[0034] In this embodiment, the anomaly monitoring result includes node anomaly information and edge anomaly information, which respectively record the numbers of the endocrine glands determined to be abnormal, the types of anomalies, and the index values of their degrees of anomalies within the current time window.

[0035] The node abnormal information of the present invention is classified as individual endocrine gland abnormality, and the judgment condition is that if the distribution value of the graph embedding vector in the standard physiological embedding space meets one of the following rules: the hormone concentration embedding component is lower than 0.2 or higher than 0.85, the change rate of the neuroendocrine signal embedding component is greater than the set dynamic fluctuation threshold, and the rhythmic parameter embedding feature presents a non-periodic characteristic pattern; The edge anomaly information is classified as an anomaly in the linkage between endocrine glands. The judgment condition is that if the graph embedding linkage component between the two target endocrine glands deviates continuously from the historical mean interval for three time windows, or the sensitivity mutation rate of the pair of endocrine glands in the collaborative perturbation sensitivity matrix exceeds the set threshold.

[0036] An endocrine health monitoring system based on artificial intelligence is used to perform an endocrine health monitoring method based on artificial intelligence, including the following modules: The physiological signal acquisition module is used to collect hormone concentrations, neuroendocrine signals and physiological rhythm parameters of multiple target endocrine glands, and perform denoising, normalization and time alignment to generate a standard physiological signal data set; A graph construction module is used to construct a preliminary physiological state graph based on a standard physiological signal dataset, extract the mutual information between gland states and the joint entropy to construct edge weights; A heterogeneous map expansion module is used to refine the preliminary physiological state map into a multi-layer heterogeneous physiological state map; Graph convolutional neural network module, which is used to fuse standard physiological signal datasets with multi-layer heterogeneous physiological state maps, extract the dynamic evolution characteristics and linkage structure characteristics of target endocrine glands, and generate graph embedding representations; The anomaly recognition module is used to identify node or edge anomalies based on graph embedding representation and preset anomaly judgment strategy, and output anomaly monitoring results.

[0037] Embodiment 1: In the third consulting room of the Endocrinology Department of A City Hospital, a female patient numbered P076 was connected to the "AI-Endocrinology Linkage Monitoring System" for continuous monitoring during a routine physical examination. The patient's name is Wang Mou, 44 years old. The main complaints are difficulty falling asleep, palpitations upon waking up, and mood swings in the past two weeks. Although the static blood test data shows that her thyroid hormones T3 (2.3 nmol / L) and TSH (3.2 μIU / mL) are within the normal range, the system still included her in the observation group. At 08:41, the on-duty nurse attached the multi-channel acquisition device numbered EN-MULTI-V2.1 to the patient's carotid artery node, sternocleidomastoid muscle area, and the left lower abdomen skin in sequence through wireless skin conductance and posterior ear nerve interface. After the device was started, it began to collect three types of real-time signals: including the diffusion intensity of T3 / T4 hormones (obtained through infrared subcutaneous spectroscopy sensor), neuroendocrine pulse signals (collected from the burst mode of vagus nerve electrical signals), and rhythm parameters (output of the HRV + skin temperature two-channel time domain model).

[0038] The patient was in a resting state, and the system completed a standard sampling. The server automatically received and recorded the original physiological signals of 6 endocrine-related glands during this period, named RawData_P076_20241106T0910.json, with a total of 1020 entries.

[0039] After the data upload was completed, the system background completed the denoising process and generated a standard signal data set. During the initial generation of the graph structure, the system found that the state co-information between the pituitary gland and the thyroid node of P076 was 0.46, which decreased significantly compared to the average value of the normal range (0.62). At the same time, the nerve signal pulse frequency of the hypothalamus-pituitary-adrenal chain suddenly increased (from the normal 0.8 Hz to 1.5 Hz). The system marked this phenomenon as a "physiological disturbance risk candidate" and recorded the initial values of the graph construction as follows: Sampling time: 2024-11-06T09:15:35; Graph nodes: 6 gland nodes, number of layers: 3 layers (hormone, nerve, rhythm); Edge weight anomaly: thyroid - pituitary (weight deviation -0.16); Response of the perturbation matrix: the sensitivity improvement rate of the hypothalamus to adrenal gland co-perturbation reached +212%.

[0040] The processing of the graph convolutional neural network was completed. The graph embedding vector of P076 deviated from the distribution center of the healthy population, with an Euclidean distance of 2.89 and Z-score = 3.27. The system automatically triggered the "red warning" mechanism and marked P076 as suspected of "latent progression of subclinical hyperthyroidism".

[0041] The abnormal monitoring module generated a report `Alert_P076_20241106T0925.pdf` in the server log, and the content is as follows: Nodal abnormality: pituitary, abnormality type = node distribution shift, abnormality score = 0.91; Side abnormalities: pituitary ↔ thyroid, the linkage speed drops suddenly (37% drop in 3 hours); Recommended actions: Repeat T3, FT4, and thyrotropin suppression test within 48 hours; Anomaly detection time: 2024-11-06 09:22:43.

[0042] The attending physician received the early warning push and called up the patient linkage status map generated by the system, confirming that the abnormal area was concentrated in the T3 neuroregulatory layer. The doctor recorded the manual intervention suggestion: "Extend the observation period + daytime hormone dynamic monitoring".

[0043] To verify the effectiveness of this system, the system compared and analyzed the traditional method of parallel access (using three static T3 values ​​for trend line regression), and the results are as follows: Table 1 Comparative data of the present invention and the traditional method

[0044] Mr. Wang suffered from sudden heart palpitations and went to the emergency department for a follow-up visit. Clinical dynamic testing confirmed that his TSH concentration had dropped sharply to 0.09μIU / mL, and he was diagnosed with "early progressive hyperthyroidism", which was completely consistent with the system's warning information two days ago.

[0045] The research team recorded the entire incident and archived it as the "Case P076 Monitoring Log (v2)" and included it in the subsequent algorithm model retraining dataset.

[0046] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An endocrine health monitoring method based on artificial intelligence, characterized in that: The steps include: S1. synchronously collecting physiological signal data sets from multiple endocrine glands using a multi-channel real-time acquisition device, preprocessing the collected physiological signal data sets, and generating a preprocessed standard physiological signal data set; S2. construct a preliminary physiological state map based on the preprocessed standard physiological signal data set, correspond each endocrine gland to a node in the physiological state map, and construct edges between nodes according to the physiological interaction relationship between the glands to form a preliminary physiological state map; S3. Expand and refine the preliminary physiological state map to construct a multi-layer heterogeneous physiological state map, where different layers represent different types of physiological interaction mechanisms, and fuse the information of nodes and edges in each layer to form a combined physiological state map that comprehensively reflects multiple physiological interaction mechanisms; S4. The combined physiological state map and the preprocessed standard physiological signal dataset are used as input and passed to the spatiotemporal fusion graph convolutional neural network model, and the output is a graph embedding representation reflecting the states of each endocrine gland and their dynamic linkage relationship; S5. Based on the graph embedding representation, the preset abnormality judgment strategy is used to analyze and judge the linkage status of each gland in the endocrine system, detect the nodes or edge relationships with abnormal patterns, and output the abnormal monitoring results, which are used as abnormal warning information.

2. The endocrine health monitoring method based on artificial intelligence according to claim 1, characterized in that: The S1 comprises the following steps: S11. Collecting raw physiological signal data from multiple endocrine glands, setting a target endocrine gland set G for collection, and for each endocrine gland in the target endocrine gland set G , set the original data of physiological signals ,in, Indicates The sampling time, Indicates at time Endocrine glands collected The hormone concentration, represents the neuroendocrine signal value, represents the physiological rhythm parameters, is the number of sampling times within the total sampling time; S12. The original sampling data of all endocrine glands are combined to form a complete physiological signal original data set, and the physiological signal original data set represents a unified data structure of multiple endocrine glands and multiple types of indicators in a continuous time series; S13. Preprocess the original data set of physiological signals, use the weighted moving average method to denoise the signals, normalize the denoised data, and use the minimum-maximum normalization method to normalize each type of signal to the interval , time-align the normalized data, and perform sampling on each endocrine gland based on a unified sampling time axis. The normalized data are linearly interpolated or sample interpolated according to time, so that the data of each endocrine gland at all time points The corresponding values ​​are obtained to obtain the standard physiological signal vector after complete time alignment. The standard physiological signal vectors of all endocrine glands at each time point are organized to form a preprocessed standard physiological signal data set with a unified structure. .

3. The endocrine health monitoring method based on artificial intelligence according to claim 2, characterized in that: The S2 comprises the following steps: S21. Based on the preprocessed standard physiological signal dataset , construct a preliminary physiological state map containing structural information and state dependency information. The preliminary physiological state map is a structured graph, including a node set, an edge set, a node feature matrix and a weighted adjacency matrix. The node set V is used to represent multiple target endocrine glands involved in monitoring. The nodes correspond to the target endocrine glands one by one. The node feature matrix X is used to represent the standard physiological state information of each target endocrine gland in the current time window. Each node feature vector consists of the normalized value of hormone concentration, the normalized value of neuroendocrine signal value and the normalized value of physiological rhythm parameter. S22. Introduce information dependency based on joint state changes to build a structured graph structure, use the dynamic mutual information of physiological states between nodes to calculate the interaction strength between endocrine glands, and construct an edge set and the weighted adjacency matrix The information dependence strength between the physiological states of each two different target endocrine glands is calculated. The information dependence strength uses the state mutual information and the state joint entropy as evaluation indicators to reflect the state linkage synergy of the two target endocrine glands in the standard physiological state feature space. After the information dependence strength calculation is completed between all target endocrine glands, the mutual information threshold is set, and only the connection relationship between the gland pairs whose information dependence strength exceeds the set threshold is retained. Finally, a set of edges reflecting high state synergy is formed, and the connection strength is recorded as the edge weight in the weighted adjacency matrix. ; S23. A weighted adjacency matrix is ​​constructed based on edge weights. The edge weights recorded in the weighted adjacency matrix are used to represent the state linkage dependency between two target endocrine glands. The larger the edge weight value, the stronger the synchronization and mutual responsiveness of the state changes between the target endocrine glands in the current time window. No connection is constructed between target endocrine gland pairs that do not exceed the threshold, and the value of the corresponding position in the adjacency matrix is ​​set to zero. S24. While constructing the preliminary physiological state map, the collaborative perturbation sensitivity matrix S is introduced to characterize the state perturbation response degree between the target endocrine glands. The collaborative perturbation sensitivity indicates the degree of synchronous change caused to the standard physiological state of other target endocrine glands when the standard physiological state of one target endocrine gland is disturbed in the current time window. The higher the collaborative perturbation sensitivity of the endocrine gland pair, the stronger its linkage reaction is, which has an indicative significance for endocrine system abnormalities. S25. Comprehensive node set , node feature matrix , weighted adjacency matrix , cooperative disturbance sensitivity matrix , constructing a preliminary physiological state map driven by state-dependent structures among multiple endocrine glands .

4. The method for endocrine health monitoring based on artificial intelligence according to claim 3, characterized in that: The S3 comprises the following steps: S31. Based on preliminary physiological state map Classify the physiological interaction mechanisms between target endocrine glands and set the physiological interaction mechanism set , a collection of physiological interaction mechanisms Including hormone regulation mechanism, neural regulation mechanism and metabolic feedback mechanism, physiological interaction mechanism collection Each type of physiological interaction mechanism in represents a different type of glandular linkage; S32. For the physiological interaction mechanism set For each type of physiological interaction mechanism in the preliminary physiological state map, the graph structure information related to the physiological interaction mechanism is extracted from the preliminary physiological state map, and the corresponding layer is constructed. Each layer contains a node set, a node feature matrix, an edge set and a weighted adjacency matrix. The node set and the node feature matrix in the layer are consistent with the preliminary physiological state map. The edge set is the target endocrine gland connection relationship set related to the physiological interaction mechanism. The edge weight recorded in the weighted adjacency matrix represents the state linkage strength between the target endocrine glands under the physiological interaction mechanism. S33. The structural information in all layers is uniformly fused, and a fusion weight vector is set. Each weight value in the fusion weight vector is used to represent the contribution of the corresponding layer in the fusion process. The weighted adjacency matrix in the layer is weighted and summed according to the fusion weight corresponding to each layer to generate a fused joint weighted adjacency matrix. The weight value of each position in the joint weighted adjacency matrix represents the joint linkage strength of the target endocrine glands under the joint action of all physiological interaction mechanisms. S34. The node sets, edge sets, node feature matrices and weighted adjacency matrices of all layers are combined with the node sets, edge sets, node feature matrices and joint weighted adjacency matrices of the fused layers to form a multi-layer heterogeneous physiological state map set , each layer reflects the target endocrine gland linkage relationship under a specific physiological interaction mechanism, and the fusion layer is used to reflect the system-level target endocrine gland linkage structure under the joint action of all physiological interaction mechanisms.

5. The method for endocrine health monitoring based on artificial intelligence according to claim 4, characterized in that: The S4 comprises the following steps: S41. Input the multi-layer heterogeneous physiological state map set and the preprocessed standard physiological signal data set as input data to the spatiotemporal fusion graph convolutional neural network model, wherein the input data includes the weighted adjacency matrix, node feature matrix and layer identification information of each layer in the multi-layer heterogeneous physiological state map; S42. Divide the preprocessed standard physiological signal data set into time windows and set the length of the continuous time window to , each target endocrine gland is divided into two time dimensions of length The time series segments are used to form a time-node joint input tensor, which represents the multi-dimensional standard physiological state change characteristics of all target endocrine glands in each time window; S43. Construct a spatiotemporal fusion graph convolutional neural network model, wherein the spatiotemporal fusion graph convolutional neural network model structure is composed of a temporal convolution module and a graph structure convolution module alternately stacked, wherein the temporal convolution module performs a one-dimensional convolution operation on the standard physiological state feature sequence of each target endocrine gland in a continuous time window based on the time-node joint input tensor, so as to extract the dynamic evolution characteristics of the target endocrine gland in the time dimension; S44. Perform graph convolution operation on the output result of the time convolution module and the structural information of each layer in the multi-layer heterogeneous physiological state atlas set. The graph structure convolution module performs multi-layer graph convolution calculation based on the weighted adjacency matrix and node feature matrix between the target endocrine glands in the graph structure and the layer identification information, extracts the linkage relationship features between the target endocrine glands under different physiological interaction mechanisms, and outputs the graph convolution output representation of each gland under the layer; S45. performing layer aggregation processing on the graph convolution output representations of all layers, and performing weighted combination on the multi-graph convolution output representations according to the layer fusion weights to form a graph embedding representation of each target endocrine gland under the comprehensive effects of all physiological interaction mechanisms; S46. Output graph embedding representation, where the graph embedding representation is a high-dimensional state representation of each target endocrine gland in the current time window, which is used to reflect the individual state characteristics of the gland and its dynamic linkage relationship under the multi-endocrine gland linkage mechanism.

6. The method for endocrine health monitoring based on artificial intelligence according to claim 5, characterized in that: The S5 comprises the following steps: S51. construct an abnormality determination input feature set based on the graph embedding representation, the abnormality determination input feature set includes a graph embedding vector of each target endocrine gland in the current time window, and the graph embedding vector comprehensively reflects the linkage relationship characteristics between the standard physiological state of the endocrine gland and other endocrine glands under multiple physiological interaction mechanisms; S52. Construct an abnormality determination strategy, which is constructed based on historical monitoring data and clinical annotation information, and obtains classification boundary conditions for determining whether the current graph embedding representation is in an abnormal state through training; S53. Perform an abnormality determination on the graph embedding vector of each target endocrine gland to determine whether there is an abnormal node pattern and an abnormal edge pattern; S54. Mark the target endocrine glands or the relationship between endocrine glands with abnormal node patterns and abnormal edge patterns, and output abnormal monitoring results, which include node abnormality information and edge abnormality information; S55. Output abnormal monitoring results, which are used as abnormal warning information for clinical decision support.

7. The method for endocrine health monitoring based on artificial intelligence according to claim 6, characterized in that: The abnormal node mode is when the distance distribution of the graph embedding vector in the high-dimensional embedding space significantly deviates from the normal training sample cluster center, or the embedding feature value exceeds the preset physiological reasonable threshold range, it is determined that the target endocrine gland has individual sexual dysfunction in the current time window; the abnormal edge mode is when the linkage representation value between a target endocrine gland and its associated endocrine glands in the graph embedding representation significantly deviates from the historical coordination pattern under normal conditions, or there is a situation where the linkage intensity decreases at a speed higher than a threshold, or increases at a speed higher than a threshold, it is determined that there is linkage abnormality or feedback abnormality between the endocrine gland pairs.

8. The method for endocrine health monitoring based on artificial intelligence according to claim 6, characterized in that: The abnormal monitoring results include node abnormality information and edge abnormality information, which respectively record the number of the endocrine gland determined to be abnormal in the current time window, the abnormality type and its abnormality degree index value.

9. The method for endocrine health monitoring based on artificial intelligence according to claim 8, characterized in that: The node abnormality information is classified as individual endocrine gland abnormality, and the judgment condition is that if the distribution value of the graph embedding vector in the standard physiological embedding space meets one of the following rules: the hormone concentration embedding component is lower than 0.2 or higher than 0.85, the change rate of the neuroendocrine signal embedding component is greater than the set dynamic fluctuation threshold, and the rhythmic parameter embedding feature presents a non-periodic characteristic pattern; The edge anomaly information is classified as an abnormal linkage between endocrine glands. The judgment condition is that if the graph embedded linkage component between two target endocrine glands deviates continuously from the historical mean interval for three time windows, or the sensitivity mutation rate of the pair of endocrine glands in the collaborative perturbation sensitivity matrix exceeds the set threshold.

10. An artificial intelligence-based endocrine health monitoring system, used to execute an artificial intelligence-based endocrine health monitoring method according to any one of claims 1 to 9, characterized in that: Includes the following modules: The physiological signal acquisition module is used to collect hormone concentrations, neuroendocrine signals and physiological rhythm parameters of multiple target endocrine glands, and perform denoising, normalization and time alignment to generate a standard physiological signal data set; A graph construction module is used to construct a preliminary physiological state graph based on a standard physiological signal dataset, extract the mutual information between gland states and the joint entropy to construct edge weights; A heterogeneous map expansion module is used to refine the preliminary physiological state map into a multi-layer heterogeneous physiological state map; Graph convolutional neural network module, which is used to fuse standard physiological signal datasets with multi-layer heterogeneous physiological state maps, extract the dynamic evolution characteristics and linkage structure characteristics of target endocrine glands, and generate graph embedding representations; The anomaly recognition module is used to identify node or edge anomalies based on graph embedding representation and preset anomaly judgment strategy, and output anomaly monitoring results.

Images