Diabetic complication prediction method based on neural network

Abstract

The invention provides a diabetic complication prediction method based on a neural network, and relates to the technical field of artificial intelligence. According to the method, firstly, through a semantic mining unit included in a target complication prediction network, a target complication interference vector is determined; secondly, through a semantic mining unit, according to the obtained target diabetic user data, performing interference removal operation on the target complication interference vector, and outputting a diabetic complication vector; and then, through a semantic analysis unit included in the target complication prediction network, performing semantic analysis operation on the diabetic complication vector, and outputting a target diabetic complication prediction result corresponding to the target diabetic user data. Based on the method, the problem that in the prior art, the reliability of diabetic complication prediction is relatively low can be solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for predicting diabetes complications based on a neural network. Background Art

[0002] The development of diabetic complications is a complex process, involving the interaction of multiple factors such as long-term hyperglycemia, insulin resistance, metabolic disorders, abnormal immune function, and abnormal blood pressure. The risk of diabetic complications can be effectively reduced through good blood sugar control, maintaining a healthy lifestyle, regular physical examinations, and management of other complications. Due to the large number of types of complications and the large number of influencing factors involved, it is difficult to make corresponding predictions of complications, making it difficult to carry out targeted management and control of complications. Based on this, it is necessary to make reliable predictions of complications based on the specific conditions of diabetic patients. However, in the prior art, it is generally based on the analysis and prediction of the corresponding experts according to the specific conditions of diabetic patients. Therefore, there is a problem of relatively low reliability. Summary of the invention

[0003] In view of this, an object of the present invention is to provide a method for predicting diabetes complications based on a neural network, so as to improve the problem of relatively low reliability of diabetes complication prediction in the prior art.

[0004] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0005] A method for predicting diabetes complications based on a neural network, comprising:

[0006] Determining a target complication interference vector through a semantic mining unit included in a target complication prediction network, wherein the target complication prediction network is a trained and optimized neural network, and the target complication interference vector is formed by fusing an interference vector on a basic vector representing the semantics of diabetes complications;

[0007] Through the semantic mining unit, according to the acquired target diabetic user data, the target complication interference vector is subjected to interference removal operation, and a diabetic complication vector is output, wherein the target diabetic user data at least includes the target diabetic user's diabetes diagnosis data, diabetes treatment data, and diabetes-related detection data;

[0008] The semantic analysis unit included in the target complication prediction network performs a semantic analysis operation on the diabetes complication vector, and outputs a target diabetes complication prediction result corresponding to the target diabetes user data.

[0009] In some preferred embodiments, in the above-mentioned neural network-based diabetes complication prediction method, the neural network-based diabetes complication prediction method further includes a step of training to form the target complication prediction network, which step includes:

[0010] Determine a training diabetes information combination, wherein the training diabetes information combination includes training diabetes user data and diabetes complication labels, and the diabetes complication labels refer to actual diabetes complications corresponding to the training diabetes user data;

[0011] The semantic mining unit included in the candidate complication prediction network performs interference information application operation on the diabetes complication label to output a training complication interference vector, and, based on the training diabetes user data, performs a first number of stages of interference removal operation on the training complication interference vector to output a training diabetes complication vector corresponding to each stage of the interference removal operation;

[0012] Determining a second number of training diabetes complication vectors capable of performing error characterization on the semantic mining unit from a first number of training diabetes complication vectors corresponding to the interference removal operations of the first number of stages, wherein the second number is smaller than the first number;

[0013] performing an error determination operation according to the second number of training diabetes complication vectors to obtain a target training error corresponding to the semantic mining unit, wherein each training diabetes complication vector in the second number of training diabetes complication vectors corresponds to determining a type of target training error;

[0014] Based on the obtained second number of target training errors, the network parameters of the semantic mining unit are updated to form the target complication prediction network.

[0015] In some preferred embodiments, in the above-mentioned neural network-based diabetes complication prediction method, the semantic mining unit included in the candidate complication prediction network performs an interference information application operation on the diabetes complication label to output a training complication interference vector, and, based on the training diabetes user data, performs a first number of stages of interference removal operations on the training complication interference vector, and outputs the training diabetes complication vector corresponding to each stage of the interference removal operation, the steps include:

[0016] Performing a semantic mining operation on the diabetes complication label by a semantic mining subunit in a semantic mining unit included in a candidate complication prediction network, and outputting a complication label semantic vector corresponding to the diabetes complication label;

[0017] Performing an interference information application operation on the complication label semantic vector corresponding to the diabetes complication label, and outputting a complication label interference vector corresponding to the complication label semantic vector;

[0018] Through the semantic mining subunit, the complication label interference vector is subjected to a vector space conversion operation to generate and output a training complication interference vector corresponding to the diabetes complication label;

[0019] Through the interference removal subunit in the semantic mining unit, the training complication interference vector is subjected to a first number of stages of interference removal operations based on the training diabetes user data, and the training diabetes complication vector corresponding to each stage of the interference removal operation is output.

[0020] In some preferred embodiments, in the above-mentioned neural network-based diabetes complication prediction method, the interference removal subunit in the semantic mining unit performs a first number of stages of interference removal operations on the training complication interference vector according to the training diabetes user data, and outputs the training diabetes complication vector corresponding to each stage of the interference removal operation, including:

[0021] In the first stage, the interference removal subunit in the semantic mining unit performs an interference removal operation on the training complication interference vector based on the diabetes user data vector corresponding to the training diabetes user data, and outputs the training diabetes complication vector corresponding to the interference removal operation of the first stage, wherein the diabetes user data vector corresponding to the training diabetes user data is formed by the semantic mining subunit performing a semantic mining operation on the training diabetes user data;

[0022] In the second stage and each subsequent stage, the interference removal subunit in the semantic mining unit performs an interference removal operation on the training diabetes complications vector corresponding to the interference removal operation in the previous stage based on the diabetes user data vector, and outputs the training diabetes complications vector corresponding to the interference removal operation in the current stage.

[0023] In some preferred embodiments, in the above-mentioned neural network-based diabetes complication prediction method, the step of determining a second number of training diabetes complication vectors capable of performing error characterization on the semantic mining unit from the first number of training diabetes complication vectors corresponding to the interference removal operations in the first number of stages includes:

[0024] Determine an ath stage and a bth stage, wherein the ath stage and the bth stage belong to one stage of the first number of stages respectively and belong to different stages;

[0025] Marking the training diabetic complication vectors corresponding to the interference removal operation of the ath stage and the training diabetic complication vectors corresponding to the interference removal operation of the bth stage in the first number of training diabetic complication vectors to obtain a second number of training diabetic complication vectors;

[0026] Among them, the ath stage is located before the bth stage, the training diabetes complications vector corresponding to the interference removal operation of the ath stage is used to determine the target training error of the first type, and the training diabetes complications vector corresponding to the interference removal operation of the bth stage is used to determine the target training error of the second type.

[0027] In some preferred embodiments, in the above-mentioned neural network-based diabetes complication prediction method, the step of performing an error determination operation based on the second number of training diabetes complication vectors to obtain a target training error corresponding to the semantic mining unit includes:

[0028] By means of the semantic analysis unit, a semantic analysis operation is performed on the training diabetes complication vector corresponding to the interference removal operation in the ath stage, and a first complication prediction result is output;

[0029] Performing a prediction result correlation analysis on the first complication prediction result and the training diabetic user data, and outputting a correlation analysis result between the first complication prediction result and the training diabetic user data;

[0030] Determining, according to the correlation analysis result, a prediction result training error corresponding to the semantic mining unit, wherein the prediction result training error belongs to a first type of training error;

[0031] According to the prediction result training error, a target training error of the interference removal operation in the a-th stage for the semantic mining unit is determined.

[0032] In some preferred embodiments, in the above-mentioned neural network-based diabetes complication prediction method, the step of performing prediction result correlation analysis on the first complication prediction result and the training diabetes user data, and outputting the correlation analysis result between the first complication prediction result and the training diabetes user data, comprises:

[0033] The semantic mining subunit in the semantic mining unit performs a semantic mining operation on the first complication prediction result, and performs a semantic space conversion operation based on the result of the semantic mining operation, and outputs a first semantic mining vector corresponding to the first complication prediction result;

[0034] The semantic mining subunit performs a semantic mining operation on the training diabetes user data, and performs a semantic space conversion operation based on the result of the semantic mining operation, and outputs a second semantic mining vector corresponding to the training diabetes user data, wherein the first semantic mining vector and the second semantic mining vector are in the same semantic space;

[0035] Determine a vector correlation parameter between the first semantic mining vector and the second semantic mining vector, and determine a correlation analysis result between the first complication prediction result and the training diabetes user data based on the vector correlation parameter;

[0036] The prediction result correlation analysis includes a semantic mining operation, a semantic space conversion operation and a vector correlation parameter determination operation.

[0037] In some preferred embodiments, in the above-mentioned neural network-based diabetes complication prediction method, the step of determining the target training error of the interference removal operation in the ath stage for the semantic mining unit based on the prediction result training error includes:

[0038] Performing a difference operation on the training diabetes complication vector corresponding to the interference removal operation of the a-th stage and the training complication interference vector, and outputting the training diabetes difference vector corresponding to the interference removal operation of the a-th stage;

[0039] Determine a corresponding first interference training error according to the difference between the training diabetes difference vector corresponding to the interference removal operation of the ath stage and the representation vector of the interference information corresponding to the training complication interference vector;

[0040] Based on the first interference training error and the prediction result training error, a target training error of the interference removal operation in the ath stage for the semantic mining unit is determined.

[0041] In some preferred embodiments, in the above-mentioned neural network-based diabetes complication prediction method, the step of performing an error determination operation based on the second number of training diabetes complication vectors to obtain a target training error corresponding to the semantic mining unit includes:

[0042] Performing a semantic analysis operation on the training diabetes complication vector corresponding to the interference removal operation in the b-th stage by a semantic analysis unit to generate a second complication prediction result;

[0043] Through a text fluency detection network formed by pre-training, the text fluency of the second complication prediction result is evaluated, the text fluency corresponding to the second complication prediction result is output, and, based on the text fluency, the text fluency training error corresponding to the semantic mining unit is determined, wherein the text fluency training error belongs to the second type of training error;

[0044] According to the text fluency training error, a target training error of the interference removal operation in the b-th stage for the semantic mining unit is determined.

[0045] In some preferred embodiments, in the above-mentioned neural network-based diabetes complication prediction method, the step of determining the target training error of the interference removal operation in the bth stage for the semantic mining unit based on the text fluency training error includes:

[0046] Performing a difference operation on the training diabetes complication vector corresponding to the interference removal operation of the b-th stage and the training complication interference vector, and outputting the training diabetes difference vector corresponding to the interference removal operation of the b-th stage;

[0047] Determine a corresponding second interference training error according to a difference between the training diabetes difference vector corresponding to the interference removal operation of the b-th stage and the representation vector of interference information corresponding to the training complication interference vector;

[0048] Based on the second interference training error and the text fluency training error, a target training error of the interference removal operation in the b-th stage for the semantic mining unit is determined.

[0049] The embodiment of the present invention provides a method for predicting diabetes complications based on a neural network. First, a target complication interference vector is determined by a semantic mining unit included in a target complication prediction network; secondly, the target complication interference vector is subjected to interference removal operation according to the acquired target diabetes user data by the semantic mining unit, and a diabetes complication vector is output; then, the diabetes complication vector is subjected to semantic analysis operation by a semantic analysis unit included in the target complication prediction network, and a target diabetes complication prediction result corresponding to the target diabetes user data is output. Based on the above method, since the prediction is performed by using a neural network, the powerful learning ability of the neural network and the prediction ability formed after learning can be fully utilized. Therefore, compared with the conventional technical solution based on expert personnel for analysis and prediction, it has higher reliability. In addition, in the process of using the neural network for prediction, the acquired user data is not directly predicted, but the basic vector representing the semantics of diabetes complications is combined, so that the semantic representation ability of the diabetes complication vector in the semantic space of complications is more reliable. In addition, since the interference vector is also fused, the overfitting problem caused by the fusion of the basic vector representing the semantics of diabetes complications can be avoided, so that the reliability of the mined diabetes complication vector can be further improved. Therefore, the problem of relatively low reliability of diabetes complication prediction existing in the prior art can be improved.

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a structural block diagram of a neural network-based diabetes complication prediction system provided in an embodiment of the present invention.

[0052] Figure 2 A flowchart of the steps of the neural network-based diabetes complication prediction method provided in an embodiment of the present invention.

[0053] Figure 3 A schematic diagram of interference removal operations at multiple stages provided by an embodiment of the present invention.

[0054] Figure 4 A specific schematic diagram of the interference removal operation provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0056] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] like Figure 1 As shown, an embodiment of the present invention provides a diabetes complication prediction system based on a neural network, wherein the diabetes complication prediction system based on a neural network may include a memory and a processor.

[0058] In detail, the memory and the processor are electrically connected directly or indirectly to achieve data transmission or interaction. For example, they can be electrically connected to each other through one or more communication buses or signal lines. The memory can store at least one software function module (computer program) that can exist in the form of software or firmware. The processor can be used to execute the executable computer program stored in the memory, so as to implement the neural network-based diabetes complication prediction method provided by the embodiment of the present invention (as described later).

[0059] Optionally, the memory may be, but not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. The processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), a system on chip (SoC), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0060] and, Figure 1 The structure shown is for illustration only. The diabetes complication prediction system based on neural network may also include Figure 1 More or fewer components as shown, or with Figure 1 The various configurations shown may, for example, include a communication unit for exchanging information with other devices.

[0061] In an alternative example, the neural network-based diabetes complication prediction system may be a server with data processing capabilities.

[0062] Combination Figure 2 The embodiment of the present invention also provides a method for predicting diabetes complications based on a neural network, which can be applied to the above-mentioned diabetes complications prediction system based on a neural network (referred to as the prediction system). The method steps defined in the process related to the method for predicting diabetes complications based on a neural network can be implemented by the diabetes complications prediction system based on a neural network.

[0063] The following will Figure 2 The specific process shown is explained in detail.

[0064] Step S110, determining a target complication interference vector through a semantic mining unit included in a target complication prediction network.

[0065] In an embodiment of the present invention, the prediction system can determine the target complication interference vector through the semantic mining unit included in the target complication prediction network. Wherein, the target complication prediction network belongs to a trained and optimized neural network, and the target complication interference vector is formed by fusing the interference vector on the basic vector that represents the semantics of diabetic complications. Exemplarily, the semantic mining unit can perform semantic mining on the description data of various diabetic complications to obtain corresponding semantic vectors, and then, each semantic vector can be spliced ​​or superimposed to obtain a basic vector that represents the semantics of diabetic complications, and then, an interference vector can be generated (the interference vector can be a random vector, wherein each parameter in the random vector can satisfy characteristics such as Gaussian distribution), and finally, the interference vector can be superimposed with the basic vector (i.e., the elements of the corresponding vector positions are added) to obtain the target complication interference vector. In addition, the specific method of semantic mining the description data can be to perform word embedding processing on the text corresponding to the diabetic complications to obtain a word embedding vector as the corresponding basic vector. For example, for "diabetic retinopathy", it has:

[0066] diabetes:

[0067] [0.12,-0.34,0.56,-0.78,0.45,-0.33,0.20,0.41,-0.29,0.65,-0.50,0.30,0.19,-0.10,0.44,0.22,-0.15,0.03,-0.02,......];

[0068] Retina:

[0069] [0.56,0.45,-0.33,0.72,-0.20,0.18,0.41,-0.14,0.33,-0.67,0.55,-0.32,0.24,0.11,-0.02,-0.19,0.10,-0.29,0.61,......];

[0070] Lesions:

[0071] [0.20,0.68,0.15,-0.34,0.50,-0.12,0.72,-0.45,0.19,-0.56,0.33,0.41,-0.23,0.11,0.29,-0.39,0.33,0.13,-0.11,......].

[0072] Then, the word embedding vectors of each word segment can be concatenated or averaged to obtain the corresponding semantic vector of diabetes complications.

[0073] Step S120, through the semantic mining unit, according to the acquired target diabetic user data, the target complication interference vector is subjected to interference removal operation, and a diabetic complication vector is output.

[0074] In an embodiment of the present invention, after obtaining the target complication interference vector, the prediction system can perform interference removal operation on the target complication interference vector through the semantic mining unit according to the acquired target diabetic user data, and output the diabetic complication vector. Wherein, the target diabetic user data at least includes the target diabetic user's diabetes diagnosis data (such as fasting blood glucose ≥ 126mg / dL (7.0mmol / L); 2-hour postprandial blood glucose ≥ 200mg / dL (≥ 11.1mmol / L); glycosylated hemoglobin ≥ 6.5%, etc.), diabetes treatment data (such as oral hypoglycemic drugs: metformin), diabetes-related test data (which can be data obtained from various related tests such as blood pressure, blood lipids, renal function, fundus, etc.), and the target diabetic user data can also include daily diet, exercise, work and rest and other habitual data.

[0075] Step S130, performing a semantic analysis operation on the diabetes complication vector through the semantic analysis unit included in the target complication prediction network, and outputting a target diabetes complication prediction result corresponding to the target diabetes user data.

[0076] In an embodiment of the present invention, after obtaining the diabetes complication vector, the prediction system can perform semantic analysis operations on the diabetes complication vector through the semantic analysis unit included in the target complication prediction network, and output the target diabetes complication prediction result corresponding to the target diabetes user data. Exemplarily, the semantic analysis unit can include a network architecture corresponding to full connection processing and classification processing, that is, after obtaining the diabetes complication vector, the diabetes complication vector can be fully connected first, and then the obtained fully connected vector can be classified and output, such as through classification functions such as SOFTMAX, to determine the prediction probability corresponding to each complication, so that the corresponding target diabetes complication prediction result can be formed, or a corresponding threshold can be set. When the corresponding prediction probability is greater than the threshold, it can be determined that the corresponding complication will appear, and when the prediction probability is not greater than the threshold, it will not appear.

[0077] Based on the above method, since the prediction is made by using a neural network, the powerful learning ability of the neural network and the prediction ability formed after learning can be fully utilized. Therefore, compared with the conventional technical solution based on expert analysis and prediction, it has higher reliability. In addition, in the process of using the neural network for prediction, the obtained user data is not predicted directly, but the basic vector representing the semantics of diabetes complications is combined, so that the semantic representation ability of the diabetes complication vector in the semantic space of complications is more reliable. In addition, since the interference vector is also fused, the overfitting problem caused by the fusion of the basic vector representing the semantics of diabetes complications can be avoided, so that the reliability of the mined diabetes complication vector can be further improved. Therefore, the problem of relatively low reliability of diabetes complication prediction in the prior art can be improved.

[0078] It can be understood that in order to make the above-mentioned target complication prediction network have reliable diabetes complication prediction function, in a specific embodiment, the diabetes complication prediction method based on neural network can also include the step of training to form the target complication prediction network. For example, it can include step S140, step S150, step S160, step S170 and step S180. The specific content of each step is described as follows.

[0079] Step S140, determining a training diabetes information combination.

[0080] In an embodiment of the present invention, a training diabetes information combination may be determined. The training diabetes information combination includes training diabetes user data and diabetes complication labels, and the diabetes complication labels refer to actual diabetes complications corresponding to the training diabetes user data. In addition, the training diabetes information combination may be multiple, and the training may be performed in batches.

[0081] Step S150, through the semantic mining unit included in the candidate complication prediction network, the diabetes complication label is subjected to interference information application operation, and a training complication interference vector is output. Furthermore, based on the training diabetes user data, the training complication interference vector is subjected to a first number of stages of interference removal operations, and the training diabetes complication vector corresponding to each stage of the interference removal operation is output.

[0082] In an embodiment of the present invention, after determining the training diabetes information combination, the semantic mining unit included in the candidate complication prediction network can be used to perform an interference information application operation on the diabetes complication label to output a training complication interference vector (that is, some interference information can be applied to the semantic information corresponding to the diabetes complication label), and, based on the training diabetes user data, the training complication interference vector is subjected to a first number of stages of interference removal operations, and a training diabetes complication vector corresponding to the interference removal operation of each stage is output. That is to say, the semantic information corresponding to the training diabetes user data can be integrated into the training complication interference vector, so that the obtained semantic information of the training diabetes complication vector carries the actual information of the diabetic user, thereby removing the interference information.

[0083] Step S160, determining a second number of training diabetes complication vectors capable of performing error characterization on the semantic mining unit from the first number of training diabetes complication vectors corresponding to the interference removal operations of the first number of stages.

[0084] In an embodiment of the present invention, after obtaining the training diabetes complication vectors, a second number of training diabetes complication vectors capable of performing error characterization on the semantic mining unit can be determined from the first number of training diabetes complication vectors corresponding to the interference removal operations of the first number of stages (i.e., one stage of interference operation corresponds to one training diabetes complication vector). The second number is less than the first number, and the specific value is configured according to demand.

[0085] Step S170: performing an error determination operation based on the second number of training diabetes complication vectors to obtain a target training error corresponding to the semantic mining unit.

[0086] In an embodiment of the present invention, after determining the second number of training diabetes complication vectors, an error determination operation can be performed based on the second number of training diabetes complication vectors to obtain a target training error corresponding to the semantic mining unit. Wherein, each training diabetes complication vector in the second number of training diabetes complication vectors corresponds to a type of target training error, so that a richer range of errors can be obtained.

[0087] Step S180, updating the network parameters of the semantic mining unit according to the obtained second number of target training errors to form the target complication prediction network.

[0088] In an embodiment of the present invention, after obtaining the target training error, the network parameters of the semantic mining unit can be updated according to the second number of target training errors obtained to form the target complication prediction network. In this way, since the target training error is richer, the update of the network parameters of the semantic mining unit can be more reliable, thereby ensuring that the accuracy of the formed target complication prediction network can be higher. In addition, in the updating process, the second number of target training errors can be summed up first, and then the network parameters can be updated in the direction of reducing the total error of the summation calculation until the total error converges. In addition, in some embodiments, after obtaining the first number of training diabetes complication vectors, the last training diabetes complication vector can be semantically analyzed by the semantic analysis unit to obtain the corresponding complication prediction result, and then the error between the complication prediction result and the diabetes complication label can be calculated, and then the sum of the error and the aforementioned total error can be calculated to obtain the target error, and finally, the network parameters of the candidate complication prediction network (i.e., the semantic mining unit and the semantic analysis unit) are updated in the direction of reducing the target error until the target error converges to obtain the corresponding target complication prediction network.

[0089] In the above implementation, for step S150, the specific method of outputting the training diabetes complication vector corresponding to the interference removal operation at each stage is not limited, and can be selected accordingly according to actual needs. For example, in a specific implementation, in order to ensure that the obtained training diabetes complication vector has better semantic representation ability, the above step S150 can further include step S151, step S152, step S153 and step S154, and the contents of each step are as follows.

[0090] Step S151, performing a semantic mining operation on the diabetes complication label through a semantic mining subunit in a semantic mining unit included in a candidate complication prediction network, and outputting a complication label semantic vector corresponding to the diabetes complication label.

[0091] In an embodiment of the present invention, the semantic mining subunit in the semantic mining unit included in the candidate complication prediction network can be used to perform semantic mining operations on the diabetes complication label, and output a complication label semantic vector corresponding to the diabetes complication label. For example, the semantic mining subunit can include a word embedding model, so that the semantic mining operation can be a word embedding process, as described above, that is, the complication label semantic vector can be a corresponding word embedding vector.

[0092] Step S152, performing an interference information application operation on the complication label semantic vector corresponding to the diabetes complication label, and outputting a complication label interference vector corresponding to the complication label semantic vector.

[0093] In an embodiment of the present invention, after obtaining the complication label semantic vector corresponding to the diabetes complication label, the complication label semantic vector corresponding to the diabetes complication label can be subjected to interference information application operation to output the complication label interference vector corresponding to the complication label semantic vector. Exemplarily, a random vector can be generated, and then the random vector can be superimposed with the complication label semantic vector to obtain the complication label interference vector.

[0094] Step S153: Performing a vector space conversion operation on the complication label interference vector through the semantic mining subunit to generate and output a training complication interference vector corresponding to the diabetes complication label.

[0095] In an embodiment of the present invention, after obtaining the complication label interference vector, the semantic mining subunit can perform a vector space conversion operation on the complication label interference vector to generate and output a training complication interference vector corresponding to the diabetes complication label. Exemplarily, the conversion matrix carried by the semantic mining subunit can be multiplied with the complication label interference vector, and then the multiplication result can be multiplied with the bias matrix carried by the semantic mining subunit, so that the corresponding training complication interference vector can be obtained.

[0096] Step S154, through the interference removal subunit in the semantic mining unit, based on the training diabetes user data, the training complication interference vector is subjected to a first number of stages of interference removal operations, and the training diabetes complication vector corresponding to each stage of the interference removal operation is output.

[0097] In an embodiment of the present invention, after obtaining the training complication interference vector, the interference removal subunit in the semantic mining unit can perform a first number of stages of interference removal operations on the training complication interference vector based on the training diabetes user data, and output the training diabetes complication vector corresponding to the interference removal operation in each stage. That is to say, by integrating the semantic information of the training diabetes user data into the training complication interference vector, the interference information can be suppressed, thereby obtaining the training diabetes complication vector.

[0098] It is understandable that in the above-mentioned implementation, for step S154, the specific manner of the interference removal operation is not limited and can be selected according to actual needs. For example, in a specific implementation, in order to ensure the reliability of the interference removal operation, so that the output training diabetes complication vector has reliable semantic information corresponding to the training diabetes user data, the above-mentioned step S154 can further include the following specific implementation contents (combined with Figure 3 shown):

[0099] First, in the first stage, the interference removal subunit in the semantic mining unit performs an interference removal operation on the training complication interference vector based on the diabetes user data vector corresponding to the training diabetes user data, and outputs the training diabetes complication vector corresponding to the interference removal operation of the first stage, wherein the diabetes user data vector corresponding to the training diabetes user data is formed by the semantic mining subunit performing a semantic mining operation (such as word embedding processing) on ​​the training diabetes user data;

[0100] Secondly, in the second stage and each subsequent stage, the interference removal sub-unit in the semantic mining unit performs an interference removal operation on the training diabetes complications vector corresponding to the interference removal operation of the previous stage based on the diabetes user data vector, and outputs the training diabetes complications vector corresponding to the interference removal operation of the current stage; that is, the diabetes user data vector is integrated into each stage, that is, the semantic information corresponding to the training diabetes user data is integrated, so that as the stage deepens, the semantic information corresponding to the training diabetes user data integrated becomes deeper, thereby fully suppressing the applied interference information. Therefore, it can be ensured that the semantic information of the training diabetes complications vectors obtained in each stage is more reliable.

[0101] It can be understood that the specific method of "performing an interference removal operation on the training diabetes complication vector corresponding to the interference removal operation of the previous stage" can be consistent with the specific method of "performing an interference removal operation on the training complication interference vector". In the embodiment of the present invention, "based on the diabetes user data vector, performing an interference removal operation on the training diabetes complication vector corresponding to the interference removal operation of the previous stage, and outputting the training diabetes complication vector corresponding to the interference removal operation of the current stage" is used as an example for explanation, which specifically includes (combined with Figure 4 shown):

[0102] First, the diabetic user data vector may be subjected to a first processing (such as multiplying it with a first matrix included in an interference removal subunit) to obtain a first vector, and the training diabetic complications vector corresponding to the interference removal operation of the previous stage may be subjected to a second processing and a third processing respectively (such as multiplying it with a second matrix and a third matrix included in the interference removal subunit respectively) to obtain a second vector and a third vector, and then, the transposed vectors of the first vector and the second vector may be multiplied to obtain corresponding vector correlation parameters, and the vector correlation parameters may be multiplied with the third vector to obtain a corresponding correlation aggregation vector, that is, the diabetic user data vector may be fused into the training diabetic complications vector corresponding to the interference removal operation of the previous stage, and further, the correlation aggregation vector may be subjected to a convolution processing (such as using the same convolution to keep the size of the vector unchanged) to obtain a corresponding convolution vector, and then, the convolution vector and the diabetic user data vector may be subjected to mean calculation to obtain a training diabetic complications vector corresponding to the interference removal operation of the current stage.

[0103] In the above-mentioned implementation, for step S160, the specific method of determining the second number of training diabetes complication vectors that can perform error characterization on the semantic mining unit is not limited and can be selected according to actual needs. For example, in a specific implementation, in order to ensure that the determined second number of training diabetes complication vectors can provide rich information and reduce the amount of calculation in the subsequent error calculation, the above-mentioned step S160 may include the following contents:

[0104] First, the ath stage and the bth stage can be determined, wherein the ath stage and the bth stage belong to one of the first number of stages respectively, and belong to different stages; illustratively, the first number can be 20, so the ath stage can be one of 10-14 stages, and the bth stage can be one of 15-19 stages, that is, they are relatively later stages to ensure better semantic representation ability;

[0105] Secondly, the training diabetic complication vectors corresponding to the interference removal operation of the a-th stage and the training diabetic complication vectors corresponding to the interference removal operation of the b-th stage in the first number of training diabetic complication vectors can be marked to obtain a second number of training diabetic complication vectors, that is, the second number of training diabetic complication vectors include the training diabetic complication vectors corresponding to the interference removal operation of the a-th stage and the training diabetic complication vectors corresponding to the interference removal operation of the b-th stage;

[0106] Among them, the ath stage is located before the bth stage, the training diabetes complications vector corresponding to the interference removal operation of the ath stage is used to determine the target training error of the first type, and the training diabetes complications vector corresponding to the interference removal operation of the bth stage is used to determine the target training error of the second type, that is, two target training errors can be determined.

[0107] It can be understood that, in the above implementation, for step S170, the specific method of obtaining the target training error corresponding to the semantic mining unit is not limited and can be selected according to actual needs. For example, in a specific implementation, in order to ensure the reliability of the determined target training error, the above step S170 can further include:

[0108] First, the semantic analysis unit may perform a semantic analysis operation on the training diabetes complication vector corresponding to the interference removal operation in the ath stage (the specific process may refer to the relevant description above), and output a first complication prediction result;

[0109] Secondly, the first complication prediction result and the training diabetic user data may be subjected to a prediction result correlation analysis, and a correlation analysis result between the first complication prediction result and the training diabetic user data may be output, wherein the correlation analysis result may be used to reflect the correlation between the first complication prediction result and the training diabetic user data;

[0110] Then, the prediction result training error corresponding to the semantic mining unit can be determined according to the correlation analysis result, wherein the prediction result training error belongs to the first type of training error; in addition, when the correlation represented by the correlation analysis result is higher, the prediction result training error can be larger, and when the correlation represented by the correlation analysis result is lower, the prediction result training error can be smaller;

[0111] Finally, the target training error of the interference removal operation in the ath stage for the semantic mining unit can be determined according to the prediction result training error.

[0112] It can be understood that, in a specific implementation, in order to ensure the reliability of the correlation analysis result, the step of performing a prediction result correlation analysis on the first complication prediction result and the training diabetic user data, and outputting the correlation analysis result between the first complication prediction result and the training diabetic user data may further include the following contents:

[0113] First, the semantic mining subunit in the semantic mining unit can perform a semantic mining operation on the first complication prediction result, and perform a semantic space conversion operation based on the result of the semantic mining operation to output a first semantic mining vector corresponding to the first complication prediction result; for example, the result of the semantic mining operation can be multiplied by the first space mapping matrix carried by the semantic mining subunit, and then the multiplication result can be added to the first space bias matrix carried by the semantic mining subunit to obtain the first semantic mining vector;

[0114] Secondly, the semantic mining subunit performs a semantic mining operation on the training diabetes user data, and performs a semantic space conversion operation based on the result of the semantic mining operation, and outputs a second semantic mining vector corresponding to the training diabetes user data, wherein the first semantic mining vector and the second semantic mining vector are in the same semantic space; for example, the result of the semantic mining operation can be multiplied by the second space mapping matrix carried by the semantic mining subunit, and then the multiplication result is added to the second space bias matrix carried by the semantic mining subunit to obtain the second semantic mining vector;

[0115] Then, a vector correlation parameter (such as cosine similarity) between the first semantic mining vector and the second semantic mining vector can be determined, and a correlation analysis result between the first complication prediction result and the training diabetes user data can be determined based on the vector correlation parameter. For example, the vector correlation parameter can be used as a correlation analysis result;

[0116] The prediction result correlation analysis includes semantic mining operations, semantic space conversion operations and vector correlation parameter determination operations, as described above.

[0117] It can be understood that, in a specific implementation, in order to ensure that the determined target training error has a high accuracy, the step of determining the target training error of the interference removal operation in the ath stage for the semantic mining unit according to the prediction result training error may further include the following specific implementation contents:

[0118] First, a difference operation can be performed on the training diabetes complication vector corresponding to the interference removal operation of the a-th stage and the training complication interference vector (the vector parameters at corresponding positions are differentiated), and the training diabetes difference vector corresponding to the interference removal operation of the a-th stage is output;

[0119] Secondly, the corresponding first interference training error can be determined based on the difference between the training diabetes difference vector corresponding to the interference removal operation of the a-th stage and the characterization vector of the interference information corresponding to the training complication interference vector (such as the aforementioned applied random vector); illustratively, the mean square error of the training diabetes difference vector and the characterization vector can be calculated to obtain the corresponding first interference training error, such as calculating the square value of the difference of the parameters corresponding to each position between the two vectors, and then calculating the sum of the square values, and the ratio between the calculated sum and the number of each position, so as to obtain the first interference training error;

[0120] Finally, based on the first interference training error and the prediction result training error, the target training error of the interference removal operation in the ath stage for the semantic mining unit can be determined; exemplarily, the sum of the first interference training error and the prediction result training error can be calculated to obtain the corresponding target training error.

[0121] It can be understood that, in the above implementation, for step S170, the specific method of obtaining the target training error corresponding to the semantic mining unit is not limited and can be selected according to actual needs. For example, in a specific implementation, in order to ensure the reliability of the determined target training error, the above step S170 can further include the following specific implementation content:

[0122] First, a semantic analysis unit may be used to perform a semantic analysis operation on the training diabetes complication vector corresponding to the interference removal operation in the bth stage to generate a second complication prediction result;

[0123] Secondly, the text fluency detection network formed by pre-training (the existing text fluency detection network can be used, and the text fluency detection network formed by learning based on the description data of diabetes complications and the corresponding fluency labels can also be used, which is not specifically limited) can be used to evaluate the text fluency of the second complication prediction result, output the text fluency corresponding to the second complication prediction result, and, based on the text fluency, determine the text fluency training error corresponding to the semantic mining unit, wherein the text fluency training error belongs to the second type of training error; illustratively, the text fluency training error can be negatively correlated with the text fluency, that is, the larger the text fluency, the smaller the text fluency training error, and the smaller the text fluency, the larger the text fluency training error;

[0124] Finally, the target training error of the interference removal operation in the bth stage for the semantic mining unit can be determined based on the text coherence training error; for example, the text coherence training error can be used as the corresponding target training error.

[0125] It can be understood that, in a specific implementation, in order to ensure the reliability of the determined target training error, the step of determining the target training error of the interference removal operation in the bth stage for the semantic mining unit according to the text fluency training error may further include the following specific implementation contents:

[0126] First, a difference operation may be performed on the training diabetes complication vector corresponding to the interference removal operation of the b-th stage and the training complication interference vector, and the training diabetes difference vector corresponding to the interference removal operation of the b-th stage may be output, as described above;

[0127] Secondly, the corresponding second interference training error can be determined according to the difference between the training diabetes difference vector corresponding to the interference removal operation of the bth stage and the representation vector of the interference information corresponding to the training complication interference vector, as described above;

[0128] Finally, based on the second interference training error and the text coherence training error, the target training error of the interference removal operation in the bth stage for the semantic mining unit can be determined; exemplarily, the sum of the second interference training error and the text coherence training error can be calculated to obtain the corresponding target training error.

[0129] In summary, the present invention provides a method for predicting diabetes complications based on a neural network. First, the target complication interference vector is determined through the semantic mining unit included in the target complication prediction network; secondly, the target complication interference vector is subjected to interference removal operation based on the acquired target diabetes user data through the semantic mining unit, and the diabetes complication vector is output; then, the diabetes complication vector is subjected to semantic analysis operation through the semantic analysis unit included in the target complication prediction network, and the target diabetes complication prediction result corresponding to the target diabetes user data is output. Based on the above method, since the neural network is used for prediction, the powerful learning ability of the neural network and the prediction ability formed after learning can be fully utilized. Therefore, compared with the conventional technical solution based on expert analysis and prediction, it has higher reliability. In addition, in the process of using neural networks for prediction, the acquired user data is not predicted directly, but the basic vectors representing the semantics of diabetes complications are combined, so that the semantic representation ability of the diabetes complications vector in the semantic space of complications is more reliable. In addition, since the interference vector is also fused, the overfitting problem caused by the fusion of the basic vectors representing the semantics of diabetes complications can be avoided, so that the reliability of the mined diabetes complications vector can be further improved. Therefore, the problem of relatively low reliability of diabetes complication prediction in the prior art can be improved.

[0130] In several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed apparatus and method can also be implemented in other ways. The apparatus and method embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the apparatus, method and computer program product according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0131] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0132] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention is essentially or partly contributed to the prior art or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk. It should be noted that, in this article, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements inherent to such a process, method, article or device. Without more constraints, an element defined by the phrase "comprising a..." does not exclude the existence of other identical elements in the process, method, article or apparatus comprising the element.

[0133] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting diabetes complications based on neural network, characterized in that: include: Determining a target complication interference vector through a semantic mining unit included in a target complication prediction network, wherein the target complication prediction network is a trained and optimized neural network, and the target complication interference vector is formed by fusing an interference vector on a basic vector representing the semantics of diabetes complications; Through the semantic mining unit, according to the acquired target diabetic user data, the target complication interference vector is subjected to interference removal operation, and a diabetic complication vector is output, wherein the target diabetic user data at least includes the target diabetic user's diabetes diagnosis data, diabetes treatment data, and diabetes-related detection data; The semantic analysis unit included in the target complication prediction network performs a semantic analysis operation on the diabetes complication vector, and outputs a target diabetes complication prediction result corresponding to the target diabetes user data.

2. The method for predicting diabetes complications based on a neural network as claimed in claim 1, characterized in that: The neural network-based diabetes complication prediction method further includes a step of training and forming the target complication prediction network, which step includes: Determine a training diabetes information combination, wherein the training diabetes information combination includes training diabetes user data and diabetes complication labels, and the diabetes complication labels refer to actual diabetes complications corresponding to the training diabetes user data; The semantic mining unit included in the candidate complication prediction network performs interference information application operation on the diabetes complication label to output a training complication interference vector, and, based on the training diabetes user data, performs a first number of stages of interference removal operation on the training complication interference vector to output a training diabetes complication vector corresponding to each stage of the interference removal operation; Determining a second number of training diabetes complication vectors capable of performing error characterization on the semantic mining unit from a first number of training diabetes complication vectors corresponding to the interference removal operations of the first number of stages, wherein the second number is smaller than the first number; performing an error determination operation according to the second number of training diabetes complication vectors to obtain a target training error corresponding to the semantic mining unit, wherein each training diabetes complication vector in the second number of training diabetes complication vectors corresponds to determining a type of target training error; Based on the obtained second number of target training errors, the network parameters of the semantic mining unit are updated to form the target complication prediction network.

3. The method for predicting diabetes complications based on a neural network as claimed in claim 2, characterized in that: The steps of applying interference information to the diabetes complication label through the semantic mining unit included in the candidate complication prediction network to output a training complication interference vector, and performing a first number of stages of interference removal operations on the training complication interference vector based on the training diabetes user data to output the training diabetes complication vector corresponding to each stage of the interference removal operation include: Performing a semantic mining operation on the diabetes complication label by a semantic mining subunit in a semantic mining unit included in a candidate complication prediction network, and outputting a complication label semantic vector corresponding to the diabetes complication label; Performing an interference information application operation on the complication label semantic vector corresponding to the diabetes complication label, and outputting a complication label interference vector corresponding to the complication label semantic vector; Through the semantic mining subunit, the complication label interference vector is subjected to a vector space conversion operation to generate and output a training complication interference vector corresponding to the diabetes complication label; Through the interference removal subunit in the semantic mining unit, the training complication interference vector is subjected to a first number of stages of interference removal operations based on the training diabetes user data, and the training diabetes complication vector corresponding to each stage of the interference removal operation is output.

4. The method for predicting diabetes complications based on a neural network as claimed in claim 3, characterized in that: The step of performing interference removal operations of a first number of stages on the training complication interference vectors according to the training diabetes user data through the interference removal subunit in the semantic mining unit, and outputting the training diabetes complication vectors corresponding to the interference removal operations of each stage, comprises: In the first stage, the interference removal subunit in the semantic mining unit performs an interference removal operation on the training complication interference vector based on the diabetes user data vector corresponding to the training diabetes user data, and outputs the training diabetes complication vector corresponding to the interference removal operation of the first stage, wherein the diabetes user data vector corresponding to the training diabetes user data is formed by the semantic mining subunit performing a semantic mining operation on the training diabetes user data; In the second stage and each subsequent stage, the interference removal subunit in the semantic mining unit performs an interference removal operation on the training diabetes complications vector corresponding to the interference removal operation in the previous stage based on the diabetes user data vector, and outputs the training diabetes complications vector corresponding to the interference removal operation in the current stage.

5. The method for predicting diabetes complications based on a neural network as claimed in claim 2, characterized in that: The step of determining a second number of training diabetes complication vectors capable of performing error characterization on the semantic mining unit from the first number of training diabetes complication vectors corresponding to the interference removal operations of the first number of stages comprises: Determine an ath stage and a bth stage, wherein the ath stage and the bth stage belong to one stage of the first number of stages respectively and belong to different stages; Marking the training diabetic complication vectors corresponding to the interference removal operation of the ath stage and the training diabetic complication vectors corresponding to the interference removal operation of the bth stage in the first number of training diabetic complication vectors to obtain a second number of training diabetic complication vectors; Among them, the ath stage is located before the bth stage, the training diabetes complications vector corresponding to the interference removal operation of the ath stage is used to determine the target training error of the first type, and the training diabetes complications vector corresponding to the interference removal operation of the bth stage is used to determine the target training error of the second type.

6. The method for predicting diabetes complications based on a neural network as claimed in claim 5, characterized in that: The step of performing an error determination operation based on the second number of training diabetes complication vectors to obtain a target training error corresponding to the semantic mining unit includes: By means of the semantic analysis unit, a semantic analysis operation is performed on the training diabetes complication vector corresponding to the interference removal operation in the ath stage, and a first complication prediction result is output; Performing a prediction result correlation analysis on the first complication prediction result and the training diabetic user data, and outputting a correlation analysis result between the first complication prediction result and the training diabetic user data; Determining, according to the correlation analysis result, a prediction result training error corresponding to the semantic mining unit, wherein the prediction result training error belongs to a first type of training error; According to the prediction result training error, a target training error of the interference removal operation in the a-th stage for the semantic mining unit is determined.

7. The method for predicting diabetes complications based on a neural network as claimed in claim 6, characterized in that: The step of performing a prediction result correlation analysis on the first complication prediction result and the training diabetic user data, and outputting the correlation analysis result between the first complication prediction result and the training diabetic user data, comprises: The semantic mining subunit in the semantic mining unit performs a semantic mining operation on the first complication prediction result, and performs a semantic space conversion operation based on the result of the semantic mining operation, and outputs a first semantic mining vector corresponding to the first complication prediction result; The semantic mining subunit performs a semantic mining operation on the training diabetes user data, and performs a semantic space conversion operation based on the result of the semantic mining operation, and outputs a second semantic mining vector corresponding to the training diabetes user data, wherein the first semantic mining vector and the second semantic mining vector are in the same semantic space; Determine a vector correlation parameter between the first semantic mining vector and the second semantic mining vector, and determine a correlation analysis result between the first complication prediction result and the training diabetes user data based on the vector correlation parameter; The prediction result correlation analysis includes a semantic mining operation, a semantic space conversion operation and a vector correlation parameter determination operation.

8. The method for predicting diabetes complications based on a neural network as claimed in claim 6, characterized in that: The step of determining a target training error of the interference removal operation in the ath stage for the semantic mining unit according to the prediction result training error includes: Performing a difference operation on the training diabetes complication vector corresponding to the interference removal operation of the a-th stage and the training complication interference vector, and outputting the training diabetes difference vector corresponding to the interference removal operation of the a-th stage; Determine a corresponding first interference training error according to the difference between the training diabetes difference vector corresponding to the interference removal operation of the ath stage and the representation vector of the interference information corresponding to the training complication interference vector; Based on the first interference training error and the prediction result training error, a target training error of the interference removal operation in the ath stage for the semantic mining unit is determined.

9. The method for predicting diabetes complications based on a neural network as claimed in claim 5, characterized in that: The step of performing an error determination operation based on the second number of training diabetes complication vectors to obtain a target training error corresponding to the semantic mining unit includes: Performing a semantic analysis operation on the training diabetes complication vector corresponding to the interference removal operation in the b-th stage by a semantic analysis unit to generate a second complication prediction result; Through a text fluency detection network formed by pre-training, the text fluency of the second complication prediction result is evaluated, the text fluency corresponding to the second complication prediction result is output, and, based on the text fluency, the text fluency training error corresponding to the semantic mining unit is determined, wherein the text fluency training error belongs to the second type of training error; According to the text fluency training error, a target training error of the interference removal operation in the b-th stage for the semantic mining unit is determined.

10. The method for predicting diabetes complications based on a neural network according to claim 9, characterized in that: The step of determining the target training error of the interference removal operation in the bth stage for the semantic mining unit according to the text fluency training error includes: Performing a difference operation on the training diabetes complication vector corresponding to the interference removal operation of the b-th stage and the training complication interference vector, and outputting the training diabetes difference vector corresponding to the interference removal operation of the b-th stage; Determine a corresponding second interference training error according to a difference between the training diabetes difference vector corresponding to the interference removal operation of the b-th stage and the representation vector of interference information corresponding to the training complication interference vector; Based on the second interference training error and the text fluency training error, a target training error of the interference removal operation in the b-th stage for the semantic mining unit is determined.

Images