State data processing method and device, computer device and storage medium

Abstract

The application relates to a state data processing method and device, computer equipment, a storage medium and a computer program product, which comprise the following steps: obtaining a historical diagnosis state set corresponding to a target object at different historical times; determining an associated diagnosis state corresponding to a historical diagnosis state in the historical diagnosis state set and a relationship type corresponding to the historical diagnosis state from a state relationship network; determining the correlation degree of a to-be-predicted state and the historical diagnosis state under the relationship type, determining the propagation feature representation for the to-be-predicted state based on the correlation degree and the associated diagnosis state; performing attention processing on the propagation feature representations under each historical diagnosis state set respectively to determine the state response feature representation of the target object; determining the object feature representation of the target object based on the state response feature representation, and determining the possibility information that the target object is in the to-be-predicted state based on the object feature representation. The method can improve the accuracy of diagnosis state determination.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a state data processing method, apparatus, computer equipment, storage medium, and computer program product. Background Technology

[0002] With the development of computer technology, computer devices are offering increasingly richer functions. In some cases, computer devices can predict the diagnostic status of a target object. For example, computer devices can predict future malfunctions of a machine based on its historical failures, and computer devices can predict future illnesses of a patient based on their past medical history.

[0003] In related technologies, the diagnostic status of a target object is usually predicted by classification. However, classification can only be performed in a few fixed categories, resulting in low accuracy when determining the diagnostic status. Summary of the Invention

[0004] Therefore, it is necessary to provide a state data processing method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy of diagnostic state determination in response to the above-mentioned technical problems.

[0005] On one hand, this application provides a state data processing method. The method includes: acquiring a set of historical diagnostic states corresponding to a target object at different historical times; determining, from a pre-established state relationship network, associated diagnostic states corresponding to the historical diagnostic states in the targeted set of historical diagnostic states, and the relationship type between the historical diagnostic states and the associated diagnostic states; the state relationship network is used to describe the association between diagnostic states; determining the correlation degree between the state to be predicted and the historical diagnostic states under the relationship type; based on the correlation degree and the associated diagnostic states, determining, under the targeted set of historical diagnostic states, the state relationship network performs attention processing on the propagation feature representation of the state to be predicted for each set of historical diagnostic states; based on the attention processing results corresponding to each set of historical diagnostic states, determining the state response feature representation of the target object for the state to be predicted; determining the object feature representation of the target object based on the state response feature representation; and determining the probability information of the target object being in the state to be predicted based on the object feature representation.

[0006] On the other hand, this application also provides a state data processing apparatus. The apparatus includes: a historical state determination module, used to acquire a set of historical diagnostic states corresponding to a target object at different historical times; an associated state determination module, used to determine, from a pre-established state relationship network, associated diagnostic states corresponding to historical diagnostic states in the targeted set of historical diagnostic states, and the relationship type between the historical diagnostic states and the associated diagnostic states; the state relationship network is used to describe the association relationships between diagnostic states; a propagation feature derivation module, used to determine the degree of association between the state to be predicted and the historical diagnostic states under the relationship type, and based on the degree of association and the associated diagnostic states, determine the propagation feature representation of the state relationship network under the targeted set of historical diagnostic states; an attention processing module for the propagation feature representation of the state to be predicted, used to perform attention processing on the propagation feature representation under each set of historical diagnostic states, and based on the attention processing results corresponding to each set of historical diagnostic states, determine the state response feature representation of the target object for the state to be predicted; an object feature determination module, used to determine the object feature representation of the target object based on the state response feature representation; and a probability prediction module, used to predict the probability information of the target object being in the state to be predicted based on the object feature representation.

[0007] On the other hand, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps described in the above-described state data processing method.

[0008] On the other hand, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps described in the above-described state data processing method.

[0009] On the other hand, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps described in the above-described state data processing method.

[0010] The aforementioned state data processing method, apparatus, computer equipment, storage medium, and computer program product acquire historical diagnostic state sets corresponding to the target object at different historical times. From a pre-established state relationship network, they determine the associated diagnostic states corresponding to the historical diagnostic states in the targeted historical diagnostic state sets, as well as the relationship types between the historical diagnostic states and the associated diagnostic states. They determine the correlation degree between the state to be predicted and the historical diagnostic states under the relationship type. Based on the correlation degree and the associated diagnostic states, they determine the propagation feature representation of the state relationship network for the state to be predicted under the targeted historical diagnostic state sets. They then perform attention processing on the propagation feature representations under each historical diagnostic state set. Based on the attention processing results, they determine the state response feature representation of the target object for the state to be predicted. Based on the state response feature representation, they determine the object feature representation of the target object. Based on the object feature representation, they predict the probability information of the target object being in the state to be predicted. This enables prediction of any state to be predicted. Since the predicted probability information is obtained using the state relationship network and the historical diagnostic state information of the target object, it can simulate the evolution of the target object's diagnostic state, thereby improving the accuracy of determining the diagnostic state. Furthermore, by applying attention processing to the propagation feature representations under each set of historical diagnostic states, the differences in the impact of diagnostic state information at different historical times on the prediction results can be taken into account, resulting in more accurate object feature representations and thus more accurate probability information. Attached Figure Description

[0011] Figure 1 This is an application environment diagram of a state data processing method in one embodiment;

[0012] Figure 2 This is a flowchart illustrating a state data processing method in one embodiment;

[0013] Figure 3 This is a flowchart illustrating the attention processing steps in one embodiment;

[0014] Figure 4 This is a flowchart illustrating the process of determining object feature representation in one embodiment;

[0015] Figure 5 This is a schematic diagram illustrating the propagation process of a state relation network in one embodiment;

[0016] Figure 6 This is an overall flowchart of a state data processing method in one embodiment;

[0017] Figure 7 This is a flowchart illustrating the classification model in one embodiment;

[0018] Figure 8This is a schematic diagram of the structure of a diagnostic state prediction model in one embodiment;

[0019] Figure 9 This is a schematic diagram of the structure of the same-stage propagation module in one embodiment;

[0020] Figure 10 This is a schematic diagram of the information integration module in one embodiment;

[0021] Figure 11 This is a flowchart illustrating an application scenario of the diagnostic state prediction model in one embodiment;

[0022] Figure 12 This is a structural block diagram of a state data processing device in one embodiment;

[0023] Figure 13 This is an internal structural diagram of a computer device in one embodiment;

[0024] Figure 14 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0027] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and intelligent transportation.

[0028] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0029] The artificial intelligence-related technologies involved in the solutions provided in this application are specifically illustrated through the following embodiments:

[0030] The state data processing method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on other servers. Terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Server 104 can be implemented using a standalone server or a server cluster composed of multiple servers. Terminal 102 can send a diagnostic state prediction request for a target object to server 104, specifying the state to be predicted. The state to be predicted can be any diagnostic state. After receiving the diagnostic state prediction request, server 104 obtains the historical diagnostic state sets corresponding to the target object at different historical times. Based on multiple historical diagnostic state sets, the state to be predicted, and a pre-established state relationship network, server 104 determines the probability information of the target object being in the state to be predicted, and then returns the probability information to terminal 102.

[0031] It should be noted that the state data processing method provided in this application embodiment can also be executed by the terminal 102 alone or by the server 104 alone. For example, after determining the target object and the state to be predicted, the terminal 102 can obtain the historical diagnostic state set corresponding to the target object at different historical times, and then determine the probability information of the target object being in the state to be predicted based on multiple historical diagnostic state sets, the state to be predicted and the pre-established state relationship network. Alternatively, the server 104 can predict the state to be predicted for each target object in the target object set according to a preset period and save the predicted probability information.

[0032] In one embodiment, such as Figure 2 As shown, a state data processing method is provided, which can be executed by a computer device, the computer device being... Figure 1 The terminal 102 or the server 104. In this embodiment, a computer device is used as the server for example, and the specific steps include:

[0033] Step 202: Obtain the set of historical diagnostic statuses of the target object at different historical times.

[0034] The target object refers to the object whose diagnostic state is to be predicted. This object can be an independent living entity or a non-living entity. Living entities can be, for example, natural persons or animals, while non-living entities can be, for example, machines or vehicles. In specific implementations, in the scenario of predicting a patient's disease, the target object can be the patient whose disease is to be predicted; in the scenario of predicting a production machine malfunction, the target object can be the production machine whose malfunction is to be predicted; and in the scenario of predicting a vehicle malfunction, the target object can be the vehicle whose malfunction is to be predicted. A historical diagnostic state set refers to a set of one or more historical diagnostic states from the same historical time period, where "multiple" means at least two. Different historical times correspond to different historical diagnostic state sets, and historical diagnostic states within the same historical diagnostic state set correspond to the same historical time. A historical diagnostic state refers to the diagnostic state of the target object at the corresponding historical time. A diagnostic state refers to a state that can be determined through diagnosis. It is understandable that the meaning of the diagnostic state is different for different types of target objects. For example, in the scenario of predicting a patient's disease, the diagnostic state can be the patient's disease; in the scenario of predicting a production machine malfunction, the diagnostic state can be the production machine's malfunction; and in the scenario of predicting a vehicle malfunction, the diagnostic state can be the vehicle's malfunction.

[0035] Specifically, the server can query the pre-established diagnostic status database based on the object identifier of the target object to obtain the set of historical diagnostic statuses of the target object at different historical times, or the server can obtain the set of historical diagnostic statuses of the target object at different historical times from other computer devices via wired or wireless means.

[0036] Step 204: Determine the associated diagnostic states corresponding to the historical diagnostic states in the target set of historical diagnostic states from the pre-established state relationship network, as well as the relationship type between the historical diagnostic states and the associated diagnostic states.

[0037] In this context, the state relation network refers to a network used to describe the relationships between diagnostic states. Each node in the state relation network represents a diagnostic state, and the edges connecting nodes represent the relationships between diagnostic states. These relationships can characterize the evolution of diagnostic states; for example, they can represent the probability of two diagnostic states occurring simultaneously, or the probability that when one diagnostic state occurs, the other will also occur. In a practical example, in the scenario of predicting a patient's disease, the relationship between diagnostic states can be a comorbidity relationship between two diseases. Comorbidity, also known as a comorbidity, refers to one or more diseases that occur simultaneously with the primary disease. The associated diagnostic states corresponding to historical diagnostic states refer to diagnostic states that are associated with historical diagnostic states. The relation type refers to the category to which the relation belongs. For example, for a relation representing the probability of two diagnostic states occurring simultaneously, the probability can be divided into ranges, with each range corresponding to a relation type.

[0038] Specifically, for each historical diagnostic state in each set of historical diagnostic states, the server can determine the associated diagnostic state corresponding to that historical diagnostic state and the relationship type between that historical diagnostic state and its corresponding associated diagnostic state from a pre-established state relationship network.

[0039] In one specific embodiment, the server can determine, from a pre-established state relationship network, the nodes directly connected to the nodes corresponding to the historical diagnostic states via connecting edges (i.e., first-order neighbor nodes). The diagnostic state represented by the node is then identified as the associated diagnostic state corresponding to the historical diagnostic state. The relationship type represented by the connecting edge between these two nodes is the relationship type between the historical diagnostic state and the associated diagnostic state. In another specific embodiment, the state relationship network can be a knowledge graph, which can be constructed based on the hierarchical information and prior information of the diagnostic states. For example, the evolutionary path between diagnostic states can be determined based on the hierarchical information and prior information of the diagnostic states, and then the knowledge graph can be constructed based on the evolutionary path.

[0040] Step 206: Determine the correlation between the state to be predicted and the historical diagnostic states under the relationship type. Based on the correlation and the associated diagnostic states, determine the propagation characteristics of the state relationship network for the state to be predicted under the target set of historical diagnostic states.

[0041] Here, the state to be predicted refers to the diagnostic state to be predicted. The correlation between the state to be predicted and historical diagnostic states under the relationship type refers to the degree of correlation between the two states under that relationship type; a higher correlation indicates a stronger correlation between them. In a specific embodiment, the correlation between the state to be predicted and historical diagnostic states under the relationship type can specifically be the similarity between them. The propagation feature representation is used to represent the feature information propagated by the state relationship network towards the state to be predicted within the target set of historical diagnostic states.

[0042] Specifically, for each historical diagnostic state in each set of historical diagnostic states, the server can calculate the correlation degree between the predicted state and the historical diagnostic state based on the diagnostic state representation features of the state to be predicted, the diagnostic state representation features of the historical diagnostic state, and the relation representation features corresponding to the relation type. When the state relation network propagates to the state to be predicted within each set of historical diagnostic states, it obtains the propagation feature representation corresponding to each set of historical diagnostic states. Since propagation in the state relation network follows paths between nodes, the propagated features can be obtained from the node features along the propagation path. The nodes along the propagation path are the nodes corresponding to the associated diagnostic states of the historical diagnostic states. Therefore, the propagation feature representation corresponding to the historical diagnostic state can be determined based on the correlation degree and the associated diagnostic states. Finally, the propagation feature representations corresponding to each historical diagnostic state in the same set of historical diagnostic states are statistically analyzed. For example, the propagation feature representations corresponding to each historical diagnostic state in the same set can be added together to obtain the propagation feature representation corresponding to each set of historical diagnostic states.

[0043] In one specific embodiment, the server can multiply the calculated correlation degree by the diagnostic status representation feature of the associated diagnostic status corresponding to the historical diagnostic status to obtain the propagation feature representation corresponding to the historical diagnostic status.

[0044] Step 208: Attention processing is performed on the propagation feature representations under each historical diagnostic state set. Based on the attention processing results corresponding to each historical diagnostic state set, the state response feature representation of the target object to the state to be predicted is determined.

[0045] The state response feature representation of the target object to the state to be predicted refers to the response information of the target object's possible diagnostic states to the state to be predicted. This state response feature representation is obtained by integrating the propagation feature representations corresponding to various historical diagnostic state sets.

[0046] Specifically, the server can perform attention processing on the propagation feature representations under each historical diagnostic state set based on the attention mechanism, and obtain the attention processing results corresponding to each historical diagnostic state set. Then, the attention processing results can be integrated. The multiple attention processing results correspond to different historical times of the target object. The integrated result can reflect the possible diagnostic state of the target object and its response information to the state to be predicted as a whole. Therefore, by integrating the attention processing results corresponding to each historical diagnostic state set, the state response feature representation of the target object to the state to be predicted can be obtained.

[0047] In a specific embodiment, integrating the attention processing results can specifically involve pooling the various attention processing results.

[0048] Step 210: Determine the object feature representation of the target object based on the state response feature representation, and determine the probability information of the target object being in the state to be predicted based on the object feature representation.

[0049] The probability information of the target object being in a state to be predicted is used to indicate the likelihood of the target object being in a state to be predicted. In one specific embodiment, the probability information may be, for example, the probability level of being in a state to be predicted; the higher the level, the greater the likelihood. In other specific embodiments, the probability information may also be the probability of being in a state to be predicted; the higher the probability, the greater the likelihood. In one specific embodiment, the probability information of the target object being in a state to be predicted can be used to indicate the likelihood that the target object will be in a state to be predicted, that is, the likelihood that the target object will be in a state to be predicted after the historical time corresponding to each set of historical diagnostic states.

[0050] Object feature representation is used to represent the features of a target object. Since the embodiments of this application are for predicting the diagnostic state of the target object, that is, to predict the probability of the target object being in the state to be predicted, the feature information of the target object about the disease to be predicted can be used as the object feature representation. Therefore, the object feature representation can be determined based on the state response feature representation of the target object to the state to be predicted.

[0051] Specifically, after determining the state response feature representation of the target object in response to the state to be predicted, the server can determine the object feature representation of the target object based on the state response feature representation, and then determine the probability information of the target object being in the state to be predicted based on the obtained object feature representation.

[0052] In one specific embodiment, the server can determine the obtained state response feature representation as the object feature representation of the target object. In other embodiments, the server can continue to perform multiple propagation calculations in the state relationship network based on the state response feature representation. Each propagation calculation will obtain the state response feature representation corresponding to that propagation stage. Finally, the server can combine the state response feature representations obtained from each propagation stage to obtain the object feature representation. Since multiple propagations have been performed, it is equivalent to simulating the diagnostic state evolution of the target object. The obtained object feature representation can better reflect the feature information of the target object, thereby predicting more accurate probability information.

[0053] In one specific embodiment, the server determines the probability information of a target object being in a state to be predicted based on the object feature representation. Specifically, this can be achieved by performing similarity calculations based on the object feature representation and the state feature representation of the state to be predicted, to obtain the probability that the target object is in the state to be predicted. The state feature representation of the state to be predicted is used to characterize the features of the state to be predicted, and can be obtained by feature mapping of the state to be predicted. Feature mapping can be done by mapping the features of the state to be predicted to a vector space, and the resulting state feature representation is a feature vector used to represent the state to be predicted.

[0054] In a specific embodiment, steps 206 to 210 can be obtained through a pre-trained diagnostic state prediction model. This diagnostic state prediction model refers to a machine learning model used to predict the probability information of a target object being in a state to be predicted. The model includes multiple propagation computation layers and prediction layers. The propagation computation layers include a same-stage propagation layer, an attention layer, and a pooling layer. In step 206, the server can obtain the state feature representation of the state to be predicted, using the initial feature representation of the state to be predicted as the input feature of the first propagation computation layer. In the same-stage propagation layer of the first propagation computation layer, the correlation between the state to be predicted and the historical diagnostic states in each historical diagnostic state set is calculated based on the input feature. Based on the correlation and the associated diagnostic states, the propagation feature representation corresponding to each historical diagnostic state is determined. In step 208, the server can... In the attention layer of the first propagation computation layer, attention processing is performed on the propagation feature representations of each historical diagnostic state set. In the pooling layer of the first propagation computation layer, the attention processing results of each historical diagnostic state set are pooled, and the output of the pooling layer is the state response feature representation of the propagation computation layer. In step 210, the server uses the state response feature representation as the input feature of the next propagation computation layer. The same steps as the first propagation computation layer are performed in one propagation computation layer, and the obtained state response feature representation is continued to be input into the next propagation computation layer until the state response feature representation of the last propagation computation layer is obtained. Finally, the object feature representation of the target object is determined by combining the state response feature representations output by each propagation computation layer. Then, in the prediction layer, prediction is performed based on the object feature representation of the target object and the state feature representation of the state to be predicted to obtain the probability information of the target object being in the state to be predicted.

[0055] In the aforementioned state data processing method, by acquiring the historical diagnostic state sets corresponding to the target object at different historical times, the associated diagnostic states corresponding to the historical diagnostic states in the targeted historical diagnostic state sets are determined from the pre-established state relationship network, as well as the relationship type between the historical diagnostic states and the associated diagnostic states. The correlation degree between the state to be predicted and the historical diagnostic states under the relationship type is determined. Based on the correlation degree and the associated diagnostic states, the propagation feature representation of the state relationship network for the state to be predicted under the targeted historical diagnostic state sets is determined. Attention processing is performed on the propagation feature representation under each historical diagnostic state set. Based on the attention processing results corresponding to each historical diagnostic state set, the state response feature representation of the target object for the state to be predicted is determined. Based on the state response feature representation, the object feature representation of the target object is determined. Based on the object feature representation, the probability information of the target object being in the state to be predicted is determined, thus enabling prediction of any state to be predicted. Since the predicted probability information is obtained using the state relationship network and the historical diagnostic state information of the target object, the evolution of the target object's diagnostic state can be simulated, thereby improving the accuracy of the diagnostic state determination.

[0056] Furthermore, by applying attention processing to the propagation feature representations under each set of historical diagnostic states, the differences in the impact of diagnostic state information at different historical times on the prediction results can be taken into account, resulting in more accurate object feature representations and thus more accurate probability information.

[0057] In one embodiment, such as Figure 3 As shown, attention processing is applied to the propagation feature representations for each set of historical diagnostic states, including:

[0058] Step 302: Obtain the historical time information corresponding to each set of historical diagnostic statuses.

[0059] Specifically, considering that the impact of historical diagnostic state sets at different times on diagnostic state prediction is different, in this embodiment, the historical time information corresponding to each historical diagnostic state set can be obtained, and the time information can be integrated into the propagation feature representation, thereby reflecting the impact of time sequence information on diagnostic state prediction.

[0060] Step 304: Obtain the time feature representation corresponding to each historical time information, and fuse the propagation feature representation and time feature representation corresponding to each historical diagnostic state set to obtain the fused feature representation of each historical diagnostic state set.

[0061] Specifically, the server can obtain the time information of each historical diagnostic state set from the database storing the historical diagnostic state set, obtain the historical time information corresponding to each historical diagnostic state set, then perform time embedding on each historical time information to obtain the time feature representation corresponding to each historical time information, and then the server can fuse the time feature representation and propagation feature representation corresponding to the same historical diagnostic state set to obtain the fused feature representation of each historical diagnostic state set.

[0062] In one specific embodiment, the server may fuse the temporal feature representation and propagation feature representation corresponding to the same historical diagnostic state set by concatenating them to obtain a fused feature representation. In other embodiments, the server may perform the fusion process by adding the temporal feature representation and propagation feature representation corresponding to the same historical diagnostic state set to obtain a fused feature representation. Here, addition refers to element-wise addition, that is, adding the elements at each corresponding position. For example, assuming the temporal feature representation corresponding to a certain historical diagnostic state set is (x1, x2, ..., xn), and the propagation feature representation corresponding to this historical diagnostic state set is (y1, y2, ..., yn), then the obtained fused feature representation could be (x1+y1, x2+y2, ..., xn+yn).

[0063] Step 306: Perform attention processing on each fusion feature representation to obtain the attention processing results corresponding to each set of historical diagnostic states.

[0064] Specifically, the server can perform self-attention processing on each fusion feature representation to obtain the attention processing result corresponding to each set of historical diagnostic states. The self-attention processing process is to take each fusion feature representation as query information, extract the information that the query information is interested in from each fusion feature representation, and finally fuse the extracted information to obtain the attention processing result of the fusion feature representation.

[0065] In the above embodiments, since the temporal feature representations corresponding to each historical time information are integrated, the temporal information can be taken into account during the attention processing, thereby obtaining a more accurate state response feature representation and further improving the accuracy of diagnostic state prediction.

[0066] In one embodiment, the propagation feature representation includes elements of multiple dimensions. Obtaining the time feature representation corresponding to each historical time information includes: determining the number of dimensions corresponding to the propagation feature representation; and determining the time feature representation corresponding to the historical time information based on the historical time information, the number of dimensions, and the order of each element in the propagation feature representation.

[0067] The propagation feature representation includes elements of multiple dimensions, each element describing the propagated feature information from a different perspective. The number of dimensions of the propagation feature representation is the number of elements it includes. In a specific embodiment, the propagation feature representation can be, for example, a multi-dimensional vector, and the number of dimensions of the propagation feature representation is the same as the dimension of the vector. For example, assuming the propagation feature representation is an n-dimensional vector (x1, x2, ..., xn), then the number of dimensions of the propagation feature representation is n.

[0068] Specifically, the server can calculate the time feature representation corresponding to the historical time information based on the historical time information, the number of dimensions, and the order of each element in the propagation feature representation. When the order of the elements in the propagation feature representation is even, the calculation is performed using a sine function; when the order of the elements in the propagation feature representation is odd, the calculation is performed using a cosine function. For details, please refer to the following formulas (1) and (2) to obtain the time feature representation:

[0069]

[0070]

[0071] in, Let t represent the temporal feature of the e-th historical diagnostic state set. e This represents the time information corresponding to the set of historical diagnostic states, where i is the i-th position number in the propagation feature representation, d represents the number of dimensions in the propagation feature representation, and ω represents the position number in the propagation feature representation. k For frequency.

[0072] In a specific embodiment, after the server obtains the time feature representation through the above formula, it can further fuse the time feature representation and the propagation feature representation corresponding to the same set of historical diagnostic states. Specifically, refer to the following formula (3), in which the fusion is achieved by adding the time feature representation and the propagation feature representation element by element:

[0073]

[0074] Where, x e,1 For the obtained fused feature representation, dim emb v is the number of dimensions of the propagation feature representation. e,1 This represents the characteristics of propagation.

[0075] In this embodiment, the time feature representation corresponding to the historical time information is determined based on the historical time information, the number of dimensions, and the order of each element in the propagation feature representation. This allows for accurate encoding of the historical time information, resulting in a time feature representation that uniquely represents the historical time information.

[0076] In one embodiment, attention processing is performed on each fusion feature representation to obtain the attention processing result corresponding to each set of historical diagnostic states. This includes: converting each fusion feature representation into a query representation, a key representation, and a value representation respectively; calculating the relevance between the query representation corresponding to the target fusion feature representation and each key representation, and determining the attention weight corresponding to each key representation based on the relevance; multiplying the attention weight corresponding to each key representation with its corresponding value representation, calculating the resulting products, and obtaining the attention processing result corresponding to the target fusion feature representation based on the statistical results.

[0077] Specifically, the server can multiply each fused feature representation by the query parameter matrix, key parameter matrix, and value parameter matrix respectively to obtain the query, key, and value representations for each fused feature representation. The attention processing for each fused feature representation can then proceed as follows: the server calculates the relevance between the query and each key for each fused feature representation. This relevance calculation can be achieved, for example, by calculating the dot product between the query and the key. The calculated relevance is then normalized using softmax. The normalized result is the attention weight for each key representation. This attention weight represents the degree of attention the fused feature representation pays to each other; higher attention allows for the extraction of more information. Specifically, the server can multiply the attention weight corresponding to each key representation with its corresponding value representation, calculate the products, and obtain the attention processing result for the fused feature representation based on the statistical results. Here, the value representation corresponding to the key representation refers to the fact that both the key representation and the value representation are derived from the same fused feature representation. In a specific embodiment, calculating the products can be done by adding the products together to obtain the sum of the products, or by averaging the products.

[0078] In one specific embodiment, the server can convert each fused feature representation into a query representation, a key representation, and a value representation using an attention processing model. The query parameter matrix, the key parameter matrix, and the value parameter matrix can be parameters that the attention processing model needs to learn during training.

[0079] In the above embodiments, by converting the fusion feature representation into a query representation, a key representation, and a value representation, the relevance of the query representation corresponding to the target fusion feature representation to each key representation can be calculated. Based on the relevance, the attention weight corresponding to each key representation can be determined. Based on the attention weight, the information of interest can be extracted from each fusion feature representation to different degrees, thereby obtaining a more accurate attention processing result.

[0080] It should be noted that the embodiments of this application aim to use the state relationship network and the historical diagnostic state information of the target object to simulate the evolution of the diagnostic state of the target object, thereby predicting the future diagnostic state of the patient. In order to make the simulated evolution of the diagnostic state more accurate, the historical diagnostic state information of the target object can be propagated multiple times through the state relationship network. The specific implementation methods of multiple propagation will be described in detail in the following embodiments.

[0081] In one embodiment, such as Figure 4 As shown, the object feature representation of the target object is determined based on the state response feature representation, including:

[0082] Step 402: Take the associated diagnostic state set composed of the associated diagnostic states corresponding to each historical diagnostic state in the target historical diagnostic state set as the current diagnostic state set, and take the state response feature representation as the current input feature.

[0083] Specifically, after obtaining the state response feature representation, the server can use this representation as the current input feature for a second propagation calculation. In the second propagation calculation, the state feature representation of the state to be predicted in the first propagation calculation is replaced with the state response feature representation, and the same steps as in the first propagation are performed. In the second propagation calculation, the associated diagnostic states of the historical diagnostic states in each historical diagnostic state set form the associated diagnostic state set of that historical diagnostic state set, thereby obtaining the current diagnostic state set for each historical diagnostic state set.

[0084] Step 404: Determine the associated diagnostic state corresponding to the current diagnostic state in the current diagnostic state set, and the current relationship type between the current diagnostic state and the corresponding associated diagnostic state; the associated diagnostic state and the current relationship type corresponding to the current diagnostic state are determined from the state relationship network.

[0085] Specifically, the associated diagnostic state and current relationship type corresponding to the current diagnostic state can be determined through the following steps: Identify the nodes directly connected to the node corresponding to the current diagnostic state through the connection edge from the state relationship network, that is, the first-order neighbor nodes of the node corresponding to the current diagnostic state. Determine the diagnostic state represented by the node as the associated diagnostic state corresponding to the current diagnostic state. The relationship type represented by the connection edge between these two nodes is the current relationship type between the current diagnostic state and the associated diagnostic state.

[0086] It should be noted that the associated diagnostic status and current relationship type corresponding to the current diagnostic status can be determined and saved before the server executes step 404. Therefore, in step 404, the server can directly query and obtain the associated diagnostic status and current relationship type corresponding to the current diagnostic status.

[0087] Step 406: Determine the current correlation between the current input feature and the current diagnostic state under the current relation type. Based on the current correlation and the associated diagnostic state corresponding to the current diagnostic state, determine the current propagation feature representation of the state relation network under the current diagnostic state set for the state response feature representation.

[0088] The current correlation between the current input feature and the current diagnostic state under the current relationship type refers to the degree of correlation between the current input feature and the state feature representation of the current diagnostic state under the current relationship type. The higher the correlation, the higher the correlation between the current input feature and the current diagnostic state under the current relationship type. In a specific embodiment, the correlation here can specifically be the similarity between the current input feature and the state feature representation of the current diagnostic state under the current relationship type.

[0089] Specifically, for each current diagnostic state in each set of current diagnostic states, the server can calculate the correlation degree based on the current input features, the diagnostic state representation features of the current diagnostic state, and the relation representation features corresponding to the current relation type. This yields the current correlation degree between the current input features and the current diagnostic state under the current relation type. Then, based on the current correlation degree and the associated diagnostic states corresponding to the current diagnostic state, the server determines the propagation feature representation of the current diagnostic state. For example, the server can multiply the current correlation degree by the state feature representation of the associated diagnostic states corresponding to the current diagnostic state to obtain the propagation feature representation of the current diagnostic state. Finally, the server statistically analyzes the propagation feature representations corresponding to each current diagnostic state in the same set of current diagnostic states. For example, the server can add the propagation feature representations corresponding to each current diagnostic state in the same set of current diagnostic states to obtain the current propagation feature representation corresponding to each set of current diagnostic states.

[0090] Step 408: Perform attention processing on the current propagation feature representation under each current diagnostic state set, and determine the state response feature representation of the target object for the state to be predicted based on the attention processing results corresponding to each current diagnostic state set.

[0091] Specifically, the server can perform attention processing on the propagation feature representations of each current diagnostic state set based on an attention mechanism, obtaining the attention processing results corresponding to each current diagnostic state set. Then, the server can integrate the multiple attention processing results to obtain the state response feature representation of the target object in response to the predicted state. This state response feature representation reflects the response information of the target object to the predicted state during this propagation process. In a specific embodiment, integrating the multiple attention processing results can specifically involve pooling the multiple attention processing results, such as average pooling or max pooling.

[0092] In a specific embodiment, after determining the state response feature representation of the target object to the state to be predicted based on the attention processing results corresponding to each current diagnostic state set, the method further includes: using the determined state response feature representation as the current input feature, determining the associated diagnostic state set composed of the associated diagnostic states corresponding to the current diagnostic state set as the current diagnostic state set, iteratively executing the step of determining the associated diagnostic state corresponding to the current diagnostic state in the current diagnostic state set, until a first preset number of iterations is reached; and determining the object feature representation of the target object based on the current state response feature representation obtained in the last iteration.

[0093] Among them, the associated diagnostic state corresponding to the current diagnostic state set refers to the associated diagnostic state corresponding to the current diagnostic state in the current diagnostic state set. For each current diagnostic state set, the associated diagnostic states corresponding to the current diagnostic state in the current diagnostic state set constitute the associated diagnostic state set corresponding to the current diagnostic state set.

[0094] Specifically, the server can use the determined state response feature representation as the current input feature to continue propagation calculation. Each propagation calculation iterates through steps 404 to 408 to obtain the state response feature representation of the target object for the state to be predicted. Then, this state response feature representation is used as the current input feature, and the associated diagnostic state set corresponding to the current diagnostic state set in this propagation calculation process is determined as the current diagnostic state set. The next propagation continues until a first preset number of iterations is reached. This first preset number of iterations can be set as needed, for example, it can be set to 5 times. Finally, the server can determine the object feature representation of the target object based on the current state response feature representation obtained in the last iteration.

[0095] In one specific embodiment, the server can directly determine the object feature representation of the target object from the current state response feature representation obtained in the last iteration. In other embodiments, the server can perform statistical analysis on all current state response feature representations, for example, by adding all current state response feature representations together to obtain the object feature representation of the target object.

[0096] In the above embodiments, by propagating the historical diagnostic state information of the target object multiple times through the state relationship network, a more accurate representation of the object features can be obtained, thereby improving the accuracy of diagnostic state prediction.

[0097] In one embodiment, determining the object feature representation of the target object based on the state response feature representation obtained in the last iteration includes: arranging the state response feature representations, which are currently input features, and the state response feature representation obtained in the last iteration in chronological order to obtain a state response feature representation sequence; performing forward encoding on the state response feature representation sequence to obtain the encoded features of each diagnostic state response feature; and determining the object feature representation of the target object based on the encoded features corresponding to the state response feature representation obtained in the last iteration.

[0098] Specifically, the server can arrange the state response feature representations, which are currently the input features, and the state response feature representations obtained in the last iteration, according to their generation time from earliest to latest, to obtain a state response feature representation sequence. For example, assuming the state response feature representation obtained from the first propagation is o1, the state response feature representation obtained from the second propagation is o2, and the state response feature representation obtained from the third propagation is o3, then the resulting state response feature representation sequence is o1 o2 o3. Further, forward encoding is performed on the state response feature representation sequence. Forward encoding refers to encoding each state response feature representation in the sequence sequentially, combining the encoding information obtained from the previous encoding with each encoding. After forward encoding, the encoded features of each diagnostic state response feature are obtained. The server can determine the object feature representation of the target object based on the encoded features corresponding to the state response feature representation obtained in the last iteration. In practical applications, the server can directly determine the encoded features corresponding to the state response feature representation obtained in the last iteration as the object feature representation of the target object, or the server can fuse the encoded features corresponding to each state response feature representation to obtain the object feature representation of the target object.

[0099] In specific embodiments, the server can use neural network structures such as LSTM (Long Short-Term Memory), RNN (Recurrent Neural Network), BiLSTM (Bi-directional Long Short-Term Memory), GRU (gated recurrent unit), and attention network to perform forward encoding on the state response feature representation sequence.

[0100] In the above embodiments, by forming a state response feature representation sequence from each state response feature representation, the information of each propagation stage can be integrated into the state response feature representation sequence, making the object feature representation of the target object more reasonable.

[0101] In one embodiment, the above-mentioned state data processing method further includes: determining the associated diagnostic state corresponding to the historical diagnostic state in the targeted historical diagnostic state set as the current diagnostic state from a pre-established state relationship network, and the current relationship type between the current diagnostic state and the corresponding associated diagnostic state; forming a state relationship combination by combining the current diagnostic state, the associated diagnostic state corresponding to the current diagnostic state, and the current relationship type, and storing it; taking the associated diagnostic state corresponding to the current diagnostic state as the current diagnostic state, iteratively executing the step of determining the associated diagnostic state corresponding to the current diagnostic state from the pre-established state relationship network until a second preset iteration number is reached; wherein, the first preset iteration number matches the second preset iteration number; determining the associated diagnostic state corresponding to the current diagnostic state in the current diagnostic state set includes: querying each stored state relationship combination to determine the associated diagnostic state corresponding to the current diagnostic state in the current diagnostic state set.

[0102] In this embodiment, the historical information of the target object, i.e., each set of historical diagnostic states, is pre-converted into multiple state relationship combinations and stored according to the state relationship network. Then, during the propagation calculation, the associated diagnostic state corresponding to the current diagnostic state in the current diagnostic state set can be determined by querying the stored state relationship combinations. It can be understood that the steps in this embodiment can be executed during step 204 of the previous embodiment. That is, the server determines the associated diagnostic state corresponding to the historical diagnostic state in the targeted historical diagnostic state set, as well as the relationship type between the historical diagnostic state and the associated diagnostic state. This can be regarded as the first propagation of the historical diagnostic state set in the state relationship network. The historical diagnostic state, the associated diagnostic state corresponding to the historical diagnostic state, and the relationship type between the historical diagnostic state and the associated diagnostic state obtained in the first propagation can also be saved as a state relationship combination. After the first propagation, the server can use the associated diagnostic state corresponding to the historical diagnostic state as the current diagnostic state and continue to propagate it multiple times in the diagnostic state network. In each propagation, the associated diagnostic state corresponding to the current diagnostic state and the current relationship type between the current diagnostic state and the corresponding associated diagnostic state are determined from the pre-established state relationship network iteratively. The current diagnostic state, the associated diagnostic state corresponding to the current diagnostic state, and the current relationship type are combined to form a state relationship combination. The number of iterations here is the second preset number of iterations. It can be understood that the combination of state relationships obtained by the iteration process here is used in the iteration process above. Therefore, the second preset number of iterations here is the same as the first preset number of iterations above.

[0103] refer to Figure 5 This is a schematic diagram illustrating the propagation process of a state relation network in a specific embodiment. The gray circle represents the current node in the graph, and the diagonally lined circle represents the node to be propagated to next. This represents the set of the e-th historical diagnostic states. Through transformation, the historical information of the target object can be represented as... Where E is the number of historical diagnostic state sets, and K is the maximum propagation order of the graph.

[0104] It should be noted that the state relationship combination obtained in each propagation can be regarded as a triple. In these triples, the current node of the graph is the head node, and the node to be propagated next is the tail node. That is, the obtained state relationship combination represents the direction of propagation. When the server determines the associated diagnostic state by querying the state relationship combination, it does so by matching the current diagnostic state with the head node in each state relationship combination. When the current diagnostic state successfully matches the head node in a certain state relationship combination, the tail node in that state relationship combination is taken as the associated diagnostic state of the current diagnostic state.

[0105] In this embodiment, by iteratively executing the steps of determining the associated diagnostic state corresponding to the current diagnostic state from the pre-established state relationship network, as well as the current relationship type between the current diagnostic state and the corresponding associated diagnostic state, and combining the current diagnostic state, the associated diagnostic state corresponding to the current diagnostic state, and the current relationship type into a state relationship combination and storing it, the historical information of the target object can be preprocessed to obtain the data required in the propagation calculation process, thereby improving the efficiency of diagnostic state prediction.

[0106] In one embodiment, the state relation network is constructed in the following way:

[0107] 1. Obtain the reference diagnostic status set corresponding to multiple reference objects at different historical times.

[0108] In this context, the reference object and the target object are objects of the same type. For example, in predicting a patient's illness, the reference object is other patients with historical medical records; in predicting a production machine malfunction, the reference object is the production machine with historical malfunction diagnosis records; and in predicting a vehicle malfunction, the reference object is the vehicle with historical malfunction diagnosis records. The reference diagnostic state set refers to the set of historical diagnostic states of the reference object at a given time.

[0109] Specifically, the server can obtain a set of historical diagnostic statuses corresponding to multiple reference objects at different historical times from the database, thus obtaining multiple sets of reference diagnostic statuses.

[0110] 2. Determine the first reference diagnostic state pair, and the two reference diagnostic states in the second reference diagnostic state pair are parallel.

[0111] In this context, the parallel relationship between the two reference diagnostic states in the first reference diagnostic state pair means that the diagnostic states included in the second reference diagnostic state pair are those that appear simultaneously during a single diagnostic process. A set of historical diagnostic states corresponds to one diagnostic session; therefore, the first reference diagnostic state pair can consist of two diagnostic states that appear within the same set of historical diagnostic states. This parallel relationship is undirected and symmetric. For example, if a production machine diagnoses both fault A and fault B simultaneously during a single diagnostic session, then fault A and fault B have a parallel relationship.

[0112] In this embodiment, the server can determine multiple first reference diagnostic state pairs, each of which includes two diagnostic states that are parallel to each other.

[0113] 3. Statistically determine the frequency of occurrence of each first reference diagnostic state pair in multiple reference diagnostic state sets, classify the frequency levels based on the statistically obtained frequency, and determine the frequency level to which the frequency of occurrence of each first reference diagnostic state pair belongs as the relationship type corresponding to each first reference diagnostic state pair.

[0114] Specifically, since the two diagnostic states in a first reference diagnostic state pair do not have a temporal order, the server can count the occurrence frequency of each first reference diagnostic state pair across multiple reference diagnostic state sets. That is, if a reference diagnostic state set includes both diagnostic states from a first reference diagnostic state pair, then that first reference diagnostic state pair appears once in that reference diagnostic state set. After obtaining the occurrence frequency of each first reference diagnostic state pair, frequency levels can be divided based on the distribution range of these frequencies. Then, the frequency level to which each first reference diagnostic state belongs can be determined, and the frequency level is used to determine the relationship type of the first reference diagnostic state.

[0115] In one specific embodiment, the server can perform quantile calculations on the occurrence frequency of each first reference diagnostic state, and classify the frequency levels based on the quantile calculation results.

[0116] 4. Construct a state relationship network based on each first reference diagnostic state pair and its corresponding relationship type.

[0117] Specifically, the server can represent each diagnostic state in the first reference diagnostic state pair as a node, and the relation type as the connection edge between the nodes, in order to construct a state relation network.

[0118] In one embodiment, the state relationship network is constructed as follows: a second reference diagnostic state pair is determined, wherein there is an order relationship between the two reference diagnostic states in the second reference diagnostic state pair; based on the set of reference diagnostic states belonging to the same reference object, the occurrence frequency of each second reference diagnostic state pair is counted, frequency level is divided based on the counted occurrence frequency, and the frequency level to which the occurrence frequency of each second reference diagnostic state belongs is determined as the relationship type corresponding to each second reference diagnostic state pair.

[0119] In this context, the sequential relationship between the two reference diagnostic states in a second reference diagnostic state pair means that one of the diagnostic states in the pair appears after the other is diagnosed. This sequential relationship is directed and asymmetric. For example, if a vehicle diagnoses fault A in the first fault diagnosis and fault B in the second, then fault A and fault B have a sequential relationship. It should be noted that because the sequential relationship is undirected and asymmetric, two diagnostic states may correspond to two second reference diagnostic state pairs. For instance, if diagnostic state A appears first and diagnostic state B appears later, then the second reference diagnostic state pair could be (diagnostic state A, diagnostic state B). If diagnostic state B appears first and diagnostic state A appears later, then the second reference diagnostic state pair could be (diagnostic state B, diagnostic state A).

[0120] Specifically, for each second reference diagnostic state pair, the server can determine whether the second reference diagnostic state pair appears in multiple reference diagnostic state sets belonging to the same reference object, and then count the number of times the second reference diagnostic state pair appears to obtain its frequency. Considering that there is an order relationship between the two diagnostic states in the second reference diagnostic state pair, the criterion for determining the appearance of the second reference diagnostic state pair is that the two diagnostic states in the second reference diagnostic state pair appear in two different historical diagnostic state sets of the same reference object, and the order of appearance is consistent with the order in the second reference diagnostic state pair. For example, suppose a second reference diagnostic state pair is (diagnostic state A, diagnostic state B), that is, diagnostic state A appears first and diagnostic state B appears later. Then, if diagnostic state A appears in a historical diagnostic state set A of a certain reference object, and diagnostic state B appears in a historical diagnostic state set B, and if the time corresponding to historical diagnostic state set A is earlier than that of historical diagnostic state set B, then the second reference diagnostic state pair is determined to have appeared once.

[0121] After statistically obtaining the occurrence frequency of each second reference diagnostic state, frequency levels can be divided based on the distribution range of these occurrence frequencies. Then, the frequency level to which each second reference diagnostic state belongs can be determined, and the frequency level can be determined as the relationship type of the second reference diagnostic state.

[0122] In one specific embodiment, the server can perform quantile calculations on the occurrence frequency of each of the second reference diagnostic states, and classify the frequency levels based on the quantile calculation results.

[0123] In this embodiment, the server can further construct a state relation network based on each first reference diagnostic state pair, the corresponding relation type of each first reference diagnostic state pair, each second reference diagnostic state pair, and the corresponding relation type of each second reference diagnostic state pair. Specifically, the server can represent each diagnostic state in the first reference diagnostic state pair as a node, and the corresponding relation type of the first reference diagnostic state pair as the connection edge between the nodes of the first reference diagnostic state pair; and represent each diagnostic state in the second reference diagnostic state pair as a node, and the corresponding relation type of the second reference diagnostic state pair as the connection edge between the nodes of the second reference diagnostic state pair, to construct the state relation network.

[0124] In the above embodiments, by statistically analyzing the frequency of occurrence of reference diagnostic state pairs and classifying the frequency of occurrence to obtain the relationship type, various complex diagnostic state relationships can be more reasonably distinguished.

[0125] In one embodiment, determining the correlation between the state to be predicted and the historical diagnostic states under the relationship type includes: for each state relationship combination, obtaining the state feature representation corresponding to the historical diagnostic states, the correlation feature representation corresponding to the associated diagnostic states, and the relationship feature representation corresponding to the relationship type in the state relationship combination; determining the correlation degree corresponding to the state relationship combination based on the state feature representation corresponding to the state to be predicted, the state feature representation corresponding to the historical diagnostic states, and the relationship feature representation; and determining the propagation feature representation of the state relationship network for the state to be predicted under the target set of historical diagnostic states based on the correlation degree and the associated diagnostic states, including: multiplying the correlation degree by the correlation feature representation to obtain the propagation feature representation corresponding to the state relationship combination; and statistically analyzing the propagation feature representation corresponding to each state relationship combination to obtain the propagation feature representation of the state relationship network for the state to be predicted under the target set of historical diagnostic states.

[0126] The state feature representations corresponding to historical diagnostic states, the association feature representations corresponding to associated diagnostic states, and the relationship feature representations corresponding to relationship types can be obtained by performing feature mapping on the historical diagnostic states, associated diagnostic states, and relationship types respectively. In a specific embodiment, feature mapping can be implemented through a machine learning model. For example, a machine learning model can map historical diagnostic states to embeddings as state feature representations of historical diagnostic states, associate diagnostic states to embeddings as state feature representations of associate diagnostic states, and relationship types to embeddings as relationship feature representations of relationship types.

[0127] Specifically, each set of historical diagnostic states includes multiple historical diagnostic states. Each historical diagnostic state, its corresponding associated diagnostic state, and relation type form a state relation combination. For each state relation combination, the server can multiply the state feature representation corresponding to the state to be predicted, the state feature representation corresponding to the historical diagnostic state, and the relation feature representation, and then perform normalization to obtain the correlation degree corresponding to the state relation combination. The correlation degree is further multiplied by the correlation feature representation corresponding to the associated diagnostic state in the state relation combination to obtain the propagation feature representation corresponding to the state relation combination. Finally, the propagation feature representations of the state relation combinations corresponding to the same set of historical diagnostic states are added together to obtain the propagation feature representation of the state relation network for the state to be predicted under each set of historical diagnostic states.

[0128] In the above embodiments, by using the state feature representation corresponding to the historical diagnostic state, the association feature representation corresponding to the associated diagnostic state, and the relationship feature representation corresponding to the relationship type, the correlation degree corresponding to the state relationship combination can be accurately and quickly determined based on the state feature representation corresponding to the state to be predicted, the state feature representation corresponding to the historical diagnostic state, and the relationship feature representation. This leads to a more accurate propagation feature representation and improves the accuracy of diagnostic state prediction.

[0129] In one embodiment, based on the attention processing results corresponding to each set of historical diagnostic states, the state response feature representation of the target object to the state to be predicted is determined, including: performing pooling processing on the attention processing results corresponding to each set of historical diagnostic states to obtain the state response feature representation of the target object to the state to be predicted; wherein, the integration processing includes max pooling or average pooling.

[0130] Specifically, in this embodiment, the server can perform average pooling or max pooling on the attention processing results corresponding to each set of historical diagnostic states, and integrate the attention processing results through pooling to obtain the state response feature representation of the target object to the state to be predicted; wherein, the integration processing includes max pooling or average pooling.

[0131] In a specific embodiment, such as Figure 6 The diagram shown is an overall flowchart of the state data processing method according to an embodiment of this application. (Reference) Figure 6In this embodiment, the state data processing method is implemented through a diagnostic state prediction model. This embodiment first constructs a state relationship network based on the entire database, and then determines the input data to the diagnostic state prediction model based on the historical information of the target object, i.e., the historical diagnostic state set, the state to be predicted, and the state relationship network. The input data is then input into the diagnostic state prediction model, and the diagnostic state prediction model is used to propagate and deduce the diagnostic state. Finally, the probability of the target object being in the state to be predicted in the future is output.

[0132] The following will provide a detailed description of this embodiment using a disease prediction scenario. It should be understood that in the disease prediction scenario, the diagnostic status refers to the disease, and the object refers to the patient. It should be noted that in this embodiment, the status data processing method can be executed by a computer device, which can be... Figure 1 The terminal 102 or the server 104. In this embodiment, a computer device is used as an example to illustrate the server.

[0133] In related technologies, multi-class classification models are typically used to classify and predict multiple diseases, treating disease prediction as a multi-class problem, with each class representing a disease. The entire model typically includes multiple single-class models, such as logistic regression and support vector machines, or uses multi-class models such as random forests, gradient boosting trees, and neural networks. After inputting the patient features required by the model, the prediction model will classify the patient's features, and the final category and corresponding disease is the predicted disease for that patient. The general process of the classification model is as follows: Figure 7 As shown. Figure 7 As shown, the patient's disease prediction result is heart disease. Multi-class disease prediction models can only predict one disease at a time, and they will always predict one disease, leading to low prediction accuracy. Furthermore, due to the inherent limitations of multi-class classification, the model can predict a limited number of diseases and cannot comprehensively cover all existing diseases.

[0134] In this application scenario, state relation network It is constructed based on the probability of two diseases occurring simultaneously, h i and t i Indicates disease, r i This represents the relationships between disease pairs. The state-relationship network consists of N combinations of disease relationships; in this application scenario, it can also be called a co-disease network. Since the database records each medical visit as a unit, and each visit may involve multiple diagnoses, disease relationships can be divided into two categories: sequential relationships r... series and parallel relationship r para (h) i ,r i series ,t i r in ) iseries Indicates when h i After being diagnosed, future medical treatment and diagnosis i The probability, (h) i ,r i para ,t i r in ) i para Indicates when h i and r i The probability of being diagnosed simultaneously during a single medical visit, h i ,r i series ,t i ) is directed asymmetric, while (h i ,r i para ,t i Undirected and symmetric. By statistically analyzing the frequency of each relation for each disease pair in the database and calculating quantiles, r is assigned according to the quantiles. series and r para They are all divided into 5 levels, with values ​​ranging from 1 to 5. series and r para The level is used as the relation type.

[0135] In this embodiment, the structure of the diagnostic state prediction model can be referred to Figure 8 By decomposing the model, the network structure of the diagnostic state prediction model can be divided into four parts: input preprocessing module, same-stage propagation module, information integration module, and probability output module. Each module is explained in detail below:

[0136] 1. Preprocessing module

[0137] The diagnostic status prediction model has three input components: the disease to be predicted (d), the patient's historical disease information, and the disease d to be predicted. and state relation network Historical disease information refers to diseases diagnosed at different historical times. Here, "historical time" refers to the time of medical visit for a historical disease, which could be a time recorded in a database. The patient to be predicted was diagnosed with this historical disease at that time, therefore, that historical time is also the diagnosis time. 'd' can be any disease the user wants to predict. In the input preprocessing module, the disease to be predicted, 'd', is first fed into the network in one-hot encoding and converted into an embedding representation (i.e., a disease feature representation) through the embedding layer (not shown in the figure) in the diagnostic state prediction model. This embedding representation is then further input into the inference module. Historical information... Based on the state relation network Convert to multiple graphs This represents the k-th propagation result of the e-th medical visit record. The current node in the k-th propagation result and the node to be propagated next can form a disease relationship combination. In this disease relationship combination, the current node is the head node, and the node to be propagated next is the tail node. The tail node is also the associated node corresponding to the head node.

[0138] Through transformation, historical information It can be represented as Where E is the maximum number of times a patient can see a doctor, and K is the maximum propagation order of the graph.

[0139] 2. Simultaneous dissemination module

[0140] The structure of the same-stage propagation module is as follows: Figure 9 As shown, in the same propagation stage This will be processed in this section as well.

[0141] This structure is divided into three parts: inference module, attention module, and pooling module. Taking the first propagation as an example, the inference module calculates the propagation feature representation for each disease relationship combination. Specifically, the inference module calculates the correlation degree based on the embedding representation of the disease to be predicted, the embedding representation of the head node in the disease relationship combination, and the feature representation of the relationship type. For details, please refer to the following formulas (4) and (5). When the relationship in the disease relationship combination is a parallel relationship, the server can calculate the correlation degree based on the disease feature representation of the disease to be predicted, the relationship feature representation of the relationship level under the parallel relationship, and the disease feature representation of the associated disease. For details, please refer to formula (4). When the relationship in the disease relationship combination is a sequential relationship, the server can calculate the correlation degree based on the disease feature representation of the disease to be predicted, the relationship feature representation of the relationship level under the sequential relationship, and the disease feature representation of the associated disease. For details, please refer to formula (5).

[0142]

[0143]

[0144] After calculating the correlation degree of each disease relationship combination, the server can further multiply the correlation degree by the embedding of the tail node in the disease relationship combination. Finally, the product of the disease relationship combinations corresponding to the same medical visit and the sequential relationship can be statistically analyzed to obtain the first statistical result, and the product of the disease relationship combinations corresponding to the same medical visit and the AND relationship can be statistically analyzed to obtain the second statistical result. The average of the first statistical result and the second statistical result is calculated to obtain the propagation feature representation of the state relationship network for the state to be predicted under the historical information of the medical visit. For details, please refer to the following formula (6):

[0145]

[0146] In formulas (4) to (6) above, (h e,i ,r e,i ,t e,i )express The i-th disease relationship combination, h e,i ,t e,i Both 'd' and 'r' represent embedding. e,i α represents the relational features of a relation type. e,i represents the correlation between the calculated i-th disease relationship combination and the disease d to be predicted, while v e,1 It is the output of the inference module, representing The graph propagation information related to d, i.e. the state relation network, is represented by the propagation characteristics of the disease to be predicted based on the historical information of the eth visit.

[0147] After passing through the inference module, the obtained {v e,1} e∈[1,E] Input attention module.

[0148] The attention processing module can weight each input based on global input information and can also process temporal information. Furthermore, compared to RNNs, which can also process temporal inputs, the attention processing module has the advantage of parallelization, achieving higher computational efficiency when E is large. The attention processing module first embeds each patient visit time into a representation, obtaining a temporal feature representation. The calculation formulas can be found in formulas (1) and (2) above. Let d represent the time embedding of the e-th medical visit, d represent the dimension of the disease embedding, and x represent the time embedding of the e-th visit. e,1Let represent the propagation feature representation corresponding to the e-th medical visit. Then, the temporal feature representation is fused with the corresponding propagation feature representation to obtain a fused feature representation. Subsequently, attention processing is performed on each fused feature representation based on a self-attention mechanism. After processing by the attention processing module, the attention processing result corresponding to the e-th medical visit is vector a. e Finally, the same-stage propagation module uses the pooling processing module to process {a} e} e∈[1,E] Max pooling is performed, which involves taking the maximum value of each vector in the attention processing results corresponding to each medical visit in their respective dimensions, and then integrating them to obtain the desired result. emb For the dimensional vector, the relevant formula can be found in the following formula (7):

[0149]

[0150] Where o1 represents all the information of the patient's historical information encoded by the same stage propagation module in the first propagation under the disease to be predicted d, that is, the state response feature representation obtained by the first propagation calculation.

[0151] Furthermore, using o1 as the input for the second propagation, the propagation calculation continues to obtain o2, ..., o K ,Depend on Figure 8 It can be seen that the diagnostic state prediction model includes multiple identical same-stage propagation modules. The state response feature representation obtained from each propagation calculation is used as the input of the next same-stage propagation module. Therefore, in the subsequent propagation calculation process, the calculation method is the same as the first propagation calculation, that is, the above formulas (4) to (7) can be referred to, except that the embedding representation of the disease to be predicted in the first propagation calculation is replaced with the current input state response feature representation.

[0152] 3. Information Integration Module

[0153] refer to Figure 10 After obtaining the information encoding of all propagation stages {o k} k∈[1,K] After that, the model will put all o k Input the LSTM model in sequence (i.e.) Figure 10 In the Long Short-Term Memory (LSTM) network, the output p of the last unit of the LSTM model is ultimately used to represent the patient's historical information, thus obtaining the patient's object feature representation. In the LSTM model, for all o... k Forward encoding is performed in chronological order. In each current encoding module (hereinafter referred to as the current module), the final output of the previous module and the input of the current module are combined. e(The state response feature representation corresponding to the eth visit) is used as input to determine the output of the forget gate, the cell memory, and the input gate respectively. Finally, the final output of the current module is determined based on the output of the forget gate, the cell memory, and the input gate. The specific calculation process can be referred to the following formulas (8) to (14):

[0154]

[0155] F e =σ(o e *U f +l e-1 *W f (9)

[0156] I e =σ(o e *U I +l e-1 *W I (10)

[0157] O e =σ(o e *U O +l e-1 *W O (11)

[0158]

[0159] l e =O e *tanh(c e (13)

[0160]

[0161] In the above formula, [W f U f W I U I W c U c W o U o ] are all corresponding weights, which are the network model parameters of LSTM; F e The output of the current module in the forget gate of the LSTM model, σ is the sigmoid function, and o e For the input of the current module, l e-1 For the final output of the previous module in the LSTM model, l e This is the final output of the current module in the LSTM model; I e The output of the input gate of the current module in the LSTM model; O e c is the output of the current module's output gate in the LSTM model.e For the cellular memory of the current module, c e-1 This is for the cellular memory of the previous module.

[0162] 4. Probability Output Module

[0163] After the first three modules, the disease embedding d (i.e., the disease feature representation of the disease to be predicted) and the patient embedding p (i.e., the patient feature representation of the patient) are obtained. Finally, the correlation between the disease feature representation and the patient feature representation is calculated in the output layer of the model to obtain the probability that the patient will suffer from the disease to be predicted in the future. The specific calculation process can be referred to in the following formulas (15) and (16):

[0164]

[0165]

[0166] in, σ For the sigmod function, This refers to the probability predicted by the model that the corresponding patient will develop the disease to be predicted in the future.

[0167] It should be noted that the above process describes the usage phase of the diagnostic status prediction model, that is, using the pre-trained diagnostic status prediction model to predict the probability of a patient developing a disease to be predicted in the future. The disease to be predicted can be any disease. During the model training phase, the training samples can be obtained from the database, with the disease from the patient's last historical record serving as the label. The remaining historical records serve as the input to the diagnostic status prediction model during training. The disease to be predicted, d, is randomly selected from the confirmed and undiagnosed diseases in the last historical record in a 1:1 ratio.

[0168] It should be noted that in the above embodiments, the feature representations of each node in each disease relationship combination can be used as parameters for model training. At the start of training, the feature representations of each node are initialized, and when the model training is completed, the feature representations of each node are obtained and stored. Of course, in other embodiments, the feature representations of these nodes can also be pre-generated using other embedding representation models, and this application does not limit this.

[0169] In this application scenario, the diagnostic status prediction model is easy to input, requiring only access to public comorbidity network data and patient history of illness. Therefore, it can be conveniently applied to physical examinations and nursing care. Users only need to manually input or retrieve the patient's history of illness from the EHR database, and the system can quickly output a list of susceptible diseases for the individual. By taking targeted measures against susceptible diseases, the efficiency of physical examinations and nursing care can be effectively improved, while reducing human and material costs. The application scenario workflow is as follows: Figure 11 As shown.

[0170] On the other hand, the mortality rate from the four major chronic non-communicable diseases is on the rise, indicating a greater disease burden from these diseases. Therefore, the prevention and control of chronic diseases is crucial. Because chronic diseases are characterized by a wide population impact, numerous complications, long duration, and complex changes, the diagnostic status prediction model provided in this application, which uses historical disease data to predict future diseases, is personalized, accurate, and efficient, and can be well applied to the prevention and control of chronic diseases or chronic complications.

[0171] In the above embodiments, by introducing an attention structure into the model, the model can process temporal information, and the model's performance and parallel computing capabilities are improved. By using an LSTM structure to integrate information from each propagation stage, the patient's embedding representation is made more reasonable, and the model is given the ability to process long sequence inputs. By classifying disease relationships into parallel and sequential relationships and performing hierarchical classification, the construction method of disease relationships can be optimized, enabling the model to more reasonably distinguish various complex disease relationships.

[0172] This application also provides another application scenario where the state data processing method is applied to predicting production machine failures. In this scenario, the object refers to the production machine, and the diagnostic state refers to the failure of the production machine. A trained diagnostic state prediction model is used to predict the probability of the production machine experiencing a future failure. Here, "future" refers to the time since the last failure record. In this embodiment, the server first constructs a fault relationship network based on a database storing historical vehicle failure records. For the target production machine to be predicted, the server can use the fault relationship network to propagate the historical failure information (i.e., historical failure sets at different historical times) of the production machine to be predicted multiple times, transforming it into a fault relationship combination. Then, for the expected failure of the production machine, the server can use the diagnostic state prediction model to predict the probability of the target production machine experiencing the expected failure in the future, based on the expected failure and the transformed fault relationship combination. The specific structure and prediction process of the diagnostic state prediction model can be found in the above embodiment.

[0173] This application also provides another application scenario where the state data processing method is applied to predict vehicle faults. In this scenario, the object refers to the vehicle, and the diagnostic state refers to the vehicle's fault. A trained diagnostic state prediction model is used to predict the probability that the vehicle will experience a future fault. Here, "future" refers to the time after the last fault record for the vehicle. In this embodiment, the server first constructs a fault relationship network based on a database storing historical vehicle fault records. For the target vehicle to be predicted, the server can use the fault relationship network to propagate the historical fault information (i.e., historical fault sets at different historical times) of the vehicle to be predicted multiple times, transforming it into a fault relationship combination. Then, for the predicted fault of the vehicle to be predicted, the server can use the diagnostic state prediction model to predict the probability that the target vehicle will experience the predicted fault in the future, based on the predicted fault and the transformed fault relationship combination. The specific structure and prediction process of the diagnostic state prediction model can be found in the above embodiment.

[0174] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0175] Based on the same inventive concept, this application also provides a state data processing apparatus for implementing the state data processing method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more state data processing apparatus embodiments provided below can be found in the limitations of the state data processing method described above, and will not be repeated here.

[0176] In one embodiment, such as Figure 12 As shown, a state data processing device 1200 is provided, comprising:

[0177] The historical status determination module 1202 is used to obtain the set of historical diagnostic statuses of the target object at different historical times.

[0178] The associated state determination module 1204 is used to determine, from the pre-established state relationship network, the associated diagnostic state corresponding to the historical diagnostic state in the target historical diagnostic state set, and the relationship type between the historical diagnostic state and the associated diagnostic state; the state relationship network is used to describe the association relationship between diagnostic states.

[0179] The propagation feature derivation module 1206 is used to determine the correlation between the state to be predicted and the historical diagnostic states under the relationship type. Based on the correlation and the associated diagnostic states, it determines the propagation feature representation of the state relationship network for the state to be predicted under the target set of historical diagnostic states.

[0180] Attention processing module 1208 is used to perform attention processing on the propagation feature representation under each historical diagnostic state set, and to determine the state response feature representation of the target object to the state to be predicted based on the attention processing results corresponding to each historical diagnostic state set.

[0181] The object feature determination module 1210 is used to determine the object feature representation of the target object based on the state response feature representation.

[0182] The probability prediction module 1212 is used to predict the probability information of a target object being in a state to be predicted based on the object's feature representation.

[0183] The aforementioned state data processing device acquires historical diagnostic state sets corresponding to the target object at different historical times, determines the associated diagnostic states corresponding to the historical diagnostic states in the targeted historical diagnostic state sets from a pre-established state relationship network, as well as the relationship type between the historical diagnostic states and the associated diagnostic states, determines the correlation degree between the state to be predicted and the historical diagnostic states under the relationship type, and, based on the correlation degree and the associated diagnostic states, determines the propagation feature representation of the state relationship network for the state to be predicted under the targeted historical diagnostic state sets. Attention processing is performed on the propagation feature representation under each historical diagnostic state set, and based on the attention processing results, the state response feature representation of the target object for the state to be predicted is determined. Based on the state response feature representation, the object feature representation of the target object is determined, and based on the object feature representation, the probability information of the target object being in the state to be predicted is predicted. This enables prediction of any state to be predicted. Since the predicted probability information is obtained using the state relationship network and the historical diagnostic state information of the target object, the evolution of the target object's diagnostic state can be simulated, thereby improving the accuracy of the diagnostic state determination. Furthermore, by applying attention processing to the propagation feature representations under each set of historical diagnostic states, the differences in the impact of diagnostic state information at different historical times on the prediction results can be taken into account, resulting in more accurate object feature representations and thus more accurate probability information.

[0184] In one embodiment, the attention processing module is further configured to: obtain historical time information corresponding to each set of historical diagnostic states; obtain time feature representations corresponding to each set of historical time information; fuse the propagation feature representations and time feature representations corresponding to each set of historical diagnostic states to obtain fused feature representations for each set of historical diagnostic states; and perform attention processing on each fused feature representation to obtain attention processing results for each set of historical diagnostic states.

[0185] In one embodiment, the attention processing module is further configured to: determine the number of dimensions corresponding to the propagation feature representation; and determine the time feature representation corresponding to the historical time information based on the historical time information, the number of dimensions, and the order of each element in the propagation feature representation.

[0186] In one embodiment, the attention processing module is further configured to convert each fusion feature representation into a query representation, a key representation, and a value representation respectively; calculate the relevance between the query representation corresponding to the target fusion feature representation and each key representation respectively, and determine the attention weight corresponding to each key representation based on the relevance; multiply the attention weight corresponding to each key representation with its corresponding value representation, calculate the resulting products, and obtain the attention processing result corresponding to the target fusion feature representation based on the statistical results.

[0187] In one embodiment, the object feature determination module is further configured to: take the associated diagnostic state set composed of the associated diagnostic states corresponding to each historical diagnostic state in the target historical diagnostic state set as the current diagnostic state set; take the state response feature representation as the current input feature to determine the associated diagnostic state corresponding to the current diagnostic state in the current diagnostic state set, and the current relationship type between the current diagnostic state and the corresponding associated diagnostic state; the associated diagnostic state corresponding to the current diagnostic state and the current relationship type are determined from the state relationship network to determine the current correlation degree between the current input feature and the current diagnostic state under the current relationship type; based on the current correlation degree and the associated diagnostic state corresponding to the current diagnostic state, determine the current propagation feature representation of the state relationship network for the state response feature representation under the current diagnostic state set; perform attention processing on the current propagation feature representation under each current diagnostic state set respectively; and determine the state response feature representation of the target object for the state to be predicted based on the attention processing results corresponding to each current diagnostic state set.

[0188] In one embodiment, the object feature determination module is further configured to: take the determined state response feature representation as the current input feature, determine the associated diagnostic state set composed of the associated diagnostic states corresponding to the current diagnostic state set as the current diagnostic state set, iteratively execute the step of determining the associated diagnostic state corresponding to the current diagnostic state in the current diagnostic state set, until a first preset number of iterations is reached; and determine the object feature representation of the target object based on the state response feature representation obtained in the last iteration.

[0189] In one embodiment, the object feature determination module is further configured to: arrange the state response feature representations, which are currently input features, and the state response feature representations obtained in the last iteration in chronological order to obtain a state response feature representation sequence; perform forward encoding on the state response feature representation sequence to obtain the encoded features of each diagnostic state response feature; and determine the object feature representation of the target object based on the encoded features corresponding to the state response feature representations obtained in the last iteration.

[0190] In one embodiment, the apparatus further includes: a graph propagation module, configured to: take the associated diagnostic states corresponding to the historical diagnostic states in the target set of historical diagnostic states as the current diagnostic state; determine the associated diagnostic states corresponding to the current diagnostic state and the current relationship type between the current diagnostic state and the corresponding associated diagnostic states from a pre-established state relationship network; combine the current diagnostic state, the associated diagnostic states corresponding to the current diagnostic state, and the current relationship type into a state relationship combination and store it; take the associated diagnostic states corresponding to the current diagnostic state as the current diagnostic state, and iteratively execute the step of determining the associated diagnostic states corresponding to the current diagnostic state from the pre-established state relationship network until a second preset iteration number is reached; wherein the first preset iteration number matches the second preset iteration number; and a probability prediction module, further configured to: query the stored state relationship combinations to determine the associated diagnostic states corresponding to the current diagnostic states in the set of current diagnostic states.

[0191] In one embodiment, the apparatus further includes a relationship network construction module, configured to: acquire a set of reference diagnostic states corresponding to multiple reference objects at different historical times; determine a first reference diagnostic state pair, wherein the two reference diagnostic states in the second reference diagnostic state pair have a parallel relationship; statistically analyze the occurrence frequency of each first reference diagnostic state pair in the multiple reference diagnostic state sets, classify the frequency levels based on the statistically obtained occurrence frequencies, and determine the frequency level to which the occurrence frequency of each first reference diagnostic state pair belongs as the relationship type corresponding to each first reference diagnostic state pair; and construct a state relationship network according to each first reference diagnostic state pair and its corresponding relationship type.

[0192] In one embodiment, the relationship network construction module is used to determine second reference diagnostic state pairs, wherein there is an order relationship between the two reference diagnostic states in the second reference diagnostic state pairs; based on the set of reference diagnostic states belonging to the same reference object, the frequency of occurrence of each second reference diagnostic state pair is counted, and frequency levels are divided based on the counted frequency, and the frequency level to which the frequency of occurrence of each second reference diagnostic state belongs is determined as the relationship type corresponding to each second reference diagnostic state pair; and a state relationship network is constructed based on each first reference diagnostic state pair, the relationship type corresponding to each first reference diagnostic state pair, each second reference diagnostic state pair, and the relationship type corresponding to each second reference diagnostic state pair.

[0193] In one embodiment, the target set of historical diagnostic states includes multiple historical diagnostic states, and each historical diagnostic state, its corresponding associated diagnostic state, and relationship type constitute a state relationship combination. The propagation feature derivation module is further configured to: for each state relationship combination, obtain the state feature representation corresponding to the historical diagnostic state, the association feature representation corresponding to the associated diagnostic state, and the relationship feature representation corresponding to the relationship type in the state relationship combination; determine the association degree corresponding to the state relationship combination based on the state feature representation corresponding to the state to be predicted, the state feature representation corresponding to the historical diagnostic state, and the relationship feature representation; multiply the association degree by the association feature representation to obtain the propagation feature representation corresponding to the state relationship combination; and statistically analyze the propagation feature representation corresponding to each state relationship combination to obtain the propagation feature representation of the state relationship network for the state to be predicted under the target set of historical diagnostic states.

[0194] In one embodiment, the attention processing module is further configured to: perform pooling processing on the attention processing results corresponding to each set of historical diagnostic states to obtain the state response feature representation of the target object in response to the state to be predicted; wherein the integration processing includes max pooling or average pooling.

[0195] Each module in the aforementioned state data processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0196] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 13As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores historical diagnostic states, feature representations, and other data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a state data processing method.

[0197] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 14 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a state data processing method. The display unit of the computer device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0198] Those skilled in the art will understand that Figure 13 , Figure 14The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0199] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described state data processing method.

[0200] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described state data processing method.

[0201] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the state data processing method described above.

[0202] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0203] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0204] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0205] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A state data processing method characterized by, The method includes: Obtain the collection of historical disease information for patients at different times of medical visit; The associated diseases corresponding to historical disease information in the targeted historical disease information set are determined from a pre-established state relationship network, as well as the relationship type between the historical disease information and the associated diseases; the state relationship network is used to describe the association between diagnostic states; the targeted historical disease information set includes multiple historical disease information, and each historical disease information, its corresponding associated disease, and relationship type constitute a state relationship combination; The disease to be predicted is fed into the relational network and converted into a disease feature representation; For each state relationship combination, obtain the state feature representation corresponding to the historical disease information, the disease feature representation corresponding to the associated disease, and the relationship feature representation corresponding to the relationship type in the state relationship combination. Based on the state feature representation corresponding to the disease to be predicted, the state feature representation corresponding to the historical disease information, and the relationship feature representation, the correlation degree corresponding to the state relationship combination is determined. Multiplying the correlation degree by the disease feature representation corresponding to the associated disease yields the propagation feature representation corresponding to the state relationship combination. By statistically analyzing the propagation feature representations corresponding to each combination of state relationships, the propagation feature representation of the state relationship network for the disease to be predicted is obtained under the target set of historical disease information. Attention processing is performed on the propagation feature representations under each set of historical disease information. Based on the attention processing results corresponding to each set of historical disease information, the state response feature representation of the patient in response to the disease to be predicted is determined. Based on the state response feature representation, the patient's object feature representation is determined, and based on the object feature representation, the probability information of the patient being in the predicted disease is determined.

2. The method of claim 1, wherein, The attention processing performed on the propagation feature representations under each historical disease information set includes: Obtain the medical visit time information corresponding to each of the aforementioned historical disease information sets; Obtain the time feature representation corresponding to each of the aforementioned medical visit time information, and fuse the propagation feature representation and time feature representation corresponding to each set of historical disease information to obtain the fused feature representation of each set of historical disease information. Attention processing is applied to each fusion feature representation to obtain the attention processing results corresponding to each set of historical disease information.

3. The method of claim 2, wherein, The propagation feature representation includes elements of multiple dimensions, and the acquisition of the time feature representation corresponding to each of the medical visit time information includes: Determine the number of dimensions corresponding to the propagation feature representation; Based on the medical visit time information, the number of dimensions, and the order of each element in the propagation feature representation, the time feature representation corresponding to the medical visit time information is determined.

4. The method of claim 2, wherein, The attention processing performed on each fused feature representation yields attention processing results for each set of historical disease information, including: Each fusion feature representation is converted into a query representation, a key representation, and a value representation, respectively; The relevance of the query representation corresponding to the target fusion feature representation and each key representation is calculated, and the attention weight corresponding to each key representation is determined based on the relevance. Each key is multiplied by its corresponding attention weight and its corresponding value, and the resulting products are statistically analyzed. Based on the statistical results, the attention processing result corresponding to the fusion feature representation is obtained.

5. The method of claim 1, wherein, The process of determining the patient's object feature representation based on the state response feature representation includes: The associated disease set, which consists of the associated diseases corresponding to each historical disease in the target historical disease information set, is taken as the current diagnostic status set, and the status response feature is taken as the current input feature. Determine the associated diseases corresponding to the current diagnostic state in the current diagnostic state set, and the current relationship type between the current diagnostic state and the corresponding associated disease; the associated diseases and the current relationship type corresponding to the current diagnostic state are determined from the state relationship network; Determine the current correlation between the current input features and the current diagnostic state under the current relationship type; based on the current correlation and the associated diseases corresponding to the current diagnostic state, determine the current propagation feature representation of the state relationship network under the current diagnostic state set, for the state response feature representation; Attention processing is performed on the current propagation feature representation under each current diagnostic state set. Based on the attention processing results corresponding to each current diagnostic state set, the state response feature representation of the patient in response to the disease to be predicted is determined.

6. The method of claim 5, wherein, After determining the patient's state response feature representation for the disease to be predicted based on the attention processing results corresponding to each current diagnostic state set, the method further includes: The determined state response features are used as the current input features. The associated disease set consisting of the associated diseases corresponding to the current diagnostic state set is determined as the current diagnostic state set. The step of determining the associated diseases corresponding to the current diagnostic state in the current diagnostic state set is iteratively executed until the first preset number of iterations is reached. Based on the state response feature representation obtained from the last iteration, the object feature representation of the patient is determined.

7. The method of claim 6, wherein, The determination of the patient's object feature representation based on the state response feature representation obtained from the last iteration includes: The state response feature representations, which are used as the current input features, and the state response feature representations obtained in the last iteration are arranged in chronological order to obtain a state response feature representation sequence. Forward encoding is performed on the state response feature representation sequence to obtain the encoded features of each diagnostic state response feature; Based on the encoded features corresponding to the state response features obtained in the last iteration, the object feature representation of the patient is determined.

8. The method of claim 6, wherein, The method further includes: The associated diseases corresponding to the historical disease information in the target set of historical disease information are used as the current diagnostic status; Determine the associated diseases corresponding to the current diagnostic state and the current relationship type between the current diagnostic state and the corresponding associated diseases from a pre-established state relationship network. Combine the current diagnostic state, the associated diseases corresponding to the current diagnostic state, and the current relationship type into a state relationship combination and store it. The associated disease corresponding to the current diagnostic state is taken as the current diagnostic state, and the step of determining the associated disease corresponding to the current diagnostic state from the pre-established state relationship network is iteratively executed until a second preset number of iterations is reached; wherein, the first preset number of iterations matches the second preset number of iterations. The process of determining the associated disease corresponding to the current diagnostic status in the current diagnostic status set includes: Query the stored combinations of state relationships to determine the associated diseases corresponding to the current diagnostic state in the current diagnostic state set.

9. The method of claim 1, wherein, The state relation network is constructed in the following way: Obtain a set of reference diagnostic states for multiple reference objects at different medical visit times; Determine a first reference diagnostic state pair, wherein the two reference diagnostic states in the first reference diagnostic state pair are in a parallel relationship; In the multiple sets of reference diagnostic states, the frequency of occurrence of each first reference diagnostic state pair is counted. Based on the counted frequency of occurrence, a frequency level is divided, and the frequency level to which the frequency of occurrence of each first reference diagnostic state pair belongs is determined as the relationship type corresponding to each first reference diagnostic state pair. A state relationship network is constructed based on each first reference diagnostic state pair and its corresponding relationship type.

10. The method of claim 9, wherein, The method further includes: Determine a second reference diagnostic state pair, wherein there is an order relationship between the two reference diagnostic states in the second reference diagnostic state pair; Based on the set of reference diagnostic states belonging to the same reference object, the frequency of occurrence of each second reference diagnostic state is statistically analyzed. Based on the frequency of occurrence obtained from statistics, frequency levels are divided, and the frequency level to which the frequency of occurrence of each second reference diagnostic state belongs is determined as the relationship type of each second reference diagnostic state to its respective counterpart. The step of constructing a state relationship network based on each first reference diagnostic state pair and its corresponding relationship type includes: A state relationship network is constructed based on each first reference diagnostic state pair, the corresponding relationship type of each first reference diagnostic state pair, each second reference diagnostic state pair, and the corresponding relationship type of each second reference diagnostic state pair.

11. The method according to any one of claims 1 to 10, characterized in that, The process of determining the patient's state response feature representation for the disease to be predicted, based on the attention processing results corresponding to each set of historical disease information, includes: Pooling is performed on the attention processing results corresponding to each set of historical disease information to obtain the patient's state response feature representation for the disease to be predicted; The pooling process includes either max pooling or average pooling.

12. A state data processing apparatus, characterized by comprising: The device includes: The historical status determination module is used to obtain the set of historical disease information corresponding to patients at different times of medical treatment; The associated state determination module is used to determine, from a pre-established state relationship network, the associated diseases corresponding to historical disease information in the targeted historical disease information set, and the relationship type between the historical disease information and the associated diseases; the state relationship network is used to describe the association relationship between diagnostic states; the targeted historical disease information set includes multiple historical disease information, and each historical disease information, its corresponding associated disease, and the relationship type constitute a state relationship combination; The propagation feature derivation module is used to input the disease to be predicted into the relationship network and convert it into a disease feature representation; for each state relationship combination, it obtains the state feature representation corresponding to the historical disease information, the disease feature representation corresponding to the associated disease, and the relationship feature representation corresponding to the relationship type in the state relationship combination; based on the state feature representation corresponding to the disease to be predicted, the state feature representation corresponding to the historical disease information, and the relationship feature representation, it determines the correlation degree corresponding to the state relationship combination; it multiplies the correlation degree by the disease feature representation corresponding to the associated disease to obtain the propagation feature representation corresponding to the state relationship combination; and it statistically analyzes the propagation feature representations corresponding to each state relationship combination to obtain the propagation feature representation of the state relationship network for the disease to be predicted under the target set of historical disease information. The attention processing module is used to perform attention processing on the propagation feature representation under each historical disease information set, and to determine the state response feature representation of the patient in response to the disease to be predicted based on the attention processing results corresponding to each historical disease information set. An object feature determination module is used to determine the object feature representation of the patient based on the state response feature representation; The probability prediction module is used to determine the probability information of the patient having the disease to be predicted based on the object feature representation.

13. The state data processing apparatus according to claim 12, characterized by The attention processing module is further configured to obtain the medical visit time information corresponding to each of the historical disease information sets; obtain the time feature representation corresponding to each of the medical visit time information sets; fuse the propagation feature representation and time feature representation corresponding to each of the historical disease information sets to obtain the fused feature representation of each of the historical disease information sets; and perform attention processing on each fused feature representation to obtain the attention processing result corresponding to each of the historical disease information sets.

14. The state data processing apparatus according to claim 13, characterized by The attention processing module is further configured to determine the number of dimensions corresponding to the propagation feature representation; and to determine the time feature representation corresponding to the medical visit time information based on the medical visit time information, the number of dimensions, and the order of each element in the propagation feature representation.

15. The state data processing apparatus according to claim 13, characterized by The attention processing module is further configured to convert each fusion feature representation into a query representation, a key representation, and a value representation respectively; calculate the relevance between the query representation corresponding to the target fusion feature representation and each key representation respectively, and determine the attention weight corresponding to each key representation based on the relevance; multiply the attention weight corresponding to each key representation with its corresponding value representation, calculate the resulting products, and obtain the attention processing result corresponding to the target fusion feature representation based on the statistical results.

16. The state data processing apparatus according to claim 12, wherein The object feature determination module is further configured to take the associated disease set composed of the associated diseases corresponding to each historical disease information in the targeted historical disease information set as the current diagnostic status set, and take the status response feature representation as the current input feature; The process involves identifying the associated diseases corresponding to the current diagnostic states in the current diagnostic state set, and determining the current relationship type between the current diagnostic state and the corresponding associated diseases; the associated diseases and current relationship type are determined from the state relationship network; determining the current correlation between the current input features and the current diagnostic state under the current relationship type; based on the current correlation and the associated diseases corresponding to the current diagnostic state, determining the current propagation feature representation of the state relationship network for the state response feature representation under the current diagnostic state set; performing attention processing on the current propagation feature representation under each current diagnostic state set; and determining the state response feature representation of the patient for the disease to be predicted based on the attention processing results corresponding to each current diagnostic state set.

17. The state data processing apparatus according to claim 16, wherein The object feature determination module is further configured to use the determined state response feature representation as the current input feature, determine the associated disease set composed of the associated diseases corresponding to the current diagnostic state set as the current diagnostic state set, iteratively execute the step of determining the associated diseases corresponding to the current diagnostic state in the current diagnostic state set until the first preset iteration number is reached; and determine the object feature representation of the patient based on the state response feature representation obtained in the last iteration.

18. The state data processing apparatus according to claim 17, wherein The object feature determination module is further configured to arrange the state response feature representations, which are the current input features, and the state response feature representations obtained in the last iteration in chronological order to obtain a state response feature representation sequence; perform forward encoding on the state response feature representation sequence to obtain the encoding features of each diagnostic state response feature; and determine the object feature representation of the patient based on the encoding features corresponding to the state response feature representations obtained in the last iteration.

19. The state data processing apparatus according to claim 17, wherein The device further includes a graph propagation module, used to take the associated diseases corresponding to the historical disease information in the targeted historical disease information set as the current diagnostic state; determine the associated diseases corresponding to the current diagnostic state and the current relationship type between the current diagnostic state and the corresponding associated diseases from a pre-established state relationship network; combine the current diagnostic state, the associated diseases corresponding to the current diagnostic state, and the current relationship type into a state relationship combination and store it; take the associated diseases corresponding to the current diagnostic state as the current diagnostic state, and iteratively execute the step of determining the associated diseases corresponding to the current diagnostic state from the pre-established state relationship network until a second preset iteration number is reached; wherein, the first preset iteration number matches the second preset iteration number; the probability prediction module is also used to query the stored state relationship combinations to determine the associated diseases corresponding to the current diagnostic state in the current diagnostic state set.

20. The state data processing apparatus according to claim 12, wherein The device further includes a relationship network construction module, used to obtain a set of reference diagnostic states corresponding to multiple reference objects at different medical visit times; determine a first reference diagnostic state pair, wherein there is a parallel relationship between the two reference diagnostic states in the first reference diagnostic state pair; count the occurrence frequency of each first reference diagnostic state pair in the multiple sets of reference diagnostic states, classify the frequency levels based on the counted occurrence frequencies, and determine the frequency level to which the occurrence frequency of each first reference diagnostic state pair belongs as the relationship type corresponding to each first reference diagnostic state pair; A state relationship network is constructed based on each first reference diagnostic state pair and its corresponding relationship type.

21. The state data processing apparatus according to claim 20, wherein The relationship network construction module is also used to determine the second reference diagnostic state pair, in which there is an order relationship between the two reference diagnostic states in the second reference diagnostic state pair; according to the set of reference diagnostic states belonging to the same reference object, the frequency of occurrence of each second reference diagnostic state pair is counted, and frequency level is divided based on the counted frequency of occurrence, and the frequency level to which the frequency of occurrence of each second reference diagnostic state belongs is determined as the relationship type of each second reference diagnostic state pair. A state relationship network is constructed based on each first reference diagnostic state pair, the corresponding relationship type of each first reference diagnostic state pair, each second reference diagnostic state pair, and the corresponding relationship type of each second reference diagnostic state pair.

22. The state data processing apparatus according to any one of claims 12 to 21, characterized by, The attention processing module is further used to perform pooling processing on the attention processing results corresponding to each set of historical disease information to obtain the patient's state response feature representation for the disease to be predicted; wherein, the pooling processing includes max pooling or average pooling. 23.A computer device, comprising a memory and a processor, wherein the memory stores a computer program. When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 11.

24. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

25. A computer program product comprising a computer program, characterised in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

Images