A clinical demand mining method and system

Abstract

The application discloses a clinical demand mining method and system, and relates to the technical field of artificial intelligence, which comprises the following steps: firstly, constructing a medical knowledge network, integrating patient clinical elements and clinical indication elements, and identifying high-connection patient clinical elements; secondly, selecting samples and mining clinical transmission chains with the samples as cores; finally, based on the description parameters of the elements on the chain, a feature encoder and an adapter are trained in a cycle to obtain a model capable of accurately mapping patient features and clinical demands. Through the integration of structured knowledge, deep association mining and dynamic model optimization, the method improves the comprehensiveness and individualization level of clinical demand mining, and is suitable for demand identification of complex patient groups.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a clinical demand mining method and system. BACKGROUND

[0002] Clinical demand mining is a key link of medical decision support, aiming to accurately identify potential clinical demands through patient characteristics. Traditional methods rely on doctor experience summary or simple statistical analysis, which has significant limitations: the association between patient characteristics and clinical demands often presents a complex network, and traditional methods lack structured knowledge integration, making it difficult to capture deep associations. SUMMARY

[0003] The purpose of the present application is to provide a clinical demand mining method and system.

[0004] In a first aspect, the embodiments of the present application provide a clinical demand mining method, comprising:

[0005] Obtain a medical knowledge network and a description parameter of each clinical element in the medical knowledge network, the medical knowledge network comprising a plurality of clinical elements and an association between two clinical elements with interactive relationship, the plurality of clinical elements comprising a plurality of patient clinical elements representing patients and a plurality of clinical indication elements representing clinical demands, the plurality of patient clinical elements comprising a plurality of high-connection patient clinical elements, the high-connection patient clinical elements having an association with a plurality of clinical indication elements and a plurality of other patient clinical elements;

[0006] Select a plurality of sample clinical elements from the medical knowledge network, for each sample clinical element, determine a plurality of clinical transmission chains with the sample clinical element as the core clinical element from the medical knowledge network, the direct association level clinical elements on a clinical transmission chain are all patient clinical elements or clinical indication elements, and the clinical transmission chain with all direct association level clinical elements being patient clinical elements is a patient clinical transmission chain, the clinical transmission chain with all direct association level clinical elements being clinical indication elements is a clinical indication transmission chain, the plurality of sample clinical elements comprising a plurality of high-connection patient clinical elements, each clinical transmission chain of the high-connection patient clinical element comprising a plurality of patient clinical transmission chains and a plurality of clinical indication transmission chains;

[0007] According to the description parameter of the clinical elements on each clinical transmission chain of each sample clinical element, the initial feature encoder and the initial feature adapter are trained in a loop to obtain a feature encoder and a feature adapter, so as to mine clinical demands according to the feature encoder and the feature adapter.

[0008] In a second aspect, the embodiments of the present application provide a server system comprising a server, the server being configured to execute the method of the first aspect.

[0009] Compared with the prior art, the present application provides the beneficial effects including: by using the clinical demand mining method and system provided by the embodiments of the present application, a medical knowledge network is constructed, patient clinical elements and clinical indication elements are integrated, and high-connection patient clinical elements are identified; secondly, samples are selected and a clinical transmission chain with the samples as the core is mined; finally, a feature encoder and an adapter are trained in a loop based on the description parameters of the elements on the chain, and a model capable of accurately mapping patient features and clinical demands is obtained. The method improves the comprehensiveness and individualization level of clinical demand mining through structured knowledge integration, deep association mining and dynamic model optimization, and is suitable for demand identification of complex patient groups. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0011] Figure 1 The step flowchart of the clinical demand mining method provided by the embodiments of the present application is shown in the figure.

[0012] Figure 2 The structural schematic block diagram of the computer device provided by the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0014] The specific embodiments of the present application will be described in detail below in combination with the drawings.

[0015] In order to solve the technical problems in the foregoing background art, Figure 1 The flowchart of the clinical demand mining method provided by the embodiments of the present application is shown in the figure, and the clinical demand mining method will be described in detail below.

[0016] In step S201, a medical knowledge network and a description parameter of each clinical element in the medical knowledge network are obtained, the medical knowledge network includes a plurality of clinical elements and an association between two clinical elements with an interaction relationship, the plurality of clinical elements include a plurality of patient clinical elements representing a patient and a plurality of clinical indication elements representing a clinical demand, and the plurality of patient clinical elements include a plurality of high-connection patient clinical elements, the high-connection patient clinical elements are associated with a plurality of clinical indication elements and a plurality of other patient clinical elements;

[0017] In step S202, a plurality of sample clinical elements are selected from the medical knowledge network, for each sample clinical element, a plurality of clinical transmission chains with the sample clinical element as a core clinical element are determined from the medical knowledge network, the direct association level clinical elements on a clinical transmission chain are all patient clinical elements or clinical indication elements, a clinical transmission chain with all direct association level clinical elements being patient clinical elements is a patient clinical transmission chain, and a clinical transmission chain with all direct association level clinical elements being clinical indication elements is a clinical indication transmission chain, the plurality of sample clinical elements include a plurality of high-connection patient clinical elements, and each clinical transmission chain of the high-connection patient clinical elements includes a plurality of patient clinical transmission chains and a plurality of clinical indication transmission chains.

[0018] In step S203, an initial feature encoder and an initial feature adapter are cyclically trained according to the description parameters of the clinical elements on each clinical transmission chain of each sample clinical element, a feature encoder and a feature adapter are obtained, and a clinical demand is mined according to the feature encoder and the feature adapter.

[0019] In the embodiments of the present application, for example, first, the server needs to build a medical knowledge network and obtain the description parameters of each clinical element in the network. The construction of the medical knowledge network is based on the integration of multiple sources of medical data, which comes from hospital information systems (such as electronic medical record systems, laboratory test systems), medical guideline databases and scientific research literature, covering patients' basic information (such as age, gender), symptoms (such as polydipsia and polyuria), signs (such as body mass index 30), test results (such as fasting blood glucose 8.2 mmol / L), treatment measures (such as subcutaneous injection of insulin), nursing needs (such as blood glucose monitoring 4 times a day) and the like. The clinical elements in the network are divided into two categories: one is the "patient clinical element" representing the individual characteristics of the patient, for example, "diagnosis of type 2 diabetes", "obesity (body mass index 30)", "history of hypertension (systolic blood pressure 150 mmHg)"; the other is the "clinical indication element" representing the clinical needs, for example, "insulin therapy (0.5 U / kg)", "diabetes diet guidance", "fundus examination (once every 6 months)". The server extracts entities (such as "fasting blood glucose" and "insulin therapy") in the electronic medical record through natural language processing technology, and uses graph analysis methods (such as graph neural networks) to mine the interaction between elements, for example, "fasting blood glucose ≥7.0 mmol / L" frequently co-occurs with "diagnosis of type 2 diabetes" (such as 90% of diabetes medical records appear at the same time), "diagnosis of type 2 diabetes" is further associated with "insulin therapy", "diet control" and other clinical indication elements and "obesity", "history of hypertension" and other patient clinical elements. On this basis, the server calculates the connectivity of the patient clinical elements (i.e. the number of associations with other elements), if a patient clinical element is associated with multiple clinical indication elements (such as more than 5) and multiple other patient clinical elements (such as more than 3) at the same time, it is marked as "highly connected patient clinical element" (for example, "diagnosis of type 2 diabetes" is connected to 12 other elements); patient clinical elements that are connected to only a small number or no indication elements (such as "simple obesity (body mass index 28)") are "low connectivity patient clinical elements". The description parameters of each clinical element need to be structured: numerical parameters (such as fasting blood glucose value) are processed by standardization (such as normalized to 0-1 range), categorical parameters (such as gender) are converted to binary encoding (such as male represented as [1, 0]), text parameters (such as complaint "polydipsia and polyuria for 1 month") are extracted by semantic analysis to obtain feature vectors, and time series parameters (such as blood glucose monitoring values in the last 3 months) are converted to statistical features (such as the average value of the last 7 days, the fluctuation range).

[0020] Subsequently, the server selects sample clinical elements from the medical knowledge network and determines a clinical transmission chain centered on each sample. The samples include high-connection patient clinical elements (such as "diagnosis of type 2 diabetes"), low-connection patient clinical elements (such as "simple obesity"), and clinical indication elements (such as "insulin therapy"). For each sample, the server extracts the associated path centered on the sample by traversing the knowledge network (such as breadth-first search) to form a "clinical transmission chain". Clinical transmission chains are divided into two categories: one is a "patient clinical transmission chain", in which all nodes on the path are patient clinical elements, for example, a patient clinical transmission chain centered on "diagnosis of type 2 diabetes" is "diagnosis of type 2 diabetes - obesity (body mass index 30) - hypertension history (systolic blood pressure 150 mmHg)" (path length 2, each layer is a patient element); the other is a "clinical indication transmission chain", in which all nodes on the path are clinical indication elements, for example, a clinical indication transmission chain centered on "diagnosis of type 2 diabetes" is "diagnosis of type 2 diabetes - insulin therapy (0.5 U / kg) - 4 times daily blood glucose monitoring" (path length 2, each layer is an indication element). The clinical transmission chain of a high-connection patient clinical element contains multiple patient transmission chains and indication transmission chains, while the transmission chain of a low-connection patient clinical element contains only patient transmission chains (as there are no indication elements associated).

[0021] Based on the above samples and transmission chains, the server cyclically trains the initial feature encoder and the initial feature adapter to optimize the model performance. The initial feature encoder is used to map the description parameters of each clinical element on the transmission chain to the feature vector (such as a 128-dimensional vector) of the core element, and its structure can adopt a multi-layer neural network (such as Transformer), with the input being the description parameters of each element on the transmission chain (such as the obesity feature and the hypertension feature of the patient transmission chain, the insulin dose feature and the monitoring frequency feature of the indication transmission chain), and the output being the clinical element feature of the core element. The initial feature adapter is used to convert the feature of the patient clinical element into a feature adapted to the clinical indication, and its structure can adopt a simple fully connected network. During training, the server optimizes the model parameters through cyclic iteration, and the overall error is composed of three parts:

[0022] The first part is the "first error parameter" of the high-connection patient clinical element, which is used to measure the deviation between its expected feature and the converted feature. The expected feature of the high-connection patient clinical element needs to integrate the information of its patient transmission chain (reflecting the patient's own characteristics) and indication transmission chain (reflecting the clinical demand association), for example, the expected feature of "diagnosis of type 2 diabetes" needs to reflect the patient's characteristics such as obesity and hypertension, as well as the indication demand such as insulin therapy and blood glucose monitoring. The converted feature is the conversion result of the adapter on the patient transmission chain feature (basic feature). The server calculates the error by comparing the difference between the expected feature and the converted feature (such as the distance between vectors) to drive the adapter to learn and integrate the indication demand information.

[0023] The second part is a "second error variable" of high connection patient clinical elements, used to ensure that the degree of adaptation with the associated indications is significantly higher than that of irrelevant indications. The server screens the indications associated with the high connection patient elements (such as "insulin therapy" and "diet control") and the irrelevant indications (such as "antibiotic therapy" and "fracture fixation") from the knowledge network, forming multiple sets of indication pairs (each set containing 1 associated indication and 1 irrelevant indication). For each indication pair, the server calculates the degree of adaptation (such as vector similarity) of the patient conversion characteristics with the associated indication characteristics and the irrelevant indication characteristics, requiring the degree of adaptation of the associated indication to be significantly higher than that of the irrelevant indication, and if not met, an error is generated to drive the model to enhance the discrimination ability of the associated indication.

[0024] The third part is a "third error variable" of patient clinical elements, used to ensure that the characteristics of patients in the same group are similar. The server initializes patient groups according to clinical guidelines (such as "uncontrolled diabetes group" and "well-controlled group"), and each group corresponds to a representative characteristic (such as "uncontrolled group" representative characteristic is high fasting blood glucose and high glycosylated hemoglobin). For each patient clinical element, the server calculates the degree of adaptation (such as similarity) of its characteristics with the representative characteristics of each group, and determines the probability of the patient belonging to each group based on the degree of adaptation (the higher the degree of adaptation, the greater the probability of belonging). By comparing the difference between the actual attribution probability and the theoretical attribution probability of the patient, the error is calculated to drive the model to map the characteristics of the same type of patient to a similar space.

[0025] During training, the server adjusts the parameters of the feature encoder and the adapter (such as through gradient descent optimization) based on the above three parts of errors, and re-divides the patient groups according to the current training results (patients are attributed to the group with the highest degree of adaptation, and the group representative characteristics are updated to the average of the patient characteristics in the group), and this process is repeated until the error is stable (such as the error change is less than the set threshold for consecutive multiple rounds).

[0026] After the training is completed, the server can perform clinical demand mining based on the trained feature encoder and the adapter. Taking "patient B (55 years old, male, body mass index 32, fasting blood glucose 8.5 mmol / L, no hypertension)" as an example: the server first extracts the patient clinical elements (such as "body mass index 32" "fasting blood glucose 8.5 mmol / L") of patient B from the electronic medical record of patient B, and is associated with the constructed medical knowledge network (such as "body mass index 32" is associated with "obesity", "fasting blood glucose 8.5 mmol / L" is associated with "diabetes may occur"). Subsequently, the server takes "diabetes may occur in patient B" as the core, extracts the patient clinical transmission chain (such as "diabetes may occur-obesity (body mass index 32)") and the transmission chain of the clinical indications to be evaluated (such as "insulin therapy-4 times of blood glucose monitoring per day" "oral hypoglycemic drugs-1 time of blood glucose monitoring per week"). The description parameters of the patient transmission chain are input into the feature encoder to obtain the feature vector of the patient clinical elements; the transmission chain parameters of each to-be-evaluated indication are input into the encoder to obtain the feature vector of the indication. Through the adapter, the patient feature vector is converted into the feature of the adapted indication, the adaptation degree (such as similarity) of each indication is calculated, and finally the indication with the highest adaptation degree (such as "insulin therapy") is selected as the target clinical demand of patient B.

[0027] In summary, the method realizes the accurate mapping from patient features to clinical demands by constructing a structured medical knowledge network, mining multiple types of clinical transmission chains, and optimizing the feature model combined with multi-task learning, and is suitable for personalized management scenarios of chronic diseases such as diabetes and cardiovascular diseases.

[0028] In the embodiment of the application, the clinical demand mining according to the feature encoder and the feature adapter can be implemented by the following examples.

[0029] A medical knowledge network is constructed and description parameters of each clinical element in the medical knowledge network are obtained, the medical knowledge network includes multiple clinical elements and there is an association between two clinical elements with interactive relationship, the multiple clinical elements include multiple patient clinical elements representing patients and multiple clinical indication elements representing to-be-evaluated clinical demands; the multiple patient clinical elements include a target patient clinical element representing a target patient;

[0030] In the medical knowledge network, a plurality of first clinical transmission chains with the target patient clinical element as the core clinical element are determined, and a plurality of second clinical transmission chains with the clinical indication element of the to-be-evaluated clinical demand as the core clinical element are determined;

[0031] inputting each first clinical transmission chain corresponding to the target patient clinical element and the description parameters of the clinical elements on each first clinical transmission chain into a feature encoder to obtain the clinical element features of the target patient clinical element, inputting each second clinical transmission chain corresponding to each to-be-evaluated clinical demand and the description parameters of the clinical elements on each second clinical transmission chain into the feature encoder to obtain the clinical element features of each to-be-evaluated clinical demand;

[0032] if all the clinical elements associated with the target patient clinical element are clinical indication elements, it is determined that each first clinical transmission chain of the target patient is a patient clinical transmission chain, then the clinical element features of the target patient clinical element are converted by a feature adapter to obtain converted features, and for each to-be-evaluated clinical demand, the clinical indication adaptation degree between the target patient and the to-be-evaluated clinical demand is obtained according to the converted features and the clinical element features of the to-be-evaluated clinical demand;

[0033] According to the clinical indication adaptation degree corresponding to each to-be-evaluated clinical demand, the target clinical demand for the target patient is determined from the to-be-evaluated clinical demands.

[0034] In the embodiments of the present application, the server can conduct clinical demand mining for a specific target patient based on the learned model capability after completing the training of the feature encoder and the feature adapter. The following takes the actual scene of "a 55-year-old male patient Li" as an example to illustrate the execution steps of the process in detail. The server first constructs a special medical knowledge network for the target patient Li. The electronic medical record of Li shows: chief complaint "polydipsia and polyuria for 2 months", fasting blood glucose 8.5 mmol / L (reference value 3.9-6.1), BMI 32 (obesity), no history of hypertension, no family history of diabetes. The server extracts "patient clinical elements" from these data: including basic information (age 55, male), symptoms (polydipsia and polyuria), signs (BMI 32), and examination results (fasting blood glucose 8.5 mmol / L). At the same time, the server extracts "to-be-evaluated clinical indication elements" from medical guidelines and historical medical records, i.e. possible clinical demands, such as "insulin subcutaneous injection (0.5 U / kg)", "metformin oral administration (500 mg bid)", "diabetes diet guidance (300 g of carbohydrates per day)", "3 times of aerobic exercise per week (30 minutes / time)", etc. The server maps the above elements to the pre-constructed global medical knowledge network to establish a correlation relationship: for example, "fasting blood glucose 8.5 mmol / L" is associated with "high possibility of diabetes" (90% of patients with fasting blood glucose ≥7.0 mmol / L in historical data are eventually diagnosed with diabetes); "BMI 32" is associated with "obesity-related risk of diabetes" (70% of obese patients have insulin resistance); "insulin subcutaneous injection" is associated with "rapid control of fasting blood glucose" (guidelines recommend using insulin when fasting blood glucose ≥8.0 mmol / L). At this time, in the medical knowledge network of the target patient, "fasting blood glucose 8.5 mmol / L" as the core "target patient clinical element" forms an interactive correlation with other patient clinical elements (such as "BMI 32" and "polydipsia and polyuria") and to-be-evaluated indication elements (such as "insulin therapy" and "diet guidance"). The server takes "fasting blood glucose 8.5 mmol / L" as the core target patient clinical element and extracts "first clinical transmission chain" (patient clinical transmission chain) by traversing the knowledge network. For example: chain 1: "fasting blood glucose 8.5 mmol / L-BMI 32 (obesity)-insulin resistance (derived according to the metabolic characteristics of people with BMI ≥28)"; chain 2: "fasting blood glucose 8.5 mmol / L-polydipsia and polyuria (symptoms)-increased extracellular osmotic pressure (pathological mechanism)". The nodes on these chains are all patient clinical elements (blood glucose value, BMI, symptoms, pathological mechanism), so they belong to patient clinical transmission chain. At the same time, the server extracts "second clinical transmission chain" (clinical indication transmission chain) for each to-be-evaluated clinical indication element.For example, taking the indication to be evaluated "insulin subcutaneous injection (0.5 U / kg)" as the core, the chain is extracted: "insulin subcutaneous injection - daily 4 times fingertip blood glucose monitoring (avoiding hypoglycemia) - adjusting the dose according to the monitoring results (dynamic management)"; taking the indication to be evaluated "metformin oral" as the core, the chain is extracted: "metformin oral - taking 30 minutes after meals (reducing gastrointestinal reactions) - monitoring liver and kidney function every month (drug safety)". The nodes on these chains are all clinical indication elements (dosing method, monitoring frequency, follow-up requirements), so they belong to the clinical indication transmission chain. The server inputs the first clinical transmission chain of the target patient and its description parameters into the trained feature encoder. For example, the description parameters of chain 1 include: fasting blood glucose 8.5 mmol / L (standardized value 2.17), BMI 32 (encoded as [0, 1, 0], corresponding to the "obesity" category), insulin resistance (text feature vector). The encoder processes these parameters through a multi-layer neural network and outputs a feature vector (such as 128 dimensions) of the target patient's clinical elements, which comprehensively reflects the patient's blood glucose level, obesity degree, and pathological mechanism. For each clinical indication element to be evaluated, the server inputs its second clinical transmission chain and description parameters into the same feature encoder. For example, the transmission chain description parameters of the indication "insulin subcutaneous injection" include: dose 0.5 U / kg (numerical encoding), monitoring frequency 4 times a day (category encoding), dose adjustment rules (text vector). The encoder outputs a feature vector (128 dimensions) of the indication, reflecting its treatment intensity, monitoring requirements, and dynamic adjustment characteristics. Since the target patient's clinical element "fasting blood glucose 8.5 mmol / L" is associated with elements in the knowledge network, including patient clinical elements (such as BMI 32) and clinical indication elements (such as insulin therapy), but according to the construction results of step 1, both patient elements and indication elements are directly associated with the clinical elements, further judgment is needed. Assuming that after verification, there is a patient element (such as BMI 32) in the directly associated elements of the target patient's clinical elements, therefore it is necessary to confirm whether all the first clinical transmission chains are patient clinical transmission chains. After traversal, all the nodes of the first transmission chains are patient clinical elements (such as blood glucose, BMI, symptoms, and pathological mechanisms in chain 1 and chain 2), so the server determines that the feature of the target patient's clinical elements needs to be converted through the feature adapter. The server calls the feature adapter to convert the feature vector of the target patient into a "converted feature" (such as 128 dimensions) that is suitable for the clinical indication. This conversion process has been learned through training how to map the patient's own features (such as high blood sugar and obesity) to a space that matches the requirements of the clinical indication. Subsequently, the server calculates the "clinical indication adaptation degree" of the converted feature and the feature of each indication to be evaluated, specifically by measuring the vector similarity (such as cosine similarity): for example, the similarity between the converted feature and the "insulin therapy" feature is 0.92, the similarity between the converted feature and the "metformin oral" feature is 0.75, and the similarity between the converted feature and the "diet guidance" feature is 0.81.The server ranks the adaptation degrees of each clinical indication to be evaluated, and selects the indication with the highest adaptation degree as the target clinical demand. In the case of Li, the adaptation degree of "insulin subcutaneous injection (0.5 U / kg)" is the highest (0.92), indicating that the indication is highly matched with the patient's characteristics of hyperglycemia, obesity and insulin resistance. Therefore, the server outputs "insulin subcutaneous injection (0.5 U / kg) combined with blood glucose monitoring 4 times a day" as the target clinical demand for Li. In summary, the server realizes the precise matching from patient characteristics to clinical demand by constructing the medical knowledge network of the target patient, mining the transmission chain, encoding the characteristics, converting the adaptation and calculating the similarity, and provides data-driven support for clinical decision-making.

[0035] In the embodiment of the present application, the overall error of the cycle training includes a first error parameter of each high-connection patient clinical element, and the first error parameter is used to represent the deviation between the expected feature representation of the corresponding high-connection patient clinical element in each cycle period and the converted feature obtained by converting the basic representation feature of the corresponding high-connection patient clinical element in each cycle period through the initial feature adapter.

[0036] The expected feature representation is a clinical element feature or an associated feature representation; the initial feature encoder is used to input the description parameters of the clinical elements on each clinical transmission chain of the sample clinical element, and output the clinical element features of the sample clinical element; if each clinical transmission chain of the sample clinical element includes multiple patient clinical transmission chains, the clinical element features include the basic representation features obtained according to the description parameters of the clinical elements on each patient clinical transmission chain; if each clinical transmission chain of the sample clinical element includes multiple clinical indication transmission chains, the clinical element features include the associated feature representations obtained according to the description parameters of the clinical elements on each clinical indication transmission chain.

[0037] In the embodiments of the present application, the server needs to constrain the accuracy of the feature representation of the high-connection patient clinical factor through the "first error parameter" in the cyclic training of the initial feature encoder and the feature adapter. Taking the high-connection patient clinical factor "diagnosis of type 2 diabetes" as an example, the server calculates the error parameter as follows. "Diagnosis of type 2 diabetes" is a high-connection patient clinical factor in the medical knowledge network, and its associated clinical factors include patient clinical factors (such as "BMI ≥ 28 (obesity)" and "history of hypertension") and clinical indication factors (such as "insulin subcutaneous injection" and "4 times of blood glucose monitoring per day"). The server extracts the clinical transmission chain of "diagnosis of type 2 diabetes" from the knowledge network: the patient clinical transmission chain (only containing patient clinical factors): for example, "diagnosis of type 2 diabetes-BMI30 (obesity)-history of hypertension (systolic blood pressure 150 mmHg)"; the clinical indication transmission chain (only containing clinical indication factors): for example, "diagnosis of type 2 diabetes-insulin subcutaneous injection (0.5 U / kg)-4 times of blood glucose monitoring per day". The server inputs the description parameters of the patient clinical transmission chain into the initial feature encoder. Taking "diagnosis of type 2 diabetes-BMI30-history of hypertension" as an example, the description parameters include: the standardized value of BMI30 (such as 0.8), the binary encoding of the history of hypertension (such as [1, 0] indicating a history), and the normalized value of systolic blood pressure 150 mmHg (such as 0.7). The encoder processes these parameters through a multi-layer neural network and outputs "basic representation features" (such as a 128-dimensional vector), which focus on the patient's own characteristics (obesity, hypertension). At the same time, the server inputs the description parameters of the clinical indication transmission chain into the same encoder. Taking "diagnosis of type 2 diabetes-insulin injection-blood glucose monitoring" as an example, the description parameters include: the numerical encoding of the insulin dose 0.5 U / kg (such as 0.6), and the category encoding of the monitoring frequency 4 times per day (such as [0, 1, 0]). The encoder outputs "associated feature representation" (a 128-dimensional vector), which focuses on the clinical indication requirements (drug intensity, monitoring density). The server sets the "expected feature representation" according to the clinical logic, which needs to integrate the association of the patient's own characteristics and the indication requirements. For example, the features of "diagnosis of type 2 diabetes" should reflect the patient's obesity, hypertension (basic representation features), and the necessity of insulin treatment and blood glucose monitoring (associated feature representation). The server generates the expected features by weighted fusion of the two: adding the basic representation features and the associated feature representation according to the weight of 0.6:0.4 (the weight is adjusted according to the clinical importance), to obtain the final expected feature vector (such as 128 dimensions). The initial feature adapter receives the basic representation features (the patient's own characteristics) and converts them into "converted features" (128 dimensions) through a fully connected layer. The goal of designing the adapter is to make the converted features as close as possible to the expected features (i.e., containing both patient characteristics and indication association information). For example, if the basic representation features only reflect "obesity", the adapter needs to learn to convert them into a vector that contains "insulin treatment" information.The server calculates a first error parameter by comparing the difference between the expected feature and the converted feature. For example, if the 50th dimension value of the expected feature is 0.7 (representing "insulin treatment necessity"), and the 50th dimension value of the converted feature is only 0.3 (not fully reflecting the indication demand), the deviation of this dimension is 0.4; the average value (such as mean square error) of the deviation of each dimension is calculated by traversing all 128 dimensions, and finally the first error parameter is obtained. The larger the error is, the more the adapter fails to effectively convert the patient's own features into features containing indication associations, and the adapter parameters (such as weights and biases) need to be adjusted through back propagation. In each round of training, the server repeats the above steps: extracting the conduction chain-encoding to obtain the basic and associated features-generating the expected feature-adapter converting the basic feature-calculating the error-updating the model parameters. For example, in the initial training, the adapter may only retain "obesity" information, resulting in a large deviation (error 0.5) between the converted feature and the expected feature; after multiple rounds of adjustment, the adapter gradually learns the association between "obesity" and "insulin treatment", and the "insulin treatment necessity" dimension value of the converted feature is increased to 0.6, and the error is reduced to 0.2, and finally converges to a stable state (error <0.1). In summary, the server constrains the feature conversion accuracy of high-connection patient clinical elements through the first error parameter, ensures that the model learns the internal association between the patient's own features and the clinical indication demand, and lays a foundation for subsequent clinical demand mining.

[0038] In the embodiment of the application, the plurality of sample clinical elements further include clinical indication elements;

[0039] The overall error of the cyclic training further includes a second error parameter of each high-connection patient clinical element;

[0040] In each cyclic training, the second error parameter of each high-connection patient clinical element is obtained through the following process, which can be implemented through the following example.

[0041] From the medical knowledge network, clinical indication elements associated with the high-connection patient clinical elements are selected as associated clinical indication elements, and clinical indication elements not associated with the high-connection patient clinical elements are selected as irrelevant clinical indication elements;

[0042] Traverse each associated clinical indication element and each irrelevant clinical indication element to obtain a plurality of clinical indication element pairs, each clinical indication element pair including one associated clinical indication element and one irrelevant clinical indication element;

[0043] For each pair of clinical indication elements, a first clinical indication fitness of the high-connectivity patient clinical element and the associated clinical indication element is obtained according to the clinical element feature of the high-connectivity patient clinical element and the associated clinical indication element in the current cycle, a second clinical indication fitness of the high-connectivity patient clinical element and the unassociated clinical indication element is obtained according to the clinical element feature of the high-connectivity patient clinical element and the unassociated clinical indication element in the current cycle, and an indication association error of the high-connectivity patient clinical element for the pair of clinical indication elements is obtained according to the inverse of the deviation between the first clinical indication fitness and the second clinical indication fitness.

[0044] A second error parameter of the high-connectivity patient clinical element is obtained according to the sum of the indication association errors of the high-connectivity patient clinical element for all pairs of clinical indication elements.

[0045] In the embodiments of the present application, the server needs to constrain the ability of the adapter to distinguish the adaptation degree of the high-connection patient clinical elements and the associated indications through a "second error variable" in addition to the first error variable when training the feature encoder and the adapter in a loop. Taking the high-connection patient clinical element of "diagnosis of type 2 diabetes" as an example, the specific process of the server calculating the second error variable is described in detail. The server first screens the clinical indication elements (referred to as "associated indications") associated with "diagnosis of type 2 diabetes" from the medical knowledge network. According to the historical medical record data, "diagnosis of type 2 diabetes" often co-occurs with "subcutaneous injection of insulin (0.5 U / kg)", "diabetes diet guidance (300 g of carbohydrates per day)", "3 times of aerobic exercise per week (30 minutes each time)", and the like (such as 90% of the medical records of diabetes containing at least one such indication), so these indications are marked as associated indications. At the same time, the server screens the clinical indication elements (referred to as "unrelated indications") unrelated to "diagnosis of type 2 diabetes". For example, "antibiotic treatment (for pneumonia)", "fracture fixation (for lower limb fracture)", "thyroid hormone replacement therapy (for hypothyroidism)", and the like rarely co-occur with "diagnosis of type 2 diabetes" (such as a co-occurrence rate of <5%), so they are marked as unrelated indications. The server traverses all combinations of the associated indications and the unrelated indications to generate "clinical indication element pairs". For example, the associated indications include "insulin therapy" and "diet guidance", and the unrelated indications include "antibiotic treatment" and "fracture fixation", so the generated element pairs are: (insulin therapy, antibiotic treatment), (insulin therapy, fracture fixation), (diet guidance, antibiotic treatment), and (diet guidance, fracture fixation). Each pair contains one associated indication and one unrelated indication for subsequent comparison. For each indication element pair, the server needs to calculate the adaptation degree of the high-connection patient clinical element ("diagnosis of type 2 diabetes") and the associated indication and the unrelated indication. Obtain the feature vector: In the current training loop, the clinical element features of "diagnosis of type 2 diabetes" have been generated by the feature encoder (such as a 128-dimensional vector (F patient )); the clinical element features of the associated indication (such as "insulin therapy") and the unrelated indication (such as "antibiotic treatment") are also generated by the encoder (F insulin ) and (F antibiotic ), respectively). Calculate the adaptation degree: the server measures the adaptation degree by vector similarity (such as cosine similarity). For example, the cosine similarity of (F patient ) and (F insulin ) is 0.85 (the first clinical indication adaptation degree), and the cosine similarity of (F patient ) and (F antibioticThe cosine similarity of the first and second adaptation degrees is 0.3 (second clinical indication adaptation degree). The server calculates the indication association error according to the deviation of the first and second adaptation degrees. The design goal is to make the adaptation degree of the associated indication significantly higher than that of the irrelevant indication, so the error is defined as "second adaptation degree - first adaptation degree + margin value" (the margin value is usually set to 0.2 to ensure sufficient discrimination), if the result is positive, an error occurs; if it is negative or zero, the error is 0. Take the element pair (insulin treatment, antibiotic treatment) as an example: the first adaptation degree is 0.85, the second adaptation degree is 0.3, the deviation is 0.3-0.85=-0.55, and after adding the margin value 0.2, it is-0.35 (≤0), so the indication association error is 0 (no adjustment is needed). If the first adaptation degree of a certain element pair (diet guidance, fracture fixation) is 0.7, the second adaptation degree is 0.6, the deviation is 0.6-0.7=-0.1, and after adding the margin value 0.2, it is 0.1 (>0), then the indication association error is 0.1 (the difference needs to be reduced through training). The server adds the indication association errors of all indication element pairs corresponding to "diagnosis of type 2 diabetes" to obtain the second error parameter of the clinical elements of the high connection patient. For example, if the errors of the four element pairs are 0, 0, 0.1, and 0.05 respectively, the second error parameter is 0.15. The second error parameter drives the model to learn to distinguish between associated and irrelevant indications. During initial training, the model may mistakenly increase the adaptation degree of the characteristics of "diagnosis of type 2 diabetes" and "antibiotic treatment" to 0.6 (close to the adaptation degree of 0.7 of "insulin treatment"), resulting in an error of 0.1 (0.6-0.7+0.2=0.1); by adjusting the feature encoder parameters (such as enhancing the association between "diabetes" and "insulin"), the adaptation degree of "insulin treatment" is increased to 0.85 and the adaptation degree of "antibiotic treatment" is reduced to 0.3 in the subsequent cycle, and the error disappears. The model finally learns to accurately identify associated indications. In summary, the server restricts the ability of the high connection patient clinical elements to distinguish the adaptation degrees of associated and irrelevant indications through the second error parameter, ensuring that the model can accurately capture the internal association between patient characteristics and clinical needs, and providing a reliable feature basis for clinical demand mining.

[0046] In the embodiments of the present application, the overall error of the cycle training further includes a third error parameter of the patient clinical elements in each sample clinical element;

[0047] The third error parameter of the patient clinical elements in each sample clinical element is obtained through the following process, which can be implemented through the following examples.

[0048] For each cycle training, the group feature identifier of the group representative point of each patient group in the current cycle and the clinical element feature of each patient clinical element in the current cycle are obtained, wherein the group feature identifier of the group representative point of each patient group in the initial group division is the initial feature setting value;

[0049] For each patient clinical factor, a clinical indication fitness between the patient clinical factor and each group representative point is identified according to a clinical factor feature of the patient clinical factor in the current cycle and a group feature of each group representative point;

[0050] According to the clinical indication fitness between each patient clinical factor and each group representative point, a group attribution probability that each patient clinical factor belongs to each patient group is obtained;

[0051] For each patient clinical factor, a third error parameter of the patient clinical factor is obtained according to a deviation between the clinical indication fitness between the patient clinical factor and each group representative point and the group attribution probability that the patient clinical factor belongs to the corresponding patient group.

[0052] In the embodiments of the present application, for example, the server in the cycle training, through the "third error parameter", the attribution consistency of the patient clinical factor between the groups is constrained, and it is ensured that the feature representation of the patients of the same kind is similar. The following takes the division of the diabetic patient group as an example to explain in detail the specific process of the server calculating the third error parameter. The server initializes the patient group according to the clinical guidelines. For example, the diabetic patients are divided into three categories: the uncontrolled group: the initial representative point feature is set as "fasting blood glucose ≥ 7.0 mmol / L, glycosylated hemoglobin (HbA1c) ≥ 7.5%" (corresponding feature vector such as [8.5, 8.0]); the well-controlled group: the initial representative point feature is set as "fasting blood glucose < 7.0 mmol / L, HbA1c < 7.0%" (corresponding feature vector such as [5.8, 6.5]); the obesity combined group: the initial representative point feature is set as "BMI ≥ 28 and diagnosed diabetes" (corresponding feature vector such as [30, 7.2]). The representative point feature of each group is the "baseline feature" initially set for subsequent fitness calculation. Taking the sample clinical factor "patient A" as an example, the clinical factor feature (generated by the feature encoder) is [fasting blood glucose 8.2 mmol / L, HbA1c 7.8%, BMI 29] (corresponding vector (F patient )). The server calculates the "clinical indication fitness" (measured by vector similarity, such as cosine similarity) of (F patient ) and each group representative point feature: the similarity with the "uncontrolled group" representative point [8.5, 8.0] is 0.9 (high matching); the similarity with the "well-controlled group" representative point [5.8, 6.5] is 0.3 (low matching); the similarity with the "obesity combined group" representative point [30, 7.2] is 0.7 (moderate matching). The server calculates the probability that "patient A" belongs to each group according to the fitness. The fitness is normalized by using the Softmax function to ensure that the probability sum is 1. For example: the uncontrolled group fitness 0.9-exponential value (e 0.9≈2.46); control good group fitness 0.3 - index value (e 0.3 ≈1.35); obesity combined group fitness 0.7 - index value (e 0.7 ≈2.01); total index sum 2.46+1.35+2.01=5.82; attribution probability: uncontrolled group (2.46 / 5.82≈0.42), control good group (1.35 / 5.82≈0.23), obesity combined group (2.01 / 5.82≈0.35). The server compares the deviation of patient fitness and attribution probability to generate a third error. The deviation is defined as the "difference between fitness and attribution probability", the larger the deviation, the greater the error. For example: uncontrolled group: fitness 0.9, attribution probability 0.42 - deviation 0.9-0.42=0.48; control good group: fitness 0.3, attribution probability 0.23 - deviation 0.3-0.23=0.07; obesity combined group: fitness 0.7, attribution probability 0.35 - deviation 0.7-0.35=0.35; total deviation (third error parameter): 0.48+0.07+0.35=0.9. After each round of training, the server adjusts the feature encoder parameters according to the error (such as enhancing the association between "high blood sugar" and "uncontrolled group" features), and re-divides the groups: patients are re-attributed to the group with the highest attribution probability, and the group representative point features are updated (such as the new representative point of the "uncontrolled group" taking the average of the patient features in the group). For example, if "patient A" finally belongs to the "uncontrolled group", the group representative point feature is updated to [8.3, 7.9] (the average blood sugar and HbA1c of patients in the group). After multiple rounds of training, the error gradually decreases (such as from 0.9 to 0.2), and the patient features are concentrated in the group to which they belong, and the feature representation of patients with the same type tends to be consistent. In summary, the server constrains the attribution consistency of patient groups through the third error parameter, ensuring that the model learns the common features of patients with the same type, and provides more accurate patient classification basis for clinical needs.

[0053] In the embodiments of the present application, the first error parameter of the high connection patient clinical element is obtained through the following process, which can be implemented through the following example.

[0054] According to the similarity measurement of each group representative point and the expected feature representation of the high connection patient clinical element in the current cycle, the group feature identifier of each group representative point is weighted and fused as the first auxiliary reinforcement feature, the expected feature representation is reinforced according to the first auxiliary reinforcement feature, and the expected feature representation after feature reinforcement is obtained as the high connection feature of the high connection patient clinical element in the current cycle;

[0055] According to a similarity measurement of each group representative point with a basic representation feature of the high-connectivity patient clinical factor in the current cycle, a group feature identifier of each group representative point is weighted and fused as a second auxiliary reinforcement feature, the basic representation feature is reinforced according to the second auxiliary reinforcement feature, a feature-reinforced basic representation feature is obtained, and the feature-reinforced basic representation feature is converted by the initial feature adapter to obtain a converted feature of the high-connectivity patient clinical factor in the current cycle.

[0056] According to a deviation between the high-connectivity feature and the converted feature of the high-connectivity patient clinical factor in the current cycle, a first error parameter of the high-connectivity patient clinical factor is obtained.

[0057] In the embodiments of the present application, the server may, for example, strengthen the feature representation of the patient group in calculating the first error parameter of the high-connectivity patient clinical factor, so as to improve the clinical relevance of the feature. The following takes the high-connectivity patient clinical factor of “diagnosis of type 2 diabetes” as an example to illustrate the specific execution process. In the current training cycle, the “expected feature” of “diagnosis of type 2 diabetes” is a feature vector generated by synthesizing the patient clinical transmission chain (such as “obesity-high blood pressure”) and the clinical indication transmission chain (such as “insulin treatment-blood glucose monitoring”) information of the patient. The vector reflects both the patient's own characteristics (such as obesity level) and the clinical needs (such as the necessity of insulin treatment). Meanwhile, the “basic representation feature” is only encoded by the patient clinical transmission chain, and mainly focuses on the patient's own characteristics (such as obesity and high blood pressure). The server needs to strengthen the expected feature in combination with the patient group information of the current cycle. Assume that the current patient group is divided into three categories: the uncontrolled group, the well-controlled group, and the obesity-merging group. The representative point of the uncontrolled group reflects the characteristics of “high blood glucose and high insulin demand” (such as more prominent fasting blood glucose and frequent monitoring). The representative point of the well-controlled group reflects the characteristics of “stable blood glucose and low insulin demand” (such as more prominent normal fasting blood glucose and low monitoring frequency). The representative point of the obesity-merging group reflects the characteristics of “severe obesity and moderate insulin demand” (such as more prominent high BMI and dietary intervention). The server first calculates the similarity between the expected feature of “diagnosis of type 2 diabetes” and the representative points of each group (for example, using the closeness between vectors as a measure). If the expected feature is very close to the representative point of the uncontrolled group (high similarity), relatively close to the representative point of the obesity-merging group (moderate similarity), and far from the representative point of the well-controlled group (low similarity), the server will weight and fuse the features of the representative points of each group according to these similarities to generate the “first auxiliary strengthened feature”. For example, the “uncontrolled group” with high similarity contributes more feature information, and the “well-controlled group” with low similarity contributes less feature information. The final auxiliary feature will highlight the clinical characteristics of the “uncontrolled group” (such as high blood glucose and high monitoring demand). The server combines the “first auxiliary strengthened feature” with the original expected feature (for example, superimposes the feature information of the two), to obtain the “expected feature after feature strengthening” (i.e., the “high-connectivity feature”). The purpose of this step is to make the expected feature more consistent with the characteristics of the actual clinical group. For example, if the feature of the “uncontrolled group” emphasizes “frequent monitoring”, the expected feature after strengthening will highlight this demand, so that the feature representation is more consistent with the clinical reality. The server processes the “basic representation feature” (the vector reflecting only the patient's own characteristics) in the same way. Calculate the similarity between the basic representation feature and the representative points of each group. If the basic representation feature is highly similar to the representative point of the obesity-merging group (such as high BMI), relatively similar to the representative point of the uncontrolled group (such as high blood glucose), and dissimilar to the representative point of the well-controlled group, then weight and fuse the features of the representative points of each group according to these similarities to generate the “second auxiliary strengthened feature”.The auxiliary feature highlights the characteristics of the "obesity combined group" (such as high BMI) and part of the characteristics of the "uncontrolled group" (such as high blood sugar). The server combines (such as superimposes) the "second auxiliary reinforced feature" with the original basic representation feature to obtain the "basic representation feature after feature reinforcement". Then, the reinforced basic feature is input into the initial feature adapter to generate the "converted feature". The role of the adapter is to convert the patient's own characteristics (such as obesity and high blood sugar) into features that adapt to clinical indications (such as reflecting the need for insulin treatment). The server compares the difference between the "high connection feature" (the reinforced expected feature) and the "converted feature" (the reinforced basic feature converted by the adapter). If the converted feature is highly similar to the high connection feature (such as both highlighting "high blood sugar and high insulin demand"), the error is small; if the converted feature fails to fully reflect the clinical needs of the high connection feature (such as only reflecting obesity and not reflecting the need for insulin treatment), the error is large. This error will drive the server to adjust the adapter parameters so that the converted feature gradually approaches the high connection feature and eventually converges to a stable state.

[0058] In summary, the server reinforces the features of the high connection patient clinical elements by combining the patient group characteristics, ensures that the feature representation is more in line with the clinical actual needs, and thus improves the accuracy and clinical relevance of model training.

[0059] In the embodiments of the present application, the group attribution probability of each patient clinical element belonging to each patient group can be implemented by the following examples according to the clinical indication adaptation degree between each patient clinical element and each group representative point.

[0060] For each group representative point, the sum of the clinical indication adaptation degrees of the sample patient clinical elements corresponding to the group representative point is taken as the group attraction of the patient group corresponding to the group representative point;

[0061] For each patient clinical element, the group correlation degree of the patient clinical element in the patient group is obtained according to the clinical indication adaptation degree between the patient clinical element and a group representative point and the group attraction of the patient group corresponding to the group representative point;

[0062] For each patient clinical element, the group correlation degree of the patient clinical element in a patient group is taken as the group attribution probability of the patient clinical element belonging to the patient group.

[0063] In the embodiments of the present application, the server combines the overall attractiveness of the group and the individual fitness of the patient when calculating the group attribution probability of the patient clinical elements, to ensure that the attribution probability reflects the matching degree of individual characteristics and the group, and also considers the overall representativeness of the group. Taking the division of the diabetic patient group as an example, the specific execution process of the server is described in detail. The server first calculates the "group attractiveness" of each patient group, that is, the sum of the fitness of all sample patient clinical elements of the group to the group representative point. Assuming that the patient groups in the current training cycle are divided into three categories: uncontrolled group (representative point characteristics: high blood sugar, high insulin demand); well-controlled group (representative point characteristics: stable blood sugar, low insulin demand); obesity combined group (representative point characteristics: severe obesity, moderate insulin demand). The server traverses all sample patient clinical elements (such as patient A, patient B, patient C), and calculates the fitness (measured by similarity, the higher the value, the more matched) of each patient to each group representative point. For example: the fitness of patient A to the uncontrolled group is 0.8, to the well-controlled group is 0.2, and to the obesity combined group is 0.6; the fitness of patient B to the uncontrolled group is 0.7, to the well-controlled group is 0.5, and to the obesity combined group is 0.9; the fitness of patient C to the uncontrolled group is 0.9, to the well-controlled group is 0.1, and to the obesity combined group is 0.7. The server sums up the fitness of each group to obtain the group attractiveness: uncontrolled group attractiveness = patient A fitness (0.8) + patient B fitness (0.7) + patient C fitness (0.9) = 2.4; well-controlled group attractiveness = patient A fitness (0.2) + patient B fitness (0.5) + patient C fitness (0.1) = 0.8; obesity combined group attractiveness = patient A fitness (0.6) + patient B fitness (0.9) + patient C fitness (0.7) = 2.2. The server calculates the "correlation degree" of each patient clinical element to each group, which combines the individual fitness of the patient and the overall attractiveness of the group. For example, taking patient A as an example: the fitness to the uncontrolled group is 0.8, the uncontrolled group attractiveness is 2.4, and the correlation degree = 0.8 ÷ 2.4 ≈ 0.33 (indicating the proportion of patient A's contribution to the overall attractiveness of the uncontrolled group); the fitness to the well-controlled group is 0.2, the well-controlled group attractiveness is 0.8, and the correlation degree = 0.2 ÷ 0.8 = 0.25 (the proportion of patient A's contribution to the well-controlled group); the fitness to the obesity combined group is 0.6, the obesity combined group attractiveness is 2.2, and the correlation degree = 0.6 ÷ 2.2 ≈ 0.27 (the proportion of patient A's contribution to the obesity combined group). The server normalizes the correlation degrees of the patient to each group (that is, calculates the proportion of each correlation degree in the total correlation degree) to obtain the probability of the patient belonging to each group. Taking patient A as an example, the correlation degrees of the patient to the three groups are 0.33, 0.25, and 0.27, respectively, and the total correlation degree is 0.33 + 0.25 + 0.27 = 0.85.Therefore: the probability of belonging to the uncontrolled group = 0.33 ÷ 0.85 ≈ 38.8%; the probability of belonging to the well-controlled group = 0.25 ÷ 0.85 ≈ 29.4%; the probability of belonging to the obesity combination group = 0.27 ÷ 0.85 ≈ 31.8%. The group attraction reflects the "representativeness" of the group in the sample. The higher the attraction, the more patients match the group, and the more prominent the clinical significance of the group. The correlation balances the relationship between the individual patient and the group as a whole, avoiding the misclassification of patients due to the high attraction of a group (such as the uncontrolled group). The normalized attribution probability more objectively reflects the "typicality" of patients in each group. In the training cycle, the server adjusts the group representative point features according to the attribution probability (such as including the patient's feature with the highest attribution probability in the update of the group representative point), and optimizes the feature encoder parameters, so that the adaptation of patients with the same type is more concentrated, and finally the accuracy of group division is improved. For example, if the patient A's attribution probability is highest in the uncontrolled group (38.8%), the server will include its features (such as high blood sugar) in the update of the uncontrolled group representative point, so that the features of this group are more in line with the actual clinical needs. In summary, the server calculates the group attraction, correlation and normalized attribution probability to ensure that the patient group division considers both individual features and the overall representativeness of the group, and provides a more reliable patient classification basis for clinical needs.

[0064] In the embodiments of the present application, each cycle training further includes the following implementation.

[0065] For each patient clinical factor, the group representative point with the highest clinical indication adaptation degree corresponding to the patient clinical factor is taken as the target group representative point, and the patient group corresponding to the target group representative point is taken as the patient group to which the patient clinical factor belongs in the next cycle;

[0066] For each patient group in the next cycle, the group feature identifier of the corresponding group representative point is updated according to the overall error of the cycle training.

[0067] In an embodiment of the present application, the server dynamically adjusts the patient group attribution and optimizes the group representative point features based on the current training state in each cycle of training to improve the accuracy of group division. Specifically, for each patient clinical element participating in the training (e.g., diabetic patients A, B, C, etc.), the server first calculates the fitness of its clinical element features to each group representative point (e.g., the representative points of the "uncontrolled group", "well-controlled group", and "obesity combined group"). The fitness is measured by the degree of similarity between the features (the higher the degree of similarity, the higher the fitness). Then, the server selects the group representative point with the highest fitness for each patient, and selects the corresponding group as the patient's attribution group in the next cycle. For example, if the features of patient A are most similar to the representative point of the "uncontrolled group" (with the highest fitness), then patient A belongs to the "uncontrolled group" in the next cycle; if the features of patient B are most similar to the representative point of the "obesity combined group", then patient B belongs to this group. After determining the patient group attribution, the server updates the representative point features of each group in combination with the overall training error of the current cycle (including feature conversion error, indication classification error, group attribution error, etc.). In specific operation, the server first aggregates the clinical element features of all patients belonging to the same group in the next cycle (e.g., the "uncontrolled group" may include patients A, D, E, etc., whose features collectively reflect the clinical characteristics of "high blood glucose and frequent monitoring"), analyzes the commonalities of these features (e.g., most patients have high blood glucose and need insulin treatment), and generates new group representative point candidate features. If the training error of a certain group in the current cycle is large (e.g., the patient features deviate significantly from the original representative point), it indicates that the original representative point cannot accurately reflect the patient features in the group, and the server will increase the weight of the common features (e.g., adjust the representative point to be closer to the average features of the patients in the group); if the error is small, it indicates that the original representative point is accurate, and only fine-tuning is needed (e.g., retain the core features of the original representative point and appropriately absorb the characteristics of new patients). Through this process, the group representative point gradually converges to the actual features of the patients in the group. For example, the representative point of the "uncontrolled group" may be based only on the guidelines (e.g., "fasting blood glucose ≥7.0 mmol / L") at the beginning of the training, but as the training progresses, if the actual fasting blood glucose of the patients in the group is generally between 8.0-9.0 mmol / L, the representative point will be updated to be closer to the features of "fasting blood glucose 8.5 mmol / L and 4 times of monitoring per day", which is more consistent with the needs of real patients. At the same time, the group attribution of the patient will tend to be fixed due to the stability of the fitness (e.g., patient A will continue to belong to the "uncontrolled group" because he continues to show high blood glucose characteristics), avoiding frequent changes in group due to single fitness fluctuations. Finally, the group division is highly consistent with the actual features of the patients, providing a more accurate classification basis for subsequent clinical demand mining.

[0068] In an embodiment of the present application, the plurality of patient clinical elements includes a plurality of low-connection patient clinical elements, and the low-connection patient clinical element is a patient clinical element that has no association with the clinical indication element.

[0069] The plurality of sample clinical elements further includes low-connection patient clinical elements, the clinical transmission chains of which are all patient clinical transmission chains, and the clinical element features are basic representation features.

[0070] In the embodiments of the present application, the server needs to specially process the "low connection patient clinical elements" when constructing the medical knowledge network and carrying out clinical demand mining training. Such patient elements have no direct association with clinical indication elements, but only have association with other patient clinical elements. The following will illustrate in detail the processing flow of the server for low connection patient clinical elements in the actual scene of "simple obesity (BMI25)" in pediatrics. The server first screens low connection patient clinical elements from the medical knowledge network. Taking pediatric patients as an example, the electronic medical record data shows that the patient clinical element of "simple obesity (BMI25, without complications such as diabetes and hypertension)" is mainly associated with other patient clinical elements such as "hyperlipidemia (high cholesterol)" and "insufficient exercise (daily activity <30 minutes)", but has no co-occurrence record (co-occurrence rate <1%) with clinical indication elements such as "insulin therapy", "blood glucose monitoring" and "hypotensive drugs". The server marks the low connection patient clinical elements by counting the number of connections between each patient clinical element and the indication element (such as the indication connection number of "simple obesity" is 0). The server takes "simple obesity (BMI25)" as the core to extract the clinical transmission chain in the knowledge network. Since it has no association with the indication element, all the transmission chains are "patient clinical transmission chains" (only containing patient clinical elements). For example: chain 1: "simple obesity (BMI25) - hyperlipidemia (total cholesterol 5.8 mmol / L) - liver enzyme elevation (ALT 45 U / L)" (path length 2, each layer is a patient element); chain 2: "simple obesity (BMI25) - insufficient exercise (daily activity 20 minutes) - decreased physical fitness (6-minute walk 300 meters)" (path length 2, all are patient elements). The server inputs the description parameters of the low connection patient clinical transmission chain into the feature encoder to generate the "basic representation feature" (a vector reflecting only the patient's own characteristics). For example, the description parameters of chain 1 include: the standardized value of BMI25 (such as 0.6), the normalized value of total cholesterol 5.8 mmol / L (such as 0.7), and the encoding of ALT 45 U / L (such as [1, 0] representing mild elevation). The encoder processes these parameters through a multi-layer neural network to output a basic representation feature (such as a 128-dimensional vector), which focuses on the patient's obesity level, blood lipid level and liver function status. In the model cycle training, the low connection patient clinical elements participate in the training as samples to help the model learn to process "non-indication-associated" patient features and improve the generalization ability. For example: generalization training of the feature encoder: the basic representation features of low connection patients need to be aligned in the same feature space with the basic representation features of high connection patients (such as "obesity" dimension feature values need to be consistent), to ensure that the model can identify the commonality (such as BMI increase) among different patients, even if their indication associations are different. Error constraint inclusiveness: when calculating the third error parameter (patient group attribution error), the basic representation features of low connection patients need to be adapted to the features of the group representative point (such as "child obesity group").For example, if the representative point feature of the "child obesity group" contains "BMI ≥ 24, lack of exercise", the base feature of the low-connection patient "simple obesity (BMI 25)" needs to have a higher degree of adaptation to the representative point (e.g., similarity 0.8), and the driving model will correctly classify it to avoid being misclassified due to no indication association. By including the clinical factors of low-connection patients, the model trained by the server can more comprehensively cover various patient features. For example, when encountering a new patient with "simple obesity (BMI 26, no complications)", the model can accurately identify the "obesity-hyperlipidemia" transmission chain based on the base representation feature of the low-connection patient, and recommend potential needs (such as "exercise intervention" and "low-fat diet guidance") that match its features, even if these needs are not directly associated with traditional indications (such as insulin therapy). This improves the model's ability to mine atypical patients and avoids missing potential needs due to sparse indication associations.

[0071] In summary, by identifying the clinical factors of low-connection patients, extracting their patient clinical transmission chains, generating base representation features, and including them in training, the server ensures that the model can handle complex indication associations for high-connection patients and capture potential needs for low-connection patients, ultimately improving the comprehensiveness and accuracy of clinical need mining.

[0072] In the embodiments of the present application, the base representation feature or the association feature representation of each sample clinical factor is obtained by the following process, which can be implemented by the following examples.

[0073] From each clinical transmission chain with the sample clinical factor as the core clinical factor, obtain each target clinical transmission chain, which is a patient clinical transmission chain or a clinical indication transmission chain;

[0074] For each target clinical transmission chain, starting from the adjacent level clinical factor in the target clinical transmission chain, according to the fusion feature of each higher level associated clinical factor connected by the clinical factor and the inherent description parameter of the clinical factor, fuse layer by layer to the core clinical factor until the fusion feature of the core clinical factor on the target clinical transmission chain is obtained; wherein the fusion feature of the most distant level associated clinical factor is determined according to the inherent description parameter of the most distant level associated clinical factor;

[0075] By graph convolution network, the fusion features of the sample clinical factors on each target clinical transmission chain are weighted and fused to obtain the features of the sample clinical factors corresponding to the target clinical transmission chain, wherein if the target clinical transmission chain is a patient clinical transmission chain, the corresponding feature is a base representation feature, and if the target clinical transmission chain is a clinical indication transmission chain, the corresponding feature is an association feature representation.

[0076] In the embodiments of the present application, the server needs to process the feature information of the clinical transmission chain of the sample clinical element through hierarchical fusion and graph convolution network when obtaining the basic representation feature or the associated feature representation of the sample clinical element. Taking "diagnosis of type 2 diabetes" (highly connected patient clinical element) and "simple obesity (BMI 25)" (lowly connected patient clinical element) as examples, the specific execution process is described in detail. The server first screens the "target clinical transmission chain" from the clinical transmission chain of the sample clinical element, and ensures that each chain only contains patient clinical elements (patient clinical transmission chain) or only contains clinical indication elements (clinical indication transmission chain). For "diagnosis of type 2 diabetes": its clinical transmission chain includes patient clinical transmission chain (such as "diagnosis of type 2 diabetes-obesity (BMI 30)-hypertension (systolic pressure 150 mmHg)") and clinical indication transmission chain (such as "diagnosis of type 2 diabetes-insulin therapy (0.5 U / kg)-4 times daily blood glucose monitoring"), and the server respectively marks the two types of chains as target transmission chains. For "simple obesity (BMI 25)": its clinical transmission chain only contains patient clinical elements (such as "simple obesity-hyperlipidemia (cholesterol 5.8 mmol / L)-elevated liver enzymes (ALT 45 U / L)"), so all the chains are target patient clinical transmission chains. The server starts from the farthest level of each target transmission chain, and fuses the features layer by layer to the core clinical element. When fusing, the inherent description parameters (such as numerical value, category, text feature) of the clinical element at the current level and the fusion features of the associated elements at the higher level are combined. Taking the patient clinical transmission chain of "diagnosis of type 2 diabetes-obesity-hypertension" as an example: the farthest level (hypertension): its inherent description parameters are "systolic pressure 150 mmHg" (numerical value feature, standardized to 0.7) and "history of hypertension" (category feature, coded as [1, 0]). The server directly takes these parameters as the fusion features of this level (such as vector [0.7, 1, 0]). The middle level (obesity): its inherent description parameters are "BMI 30" (numerical value feature, standardized to 0.8) and "abdominal circumference 100 cm" (numerical value feature, standardized to 0.6). The server fuses the inherent description parameters of obesity with the fusion features of the next level (hypertension) ([0.7, 1, 0]) (such as addition or splicing), to obtain the fusion features of the middle level (such as [0.8+0.7, 0.6+1, 0]-[1.5, 1.6, 0]). The core level (diagnosis of type 2 diabetes): its inherent description parameters are "fasting blood glucose 8.5 mmol / L" (standardized to 0.9) and "glycosylated hemoglobin 7.8%" (standardized to 0.8). The server fuses the inherent description parameters of the core level with the fusion features of the middle level (obesity) ([1.5, 1.6, 0]) to obtain the fusion features on this transmission chain (such as [0.9+1.5, 0.8+1.6, 0]-[2.4, 2.4, 0]).Take the clinical indication chain of "diagnosis of type 2 diabetes - insulin treatment - daily 4 times blood glucose monitoring" as an example: the farthest level (blood glucose monitoring): the inherent description parameter is "monitoring frequency 4 times a day" (category feature, encoded as [0, 1, 0]), "monitoring time before and after meals" (text feature, vector [0.3, 0.5]). The fusion feature is [0, 1, 0, 0.3, 0.5]. The middle level (insulin treatment): the inherent description parameter is "dose 0.5U / kg" (numerical feature, standardized to 0.6), "injection site abdomen" (category feature, encoded as [1, 0]). Fuse the inherent parameters of the middle level with the fusion features of the next level (blood glucose monitoring) ([0, 1, 0, 0.3, 0.5]), resulting in [0.6+0, 1+1, 0+0, 0.3+0.3, 0.5+0.5]-[0.6, 2, 0, 0.6, 1.0]. The core level (diagnosis of type 2 diabetes): the inherent description parameter is "need to quickly control blood sugar" (text feature, vector [0.7, 0.9]). Fuse the core level parameters with the fusion features of the middle level (insulin treatment) ([0.6, 2, 0, 0.6, 1.0]), resulting in the fusion feature on this chain (such as [0.6+0.7, 2+0.9, 0, 0.6, 1.0]-[1.3, 2.9, 0, 0.6, 1.0]). The server uses graph convolution network (GCN) to weight the fusion features of the sample clinical elements on each target chain, and the weight is dynamically adjusted according to the importance of the chain (such as chain length, clinical relevance). For "diagnosis of type 2 diabetes": the fusion features of the patient's clinical chain ([2.4, 2.4, 0]) and the clinical indication chain ([1.3, 2.9, 0, 0.6, 1.0]) are input into GCN. GCN assigns weights according to the clinical relevance of the chain (such as the patient's chain reflecting pathological mechanism, the indication chain reflecting treatment needs) (such as the patient's chain weight 0.6, the indication chain weight 0.4), and the weighted fusion results in the basic representation feature (patient chain fusion result) and the associated feature representation (indication chain fusion result). For "simple obesity (BMI 25)": all its target chains are patient clinical chains (such as the fusion feature of "obesity - hyperlipidemia - liver enzyme elevation" [1.8, 1.2, 0.5]), and GCN weights each chain feature (such as chain 1 weight 0.7, chain 2 weight 0.3). Finally, the basic representation feature is output (such as [1.8x0.7+1.5x0.3, 1.2x0.7+1.0x0.3, 0.5x0.7+0.4x0.3]-[1.71, 1.14, 0.47]). Through level fusion and graph convolution weighting, the server ensures that the features of the sample clinical elements not only contain their own inherent information (such as blood glucose value, BMI), but also integrate the associated information of other elements in the chain (such as hypertension, insulin treatment), making the feature representation more comprehensive and more consistent with clinical logic.For example, the base of "diagnosis of type 2 diabetes" indicates that the feature not only reflects the blood glucose level, but also includes the influence of obesity and high blood pressure; the associated feature indicates that not only the insulin dose, but also the need for monitoring frequency is integrated. This process provides a more accurate feature base for subsequent model training and clinical demand mining.

[0077] In the embodiment of the present application, the step of fusing each level in the core clinical element fusion by layer includes:

[0078] For any clinical element currently processed for level fusion, according to the description parameters of the clinical element, the original base feature of the clinical element is obtained;

[0079] For any clinical element currently processed for level fusion, each higher level associated clinical element associated with the clinical element is taken as a target clinical element, and the fusion feature of the target clinical element is obtained;

[0080] For any clinical element currently processed for level fusion, according to the original base feature of the clinical element and the fusion feature of each target clinical element, the correlation between the clinical element and each target clinical element is obtained respectively;

[0081] For any clinical element currently processed for level fusion, according to the corresponding correlation of each target clinical element, the fusion feature of each target clinical element is weighted and fused, and according to the result of weighted fusion and the original base feature of the clinical element, the fusion feature of the clinical element is obtained.

[0082] In the embodiments of the present application, the server needs to start from the most remote level of the clinical elements when processing the hierarchical feature fusion of the clinical transmission chain, and fuse the features from the core clinical elements layer by layer to ensure that the features of each level not only contain their own information, but also integrate the influence of the associated elements of the higher level. Taking the patient clinical transmission chain of "diagnosis of type 2 diabetes - obesity (BMI 30) - hypertension (systolic blood pressure 150 mmHg)" as an example, the specific execution process of the server is described in detail. The server first processes the most remote level of the transmission chain, "hypertension (systolic blood pressure 150 mmHg)". The description parameters of this clinical element include: systolic blood pressure value (150 mmHg, standardized to 0.7), whether there is a history of hypertension ("yes", category code [1, 0]). The server converts these parameters into "original basic features" (such as vector [0.7, 1, 0]), which only reflect the properties of "hypertension" itself. "Hypertension" belongs to the next level of "obesity" in the transmission chain, so "obesity (BMI 30)" is the "higher level associated clinical element" of "hypertension". The server needs to obtain the fusion features of "obesity" (at this time, the fusion features of "obesity" have not been calculated, and the "hypertension" level needs to be processed first, and then the "obesity" level). The server calculates the correlation between "hypertension" and "obesity". The basis of the correlation can be the co-occurrence frequency in the clinical data (such as 80% of obese patients with hypertension in the historical medical records) or the feature similarity (such as the association between the BMI feature of obesity and the systolic blood pressure feature of hypertension in the pathological mechanism). For example, if the co-occurrence frequency is 80%, the correlation value is set to 0.8 (the higher the value, the stronger the association). The server fuses the fusion features of the higher level element (obesity) and the original basic features of the current level (hypertension) with the correlation as the weight. Since the fusion features of "obesity" need to be calculated after "hypertension", this paper takes the hypothetical "obesity" fusion features ([0.8, 0.6], reflecting the standardized values of BMI 30 and abdominal circumference 100 cm) as an example: higher level (obesity) fusion features: [0.8, 0.6]; correlation weight: 0.8; weighted fusion result: 0.8 x [0.8, 0.6] = [0.64, 0.48]; current level (hypertension) fusion features: original basic features [0.7, 1, 0] + weighted fusion result [0.64, 0.48, 0] = [1.34, 1.48, 0] (here, the addition is only an example, and in actual use, splicing or other methods can also be used). After completing the fusion of "hypertension" and "obesity" levels, the server processes the core level, "diagnosis of type 2 diabetes". Its original basic features include: fasting blood glucose 8.5 mmol / L (standardized to 0.9), glycosylated hemoglobin 7.8% (standardized to 0.8). The higher level associated element is "obesity" (fusion features [1.34, 1.48, 0]).The correlation between "diagnosis of type 2 diabetes" and "obesity" is calculated (for example, based on the co-occurrence rate in the medical record, 90%, set to 0.9); the fused features of "obesity" are weighted and fused: 0.9 x [1.34, 1.48, 0] = [1.206, 1.332, 0]; the core level fused features: the original basic features [0.9, 0.8] + the weighted fusion result [1.206, 1.332] = [2.106, 2.132] (integrating obesity, hypertension and self blood glucose features). Through this layer-by-layer fusion process, the server ensures that the fused features of the core clinical elements (such as "diagnosis of type 2 diabetes") not only contain their own blood glucose and glycosylated hemoglobin information, but also integrate the influence of related elements such as BMI of "obesity" and systolic pressure of "hypertension" on the transmission chain. For example, the final fused features [2.106, 2.132] not only reflect the patient's high blood glucose state, but also imply the driving effect of obesity and hypertension on diabetes, making the feature representation more consistent with the clinical pathological logic. This process is repeated in each clinical transmission chain (such as the patient clinical transmission chain and the clinical indication transmission chain), and finally the fused features of each chain are weighted and fused by the graph convolution network to generate the basic representation features of the sample clinical elements (patient chain) or the associated feature representation (indication chain), providing a more comprehensive and accurate feature basis for subsequent model training and clinical demand mining.

[0083] The embodiment of the present application provides a computer device 100, which comprises a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned clinical demand mining method. As shown in the figure, Figure 2 Figure 2 The structural block diagram of the computer device 100 provided by the embodiment of the present application is shown. The computer device 100 comprises a memory 111, a processor 112 and a communication unit 113. In order to realize the transmission or interaction of data, the memory 111, the processor 112 and the communication unit 113 are directly or indirectly electrically connected to each other.

[0084] The foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Many modifications and variations are possible in light of the above teachings. The embodiments are chosen and described in order to best explain the principles of the disclosure and its practical application to thereby enable others skilled in the art to best utilize the disclosure and various embodiments with various modifications as are suited to the particular use contemplated.​

Claims

1. A clinical demand mining method, characterized in that, The method comprises the following steps: obtaining a medical knowledge network and a description parameter of each clinical element in the medical knowledge network, the medical knowledge network comprising a plurality of clinical elements and an association between two clinical elements having an interaction relationship, wherein the association is established based on a medical diagnosis standard or a pathophysiological mechanism, the plurality of clinical elements comprising a plurality of patient clinical elements representing a patient and a plurality of clinical indication elements representing a clinical requirement, and the plurality of patient clinical elements comprising a plurality of high-connection patient clinical elements, the high-connection patient clinical elements having an association with a plurality of clinical indication elements and a plurality of other patient clinical elements; wherein the description parameter is processed by the following method: the numerical parameter is normalized to a preset range by standardization processing, the category parameter is converted into binary coding, and the text parameter is extracted into a feature vector by semantic analysis; selecting a plurality of sample clinical elements from the medical knowledge network, for each sample clinical element, determining a plurality of clinical transmission chains with the sample clinical element as the core clinical element from the medical knowledge network, the direct association level clinical elements on a clinical transmission chain being either patient clinical elements or clinical indication elements, and the clinical transmission chain with the direct association level clinical elements being all patient clinical elements being a patient clinical transmission chain, and the clinical transmission chain with the direct association level clinical elements being all clinical indication elements being a clinical indication transmission chain, the plurality of sample clinical elements comprising a plurality of high-connection patient clinical elements, and each clinical transmission chain of the high-connection patient clinical elements comprising a plurality of patient clinical transmission chains and a plurality of clinical indication transmission chains; recurrently training an initial feature encoder and an initial feature adapter according to the description parameters of the clinical elements on each clinical transmission chain of each sample clinical element to obtain a feature encoder and a feature adapter, wherein the recurrent training minimizes an overall error, and the overall error comprises: a first error parameter for constraining the converted features of the high-connection patient clinical elements to contain medical requirement information of the clinical indication elements, specifically characterized by the deviation between the expected feature representation of the corresponding high-connection patient clinical element at each cycle and the converted features obtained by the initial feature adapter according to the basic representation features; a second error parameter for ensuring that the first clinical indication adaptation degree of the high-connection patient clinical elements and the associated clinical indication elements is significantly higher than the second clinical indication adaptation degree of the high-connection patient clinical elements and the irrelevant clinical indication elements, specifically obtained by calculating the sum of the indication association errors of the clinical indication element pairs of the associated clinical indication elements and the irrelevant clinical indication elements; a third error parameter for maintaining clinical classification consistency through patient group division, specifically characterized by the deviation between the clinical indication adaptation degree and the group belonging probability between each patient clinical element and the group representative point; performing clinical requirement mining according to the feature encoder and the feature adapter, wherein the clinical requirement mining comprises inputting the clinical transmission chain of a target patient clinical element into the feature encoder to obtain clinical element features, and calculating the clinical indication adaptation degree with the clinical element features of the clinical requirement to be evaluated after converting the converted features by the feature adapter, so as to determine the target clinical requirement.

2. The method of claim 1, wherein, The clinical demand mining according to the feature encoder and the feature adapter comprises: constructing a medical knowledge network and obtaining a description parameter of each clinical element in the medical knowledge network, the medical knowledge network comprising a plurality of clinical elements and an association between two clinical elements having an interaction relationship, the plurality of clinical elements comprising a plurality of patient clinical elements representing a patient and a plurality of clinical indication elements representing a clinical demand to be evaluated, and the plurality of patient clinical elements comprising a target patient clinical element representing a target patient; determining a plurality of first clinical transmission chains with the target patient clinical element as a core clinical element and a plurality of second clinical transmission chains with a clinical indication element representing a clinical demand to be evaluated as a core clinical element in the medical knowledge network; inputting each first clinical transmission chain corresponding to the target patient clinical element and a description parameter of a clinical element on each first clinical transmission chain into a feature encoder to obtain a clinical element feature of the target patient clinical element, and inputting each second clinical transmission chain corresponding to each clinical demand to be evaluated and a description parameter of a clinical element on each second clinical transmission chain into the feature encoder to obtain a clinical element feature of each clinical demand to be evaluated; if all the clinical elements associated with the target patient clinical element are clinical indication elements, determining that each first clinical transmission chain of the target patient is a patient clinical transmission chain, converting the clinical element feature of the target patient clinical element through a feature adapter to obtain a converted feature, and obtaining a clinical indication adaptation degree between the target patient and each clinical demand to be evaluated according to the converted feature and the clinical element feature of the clinical demand to be evaluated for each clinical demand to be evaluated; determining a target clinical demand for the target patient from each clinical demand to be evaluated according to the clinical indication adaptation degree corresponding to each clinical demand to be evaluated.

3. The method of claim 1, wherein, The expected feature is represented as a clinical element feature or an association feature representation; the initial feature encoder is used to input a description parameter of a clinical element on each clinical transmission chain of a sample clinical element to output the clinical element feature of the sample clinical element; if each clinical transmission chain of the sample clinical element comprises a plurality of patient clinical transmission chains, the clinical element feature comprises the basic representation feature obtained according to the description parameter of the clinical element on each patient clinical transmission chain, and if each clinical transmission chain of the sample clinical element comprises a plurality of clinical indication transmission chains, the clinical element feature comprises the association feature representation obtained according to the description parameter of the clinical element on each clinical indication transmission chain.

4. The method of claim 1, wherein, The plurality of sample clinical elements further comprise clinical indication elements; In each cycle of training, the second error parameter of each high-connection patient clinical element is obtained through the following process, comprising: screening, from the medical knowledge network, a clinical indication element associated with the high-connection patient clinical element as an associated clinical indication element and screening, from the medical knowledge network, a clinical indication element not associated with the high-connection patient clinical element as an irrelevant clinical indication element; obtaining a plurality of pairs of clinical indicator elements, each pair of clinical indicator elements comprising one associated clinical indicator element and one unassociated clinical indicator element; for each pair of clinical indicator elements, obtaining a first clinical indicator fitness between the high-connectivity patient clinical element and the associated clinical indicator element in the pair of clinical indicator elements according to the clinical element feature of the high-connectivity patient clinical element in the current cycle, obtaining a second clinical indicator fitness between the high-connectivity patient clinical element and the unassociated clinical indicator element in the pair of clinical indicator elements according to the clinical element feature of the high-connectivity patient clinical element in the current cycle, and obtaining an indicator association error of the high-connectivity patient clinical element for the pair of clinical indicator elements according to the inverse of the deviation between the first clinical indicator fitness and the second clinical indicator fitness; obtaining a second error parameter of the high-connectivity patient clinical element according to the sum of the indicator association errors of the high-connectivity patient clinical element for all pairs of clinical indicator elements.

5. The method of claim 1, wherein, The third error parameter of the patient clinical element in each sample clinical element is obtained through the following process, comprising: for each cycle training, obtaining the group feature identifier of the group representative point of each patient group in the current cycle and the clinical element feature of each patient clinical element in the current cycle, wherein the group feature identifier of the group representative point in the initial group division is set as an initial feature value; for each patient clinical element, obtaining the clinical indicator fitness between the patient clinical element and each group representative point according to the clinical element feature of the patient clinical element in the current cycle and the group feature identifier of each group representative point; obtaining the group attribution probability of each patient clinical element belonging to each patient group according to the clinical indicator fitness between each patient clinical element and each group representative point respectively; for each patient clinical element, obtaining the third error parameter of the patient clinical element according to the deviation between the clinical indicator fitness between the patient clinical element and each group representative point and the group attribution probability of the patient clinical element belonging to the corresponding patient group.

6. The method of claim 5, wherein, The first error parameter of the high-connectivity patient clinical element is obtained through the following process, comprising: weighting and fusing the group feature identifier of each group representative point as a first auxiliary reinforcement feature according to the similarity measure between each group representative point and the expected feature representation of the high-connectivity patient clinical element in the current cycle, performing feature reinforcement on the expected feature representation according to the first auxiliary reinforcement feature to obtain a feature-reinforced expected feature representation, and taking the feature-reinforced expected feature representation as the high-connectivity feature of the high-connectivity patient clinical element in the current cycle. According to a similarity measurement of each group representative point and a basic representation feature of the high-connectivity patient clinical element in the current cycle, a group feature identifier of each group representative point is weighted and fused as a second auxiliary reinforcement feature, the basic representation feature is reinforced according to the second auxiliary reinforcement feature, and a feature-reinforced basic representation feature is obtained; the feature-reinforced basic representation feature is converted by the initial feature adapter, and a converted feature of the high-connectivity patient clinical element in the current cycle is obtained. According to a deviation between the high-connectivity feature and the converted feature of the high-connectivity patient clinical element in the current cycle, a first error parameter of the high-connectivity patient clinical element is obtained.

7. The method of claim 5, wherein, According to a clinical indication adaptation degree between each patient clinical element and each group representative point, a group attribution probability of each patient clinical element belonging to each patient group is obtained, including: For each group representative point, a sum of clinical indication adaptation degrees of sample patient clinical elements corresponding to the group representative point is taken as a group attraction of a patient group corresponding to the group representative point; For each patient clinical element, a group association degree of the patient clinical element in a patient group is obtained according to a clinical indication adaptation degree between the patient clinical element and a group representative point and a group attraction of a patient group corresponding to the group representative point; For each patient clinical element, a group association degree of the patient clinical element in a patient group is taken as a group attribution probability of the patient clinical element belonging to the patient group.

8. The method of claim 5, wherein, Each cycle training further includes: For each patient clinical element, a group representative point corresponding to a highest clinical indication adaptation degree of the patient clinical element is taken as a target group representative point, and a patient group corresponding to the target group representative point is taken as a patient group to which the patient clinical element belongs in a next cycle; For each patient group in the next cycle, a group feature identifier of a corresponding group representative point is updated according to an overall error of the cycle training.

9. The method of claim 1, wherein, The plurality of patient clinical elements includes a plurality of low-connectivity patient clinical elements, the low-connectivity patient clinical element being a patient clinical element that does not exist in association with a clinical indication element; the plurality of sample clinical elements further includes a low-connectivity patient clinical element, a clinical transmission chain of the low-connectivity patient clinical element being a patient clinical transmission chain, and a clinical element feature being a basic representation feature.

Images