Diagnostic model training method and device, medium and electronic equipment

Abstract

The invention belongs to the technical field of medical health, and relates to a diagnosis model training method and device, a medium and electronic equipment. The method comprises the following steps: acquiring an initialization model for diagnosing cardiovascular diseases and initial parameters of the initialization model, and pruning the initialization model according to the initial parameters to obtain a first diagnosis model; performing pre-training processing on the first diagnosis model to obtain a second diagnosis model, and performing fine tuning processing on the second diagnosis model to obtain a third diagnosis model; performing model training on the third diagnosis model to obtain a fourth diagnosis model, and performing sparsity analysis on the fourth diagnosis model to obtain target parameters of the fourth diagnosis model; and performing weighted aggregation processing on the fourth diagnosis model by using the target parameter to obtain a target diagnosis model, and deploying the target diagnosis model on the electrocardiogram equipment to diagnose cardiovascular diseases. The diagnosis model can be deployed on portable electrocardiogram equipment and the like, and the generalization ability and accuracy of the model are improved.

Description

Technical Field

[0001] The present disclosure belongs to the field of medical health technology, and in particular, relates to a diagnostic model training method, a diagnostic model training device, a computer-readable storage medium, and an electronic device. Background Art

[0002] Intelligent diagnosis of cardiovascular disease is one of the key technologies in the field of healthcare. With the development of mobile medical and remote monitoring technologies, portable electrocardiogram (ECG) devices have become increasingly popular due to their convenience and low cost. These devices usually have limited computing and storage resources, so lightweight models are needed to achieve fast and accurate diagnosis.

[0003] However, existing deep learning models are difficult to deploy directly on these devices due to their complex structures and large number of parameters. In addition, the ECG devices used by different medical institutions are different, which poses a challenge to the generalization and adaptability of the model. Summary of the invention

[0004] In order to overcome the problems existing in the related art, the present disclosure provides a diagnostic model training method, a diagnostic model training device, a computer-readable storage medium and an electronic device.

[0005] According to a first aspect of an embodiment of the present disclosure, a method for training a diagnostic model is provided, the method comprising:

[0006] Acquire an initialization model for diagnosing cardiovascular diseases and initial parameters of the initialization model, and perform a pruning operation on the initialization model according to the initial parameters to obtain a first diagnostic model;

[0007] Pre-training the first diagnostic model to obtain a second diagnostic model, and fine-tuning the second diagnostic model to obtain a third diagnostic model;

[0008] Performing model training on the third diagnostic model to obtain a fourth diagnostic model, and performing sparsity analysis on the fourth diagnostic model to obtain target parameters of the fourth diagnostic model;

[0009] The target parameter is used to perform weighted aggregation processing on the fourth diagnostic model to obtain a target diagnostic model, which is deployed on an electrocardiogram device for diagnosis of cardiovascular diseases.

[0010] Optionally, performing a pruning operation on the initialized model according to the initial parameters to obtain a first diagnostic model includes:

[0011] Calculating the initial parameters to obtain performance parameters of the electrocardiogram device, and obtaining parameter weights of the performance parameters;

[0012] Calculating the performance parameters and the parameter weights to obtain pruning parameters of the initialization model, and determining a plurality of pruning algorithms corresponding to the pruning parameters;

[0013] A target algorithm is determined among the multiple pruning algorithms according to the pruning parameters, and a pruning operation is performed on the initialization model according to the target algorithm to obtain a first diagnostic model.

[0014] Optionally, performing a pruning operation on the initialization model according to the target algorithm to obtain a first diagnostic model includes:

[0015] Obtaining an influence coefficient corresponding to the initial parameter, and using the influence coefficient to perform an initial screening process on the initialization model to obtain a screened initialization model;

[0016] Calculating the influence coefficient to obtain a coefficient threshold, and determining a parameter to be screened in the initial parameters of the screened initialization model according to the coefficient threshold;

[0017] The coefficient correlation is determined according to the parameters to be screened, and the screened initialization model is subjected to secondary screening processing according to the coefficient correlation to obtain a first diagnostic model.

[0018] Optionally, performing a pruning operation on the initialization model according to the target algorithm to obtain a first diagnostic model includes:

[0019] Obtaining an influence coefficient corresponding to the initial parameter, and calculating the influence coefficient to obtain a global standard value and a local standard value;

[0020] Performing preliminary fine-cutting processing on the initialization model according to the global standard value and the local standard value to obtain coefficients to be reconstructed;

[0021] The coefficients to be reconstructed are reconstructed to obtain reconstruction parameters, and the initialized model after preliminary fine-cutting is dynamically fine-cut according to the reconstruction parameters to obtain a first diagnostic model.

[0022] Optionally, performing a pruning operation on the initialization model according to the target algorithm to obtain a first diagnostic model includes:

[0023] Acquire the information dimension of the initial parameter and the dimension threshold of the information dimension, and perform preliminary screening processing on the initialization model according to the dimension threshold to obtain a screened initialization model;

[0024] Performing entropy spectrum analysis on the screened initialization model to obtain an entropy value, and determining a gain adjustment factor according to the entropy value;

[0025] The entropy value and the gain adjustment factor are calculated to obtain an entropy value threshold, and a fine pruning operation is performed on the screened initialization model according to the entropy value threshold to obtain a first diagnostic model.

[0026] Optionally, performing a pruning operation on the initialized model according to the initial parameters to obtain a first diagnostic model includes:

[0027] Performing a pruning operation on the initialization model according to the initial parameters to obtain a pruned initialization model and a pruning evaluation parameter, and obtaining a pruning threshold corresponding to the pruning evaluation parameter;

[0028] The pruned initialization model is determined as a first diagnostic model according to the pruning evaluation parameter and the pruning threshold.

[0029] Optionally, the pre-training the first diagnostic model to obtain the second diagnostic model includes:

[0030] Acquire a data set stored on an electrocardiogram device, and use the data set to pre-train the first diagnostic model to obtain a fifth diagnostic model;

[0031] The fifth diagnostic model is subjected to distillation learning to obtain a second diagnostic model.

[0032] According to a second aspect of an embodiment of the present disclosure, there is provided a diagnostic model training device, comprising:

[0033] A model pruning module is configured to obtain an initialization model for diagnosing cardiovascular diseases and initial parameters of the initialization model, and perform a pruning operation on the initialization model according to the initial parameters to obtain a first diagnostic model;

[0034] A model fine-tuning module is configured to perform a pre-training process on the first diagnostic model to obtain a second diagnostic model, and to perform a fine-tuning process on the second diagnostic model to obtain a third diagnostic model;

[0035] a sparse analysis model configured to perform model training on the third diagnostic model to obtain a fourth diagnostic model, and perform sparsity analysis on the fourth diagnostic model to obtain target parameters of the fourth diagnostic model;

[0036] The deployment diagnosis module is configured to use the target parameters to perform weighted aggregation processing on the fourth diagnosis model to obtain a target diagnosis model, so as to be deployed on an electrocardiogram device for the diagnosis of cardiovascular diseases.

[0037] According to a third aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored. When the program instructions are executed by a processor, the steps of the diagnostic model training method provided in any one of the first aspects of the present disclosure are implemented.

[0038] According to a fourth aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0039] processor;

[0040] a memory for storing processor-executable instructions;

[0041] Wherein, the processor is configured to: execute the executable instructions to implement the steps of any one of the diagnostic model training methods provided in the first aspect of the present disclosure.

[0042] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects:

[0043] In the method and apparatus provided by the exemplary embodiments of the present disclosure, the initialization model is pruned according to the initial parameters, and different degrees of model pruning can be adaptively performed according to the actual conditions and computing resources of different electrocardiogram devices, so as to better adapt to the performance of different electrocardiogram devices, improve the execution efficiency of the model on each electrocardiogram device, speed up the training process of the model, make the model more efficient, improve the generalization ability and accuracy of the model, and lay a foundation for the deployment of the model on a portable electrocardiogram device. Further, in the process of pre-training and model training, it is proposed to fine-tune the diagnostic model so that the model can better adapt to the characteristics of local data, generate personalized models on different electrocardiogram devices, improve the performance of the model on the electrocardiogram device in a targeted manner, enable the model to continuously adapt to new data and scenarios, maintain the update degree of the model, and further improve the accuracy and generalization ability of the model. Furthermore, the trained model is subjected to sparsity analysis and weighted aggregation processing to ensure that the aggregation effect can accurately reflect the characteristics and structure of the pruned model, and more evenly consider the heterogeneity of different resources of the electrocardiogram device and the heterogeneity of the data structure, so as to improve the generalization and accuracy of the model.

[0044] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0046] Figure 1A schematic diagram schematically shows a flow chart of a method for training a diagnostic model in an exemplary embodiment of the present disclosure;

[0047] Figure 2 A schematic diagram of a process of performing a pruning operation on an initialized model in an exemplary embodiment of the present disclosure is schematically shown;

[0048] Figure 3 A schematic diagram of a process of performing a pruning operation on an initialized model in an exemplary embodiment of the present disclosure is schematically shown;

[0049] Figure 4 A schematic diagram of a process of another method for performing a pruning operation on an initialized model in an exemplary embodiment of the present disclosure is schematically shown;

[0050] Figure 5 A schematic diagram of a process of performing a pruning operation on an initialized model in an exemplary embodiment of the present disclosure is schematically shown;

[0051] Figure 6 A schematic diagram of a process of obtaining a first diagnostic model according to a pruning operation in an exemplary embodiment of the present disclosure is schematically shown;

[0052] Figure 7 A schematic diagram schematically shows a flow chart of a method for pre-training a first diagnostic model in an exemplary embodiment of the present disclosure;

[0053] Figure 8 A structural framework diagram schematically illustrates a method for training a diagnostic model in an application scenario in an exemplary embodiment of the present disclosure;

[0054] Fig. 9 A flowchart schematically illustrates a method for training a diagnostic model in an application scenario in an exemplary embodiment of the present disclosure;

[0055] Fig.10 A schematic diagram schematically shows a structure of a training device for a diagnostic model in an exemplary embodiment of the present disclosure;

[0056] Fig.11 A schematic diagram schematically shows the structure of another diagnostic model training device in an exemplary embodiment of the present disclosure;

[0057] Fig.12 The structural diagram of another diagnostic model training device in the exemplary embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0058] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0059] It should be noted that all actions of acquiring signals, information or data in the present disclosure are carried out in compliance with the relevant data protection laws and policies of the country where the device is located and with the authorization given by the owner of the corresponding device.

[0060] Intelligent diagnosis of cardiovascular disease is one of the key technologies in the field of healthcare. With the development of mobile medical and remote monitoring technologies, portable electrocardiogram devices are gaining more and more attention due to their convenience and low cost. These devices usually have limited computing and storage resources, so lightweight models are needed to achieve fast and accurate diagnosis.

[0061] To improve the intelligent diagnosis of cardiovascular diseases, researchers have developed a variety of deep learning-based methods, which are usually trained on large data sets and can achieve high-precision diagnosis.

[0062] However, these models often require powerful computing resources and are not suitable for direct application to portable devices. In other words, existing deep learning models are difficult to deploy directly on these devices due to their complex structure and large number of parameters. In addition, the ECG equipment used by different medical institutions is different, which poses a challenge to the generalization and adaptability of the model.

[0063] In view of the problems existing in the related art, the present disclosure provides a training method for a diagnostic model. Figure 1 is a flow chart of a method for training a diagnostic model according to an exemplary embodiment. Figure 1 As shown, the method may at least include the following steps:

[0064] Step S110: Acquire an initialization model for diagnosing cardiovascular diseases and initial parameters of the initialization model, and perform a pruning operation on the initialization model according to the initial parameters to obtain a first diagnostic model.

[0065] Step S120: pre-train the first diagnostic model to obtain a second diagnostic model, and fine-tune the second diagnostic model to obtain a third diagnostic model.

[0066] Step S130: Perform model training on the third diagnostic model to obtain a fourth diagnostic model, and perform sparsity analysis on the fourth diagnostic model to obtain target parameters of the fourth diagnostic model.

[0067] Step S140: Perform weighted aggregation processing on the fourth diagnostic model using the target parameters to obtain a target diagnostic model, which is deployed on an electrocardiogram device for diagnosis of cardiovascular diseases.

[0068] In an exemplary embodiment of the present disclosure, the initialization model is pruned according to the initial parameters, and different degrees of model pruning can be adaptively performed according to the actual conditions and computing resources of different electrocardiogram devices, so as to better adapt to the performance of different electrocardiogram devices, improve the execution efficiency of the model on each electrocardiogram device, speed up the training process of the model, make the model more efficient, improve the generalization ability and accuracy of the model, and lay a foundation for the deployment of the model on a portable electrocardiogram device. Further, in the process of pre-training and model training, it is proposed to fine-tune the diagnostic model so that the model can better adapt to the characteristics of local data, generate personalized models on different electrocardiogram devices, improve the performance of the model on the electrocardiogram device in a targeted manner, enable the model to continuously adapt to new data and scenarios, maintain the update degree of the model, and further improve the accuracy and generalization ability of the model. Furthermore, the trained model is subjected to sparsity analysis and weighted aggregation processing to ensure that the aggregation effect can accurately reflect the model characteristics and structure after pruning, and more evenly consider the heterogeneity of different resources of the electrocardiogram device and the heterogeneity of the data structure, so as to improve the generalization and accuracy of the model.

[0069] The following is a detailed description of each step of the diagnostic model training method.

[0070] In step S110, an initialization model for diagnosing cardiovascular diseases and initial parameters of the initialization model are obtained, and a pruning operation is performed on the initialization model according to the initial parameters to obtain a first diagnostic model.

[0071] In an exemplary embodiment of the present disclosure, an intelligent diagnosis model for cardiovascular disease is initialized as an initialization model according to actual needs.

[0072] For example, when the initialization model is a ResNet (deep residual network) model, the structure of the ResNet model and the initial parameters of the initialization model, such as the input size, the type of output disease, etc., can be designed accordingly.

[0073] In an alternative embodiment, Figure 2 A flow chart of a method for pruning an initialized model is shown, such as Figure 2 As shown, the method may at least include the following steps: In step S210, the initial parameters are calculated to obtain the performance parameters of the electrocardiogram device, and the parameter weights of the performance parameters are obtained.

[0074] In general, the initial parameters include not only the input size and the type of output disease, but also the computing power C of the ECG device, the memory capacity M, the delay requirement L, and the maximum delay time L. max , the computing power CC required by the model and the memory MM required by the model, etc.

[0075] Furthermore, the computing power of the ECG equipment is CM is the computing power required for model estimation. Computing power is usually expressed in floating-point operations per second. The higher the CC value, the more computing power the server has to run the model. If CC is low, model pruning is needed to reduce the computing requirements of the model.

[0076] Calculate the memory capacity ratio of the ECG device MM is the estimated memory required to run the model. The higher the MC, the more memory the server has to load and run the model. If the MC is low, you need to prune the model to reduce its memory usage.

[0077] Calculating latency ratio for ECG equipment L max is the maximum delay time that can be accepted in the actual application of cardiovascular disease diagnosis, L is the estimated delay time of the current model on the server, and PL reflects the user experience and system response time.

[0078] Therefore, performance parameters such as computing power ratio, memory capacity ratio and delay ratio can be calculated based on the initial parameters, and weight coefficients ω1, ω2, ω3, ω4 assigned to each performance parameter can also be obtained.

[0079] In step S220, the performance parameters and parameter weights are calculated to obtain pruning parameters of the initialization model, and multiple pruning algorithms corresponding to the pruning parameters are determined.

[0080] Generally, the pruning degree P can be calculated as a pruning parameter using P=ω1*(1-T)+ω2*CC+ω3*MC+ω4*PL, where T is the estimated compression rate of the electrocardiogram device.

[0081] In the absence of a pruned model size, the compression rate T is used as the target, and the degree of pruning is estimated by 1-T. At the same time, the three performance parameters of computing power ratio (CC), memory capacity ratio (MC) and latency ratio (PL) are comprehensively considered, so that the computing and memory resources of the ECG device can be adapted while meeting the real-time requirements.

[0082] Furthermore, multiple pruning algorithms corresponding to pruning parameters can be determined as shown in Table 1:

[0083] Degree of pruning Pruning Algorithm P<10 No pruning algorithm is used 10≤P≤20 IPFS 40<P≤70 DCOP P>70 DMIOP

[0084] Table 1

[0085] The IPFS (Intelli-Prune Feature Selection) algorithm focuses on feature-level pruning through feature coefficient derivation and zero coefficient screening, as well as feature importance evaluation. This method achieves accurate model pruning by identifying and retaining the features that have the greatest impact on the prediction target.

[0086] The precise pruning strategy of the IPFS algorithm is suitable for a pruning degree of 10-20%. This pruning degree moderately reduces the number of parameters of the model while retaining the structure and performance of the model to the greatest extent.

[0087] The DCOP (Dynamic Covariance Optimization Pruning) algorithm focuses on the interdependencies between weights and achieves medium-depth pruning by mapping the covariance structure and setting global and local importance boundaries. This method can accurately identify and prune weights that contribute less to the model by dynamically adjusting the threshold.

[0088] The DCOP algorithm is suitable for achieving a 40-70% pruning degree. This pruning degree reduces the model size and computational requirements while maintaining moderate model complexity and high performance.

[0089] The DMIOP (Dynamic Multi-factor Information Optimization Pruning) algorithm can perform deep feature and weight pruning through information dimension baseline mapping and dynamic adjustment, as well as entropy spectrum construction and analysis. This comprehensive evaluation not only identifies key features, but also goes deep into the model to identify and prune redundant or inefficient information dimensions.

[0090] The DMIOP algorithm is suitable for achieving a pruning degree greater than 70%. This pruning degree significantly reduces the number of model parameters and computational requirements while preserving the key performance of the model as much as possible.

[0091] In step S230, a target algorithm is determined from among a plurality of pruning algorithms according to the pruning parameters, and a pruning operation is performed on the initialized model according to the target algorithm to obtain a first diagnostic model.

[0092] According to the calculated size of the pruning parameters and the corresponding relationship between the pruning parameters and the pruning algorithms shown in Table 1, a target algorithm can be determined from a plurality of pruning algorithms to further perform pruning operations on the initialized model according to the target algorithm.

[0093] In an alternative embodiment, Figure 3 A flow chart of a method for pruning an initialized model is shown, such as Figure 3 As shown, the method may at least include the following steps: In step S310, the influence coefficient corresponding to the initial parameter is obtained, and the influence coefficient is used to perform initial screening processing on the initialization model to obtain the screened initialization model.

[0094] The influence coefficient is determined based on the degree of influence of each initial parameter on the prediction target, which facilitates the identification of features that have less influence on the model prediction.

[0095] The initial parameter matrix T and target vector x of the initialization model are obtained, and the objective function is proposed as shown in formula (1):

[0096]

[0097] Among them, β is the feature coefficient, p is the number of features, λ is the regularization weight, which controls the strength of L1 regularization, and n is the number of samples. Through the Lasso regression model, the influence coefficient vector W of the feature can be obtained.

[0098] Furthermore, the initial parameters that have little or no effect on the prediction target are screened and removed.

[0099] Specifically, traverse the influence coefficient vector W of the feature, for each influence coefficient w i , check whether it is zero. If it is zero, remove the corresponding feature T from the feature matrix T i , to obtain the screened initialization model.

[0100] In step S320, the influence coefficient is calculated to obtain a coefficient threshold, and the parameters to be screened are determined from the initial parameters of the screened initialized model according to the coefficient threshold.

[0101] To provide a quantitative decision boundary for further feature selection, we can more accurately define which features are key variables in model prediction. Specifically, we calculate the mean μ of the features with non-zero coefficients. non-zero and standard deviation σ non-zero , and the dynamic threshold obtained by adjusting the hyperparameter k is used as the coefficient threshold, as shown in formula (2):

[0102] θ=μ non-zero +kσ non-zero (2)

[0103] Furthermore, identifying features that are retained after the first pruning but may contribute less to the model will provide a basis for the next step of fine pruning.

[0104] Traverse the feature influence coefficient vector W, for each coefficient wi , will satisfy 0 <w i The features <θ constitute the secondary screening parameters and enter the next screening process.

[0105] In step S330, the coefficient correlation is determined according to the parameters to be screened, and the screened initialization model is subjected to secondary screening processing according to the coefficient correlation to obtain a first diagnostic model.

[0106] Evaluate the relationship between the secondary screening parameters and the target variable. For the secondary screening feature T i , calculate its feature mutual correlation (MCR) with the target variable x as shown in formula (3):

[0107]

[0108] Among them, α is the MCR weight coefficient, p tx is the joint probability distribution of the feature and the target variable; p t 、p x are the marginal probability distributions of the feature and target variables respectively; Cov(t,x) is the covariance of the feature and target variable; σ t , σ x are the standard deviations of the feature and target variables.

[0109] For each feature T i , evaluate whether its MCR is lower than the threshold θ MCR Among them, the threshold θ MCR Take the last third of the MCR corresponding to all features. If it is lower than the threshold θ MCR , then remove the corresponding feature T from the feature matrix T i , to obtain the first diagnostic model.

[0110] Thus, the core feature matrix is ​​determined to form the final feature set for establishing a prediction model, summarizing all unpruned features, verifying the consistency and completeness of the remaining features, and outputting the final core feature set for subsequent training and prediction of the first diagnostic model.

[0111] In an alternative embodiment, Figure 4 A flow chart of another method for pruning an initialized model is shown, Figure 4 As shown, the method may at least include the following steps: In step S410, the influence coefficient corresponding to the initial parameter is obtained, and the influence coefficient is calculated to obtain a global standard value and a local standard value.

[0112] First, the weight matrix is ​​initialized. Specifically, the initial weight set W of the model is defined as the influence coefficient to lay the foundation for the pruning process.

[0113] Then, according to formula (4), the weight pair (ω i ,ω j ) Calculate the covariance C between weights ij :

[0114] C ij = Cov(ω i ,ω j ) (4)

[0115] Then, a standard is set to distinguish the importance of weights as the basis for preliminary pruning. Calculate the median C of the covariance between all weight pairs med , average value C mean and variance σ C , calculate the covariance equilibrium standard value as shown in formula (5):

[0116] τ base =αC med +(1-α)C mean (5)

[0117] Among them, α is the standard value influencing factor, and then the global importance limit threshold is calculated according to formula (6) as the global standard value:

[0118]

[0119] Among them, β1 is the global threshold influencing factor, which is limited according to actual conditions.

[0120] Further refinement within the global importance boundary provides guidance for secondary pruning. The local importance is calculated as the local standard value according to formula (7):

[0121]

[0122] Among them, β2 is the local limit threshold influencing factor, which is limited according to actual conditions.

[0123] In step S420, the initialization model is preliminarily fine-cut according to the global standard value and the local standard value to obtain the coefficients to be reconstructed.

[0124] Specifically, traverse all covariances, if the weight pair (ω i ,ω j ) ij Less than the global standard value τ g , then remove the corresponding weight, that is, W ij =0; if τ g <C ij <τ l , and the corresponding weights constitute a weight set to be reconstructed.

[0125] Furthermore, the weight set to be reconstructed is reparameterized, and new learnable parameters are introduced to maintain the model performance. According to formula (8), the weights that have been initially simplified are reparameterized to obtain the coefficients to be reconstructed:

[0126] w i ′ j =w ij +σ C ⊙∈(8)

[0127] Among them, w i ' j is the updated weight, σ C is the standard deviation, ∈ is the weight vector, satisfying is a normal distribution.

[0128] In step S430, the coefficients to be reconstructed are reconstructed to obtain reconstruction parameters, and the initialization model after the preliminary fine-cutting is dynamically fine-cut according to the reconstruction parameters to obtain the first diagnosis model.

[0129] A new global importance limit threshold is recalculated based on the reconstructed coefficients to be reconstructed. If the reconstructed coefficients to be reconstructed are less than the new global importance limit threshold, a second weight deletion is performed to obtain the first diagnostic model.

[0130] After pruning, the model is calibrated to ensure that the model performance meets expectations.

[0131] In an alternative embodiment, Figure 5 A flow chart of another method for pruning an initialized model is shown, Figure 5 As shown, the method may include at least the following steps: In step S510, the information dimension of the initial parameter and the dimension threshold of the information dimension are obtained, and the initialization model is preliminarily screened according to the dimension threshold to obtain a screened initialization model.

[0132] For each channel feature map of the initial parameters, the corresponding information dimension can be calculated according to formula (9):

[0133] InfoDim(F c )=rank(F c ) (9)

[0134] Among them, F c is the feature map of channel c, rank(F c ) is F c rank.

[0135] Furthermore, the average value of all information dimensions InfoDim is calculatedmean and standard deviation σ i-d .

[0136] Calculate the dynamic adjustment coefficient λ=f(X,F c ,P), where X is the entire model parameter and P is the model performance index, so that the threshold of the information dimension is obtained as the dimension threshold according to formula (10):

[0137] θ i-d =InfoDim mean +λσ i-d (10)

[0138] Traverse all information dimensions and remove the channels corresponding to information dimensions that are smaller than the dimension threshold to obtain the screened initialization model.

[0139] In step S520, entropy spectrum analysis is performed on the screened initialization model to obtain an entropy value, and a gain adjustment factor is determined according to the entropy value.

[0140] According to the formula Calculate the entropy value H(F c ), where p i is the probability of the i-th value appearing in the feature map.

[0141] Furthermore, the gain adjustment factor is calculated according to formula (11):

[0142] ε=f(H t ,H v ,P) (11)

[0143] Among them, H t is the sum of the entropy values ​​of the training set, H v is the sum of the entropy values ​​of the validation set, and P is the model performance indicator.

[0144] In step S530, the entropy value and the gain adjustment factor are calculated to obtain an entropy value threshold, and a fine pruning operation is performed on the screened initialization model according to the entropy value threshold to obtain a first diagnostic model.

[0145] Calculate the average value of all entropy values ​​H mean and standard deviation σ h Afterwards, the dynamic equilibrium threshold of the entropy value is calculated according to formula (12) as the entropy value threshold:

[0146] θ h =H mean +εσ h (12)

[0147] For each channel c, if its entropy spectrum H(F c) is less than the dynamic balance threshold of the entropy value, channel c is pruned to obtain the first diagnostic model.

[0148] It is worth noting that in Figure 3-Figure 5 In the process of pruning the initialization model shown, whether the pruned model can be used as the first diagnostic model can be determined after evaluating the pruning effect.

[0149] In an alternative embodiment, Figure 6 A flow chart of a method for obtaining a first diagnostic model according to a pruning operation is shown, as Figure 6 As shown, the method may at least include the following steps: in step S610, a pruning operation is performed on the initialization model according to the initial parameters to obtain a pruned initialization model and a pruning evaluation parameter, and a pruning threshold corresponding to the pruning evaluation parameter is obtained.

[0150] The way to prune the initialized model according to the initial parameters is as follows Figure 3-Figure 5 After the pruning operation, in addition to obtaining the pruned initialization model, the pruning evaluation parameter Q can also be calculated according to formula (13):

[0151] Q=||W p -W o ||2 (13)

[0152] Among them, W p is the model parameter after pruning, W o is the initial parameter before pruning.

[0153] Furthermore, the pruning threshold of the model can be proposed according to the actual situation of the model.

[0154] In step S620, the pruned initialization model is determined as the first diagnosis model according to the pruning evaluation parameter and the pruning threshold.

[0155] when , the model pruning process ends, and the pruned initialization model is determined as the first diagnostic model.

[0156] In addition, when When , the initialized model can be pruned again; when This indicates that the client has over-pruned. Return to the last pruning model, adjust the pruning rate, and perform pruning again.

[0157] In step S120, the first diagnostic model is pre-trained to obtain a second diagnostic model, and the second diagnostic model is fine-tuned to obtain a third diagnostic model.

[0158] In an exemplary embodiment of the present disclosure, after the first diagnostic model is determined, training processing can be performed on this basis to compensate for the performance loss that may be caused by the pruning operation and improve the adaptability of the model to local data.

[0159] In an alternative embodiment, Figure 7 A schematic diagram of a process for pre-training a first diagnostic model is shown. Figure 7 As shown, the method may at least include the following steps: In step S710, a data set stored on an electrocardiogram device is acquired, and the data set is used to pre-train a first diagnostic model to obtain a fifth diagnostic model.

[0160] A portion of data is selected or sampled from the local data set of the electrocardiogram device for pre-training, and partial data extraction processing can be performed according to the degree of pruning and model requirements. For example, when the degree of pruning is large, more data can be extracted for pre-training to compensate for performance loss.

[0161] For the first diagnostic model, pre-training hyperparameters such as learning rate, batch size, number of training rounds, etc. can be set, and the first diagnostic model can be pre-trained using the data set. The performance changes of the model during the pre-training process can also be tracked in real time to obtain the fifth diagnostic model.

[0162] In step S720, distillation learning is performed on the fifth diagnostic model to obtain a second diagnostic model.

[0163] After the fifth diagnostic model is obtained, the accuracy and complexity of the model can be verified. When the fifth diagnostic model meets the requirements, the fifth diagnostic model of each electrocardiogram device can be subjected to knowledge distillation learning to obtain a second diagnostic model.

[0164] Furthermore, a loss function corresponding to the fine-tuning of the second diagnostic model of the electrocardiogram device is designed according to formula (14):

[0165]

[0166] Among them, w is the weight parameter of the model, N is the number of samples, and y i is the true label of the i-th sample, is the probability predicted by the model, and λ1 and λ2 are the weights of the regularization term.

[0167] Furthermore, the fine-tuning evaluation index W is calculated according to formula (15):

[0168] W=β1*Acc+β2*Ms (15)

[0169] Among them, Acc is the accuracy of the model, Ms is the size of the model, which can be measured by indicators such as the number of parameters or storage space, and β1 and β2 are weight parameters. The weights λ1 and λ2 of the client regularization term are determined according to the fine-tuning evaluation indicator W.

[0170] When the federated learning is performed for the first time, the input client model structure and parameters are input into each electrocardiogram device; when it is not the first time for the federated learning, the second diagnostic model is fine-tuned according to the fine-tuning evaluation index W to obtain the third diagnostic model.

[0171] In step S130, model training is performed on the third diagnostic model to obtain a fourth diagnostic model, and sparsity analysis is performed on the fourth diagnostic model to obtain target parameters of the fourth diagnostic model.

[0172] In an exemplary embodiment of the present disclosure, after the third diagnostic model is obtained, the third diagnostic model is trained using a local data set of the electrocardiogram device to obtain a trained fourth diagnostic model.

[0173] Since the fourth diagnostic model is obtained by performing pruning operations and other steps on the initialization model, the fourth diagnostic model and the initialization model may be subjected to sparsity analysis.

[0174] Specifically, the structure of the fourth diagnostic model is compared with the structure of the initialization model to determine the model parameters pruned from the fourth diagnostic model compared to the initialization model, and the pruned model parameters are padded with zeros to obtain the sparse structure of the fourth diagnostic model as the target parameters.

[0175] In step S140, the fourth diagnostic model is weighted and aggregated using the target parameters to obtain a target diagnostic model, which is deployed on an electrocardiogram device for diagnosis of cardiovascular diseases.

[0176] In an exemplary embodiment of the present disclosure, a target diagnostic model is obtained by performing weighted aggregation processing on non-zero parameters in the target parameters according to formula (16):

[0177]

[0178] Where N is the total number of ECG device clients, M is the number of model parameters, is the jth parameter of the i-th ECG device client in round t+1, is the parameter weight, n is the total number of samples of N ECG device clients for parameter training, n i is the sample size of ECG device client i.

[0179] When the target diagnosis model meets the requirements or reaches the number of iterations, the target diagnosis model can be deployed on each electrocardiogram device so that each client can obtain the local target diagnosis model to diagnose cardiovascular diseases.

[0180] The following is a detailed description of the training method of the diagnostic model in the embodiment of the present disclosure in conjunction with an application scenario.

[0181] Figure 8 The structural framework diagram of the training method of the diagnostic model in the application scenario is shown, such as Figure 8 As shown, the framework includes a pre-trained central server, a central server, a pruning module, a fine-tuning module, and a client of an ECG device.

[0182] The pre-trained central server and the central server may be the same server or different servers, and this exemplary embodiment does not impose any special limitation on this.

[0183] The pruning module and the fine-tuning module can be deployed on the pre-training central server and / or the central server.

[0184] Fig. 9 A flow chart of the training method of the diagnostic model in the application scenario is shown, such as Fig. 9 As shown, in step S901, the pre-training central server initializes the cardiovascular disease diagnosis model according to actual needs.

[0185] Initialize a cardiovascular disease intelligent diagnosis model according to actual needs. Specifically, the pre-training central server designs the corresponding model structure and initializes the pre-training model parameters.

[0186] In step S902, the initial pre-trained model structure is sent to the pruning module.

[0187] The initialized pre-trained model structure and model parameters are sent to the pruning module.

[0188] In step S903, the pruning module performs pruning operations on the initial pre-trained model according to the actual situation of each client, and evaluates the pruned model of each client according to the evaluation index of the pruning evaluation module, so as to further adjust the pruned model.

[0189] The pruning module performs model pruning to reduce the complexity of the model and improve its performance on resource-constrained devices.

[0190] This module consists of two main units, namely the pruning unit and the pruning evaluation unit. The pruning unit is used to design the degree of pruning of the model according to the storage capacity, computing power, memory capacity, latency requirements, etc. of each client, and remove unimportant parts of the model, such as weights or neurons, to reduce the size and computing requirements of the model, and finally obtain a pruned model suitable for each client; the pruning evaluation unit is mainly used to evaluate the performance of the pruned model to ensure that the pruning operation does not damage the key functions of the model.

[0191] The pruning unit implements the corresponding model pruning algorithm according to the actual situation of different clients and performs pruning operations on the pre-trained model to a corresponding degree.

[0192] Specifically, the input parameters are obtained, including the client's computing power C, memory capacity M, delay requirement L, and maximum delay time L. max , the computing power CC required by the model and the memory MM required by the model, etc.

[0193] Furthermore, the computing power of the client is calculated C is the computing power of the client, and CM is the computing power required for model estimation. Computing power is usually expressed in floating-point operations per second. The higher the CC value, the more computing power the server has to run the model. If CC is low, model pruning is needed to reduce the computing requirements of the model.

[0194] Calculate the client's memory capacity ratio M is the client's memory capacity, and MM is the estimated memory required to run the model. The higher the MC, the more memory the server has to load and run the model; if the MC is low, you need to prune the model to reduce its memory usage.

[0195] Calculate the client's latency ratio L max is the maximum delay time that can be accepted in the actual application of cardiovascular disease diagnosis, L is the estimated delay time of the current model on the server, and PL reflects the user experience and system response time.

[0196] According to actual needs, weight coefficients ω1, ω2, ω3, ω4 are assigned to each factor, so as to calculate the pruning degree P = ω1*(1-T)+ω2*CC+ω3*MC+ω4*PL of the corresponding client.

[0197] Among them, T is the estimated compression rate of the client. In the absence of pruned model size, the compression rate T is used as the target, and the degree of pruning is estimated by 1-T. At the same time, the three influencing parameters of computing power ratio, memory capacity ratio and delay ratio are comprehensively considered, so that the computing and memory resources of the server can be adapted while meeting the real-time requirements.

[0198] According to the corresponding relationship between the pruning degree and the pruning algorithm shown in Table 1, the pruning operation is performed on the client.

[0199] Different degrees of model pruning have a significant impact on adapting to servers with different performance.

[0200] For example, the pruning algorithm corresponding to the pruning degree, lightweight pruning reduces memory usage (MM) and computing requirements (CC), which is suitable for servers with sufficient memory and strong computing power, but has limited improvement on latency.

[0201] Moderate pruning further reduces model complexity, significantly reduces memory (MM) and computing requirements (CC), is suitable for medium-performance servers, and improves latency ratio.

[0202] Deep pruning greatly compresses the model size and is suitable for servers with limited computing power (CC) and memory (MM). It greatly improves the latency ratio and meets real-time requirements.

[0203] By adjusting the pruning strategy, you can optimize the efficiency of the model running on a specific server while maintaining its performance.

[0204] Specifically, the IPFS algorithm innovatively achieves precise feature pruning through adaptive dual thresholds and comprehensive feature importance indicators.

[0205] First, the feature coefficient is derived: determine the influence of each feature on the prediction target, so as to identify the features with less influence on the model prediction. The feature matrix T and target vector x of the model are obtained, and the objective function is proposed as shown in formula (1).

[0206] Then, zero coefficient screening: Screen and remove features that have little or no effect on the predicted target. Traverse the coefficient vector W of the feature, for each coefficient w i , check whether it is zero. If it is zero, remove the corresponding feature T from the feature matrix T i .

[0207] Dynamic threshold setting: Provides a quantitative decision boundary for further feature screening, which can more accurately define which features are key variables in model prediction. Calculate the mean μ of the features with non-zero coefficients non-zero and standard deviation σ non-zero , and the dynamic threshold is obtained by adjusting the hyperparameter k, as shown in formula (2).

[0208] Secondary screening feature identification: Identify those features that are retained after the first pruning but may contribute less to the model. This will provide a basis for the next step of fine pruning. Traverse the coefficient vector W of the feature, for each coefficient w i , will satisfy 0 <w iThe features <θ constitute the secondary features to be screened and enter the next screening process.

[0209] Feature importance evaluation: Evaluate the relationship between the secondary to-be-screened features and the target variable. i , and calculate its feature mutual correlation (MCR) with the target variable x as shown in formula (3).

[0210] Feature fine pruning: For each feature T i , evaluate whether its MCR is lower than the threshold θ MCR Among them, the threshold θ MCR Take the last third of the MCR corresponding to all features. If it is lower than the threshold θ MCR , then remove the corresponding feature T from the feature matrix T i .

[0211] Finally, the core feature matrix is ​​determined to form the final feature set, which is used to establish the prediction model, summarize all the features that have not been pruned, verify the consistency and completeness of the remaining features, and output the final core feature set for subsequent model training and prediction.

[0212] DCOP adds covariance analysis and dynamic dual threshold adjustment to make the pruning process more refined and adaptive.

[0213] Specifically, weight matrix initialization: define the initial weight set W of the model, laying the foundation for the pruning process.

[0214] Covariance structure mapping: for weight pairs (ω i ,ω j ) Calculate the covariance C between weights ij As shown in formula (4), the covariance between all weights is calculated.

[0215] Global importance limit setting: Set a standard to distinguish the importance of weights as a basis for preliminary pruning. Calculate the median C of the covariance between all weight pairs med , average value C mean and variance σ C , the covariance equalization standard value is calculated as shown in formula (5).

[0216] Next, the global importance limit threshold is calculated according to formula (6).

[0217] The local importance limit is set, and further refined within the global importance limit to provide guidance for secondary pruning. The local importance is calculated as shown in formula (7).

[0218] Preliminary fine-pruning: remove weights that show low interdependence in the covariance structure. Go through all covariances and if the weight pair (ω i ,ω j ) ij Less than the global importance limit threshold τ g , then remove the corresponding weight, that is, W ij =0; if τ g <C ij <τ l The corresponding weights constitute a weight set to be reconstructed.

[0219] Weight reconstruction: Reparameterize the weight set to be reconstructed and introduce new learnable parameters to maintain model performance. Reparameterize the weights that have been initially streamlined, as shown in formula (8).

[0220] Dynamic pruning: further prunes the model by removing those weights that are still considered unimportant after parameter reconstruction. Recalculate the new global importance limit threshold based on the reconstructed parameters. If the reconstructed weight parameter is less than the new global importance limit threshold, perform a second weight removal.

[0221] Performance optimization calibration: After pruning, perform a final calibration on the model to ensure that the model performance meets expectations.

[0222] DMOIP integrates dynamic threshold adjustment and multi-dimensional feature evaluation to achieve adaptive and detailed deep learning model pruning.

[0223] Specifically, information dimension baseline mapping: determine the information dimension baseline of each channel feature map, and calculate the information dimension of each passing feature map as shown in formula (9).

[0224] Dynamic adjustment of information dimension: Dynamically adjust the rank threshold according to the statistical characteristics of the information dimension. Calculate the average value of all information dimensions InfoDim mean and standard deviation σ i-d .

[0225] Preliminary information dimension screening: Perform preliminary pruning based on the dynamically adjusted information dimension threshold. Calculate the dynamic adjustment coefficient λ = f(X,F c ,P), where X is the entire model parameter and P is the model performance indicator, so the threshold of the information dimension is obtained as shown in formula (10).

[0226] Traverse all information dimensions and eliminate the channels corresponding to information dimensions that are smaller than the threshold.

[0227] Entropy spectrum construction and analysis: Entropy spectrum analysis is performed on the retained channels. According to the formula Calculate the entropy value H(F) of channel C c ), where p i is the probability of the i-th value appearing in the feature map.

[0228] Gain dynamic evaluation: Evaluate and determine the gain adjustment factor. Calculate the gain adjustment factor according to formula (11).

[0229] Entropy value dynamic balance setting: Set the dynamic balance threshold of entropy value. Calculate the average value H of all entropy values mean and standard deviation σ h The dynamic equilibrium threshold for calculating the entropy value is shown in formula (12).

[0230] Fine entropy balance pruning: Fine pruning is performed based on the dynamic balance threshold of entropy. For each channel c, if its entropy spectrum H(F c ) is less than the dynamic balance threshold of the entropy value, channel c is pruned.

[0231] The model is pruned according to the corresponding pruning algorithm, and the pruning evaluation index Q of the pruning model is calculated based on the weight of the pruning model according to formula (13).

[0232] Propose a model pruning threshold based on the actual situation of the model like Then return to the step of pruning the client and make further pruning adjustments; if The model pruning process ends and the next process begins, starting client pre-training. This indicates that the client has been over-pruned. Return to the last pruned model, adjust the pruning rate, return to the steps of pruning the client, and prune the model again.

[0233] In step S904, the pruned model is sent to each client, and each client creates a local pre-trained model and performs pre-training based on a local part of the cardiovascular data set.

[0234] Based on the pruned model, further pre-training is performed using part of the client's local data set to compensate for the performance loss that may be caused by pruning and improve the model's adaptability to local data.

[0235] The client selects or samples a portion of data from the local dataset for pre-training.

[0236] Depending on the degree of pruning and model requirements, a data selection unit is used to extract part of the data from the local dataset. When the degree of pruning is large, more data may be needed to compensate for the performance loss.

[0237] Then, set the hyper parameters of pre-training, such as learning rate, batch size, number of training rounds, etc. Use the training execution unit to pre-train the model on the client. Use the performance monitoring unit to track the performance changes of the model in real time during the pre-training process.

[0238] In step S905, the pre-training result is sent to the pre-training central server.

[0239] Feedback the pre-training results and model parameters to the central server.

[0240] In step S906, the central server designs the initial model of each client according to the determined model and pre-trained model parameters and aggregates the models of each client.

[0241] The client sends the pre-training results to the pre-training central server, which checks each client model and training results to verify whether the client model meets the requirements, such as verifying the model accuracy, complexity, etc.

[0242] If the model meets the requirements, the model of each client is subjected to knowledge distillation learning to obtain the model structure and initial parameters of each client, and the model structure and initial parameters of the corresponding client are sent to the central server; if the model does not meet the requirements, step S902 is entered and the model pruning module is input to perform further model pruning operations, etc.

[0243] In step S907, the initial training model is input into the fine-tuning module.

[0244] The central server sends the model structure and initial parameters of each client to the fine-tuning module to start federated learning for cardiovascular disease diagnosis.

[0245] In step S908, the fine-tuning module fine-tunes and evaluates the training model of each client according to the actual situation and model results, and further fine-tunes the training model according to the evaluation indicators.

[0246] Fine-tune the model in federated learning for cardiovascular disease diagnosis to improve model performance and adapt to the training of cardiovascular disease diagnosis models for each client.

[0247] In this module, a loss function corresponding to model fine-tuning of each client is designed as shown in formula (14).

[0248] Furthermore, the fine-tuning evaluation index W is calculated according to formula (15).

[0249] When federated learning is performed for the first time, the fine-tuning module directly inputs the client model structure and parameters input by the central server to each client; when it is not the first time for federated learning, a model fine-tuning step is performed.

[0250] In step S909, the training model and parameters are sent to each client, and each client trains the cardiovascular intelligent diagnosis model according to the model structure and the local cardiovascular data set.

[0251] According to the designed model structure and parameters, the local cardiovascular disease intelligent diagnosis model training is started using the local cardiovascular data to obtain the corresponding client model parameters.

[0252] In step S910, the training model parameters are sent to the central server.

[0253] Each client uploads the local model parameters to the corresponding central server.

[0254] Furthermore, central server federated model aggregation: the central server collects the model parameters of each client and proposes an aggregation method suitable for pruning models.

[0255] In this aggregation method, the sparsity analysis of the models uploaded by each client is first performed to obtain the sparse structure of each client. This aggregation process only performs weighted aggregation on the non-zero weights in the model parameter weights. This method can ensure that the aggregation effect accurately reflects the characteristics and structure of the pruned model. The aggregation method is shown in formula (16).

[0256] In step S911, whether the iteration is completed.

[0257] Determine whether the iteration termination condition is met. The termination condition can be that the termination number of iterations has been reached or the model has met the requirements. If the termination condition is met, the central aggregation model parameters are sent to each client, and each client obtains a local intelligent diagnosis model for cardiovascular disease; if the termination condition is not met, return to the previous step and start a new round of federated learning training of the intelligent diagnosis model for cardiovascular disease.

[0258] In an exemplary embodiment of the present disclosure, a model pre-training pruning strategy is proposed, and the model is adaptively pruned through a pre-training process according to the actual situation of each client. The model pruning operation enables the model to adapt to the differences in computing resources of different devices, so that the cardiovascular model can be deployed on a portable electrocardiogram device, and the high-precision cardiovascular intelligent diagnosis model of tertiary hospitals and other hospitals can be adaptively applied to the portable electrocardiogram device, thereby improving the accuracy of the model.

[0259] Then, a pruning strategy for different computing resource capabilities was proposed. Based on the actual situation of the device, different degrees of model pruning were adaptively performed to better adapt to the performance of different devices and improve the execution efficiency of the model on each device; speed up the model training process, while making the model more efficient and improving the generalization ability and accuracy of the model.

[0260] Furthermore, a model fine-tuning strategy for the federated learning process is proposed. By adjusting the model according to actual conditions during the training process, the model can better adapt to the characteristics of local data, generate personalized models, and improve the performance of the model on local devices; help the model continuously learn new knowledge and patterns, achieve incremental learning, and enable the model to continuously adapt to new data and scenarios, keeping the model up to date; through fine-tuning, the model parameters can be further optimized to improve the accuracy and generalization ability of the model.

[0261] Furthermore, a federated aggregation strategy based on model pruning is proposed to ensure that the aggregation effect accurately reflects the characteristics and structure of the pruned model.

[0262] Finally, based on the above strategies, federated learning is improved to more evenly consider the impact of heterogeneity of different client resources and data heterogeneity, thereby improving the generalization and accuracy of the model.

[0263] In addition, in an exemplary embodiment of the present disclosure, a training device for a diagnosis model is also provided. Fig.10 A schematic diagram of the structure of the training device of the diagnostic model is shown, Fig.10 As shown, the diagnostic model training device 1000 may include: a model pruning module 1010, a model fine-tuning module 1020, a sparse analysis model 1030 and a deployment diagnosis module 1040. Among them:

[0264] The model pruning module 1010 is configured to obtain an initialization model for diagnosing cardiovascular diseases and initial parameters of the initialization model, and perform a pruning operation on the initialization model according to the initial parameters to obtain a first diagnostic model;

[0265] A model fine-tuning module 1020 is configured to perform a pre-training process on the first diagnostic model to obtain a second diagnostic model, and to perform a fine-tuning process on the second diagnostic model to obtain a third diagnostic model;

[0266] The sparse analysis model 1030 is configured to perform model training on the third diagnostic model to obtain a fourth diagnostic model, and perform sparsity analysis on the fourth diagnostic model to obtain target parameters of the fourth diagnostic model;

[0267] The deployment diagnosis module 1040 is configured to use the target parameters to perform weighted aggregation processing on the fourth diagnosis model to obtain a target diagnosis model, so as to be deployed on an electrocardiogram device to diagnose cardiovascular diseases.

[0268] In some embodiments of the present disclosure, the model pruning module 1010 is configured to:

[0269] Calculating the initial parameters to obtain performance parameters of the electrocardiogram device, and obtaining parameter weights of the performance parameters;

[0270] Calculating the performance parameters and the parameter weights to obtain pruning parameters of the initialization model, and determining a plurality of pruning algorithms corresponding to the pruning parameters;

[0271] A target algorithm is determined among the multiple pruning algorithms according to the pruning parameters, and a pruning operation is performed on the initialization model according to the target algorithm to obtain a first diagnostic model.

[0272] In some embodiments of the present disclosure, the model pruning module 1010 is configured to:

[0273] Obtaining an influence coefficient corresponding to the initial parameter, and using the influence coefficient to perform an initial screening process on the initialization model to obtain a screened initialization model;

[0274] Calculating the influence coefficient to obtain a coefficient threshold, and determining a parameter to be screened in the initial parameters of the screened initialization model according to the coefficient threshold;

[0275] The coefficient correlation is determined according to the parameters to be screened, and the screened initialization model is subjected to secondary screening processing according to the coefficient correlation to obtain a first diagnostic model.

[0276] In some embodiments of the present disclosure, the model pruning module 1010 is configured to:

[0277] Obtaining an influence coefficient corresponding to the initial parameter, and calculating the influence coefficient to obtain a global standard value and a local standard value;

[0278] Performing preliminary fine-cutting processing on the initialization model according to the global standard value and the local standard value to obtain coefficients to be reconstructed;

[0279] The coefficients to be reconstructed are reconstructed to obtain reconstruction parameters, and the initialized model after preliminary fine-cutting is dynamically fine-cut according to the reconstruction parameters to obtain a first diagnostic model.

[0280] In some embodiments of the present disclosure, the model pruning module 1010 is configured to:

[0281] Acquire the information dimension of the initial parameter and the dimension threshold of the information dimension, and perform preliminary screening processing on the initialization model according to the dimension threshold to obtain a screened initialization model;

[0282] Performing entropy spectrum analysis on the screened initialization model to obtain an entropy value, and determining a gain adjustment factor according to the entropy value;

[0283] The entropy value and the gain adjustment factor are calculated to obtain an entropy value threshold, and a fine pruning operation is performed on the screened initialization model according to the entropy value threshold to obtain a first diagnostic model.

[0284] In some embodiments of the present disclosure, the model pruning module 1010 is configured to:

[0285] Performing a pruning operation on the initialization model according to the initial parameters to obtain a pruned initialization model and a pruning evaluation parameter, and obtaining a pruning threshold corresponding to the pruning evaluation parameter;

[0286] The pruned initialization model is determined as a first diagnostic model according to the pruning evaluation parameter and the pruning threshold.

[0287] In some embodiments of the present disclosure, the model fine-tuning module 1020 is configured to:

[0288] Acquire a data set stored on an electrocardiogram device, and use the data set to pre-train the first diagnostic model to obtain a fifth diagnostic model;

[0289] The fifth diagnostic model is subjected to distillation learning to obtain a second diagnostic model.

[0290] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0291] The present disclosure also provides a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, implement the steps of the diagnostic model training method provided by the present disclosure.

[0292] Fig.11 1 is a block diagram of another training device 1100 for a diagnostic model according to an exemplary embodiment. For example, the device 1100 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0293] Reference Fig.11 , the device 1100 may include one or more of the following components: a processing component 1102 , a memory 1104 , a power component 1106 , a multimedia component 1108 , an audio component 1110 , an input / output interface 1112 , a sensor component 1114 , and a communication component 1116 .

[0294] The processing component 1102 generally controls the overall operation of the device 1100, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 1102 may include one or more processors 1120 to execute instructions to complete all or part of the steps of the above method. In addition, the processing component 1102 may include one or more modules to facilitate the interaction between the processing component 1102 and other components. For example, the processing component 1102 may include a multimedia module to facilitate the interaction between the multimedia component 1108 and the processing component 1102.

[0295] The memory 1104 is configured to store various types of data to support the operation of the device 1100. Examples of such data include instructions for any application or method operating on the device 1100, contact data, phone book data, messages, pictures, videos, etc. The memory 1104 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0296] The power supply component 1106 provides power to the various components of the device 1100. The power supply component 1106 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device 1100.

[0297] The multimedia component 1108 includes a screen that provides an output interface between the device 1100 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundaries of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia component 1108 includes a front camera and / or a rear camera. When the device 1100 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.

[0298] The audio component 1110 is configured to output and / or input audio signals. For example, the audio component 1110 includes a microphone (MIC), and when the device 1100 is in an operating mode, such as a call mode, a recording mode, and a speech recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in the memory 1104 or sent via the communication component 1116. In some embodiments, the audio component 1110 also includes a speaker for outputting audio signals.

[0299] The input / output interface 1112 provides an interface between the processing component 1102 and the peripheral interface modules, which may be keyboards, click wheels, buttons, etc. These buttons may include but are not limited to: a home button, a volume button, a start button, and a lock button.

[0300] The sensor assembly 1114 includes one or more sensors for providing various aspects of the status assessment of the device 1100. For example, the sensor assembly 1114 can detect the open / closed state of the device 1100, the relative positioning of components, such as the display and keypad of the device 1100, the sensor assembly 1114 can also detect the position change of the device 1100 or a component of the device 1100, the presence or absence of user contact with the device 1100, the orientation or acceleration / deceleration of the device 1100, and the temperature change of the device 1100. The sensor assembly 1114 can include a proximity sensor configured to detect the presence of a nearby object without any physical contact. The sensor assembly 1114 can also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 1114 can also include an accelerometer, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0301] The communication component 1116 is configured to facilitate wired or wireless communication between the device 1100 and other devices. The device 1100 can access a wireless network based on a communication standard, such as WiFi, 2G or 3G, or a combination thereof. In an exemplary embodiment, the communication component 1116 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 1116 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0302] In an exemplary embodiment, the device 1100 can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above methods.

[0303] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 1104 including instructions, and the instructions can be executed by the processor 1120 of the device 1100 to perform the above method. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0304] In addition to being an independent electronic device, the above-mentioned device can also be a part of an independent electronic device. For example, in one embodiment, the device can be an integrated circuit (IC) or a chip, wherein the integrated circuit can be an IC or a collection of multiple ICs; the chip can include but is not limited to the following types: GPU (Graphics Processing Unit), CPU (Central Processing Unit), FPGA (Field Programmable Gate Array), DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), SOC (System on Chip, SoC, system on chip or system-level chip), etc. The above-mentioned integrated circuit or chip can be used to execute executable instructions (or codes) to implement the above-mentioned diagnostic model training method. The executable instructions can be stored in the integrated circuit or chip, or can be obtained from other devices or equipment, such as the integrated circuit or chip includes a processor, a memory, and an interface for communicating with other devices. The executable instructions may be stored in the memory, and when the executable instructions are executed by the processor, the above-mentioned diagnostic model training method may be implemented; alternatively, the integrated circuit or chip may receive the executable instructions through the interface and transmit them to the processor for execution, so as to implement the above-mentioned method.

[0305] In another exemplary embodiment, a computer program product is also provided. The computer program product includes a computer program executable by a programmable device, and the computer program has a code portion for executing the above-mentioned training method of the diagnosis model when executed by the programmable device.

[0306] Fig.12 1 is a block diagram of another training device 1200 for a diagnostic model according to an exemplary embodiment. For example, the device 1200 may be provided as a server. Fig.12 The apparatus 1200 includes a processing component 1222, which further includes one or more processors, and a memory resource represented by a memory 1232 for storing instructions executable by the processing component 1222, such as an application. The application stored in the memory 1232 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 1222 is configured to execute instructions to perform the above-mentioned diagnostic model training method.

[0307] The device 1200 may also include a power supply component 1226 configured to perform power management of the device 1200, a wired or wireless network interface 1250 configured to connect the device 1200 to a network, and an input / output interface 1258. The device 1200 may operate based on an operating system stored in the memory 1232.

[0308] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the present disclosure. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0309] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A diagnostic model training method, characterized in that: include: Acquire an initialization model for diagnosing cardiovascular diseases and initial parameters of the initialization model, and perform a pruning operation on the initialization model according to the initial parameters to obtain a first diagnostic model; Pre-training the first diagnostic model to obtain a second diagnostic model, and fine-tuning the second diagnostic model to obtain a third diagnostic model; Performing model training on the third diagnostic model to obtain a fourth diagnostic model, and performing sparsity analysis on the fourth diagnostic model to obtain target parameters of the fourth diagnostic model; The target parameter is used to perform weighted aggregation processing on the fourth diagnostic model to obtain a target diagnostic model, which is deployed on an electrocardiogram device for diagnosis of cardiovascular diseases.

2. The diagnostic model training method according to claim 1, characterized in that: The step of performing a pruning operation on the initialization model according to the initial parameters to obtain a first diagnostic model includes: Calculating the initial parameters to obtain performance parameters of the electrocardiogram device, and obtaining parameter weights of the performance parameters; Calculating the performance parameters and the parameter weights to obtain pruning parameters of the initialization model, and determining a plurality of pruning algorithms corresponding to the pruning parameters; A target algorithm is determined among the multiple pruning algorithms according to the pruning parameters, and a pruning operation is performed on the initialization model according to the target algorithm to obtain a first diagnostic model.

3. The diagnostic model training method according to claim 2, characterized in that: The step of performing a pruning operation on the initialization model according to the target algorithm to obtain a first diagnostic model includes: Obtaining an influence coefficient corresponding to the initial parameter, and using the influence coefficient to perform an initial screening process on the initialization model to obtain a screened initialization model; Calculating the influence coefficient to obtain a coefficient threshold, and determining a parameter to be screened in the initial parameters of the screened initialization model according to the coefficient threshold; The coefficient correlation is determined according to the parameters to be screened, and the screened initialization model is subjected to secondary screening processing according to the coefficient correlation to obtain a first diagnostic model.

4. The diagnostic model training method according to claim 2, characterized in that: The step of performing a pruning operation on the initialization model according to the target algorithm to obtain a first diagnostic model includes: Obtaining an influence coefficient corresponding to the initial parameter, and calculating the influence coefficient to obtain a global standard value and a local standard value; Performing preliminary fine-cutting processing on the initialization model according to the global standard value and the local standard value to obtain coefficients to be reconstructed; The coefficients to be reconstructed are reconstructed to obtain reconstruction parameters, and the initialized model after preliminary fine-cutting is dynamically fine-cut according to the reconstruction parameters to obtain a first diagnostic model.

5. The diagnostic model training method according to claim 2, characterized in that: The step of performing a pruning operation on the initialization model according to the target algorithm to obtain a first diagnostic model includes: Acquire the information dimension of the initial parameter and the dimension threshold of the information dimension, and perform preliminary screening processing on the initialization model according to the dimension threshold to obtain a screened initialization model; Performing entropy spectrum analysis on the screened initialization model to obtain an entropy value, and determining a gain adjustment factor according to the entropy value; The entropy value and the gain adjustment factor are calculated to obtain an entropy value threshold, and a fine pruning operation is performed on the screened initialization model according to the entropy value threshold to obtain a first diagnostic model.

6. The diagnostic model training method according to claim 1, characterized in that: The step of performing a pruning operation on the initialization model according to the initial parameters to obtain a first diagnostic model includes: Performing a pruning operation on the initialization model according to the initial parameters to obtain a pruned initialization model and a pruning evaluation parameter, and obtaining a pruning threshold corresponding to the pruning evaluation parameter; The pruned initialization model is determined as a first diagnostic model according to the pruning evaluation parameter and the pruning threshold.

7. The diagnostic model training method according to claim 1, characterized in that: The pre-training of the first diagnostic model to obtain a second diagnostic model includes: Acquire a data set stored on an electrocardiogram device, and use the data set to pre-train the first diagnostic model to obtain a fifth diagnostic model; The fifth diagnostic model is subjected to distillation learning to obtain a second diagnostic model.

8. A diagnostic model training device, characterized in that: include: A model pruning module is configured to obtain an initialization model for diagnosing cardiovascular diseases and initial parameters of the initialization model, and perform a pruning operation on the initialization model according to the initial parameters to obtain a first diagnostic model; A model fine-tuning module is configured to perform a pre-training process on the first diagnostic model to obtain a second diagnostic model, and to perform a fine-tuning process on the second diagnostic model to obtain a third diagnostic model; a sparse analysis model configured to perform model training on the third diagnostic model to obtain a fourth diagnostic model, and perform sparsity analysis on the fourth diagnostic model to obtain target parameters of the fourth diagnostic model; The deployment diagnosis module is configured to use the target parameters to perform weighted aggregation processing on the fourth diagnosis model to obtain a target diagnosis model, so as to be deployed on an electrocardiogram device for the diagnosis of cardiovascular diseases.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

10. An electronic device, characterized in that: include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the steps of the method according to any one of claims 1 to 7.

Images