Medical data amplification method, training method of sepsis concurrent DIC prediction model and related device

Abstract

The invention discloses a medical data amplification method, a training method of a sepsis concurrent DIC prediction model and a related device. Acquiring a historical clinical index data set; any historical clinical index data sample is extracted from the historical clinical index data set to serve as a target sample, and similar samples are determined for the target sample based on the detection index difference; and randomly generating a new clinical index data sample based on the detection index of the target sample, the respective detection indexes of the similar samples and the diagnosis result index of the target sample. The new samples are regenerated by means of the target samples and the similar samples thereof, and the generated samples have randomness, so that the problems of over-fitting and poor generalization of the trained model caused by few patients with some diseases and insufficient historical data sample stock are solved. The accuracy of model prediction can be improved through richer samples, and the improvement of the prediction performance of the prediction model in the medical scene better assists medical workers in improving the working efficiency and quality.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a medical data amplification method, a training method for a sepsis-complicated DIC prediction model, and related devices. Background Art

[0002] In the medical field, many diseases require laboratory test results and doctors' rich clinical experience to be diagnosed, and the diagnosis cost is high. If the doctor's clinical experience is not rich enough, or there are personal misjudgments, it is easy to cause medical accidents due to misdiagnosis. At present, artificial intelligence technology is expected to assist medical workers in the medical scene and reduce the cost of disease diagnosis. The number of patients with some diseases is small, so the historical data sample inventory is insufficient, and there is a lack of rich samples, making it difficult to train an effective prediction model.

[0003] Sepsis is a systemic inflammatory response caused by infection, which often leads to complications such as disseminated intravascular coagulation (DIC). Early and accurate diagnosis of DIC is crucial to reducing mortality. Training a neural network model for predicting sepsis complicated by DIC is also very important for assisting medical workers in diagnosing DIC. However, due to the small number of DIC patient samples, the model for predicting sepsis complicated by DIC trained on a small number of samples has problems of overfitting, poor generalization, and insufficient prediction accuracy. Combined with the above application scenario examples, it can be seen that too small a sample size is a key issue in training medical prediction models and improving their prediction performance. Summary of the invention

[0004] Based on the above problems, the present application provides a medical data augmentation method, a training method for a sepsis-complicated DIC prediction model and related devices, the purpose of which is to increase the sample size to improve the predictive performance of the medical prediction model.

[0005] The embodiments of the present application disclose the following technical solutions:

[0006] In a first aspect, the present application provides a medical data augmentation method, the method comprising:

[0007] Acquire a historical clinical indicator data set; the historical clinical indicator data set includes a plurality of historical clinical indicator data samples formed by examining a target disease; the historical clinical indicator data samples include a diagnosis result indicator and a plurality of detection indicators, the diagnosis result indicator being used to indicate whether the target disease is diagnosed;

[0008] Extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and determine K similar samples for the target sample based on the difference between the detection index of other historical clinical indicator data samples and the corresponding detection index of the target sample; K is a positive integer greater than 1;

[0009] Based on the detection index of the target sample, the detection index of each of the K similar samples and the diagnosis result index of the target sample, K new clinical indicator data samples are randomly generated; the diagnosis result index of the K new clinical indicator data samples is consistent with the diagnosis result index of the target sample.

[0010] In an optional implementation of the first aspect of the present application, based on the detection index of the target sample, the detection index of each of the K similar samples, and the diagnosis result index of the target sample, K new clinical indicator data samples are randomly generated, including:

[0011] For the kth similar sample among the K similar samples, the difference between the target sample and the kth similar sample is calculated one by one for each detection index to obtain the absolute value of the difference between each detection index; k is an integer from 1 to K;

[0012] Generate a random number in the range of 0 to 1, and multiply the random number by the absolute value to obtain a product;

[0013] The product is added to the corresponding detection index of the target sample, and the resulting addition results are used as multiple detection indexes in the kth new clinical indicator data sample, and the diagnosis result index of the target sample is used as the diagnosis result index of the kth new clinical indicator data sample.

[0014] In an optional implementation of the first aspect of the present application, the multiple detection indicators are specifically N detection indicators, where N is an integer greater than 1; the N detection indicators include one or more initial key indicators;

[0015] Extracting any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and determining K similar samples for the target sample based on the difference between the detection index of other historical clinical indicator data samples and the corresponding detection index of the target sample, including:

[0016] Obtaining similarity calculation weights corresponding to the N detection indicators respectively; wherein the similarity calculation weights corresponding to the initial key indicator are higher than the similarity calculation weights corresponding to the remaining detection indicators;

[0017] Extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and obtain similarity representation values ​​between other historical clinical indicator data samples and the target sample based on differences between detection indicators of other historical clinical indicator data samples and detection indicators corresponding to the target sample, and similarity calculation weights corresponding to the N detection indicators respectively;

[0018] The K historical clinical indicator data samples with the highest similarity represented by the similarity representation values ​​are determined as K similar samples of the target sample.

[0019] In an optional implementation of the first aspect of the present application, before determining K similar samples for the target sample, the method further includes:

[0020] Dividing a plurality of historical clinical indicator data samples in the historical clinical indicator data set into a first training set and a first test set according to a first preset ratio; the target sample and K similar samples of the target sample are both taken from the first training set;

[0021] The medical data augmentation method provided in the first aspect may further include:

[0022] Traversing the first training set until each historical clinical indicator data sample in the first training set is sequentially used as a target sample, and K new clinical indicator data samples are generated multiple times;

[0023] Construct a backup new sample data set based on the K new clinical indicator data samples generated previously;

[0024] A plurality of new clinical indicator data samples are randomly extracted from the standby new sample data set according to a proportional coefficient w, and an available data set is constructed based on the extracted plurality of new clinical indicator data samples and the first training set; w is in the range of 0 to 1;

[0025] Using the available data set to train a prediction model for a target disease, and using the first test set to test the accuracy of the prediction model;

[0026] Dynamically update the w in the range of 0 to 1 according to a preset step size, so as to construct available data sets corresponding to different w for multiple times and obtain the accuracy rates corresponding to different w;

[0027] The w corresponding to the highest accuracy rate is used as the optimal proportional coefficient corresponding to the spare new sample data set, and the available data set constructed according to the optimal proportional coefficient is used as the historical clinical indicator expansion data set.

[0028] A second aspect of the present application provides a training method for a sepsis-complicated DIC prediction model, the method comprising:

[0029] Acquire a historical clinical indicator amplification data set, and determine M feature selection methods; the historical clinical indicator amplification data set is constructed by the medical data amplification method introduced in the first aspect; M is an integer greater than 1; the target disease is sepsis complicated by DIC;

[0030] Using the M feature selection methods respectively, perform feature selection in the historical clinical indicator amplification data set to obtain M groups of detection indicator importance data, wherein one group of detection indicator importance data corresponds to one feature selection method, and one group of detection indicator importance data includes feature importance representation values ​​corresponding to N detection indicators respectively;

[0031] For each set of detection index importance data, the indexes are screened based on the feature selection criteria of the corresponding feature selection method to generate M feature subsets;

[0032] Using the initial optimal subset and the M feature subsets, the detection index of each sample in the historical clinical index amplification data set is screened to form a data set after index screening; the initial optimal subset includes initial key indicators;

[0033] Whether or not DIC was diagnosed was used as the target indicator, and the prediction model for sepsis complicated with DIC was trained using the data set after indicator screening.

[0034] In an optional implementation of the second aspect of the present application, the initial optimal subset and the M feature subsets are used to screen the detection index of each sample in the historical clinical index amplification data set to form an index-screened data set, including:

[0035] Taking the union of each of the M feature subsets and the initial optimal subset as the expansion set corresponding to each of the M feature subsets;

[0036] Taking the intersection of the M feature subsets as the minimum feature subset, and taking the union of the initial optimal subset and the minimum feature subset as the extended set of the minimum feature subset;

[0037] Based on the feature importance representation values ​​corresponding to the respective detection indicators in the M sets of detection indicator importance data, a maximum feature subset is created, and a union of the initial optimal subset and the maximum feature subset is used as an extended set of the maximum feature subset;

[0038] The detection indicators of each sample in the historical clinical indicator amplification data set are screened by using the initial optimal subset, the M feature subsets, the extended sets corresponding to the M feature subsets, the minimum feature subset, the extended set of the minimum feature subset, the maximum feature subset and the extended set of the maximum feature subset to form a data set after indicator screening.

[0039] In an optional implementation of the second aspect of the present application, the initial optimal subset, the M feature subsets, the extended sets corresponding to the M feature subsets, the minimum feature subset, the extended set of the minimum feature subset, the maximum feature subset, and the extended set of the maximum feature subset are used to screen the detection indicators of each sample in the historical clinical indicator amplification data set to form an indicator-screened data set, including:

[0040] Dividing the plurality of samples in the historical clinical indicator amplification data set into a second training set and a second test set according to a second preset ratio;

[0041] Select any one indicator set from the initial optimal subset, the M feature subsets, the extended sets corresponding to the M feature subsets, the minimum feature subset, the extended set of the minimum feature subset, the maximum feature subset, and the extended set of the maximum feature subset;

[0042] For the samples in the second training set and the second test set, multiple detection indexes are extracted according to the selected index sets, and diagnostic result indicators are extracted to form a third training set and a third test set corresponding to each selected index set.

[0043] In an optional implementation of the second aspect of the present application, whether or not DIC is diagnosed is used as a target indicator, and a prediction model for sepsis complicated with DIC is obtained by training a data set after indicator screening, including:

[0044] Whether or not DIC was diagnosed was taken as the target indicator, and the third training set and the third test set corresponding to each selected indicator set were used for model training and testing respectively. The indicator set corresponding to the model with the highest test evaluation score was taken as the optimal feature subset, and the model was used as a prediction model for sepsis complicated by DIC.

[0045] In an optional implementation of the second aspect of the present application, creating a maximum feature subset based on a feature importance representation value corresponding to each detection indicator in the M sets of detection indicator importance data in the M feature subsets includes:

[0046] For the sth feature subset among the M feature subsets, based on the feature importance representation value corresponding to each detection indicator in the sth feature subset in the sth group of detection indicator importance data, the standardized weight of each detection indicator in the sth feature subset in the sth feature subset is calculated; s takes a value between 1 and M, and s is an integer;

[0047] The average value of the standardized weights of the same detection indicator in different feature subsets is calculated to obtain the average value of the standardized weights corresponding to each detection indicator;

[0048] According to the standardized weight average values ​​corresponding to each detection index and the preset feature selection criteria, feature screening is performed on each detection index in the M feature subsets, and the screened detection indexes are used to construct the largest feature subset.

[0049] The third aspect of the present application provides a medical data amplification device, the device comprising:

[0050] A historical data set acquisition module, used to acquire a historical clinical indicator data set; the historical clinical indicator data set includes a plurality of historical clinical indicator data samples formed by examining a target disease; the historical clinical indicator data samples include a diagnosis result indicator and a plurality of detection indicators, the diagnosis result indicator is used to indicate whether the target disease is diagnosed;

[0051] A similar sample determination module, used to extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and determine K similar samples for the target sample based on the difference between the detection index of other historical clinical indicator data samples and the corresponding detection index of the target sample; K is a positive integer greater than 1;

[0052] A new sample generation module is used to randomly generate K new clinical indicator data samples based on the detection index of the target sample, the detection index of each of the K similar samples and the diagnosis result index of the target sample; the diagnosis result index of the K new clinical indicator data samples is consistent with the diagnosis result index of the target sample.

[0053] In a fourth aspect, the present application provides a training device for a sepsis-complicated DIC prediction model, the device comprising:

[0054] An augmented data set acquisition module, used to acquire an augmented data set of historical clinical indicators and determine M feature selection methods; the augmented data set of historical clinical indicators is constructed by the medical data augmentation method introduced in the first aspect; M is an integer greater than 1;

[0055] A feature selection module, used to respectively use the M feature selection methods to perform feature selection in the historical clinical indicator amplification data set to obtain M groups of detection indicator importance data, wherein a group of detection indicator importance data corresponds to a feature selection method, and a group of detection indicator importance data includes feature importance representation values ​​corresponding to N detection indicators respectively;

[0056] An indicator screening module is used to screen indicators for each set of detection indicator importance data based on the feature selection criteria of the corresponding feature selection method to generate M feature subsets;

[0057] A data set screening module, used to screen the detection index of each sample in the historical clinical index amplification data set using the initial optimal subset and the M feature subsets to form a data set after index screening; the initial optimal subset includes initial key indicators;

[0058] The model training module is used to take whether the patient is diagnosed with DIC as the target indicator and train the sepsis-complicated DIC prediction model using the data set after indicator screening.

[0059] A fifth aspect of the present application provides a device for predicting sepsis complicated with DIC, the device comprising:

[0060] An indicator data acquisition module, used to acquire the indicator data of the target object according to the detection indicators in the indicator-screened data set; the indicator-screened data set is formed after screening according to the training method of the sepsis complicated by DIC prediction model introduced in any implementation of the second aspect;

[0061] The DIC prediction module is used to call the sepsis complicated by DIC prediction model to predict the probability that the target object suffers from sepsis complicated by DIC based on the indicator data of the target object; the sepsis complicated by DIC prediction model is obtained after training based on the training method of the sepsis complicated by DIC prediction model introduced in any implementation method of the second aspect.

[0062] In a sixth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store program code, wherein the program code is used to execute the steps of the medical data augmentation method as described in any implementation of the first aspect, or the program code is used to execute the steps of the training method for the sepsis complicated by DIC prediction model as described in any implementation of the second aspect.

[0063] In a seventh aspect, the present application provides a processor, which is used to run a computer program. When the computer program is running, the steps of the medical data augmentation method introduced in any implementation of the first aspect are executed, or when the computer program is running, the steps of the training method of the sepsis complicated by DIC prediction model introduced in any implementation of the second aspect are executed.

[0064] Compared with the prior art, this application has the following beneficial effects:

[0065] The present application discloses a medical data amplification method, a training method for a sepsis-complicated DIC prediction model, and related devices. Obtain a historical clinical indicator data set; extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and determine a similar sample for the target sample based on the difference in detection indicators; randomly generate new clinical indicator data samples based on the detection indicators of the target sample, the detection indicators of its similar samples, and the diagnostic result indicators of the target sample. In the present application, new samples are regenerated with the help of the target sample and its similar samples. The generated samples are random, which solves the problem of overfitting and poor generalization of the trained model due to the small number of patients with some diseases and insufficient stock of historical data samples. The accuracy of model prediction can be improved through more abundant samples, and the improvement of the prediction performance of the prediction model in medical scenarios can better assist medical workers in improving work efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0067] Figure 1 A flowchart of a medical data augmentation method provided in an embodiment of the present application;

[0068] Figure 2 This is an example implementation flowchart of determining similar samples in an embodiment of the present application;

[0069] Figure 3 A flowchart of another medical data augmentation method provided in an embodiment of the present application;

[0070] Figure 4 A flowchart of a method for training a sepsis-complicated DIC prediction model provided in an embodiment of the present application;

[0071] Figure 5 A flowchart of a method of filtering detection indicators by constructing a feature subset and forming a filtered data set provided in an embodiment of the present application;

[0072] Figure 6 A schematic diagram of the structure of a medical data amplification device provided in an embodiment of the present application;

[0073] Figure 7 A schematic diagram of the structure of a training device for a sepsis-complicated DIC prediction model provided in an embodiment of the present application;

[0074] Figure 8A schematic diagram of the structure of a device for predicting sepsis complicated by DIC provided in an embodiment of the present application. DETAILED DESCRIPTION

[0075] The use of artificial intelligence technology in the medical field can assist medical workers in their work. For example, training a prediction model through artificial intelligence technology can shorten the time spent on manual analysis of medical data and correct errors or loopholes in human judgment. If the model has a sufficient number of samples, the model can be trained and tested with sufficient data, and the trained model can also have more accurate prediction performance; if the model lacks sample data, the trained model is likely to have problems of overfitting and poor generalization performance, which will lead to the model's prediction performance failing to meet expectations.

[0076] To solve the above problems, the present application discloses a medical data amplification method, a training method for a sepsis-complicated DIC prediction model, and related devices. New clinical indicator data samples are randomly generated based on the detection indicators of the target sample in the historical clinical indicator data set, the detection indicators of the similar samples determined for the target sample, and the diagnostic result indicators of the target sample. New samples with randomness are regenerated using the target sample and its similar samples, which solves the problem of overfitting and poor generalization of the trained model due to the small number of patients with some diseases and insufficient stock of historical data samples. The accuracy of model prediction can be improved through more abundant samples, and the improvement of the prediction performance of the prediction model in medical scenarios can better assist medical workers in improving work efficiency and quality.

[0077] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0078] See also Figure 1 , which is a flow chart of a medical data augmentation method provided in an embodiment of the present application. Figure 1 ,Medical data augmentation methods include:

[0079] S101. Obtain a historical clinical indicator dataset.

[0080] The technical solution of the present application is mainly aimed at expanding the number of samples for diseases with a small number of samples, so that these newly expanded samples can serve to implement the training of prediction models in medical scenarios. In the present application, in order to achieve the expansion of sample data, a historical clinical indicator data set can be first obtained in this step. It should be noted that the historical clinical indicator data set here is oriented to a certain disease. For the sake of convenience, the disease corresponding to the historical clinical indicator data set is defined as the target disease in the technical solution of the present application. The target disease can be any disease with a shortage of samples. As an example, the target disease can be sepsis complicated by DIC introduced above.

[0081] The historical clinical indicator data set includes multiple historical clinical indicator data samples formed by examining the target disease. In the embodiment of the present application, each historical clinical indicator data sample in the historical clinical indicator data set includes a diagnosis result indicator and multiple detection indicators. Among them, the diagnosis result indicator is used to indicate whether it is diagnosed as the target disease. In some possible implementations, the diagnosis result indicator is collected from the patient's clinical diagnosis record. The detection indicator can be medical data formed by examinations, tests, etc. during the patient's visit to the relevant department of the hospital or hospitalization. In practical applications, each historical clinical indicator data sample in the historical clinical indicator data set corresponds to a patient.

[0082] Taking sepsis complicated with DIC as an example, the diagnostic result indicator in the historical clinical indicator data sample indicates whether the patient suffers from sepsis complicated with DIC. For example, a diagnostic result indicator of 1 indicates that the patient suffers from sepsis complicated with DIC, and a diagnostic result indicator of 0 indicates that the patient does not suffer from sepsis complicated with DIC.

[0083] Taking sepsis complicated by DIC as an example, the detection indicators in the historical clinical indicator data sample may include but are not limited to: platelets, antithrombin III, prothrombin time, international normalized ratio, activated partial thromboplastin time, fibrinogen, thrombin time, D-dimer, fibrin degradation products, absolute white blood cell count, absolute lymphocyte count, absolute neutrophil count, absolute monocyte count, lymphocyte percentage, neutrophil percentage, monocyte percentage, CD4+T cells, absolute CD4+T cell count, CD8+T cells, absolute CD8+T cell count, CD4 / CD8 ratio, CD3+T cells, absolute CD3+T cell count, natural killer cell percentage, absolute natural killer cell count, B cell percentage, absolute B cell count, monocyte HLA-DR, regulatory T cells, neutrophil CD64, procalcitonin, C-reactive protein, interleukin-6, interleukin-10 , dendritic cells, plasmacytoid dendritic cells, myeloid dendritic cells, CD3+CD14+ T cells, CD56bright natural killer cells, CD56dim natural killer cells, CD56-CD16+ natural killer cells, non-classical monocytes, intermediate monocytes, classical monocytes, natural killer T cells, follicular helper T cells, follicular helper T cells, follicular helper T cells, follicular helper T cells, follicular helper T cells, Th1 helper T cells, Th1 / Th17 ratio, Th2 helper T cells, Th17 helper T cells, Tc1 cells, Tc1 / Tc17 ratio, Tc2 cells, -Tc17 cells, central memory T cells, naive T cells, effector memory T cells, effector memory RA+ T cells, effector T cells, regulatory T cells, central memory Tc cells, naive Tc cells, effector memory Tc cells, effector memory RA+ Tc cells, effector Tc cells, regulatory Tc cells.

[0084] It is understandable that in the specific implementation of the present application, historical clinical indicator data samples are involved. When the above embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0085] S102, extracting any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and determining K similar samples for the target sample based on the differences between the detection indicators of other historical clinical indicator data samples and the corresponding detection indicators of the target sample.

[0086] As mentioned above, the number of samples of the target disease may be in short supply. For example, the historical clinical indicator data set contains 70 historical clinical indicator data samples. Obviously, the sample size of this order of magnitude is difficult to support the training of a prediction model for the target disease with good prediction performance. To this end, the technical solution of the present application needs to perform data amplification on the basis of the historical clinical indicator data set. In this step, multiple similar samples are first determined for each sample in the historical clinical indicator data set. For ease of explanation, the following takes a target sample as an example to introduce an example implementation process for determining K similar samples for it. K is a positive integer greater than 1. As an example, K=20, K=30, etc. The value of K can be determined according to actual needs, and the number of K is less than or equal to the number of historical clinical indicator data samples in the historical clinical indicator data set. Any historical clinical indicator data sample in the historical clinical indicator data set can be used as a target sample. In other words, similar samples of any historical clinical indicator data sample can be determined in the following manner.

[0087] In an optional implementation, the multiple detection indicators are specifically N detection indicators, where N is an integer greater than 1; the N detection indicators include one or more initial key indicators. Taking the target disease as sepsis complicated with DIC as an example, the N indicators may include the platelets, antithrombin III, prothrombin time, international normalized ratio, activated partial thromboplastin time, fibrinogen and other detection indicators introduced above. As an example, there are two initial key indicators, namely platelets and procalcitonin.

[0088] It should be noted that the use of platelets and procalcitonin as the initial key indicators for sepsis complicated by DIC is only an example of the initial key indicators selected for the disease of sepsis complicated by DIC. In the literature on the disease of sepsis complicated by DIC, other detection indicators may also be identified as the initial key indicators of sepsis complicated by DIC, or other numbers of initial key indicators (such as 3 or 4) may be selected, which is not limited here.

[0089] Figure 2 An example implementation process of determining similar samples in an embodiment of the present application is shown. Figure 2 As shown, in a possible implementation, any historical clinical indicator data sample is extracted from the historical clinical indicator data set as a target sample, and K similar samples are determined for the target sample based on the difference between the detection index of other historical clinical indicator data samples and the corresponding detection index of the target sample, including:

[0090] S1021. Obtain similarity calculation weights corresponding to N detection indicators.

[0091] In an embodiment of the present application, when performing similarity calculation between samples, each detection index is set with a weight for similarity calculation. Since the initial key index is selected based on the clinical experience of the target disease, the important effect of the initial key index on the early warning of the target disease is understood through long-term clinical practice experience. Therefore, in this solution, the similarity calculation weight corresponding to the initial key index is set, which is higher than the similarity calculation weight corresponding to the remaining detection indicators. For example, the similarity calculation weight of other indicators other than the initial key index in the N detection indicators is 1, and the similarity calculation weight of the initial key index is 1.5. It should be noted that if the number of initial key indicators contained in the N detection indicators is greater than 1, that is, there are multiple different initial key indicators, then the similarity calculation weights of different initial key indicators can be the same or different. It can be set specifically according to actual needs.

[0092] S1022. Extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and obtain similarity representation values ​​between other historical clinical indicator data samples and the target sample based on the differences between the detection indicators of other historical clinical indicator data samples and the corresponding detection indicators of the target sample, and the similarity calculation weights corresponding to the N detection indicators.

[0093] The following formula shows how to obtain the similarity representation value:

[0094]

[0095] In this formula, d m is the similarity characterization value, x represents the detection index, where the superscript of x is used to distinguish different samples to which the detection index belongs. For example, a superscript of 0 represents a target sample, and a superscript of m represents a sample of other historical clinical index data. The subscript of x is used to distinguish different detection indicators. As can be seen from the formula, when calculating the similarity characterization value, the numerical values ​​of the same detection index of different samples are subtracted and then squared to measure the difference (or similarity) between the sample with the superscript m and the target sample on the detection index. It can be understood that the larger the square value, the greater the difference and the lower the similarity; conversely, if the square value is smaller, the smaller the difference and the higher the similarity. Combined with the above formula, the similarity characterization value d m The larger the value is, the greater the difference between some other historical clinical indicator data sample and the target sample is, and accordingly, the lower the similarity between the two samples is.

[0096] In the above formula, the parameter before the bracket represents the similarity calculation weight. In the example of this formula, the subscripts x and n represent two initial key indicators. Correspondingly, the similarity calculation weights of these two indicators are w and n respectively. 1 and w 2In this example, the similarity calculation weights of other detection indicators are set to 1 by default, while w 1 and w 2 is a value greater than 1, w 1 and w 2 The specific values ​​of can be the same or different.

[0097] S1023. Determine the K historical clinical indicator data samples with the highest similarity represented by the similarity representation values ​​as K similar samples of the target sample.

[0098] By traversing each non-target sample in the historical clinical indicator data set and calculating the similarity between it and the target sample, the similarity characterization value between each historical clinical indicator data sample and the target sample can be obtained. Thus, one or more samples that are most similar to the target sample can be analyzed and determined as similar samples of the target sample. For example, there are 80 samples in the historical clinical indicator data set. The first sample is used as the target sample. The similarity characterization values ​​of the remaining 79 samples and the target sample are calculated. The smaller the similarity characterization value, the higher the similarity between the two samples. Take the K samples with the smallest similarity characterization value as the K similar samples of the target sample.

[0099] Similarly, based on the above implementation, each historical clinical indicator data sample in the historical clinical indicator data set is traversed and taken as the target sample in turn, so that K similar samples corresponding to each historical clinical indicator data sample can be determined.

[0100] S103. Randomly generate K new clinical indicator data samples based on the detection index of the target sample, the detection index of each of the K similar samples, and the diagnosis result index of the target sample.

[0101] The diagnostic result indicators of the K new clinical indicator data samples are consistent with the diagnostic result indicators of the target samples.

[0102] In an optional implementation, based on the detection index of the target sample, the detection index of each of the K similar samples, and the diagnostic result index of the target sample, K new clinical index data samples are randomly generated, including:

[0103] For the kth similar sample among the K similar samples, the difference between the target sample and the kth similar sample is calculated one by one for the detection index, and the absolute value of the difference of each detection index is obtained; wherein k is an integer from 1 to K, that is, the kth similar sample can be any similar sample among the K similar samples of the target sample. Next, a random number in the range of 0 to 1 is generated, and the random number is multiplied by the absolute value to obtain the product. The random number can be regarded as a random coefficient of the absolute value, so that the calculated product has a certain randomness. The product is added to the corresponding detection index of the target sample, and the various addition calculation results are used as multiple detection indicators in the kth new clinical indicator data sample, and the diagnostic result index of the target sample is used as the diagnostic result index of the kth new clinical indicator data sample.

[0104] It can be understood that, since the diagnosis result index of the kth new clinical indicator data sample depends on the diagnosis result index of the target sample, the diagnosis result indexes of the K new clinical indicator data samples of the target sample are consistent with the diagnosis result index of the target sample.

[0105] The implementation method of randomly generating K new clinical indicator data samples introduced above can be understood in combination with the following formula:

[0106] x 0,k =x 0 +rand(0,1)*|x 0 -x k

[0107] In this formula, x 0 Data representing the detection index of the target sample, x 0 The data representing the detection index of the kth similar sample of the target sample, rand(0,1) represents the generation of a random number in the range of 0 to 1. 0,k Represents the data of the detection indicators in the kth new clinical indicator data sample generated based on the data of the detection indicators of the target sample and the kth similar sample to the target sample.

[0108] Since the above formula is used to realize the construction of the detection indicators in the new clinical indicator data sample, and the diagnostic result indicators of the new clinical indicator data sample can be determined based on the target sample, it is equivalent to reconstructing a new clinical indicator data sample with complete data for the detection indicators and complete data for the diagnostic result indicators. Since the target sample has K similar samples, it is equivalent to reconstructing K new clinical indicator data samples. In addition, since the historical clinical indicator data set itself contains several historical clinical indicator data samples, the historical clinical indicator data samples therein are traversed in sequence as target samples, that is, K new clinical indicator data samples can be created based on each historical clinical indicator data sample and its K similar samples.

[0109] In the above embodiments, the implementation process of the medical data amplification method is introduced. First, a historical clinical indicator data set is obtained; then, any historical clinical indicator data sample is extracted from the historical clinical indicator data set as a target sample, and a similar sample is determined for the target sample based on the difference in detection indicators; then, based on the detection indicators of the target sample, the detection indicators of its similar samples, and the diagnostic result indicators of the target sample, a new clinical indicator data sample is randomly generated. In this application, new samples are regenerated with the help of target samples and their similar samples. The generated samples are random, which solves the problem of overfitting and poor generalization of the trained model due to the small number of patients with some diseases and insufficient stock of historical data samples. The accuracy of model prediction can be improved through more abundant samples, and the improvement of the prediction performance of the prediction model in medical scenarios can better assist medical workers in improving work efficiency and quality.

[0110] In practical applications, in order to better solve the problem of overfitting in model training, the inventor proposes that, on the basis of solving the problem of insufficient sample data, the model data set can be optimized through dynamic control. Figure 3 The process shown is further explained. Figure 3 FIG. 1 is a flow chart of another medical data augmentation method provided in an embodiment of the present application. Figure 3 The medical data augmentation method shown comprises:

[0111] S301. Obtain a historical clinical indicator data set, and divide a plurality of historical clinical indicator data samples in the historical clinical indicator data set into a first training set and a first test set according to a first preset ratio.

[0112] As an example, in the historical clinical indicator dataset D 0 Among the multiple historical clinical indicator data samples, 20% are randomly extracted as the first test set D 0_test , the remaining 80% is used as the first training set D 0_train .

[0113] S302: Obtain similarity calculation weights corresponding to N detection indicators.

[0114] S303. Extract any historical clinical indicator data sample from the first training set as a target sample, and obtain similarity representation values ​​between other historical clinical indicator data samples and the target sample based on the differences between the detection indicators of other historical clinical indicator data samples in the first training set and the corresponding detection indicators of the target sample, and the similarity calculation weights corresponding to the N detection indicators.

[0115] S304: Determine the K historical clinical indicator data samples in the first training set that have the highest similarity represented by the similarity representation value with the target sample as K similar samples of the target sample.

[0116] The implementation of S302 to S304 is basically the same as that of S1021 to S1023, and will not be described here. The corresponding implementation details can refer to the introduction in the above embodiment. Different from the above embodiment, in this embodiment, it is proposed that the target sample and similar samples of the target sample are selected from the first training set.

[0117] S305. Randomly generate K new clinical indicator data samples based on the detection index of the target sample, the detection index of each of the K similar samples, and the diagnosis result index of the target sample.

[0118] S306, determine whether the traversal of the first training set is completed, if yes, proceed to S307; if not, return to S303.

[0119] When the traversal of the first training set is completed, it means that each historical clinical indicator data sample in the first training set has been regarded as a target sample, and the corresponding K new clinical indicator data samples have been generated. That is, the medical data expansion process based on the first training set is completed.

[0120] Overfitting means that the data works very well on the training set but very poorly on the test set. A large amount of data can effectively reduce overfitting. Therefore, the technical solution of the present application chooses data augmentation. In addition, the embodiment of the present application proposes to use interpolation method for augmentation. Please refer to the implementation process and formula of randomly generating K new clinical indicator data samples introduced above for details. What is obtained by interpolation amplification is "false data" (non-real data). If there is too much false data, it will affect the performance of the model. This is because the test set is all real data, and if there is too much false data in the training set, it will affect the prediction of real data, so there may be an overfitting problem with good effect on the training set but poor effect on the test set. The technical solution of the present application proposes to use dynamic control to find the optimal proportion of amplified data to reduce the overfitting phenomenon. See steps S307 to S311 of the dynamic control below for details.

[0121] S307: Construct a backup new sample data set based on the K new clinical indicator data samples generated previously.

[0122] It is understandable that before entering S307, the process of using each historical clinical indicator data sample in the first training set as the target sample in turn to generate K new clinical indicator data samples multiple times has been completed. On this basis, a large number of new clinical indicator data samples formed by different historical clinical indicator data samples as target samples can be collected to form a new data set, which is referred to as a spare new sample data set in the technical solution of this application. It is understandable that each sample in the spare new sample data set is created by random interpolation based on the target sample and similar samples. Therefore, the spare new sample data set can also be regarded as a false sample data set.

[0123] The embodiment of the present application realizes dynamic control of the number of false samples used for model training through the proportional coefficient w. The following is an explanation in conjunction with S308 to S310.

[0124] S308. Randomly extract a plurality of new clinical indicator data samples from the spare new sample data set according to the proportion coefficient w, and construct an available data set based on the extracted plurality of new clinical indicator data samples and the first training set.

[0125] In an embodiment of the present application, a proportional coefficient w is set, and a plurality of new clinical indicator data samples are randomly extracted from the spare new sample data set according to the proportional coefficient w. For example, if the proportional coefficient w=0.2, it means that 20% of the new clinical indicator data samples are extracted from the spare new sample data set. If the spare new sample data set contains 3000 new clinical indicator data samples, according to the proportional coefficient w=0.2, it means that 600 new clinical indicator data samples are randomly eliminated. The extracted 600 new clinical indicator data samples are mixed with the first training set introduced above, the real samples and the false samples to construct an available data set. The samples in the available data set are used to train the model. If the proportion of false samples mixed in is too high, it may also cause the trained model to overfit and perform poorly on the real data. For this reason, w takes values ​​in the range of 0 to 1. Since w takes different values, the performance of the trained model on the real data will be different accordingly. Therefore, the execution purpose of the technical solutions S308 to S310 of the present application is to determine a suitable proportional coefficient in the range of 0 to 1. Thus, it is possible to enrich the samples and avoid overfitting of the model. In specific implementation, the step size of w can be set, and different w values ​​can be simulated from the initial value of w to build a usable data set, and then the model can be trained, and the trained model can be tested with the first test set to determine the prediction performance of the model trained with different w values.

[0126] The composition of the available dataset can be expressed as: new =D 0_train +wD g Among them, D 0_trainrepresents the first training set, w represents the proportional coefficient, D g Denotes the spare new sample data set, D new Represents the available datasets for the build.

[0127] S309: Use the available data set to train a prediction model for the target disease, and use the first test set to test the accuracy of the prediction model.

[0128] Since the historical clinical indicator data samples included in the first test set are real samples, this step actually trains the model through the available data set mixed with real and fake data, and uses real samples to test the accuracy of the model. As an example, through this step, a corresponding relationship between the w value and the accuracy of the tested prediction model can be established.

[0129] S310 , dynamically update w in the range of 0 to 1 according to a preset step size, so as to construct available data sets corresponding to different w for multiple times, and obtain accuracy rates corresponding to different w.

[0130] As an example, the proportional coefficient w increases or decreases gradually according to a preset step size. For example, if the initial value of the proportional coefficient w is set close to 0, it can be gradually increased according to the preset step size; if the initial value of the proportional coefficient w is set close to 1, it can be gradually decreased according to the preset step size. As an example, the initial value of the proportional coefficient w is 0.1, and the step size is set to 0.05. When the proportional coefficient w=0.1, the accuracy corresponding to 0.1 is obtained. When w is dynamically increased to 0.15, the accuracy corresponding to 0.15 can be obtained. When w is dynamically increased by 0.2, the accuracy corresponding to 0.2 can be obtained. The end condition of the dynamic change of the proportional coefficient w is: w exceeds the range of 0 to 1.

[0131] If w is updated incrementally, it can be expressed as w=w+s; if w is updated decrementally, it can be expressed as w=ws. In this expression, the w on the right side of the equal sign is the w before the update, and the w on the left side of the equal sign is the w after the update.

[0132] S311. Use w corresponding to the highest accuracy rate as the optimal proportional coefficient corresponding to the backup new sample data set, and use the available data set constructed according to the optimal proportional coefficient as the historical clinical indicator expansion data set.

[0133] Since the accuracy of the trained model corresponding to different w values ​​has been obtained through the above steps, the highest accuracy can be determined by analysis, and the value of the proportional coefficient w corresponding to the model with the highest accuracy is taken as the optimal proportional coefficient. This means that when w is taken as the optimal proportional coefficient, the prediction performance (accuracy) of the trained model reaches its peak.

[0134] For example, the optimal ratio coefficient is finally determined to be 0.35, which means that 35% of false samples are randomly extracted from the spare new sample data set and mixed with 100% of the real samples of the first training set. The resulting usable data set can be used to train a prediction model with better prediction performance for the target disease. The usable data set constructed by this optimal ratio coefficient is used as the historical clinical indicator expansion data set.

[0135] In the embodiment of the present application, the ratio coefficient of randomly extracting new clinical indicator data samples from the spare new sample data set is dynamically controlled, and then the optimal ratio coefficient is determined, thereby optimizing the composition of the historical clinical indicator expansion data set. Moreover, through the historical clinical indicator expansion data set, the prediction performance of the trained model can be improved, and a more accurate prediction utility can be exerted in the medical field.

[0136] Based on the medical data amplification method introduced in the above embodiment, the present application also proposes a training method for a sepsis-complicated DIC prediction model. The implementation premise of this method is to construct a historical clinical indicator amplification data set based on the method provided in the above embodiment. The historical clinical indicator amplification data set is a data set determined based on a dynamically adjusted proportional coefficient. This data set not only solves the problem of insufficient historical sample inventory, but also has been tested by the first test set, and can help train a prediction model with a higher prediction accuracy. Therefore, the sepsis-complicated DIC model is trained based on this data set. The following is combined with Figure 4 The implementation process of the training method for the sepsis-complicated DIC prediction model is explained.

[0137] like Figure 4 As shown in FIG. 1 , the training method of the sepsis complicated with DIC prediction model includes:

[0138] S401. Obtain a historical clinical indicator expansion data set and determine M feature selection methods.

[0139] The historical clinical indicator augmentation dataset is constructed using the medical data augmentation method introduced in the first aspect.

[0140] In the medical data amplification method described above, a series of indicators can be extracted from the historical data of sepsis complicated by DIC. Some of these indicators, based on clinical experience, can be used to predict sepsis complicated by DIC. For example, the initial key indicators of platelets and procalcitonin are selected in this scheme. For the remaining other indicators, it is impossible to know whether they play a role in predicting whether sepsis is complicated by DIC based on clinical experience alone. If a prediction model is constructed using an indicator with a smaller effect, this indicator will become noise and reduce the accuracy of the model. Therefore, in the embodiment of the present application, it is necessary to further screen many detection indicators, narrow the range of indicators, and more effectively improve the accuracy of the trained prediction model. In this regard, the embodiment of the present application proposes to determine M kinds of feature selection methods to assist in the screening of detection indicators. M is an integer greater than 1. As an example, M=2. The use of multiple feature selection methods can take into account a variety of situations more comprehensively, avoiding the lack of rigor and accuracy of indicator screening due to the defects of a single feature selection method.

[0141] In a possible implementation, the two feature selection methods are Lasso regression algorithm and Boruta algorithm. The characteristics of these two feature selection methods are described below.

[0142] (1) Lasso regression algorithm (abbreviated as: Lasso algorithm): It has the advantages of high computational efficiency, strong feature selection ability and suitability for linear models.

[0143] The Lasso algorithm can achieve fast model training and feature selection by adding L1 regularization terms, and is particularly suitable for large-scale data sets. By compressing the weights of unimportant features to 0, feature selection can be achieved, which can effectively reduce the complexity of the model and improve the interpretability and predictive ability of the model. The Lasso algorithm is suitable for linear regression models, can process linear relationship data, and performs well when processing data with a large number of features.

[0144] (2) Boruta algorithm: It has high stability, is suitable for high-dimensional data, and is based on random forests.

[0145] The Boruta algorithm can effectively evaluate the importance of features and reduce the risk of overfitting through multiple self-service resampling and feature shuffling, thereby improving the stability and generalization ability of the model. When processing high-dimensional data, the Boruta algorithm can effectively screen out important features and remove irrelevant features, thereby simplifying the model and improving computational efficiency. The Boruta algorithm is based on the random forest model and can capture the interactions between features. It is suitable for data sets with complex feature interactions.

[0146] In general, the Boruta algorithm is suitable for situations where there is a complex relationship between features and prediction targets, while the Lasso algorithm is more suitable for simple relationships. However, there is no clear definition in the field as to whether the relationship between features and prediction targets is simple or complex. Therefore, in order to better cope with various situations encountered in actual use, this solution can use the various subsets obtained by the two methods to test them separately and select the best one.

[0147] S402. Use M feature selection methods respectively to perform feature selection in the historical clinical indicator amplification data set to obtain M groups of detection indicator importance data.

[0148] In the example of M=2, the Lasso algorithm is used to select features in the historical clinical index amplification data set to obtain the importance data of the first group of detection indicators, which is expressed as The Boruta algorithm is also used to perform feature selection in the historical clinical index amplification data set to obtain the importance data of the second group of detection indicators, which is expressed as

[0149] exist and In the above example, the same subscript represents the same detection index, and the superscripts 1 and 2 correspond to different feature selection methods. That is, each set of detection index importance data corresponds to a feature selection method. A set of detection index importance data includes the feature importance representation values ​​corresponding to each detection index in the N detection indicators. In the above example, N=n+1.

[0150] As an example, the subscripts 0…n represent {platelets, antithrombin III, prothrombin time, international normalized ratio, neutrophil CD64, …, C-reactive protein, effector Tc cells, regulatory Tc cells} respectively. It should be noted that the N detection indicators here include the two initial key indicators of "platelets" and "procalcitonin".

[0151] The feature importance representation value can be reflected by a weight in the range of 0 to 1. Among many detection indicators, the larger the feature importance representation value of a detection indicator is, the higher the feature importance of the detection indicator determined by the feature selection method is. Different feature selection methods may assign the same or different feature importance representation values ​​to the same detection indicator.

[0152] S403 . For each set of detection index importance data, perform index screening based on the feature selection criteria of the corresponding feature selection method to generate M feature subsets.

[0153] Each feature selection method may have different corresponding feature selection criteria. For example, in the Lasso regression algorithm, the feature importance representation value of unimportant detection indicators is set to 0, so the feature selection criterion can be to select detection indicators with non-zero feature importance representation values ​​to construct the first feature subset D corresponding to the Lasso regression algorithm. 2 The feature selection method corresponding to the Boruta algorithm can screen indicators according to a certain threshold. For example, if a threshold of 0.7 is set, if the feature importance representation value is greater than 0.7, it is selected as an important indicator and added to the second feature subset D. 3 .

[0154] S404, using the initial optimal subset and the M feature subsets, screening the detection index of each sample in the historical clinical index amplification data set to form a data set after index screening.

[0155] The initial optimal subset includes the initial key indicators. As an example, the initial optimal subset is represented by D 1 , D 1 ={platelets, procalcitonin}. In this application, for the example case of M=2, using D 1 , D 2 and D 3 The detection indicators of each sample in the historical clinical indicator amplification data set are screened to form a data set after indicator screening. In the data set formed after screening, the detection indicators in each sample are screened, and the indicator dimension is reduced. For example, in the historical clinical indicator amplification data set, each sample contains N detection indicators + 1 diagnostic result indicator. After screening the detection indicators, each sample in the data set after indicator screening contains N / 2 detection indicators and 1 diagnostic result indicator. The specific screening intensity and the number of detection indicators after screening can be set according to actual needs, and there is no restriction here.

[0156] For ease of understanding, the following is a specific implementation process of this step. Figure 5 As shown, this figure is a flow chart of a method for screening detection indicators by constructing feature subsets and forming a screened data set provided by an embodiment of the present application. Using the initial optimal subset and M feature subsets, the detection indicators of each sample in the historical clinical indicator amplification data set are screened to form a data set after indicator screening, which may specifically include:

[0157] S4041. Taking the union of each of the M feature subsets and the initial optimal subset as the expansion set corresponding to each of the M feature subsets.

[0158] This process of this step can be expressed as: new_i =D i +D 1. Where + represents the union. Different values ​​of i represent different feature subsets. For example, when i = 2, D new_2 Denotes the first feature subset D 2 Corresponding extended set; when i=3, D new_3 Denotes the second feature subset D 3 Corresponding extended set. It can be understood that if the first feature subset D 2 Or the second feature subset D 3 It does not contain any initial key indicator or does not contain all the initial key indicators. 1 The union is calculated to form an extended set corresponding to each feature subset, so that the formed extended set can cover all the initial key indicators.

[0159] S4042. The intersection of the M feature subsets is taken as the minimum feature subset, and the union of the initial optimal subset and the minimum feature subset is taken as the extended set of the minimum feature subset.

[0160] The intersection of M feature subsets is actually to determine the important indicators of multiple different feature selection methods. The detection indicators selected by the consensus of these multiple feature selection methods are added to the minimum feature subset. The method for obtaining the extended set of the minimum feature subset is similar to the method for obtaining the extended set in S4031, which is the same as the initial optimal subset D. 1 Find the union.

[0161] If M = 2, the expression of the minimum feature subset is: D min =D 2 ∩D 3 . Where ∩ means finding the intersection. The extended set of the minimum feature subset can be expressed as: D min_new =D min +D 1 .

[0162] S4043. Create a maximum feature subset based on the feature importance representation values ​​corresponding to each detection indicator in the M groups of detection indicator importance data, and use the union of the initial optimal subset and the maximum feature subset as the extended set of the maximum feature subset.

[0163] In the present application, a maximum feature subset is further created. In the process of constructing the maximum feature subset, each detection indicator in the M feature subsets is taken into account, wherein the corresponding feature importance characterization value in the M group of detection indicator importance data. In an optional implementation of this step, based on the corresponding feature importance characterization value of each detection indicator in the M group of detection indicator importance data, the maximum feature subset is created, including:

[0164] For the sth feature subset among the M feature subsets, based on the feature importance representation value corresponding to each detection indicator in the sth feature subset in the sth group of detection indicator importance data, the standardized weight of each detection indicator in the sth feature subset is calculated. s takes a value between 1 and M, and s is an integer. Taking a certain feature subset as an example, the process of calculating the standardized weight of the detection indicator can be expressed as:

[0165]

[0166] In the formula, w i_sta represents the standardized weight of the i-th indicator in the feature subset, |W| max represents the maximum absolute value of each feature importance representation value in the set of detection index importance data W. Similarly, |W| min The minimum absolute value of each feature importance representation value in the set of detection index importance data W is represented by the difference between the maximum and minimum absolute values ​​as the denominator for calculating the standardized weight. The feature importance representation value corresponding to the i-th indicator in the set of detection index importance data is represented by w i , whose absolute value is used as the numerator for calculating the normalized weight.

[0167] Next, the standardized weights of the same detection indicator in different feature subsets are averaged to obtain the standardized weight averages corresponding to each detection indicator. The process of averaging can be seen in the following formula:

[0168]

[0169] In this formula, different values ​​of m represent different feature subsets in the M feature subsets. M is the total number of feature subsets formed by M feature selection methods. The result w obtained by averaging i_max This is the standardized weighted average value of the detection index. When constructing the largest feature subset, refer to the standardized weighted average value of each detection index.

[0170] Specifically, according to the standardized weighted average values ​​corresponding to each detection indicator and the preset feature selection criteria, feature screening can be performed on each detection indicator in the M feature subsets, and the screened detection indicators can be used to construct the maximum feature subset. For example, if the preset feature selection criterion is to select if it is not 0, then all detection indicators with standardized weighted average values ​​other than 0 can be selected into the maximum feature subset. In another possible example, if the preset feature selection criterion is to select if it exceeds the 0.5 threshold, then all detection indicators with standardized weighted average values ​​exceeding 0.5 can be selected into the maximum feature subset. It should be noted that, in combination with actual conditions, the feature selection criteria for the maximum feature subset can be set, and the selection criteria are not limited here.

[0171] The construction method of the extended set of the largest feature subset is similar to the construction method of the extended set of each subset mentioned above: the union of the initial optimal subset and the largest feature subset is used as the extended set of the largest feature subset. This process can be expressed as: max_new =D max +D 1 . Where D max represents the largest subset of features, D max_new The extended set representing the largest subset of features.

[0172] S4044. Filter the detection indicators of each sample in the historical clinical indicator amplification data set by using the initial optimal subset, M feature subsets, the extended sets corresponding to the M feature subsets, the minimum feature subset, the extended set of the minimum feature subset, the maximum feature subset and the extended set of the maximum feature subset to form a data set after indicator screening.

[0173] The multiple samples in the historical clinical index amplification data set are divided into a second training set and a second test set according to a second preset ratio. As an example, 30% can be extracted and divided into the second test set, and the other 70% can be divided into the training set. Through the above steps, the initial optimal subset D is gradually constructed. 1 , M feature subsets (e.g. D 2 and D 3 ), the corresponding expansion sets of M feature subsets (for example, D new_2 and D new_3 ), the smallest feature subset D min , the extended set D of the smallest feature subset min_new , the largest feature subset D max and the extended set D of the largest subset of features max_new Next, select any one indicator set from these sets, perform multi-detection indicator extraction for the samples in the second training set and the second test set according to the selected indicator set, and extract the diagnostic result indicators to form the third training set and the third test set corresponding to each selected indicator set. In this way, the sample indicator screening for the training set and the test set is completed based on the important indicators contained in a certain indicator set.

[0174] S405. Whether or not the patient is diagnosed with DIC is used as a target indicator, and a prediction model for sepsis complicated with DIC is obtained by training the data set after the indicator screening.

[0175] Whether or not the patient is diagnosed with DIC is used as the target indicator, and the model is trained and tested using the third training set and the third test set corresponding to each selected indicator set, respectively, to obtain the test evaluation score of the model. In the embodiment of the present application, "F1-sorce" can be selected as the evaluation indicator. The calculation method is:

[0176]

[0177] Wherein, P represents precision, and R represents recall. "F1-sorce" is a commonly used indicator in machine learning and will not be elaborated here. Since each selected indicator set can form a corresponding third training set and a third test set, accordingly, the test evaluation score formed by each model trained by different third training sets after passing through the third test set can also be obtained, such as the value of "F1-sorce". Based on this, the values ​​of the test evaluation scores can be analyzed and compared. After each of the above indicator sets has been selected once, the model test evaluation score corresponding to each indicator set can be obtained. From this, the selected indicator set corresponding to the model with the highest test evaluation score is determined as the optimal feature subset. Using this model, the highest test evaluation score can be obtained, so the prediction model trained by the third training set screened by the optimal feature subset can be used as a prediction model for sepsis complicated by DIC.

[0178] In one example, the optimal feature subset finally determined based on the above method includes: {platelets, procalcitonin, fibrinogen, interleukin-10, CD4+T cells}.

[0179] In the above embodiment, M feature subsets are constructed based on a variety of different feature selection methods, and further combined with the initial optimal subset to gradually derive a larger number of indicator sets. Based on these indicator sets, by screening and optimizing the data set, a prediction model based on different training sets can be created, and the prediction model can be tested accordingly. By comparing the test performance of the prediction model horizontally, the feature subset is optimized, and other key indicators related to sepsis complicated by DIC other than the initial key indicators are determined, which assists in denoising the indicator data, reduces the difficulty of model training, and improves the prediction performance of the model.

[0180] The following describes the use of the model. The sepsis-complicated DIC prediction model trained above can be used to diagnose the patient's DIC symptoms.

[0181] Specifically, the index data of the target object (for example, a patient visiting a hospital) can be obtained according to the detection index in the data set after index screening. Among them, the data set after index screening is formed after screening according to the training method of the sepsis complicated by DIC prediction model introduced in the above embodiment. In an optional implementation, the data set after index screening referred to above can specifically be an optimal feature subset. Next, the sepsis complicated by DIC prediction model is called to predict the probability of the target object suffering from sepsis complicated by DIC based on the index data of the target object; the sepsis complicated by DIC prediction model is obtained after training based on the training method of the sepsis complicated by DIC prediction model.

[0182] The output of the sepsis-complicated DIC prediction model is a probability, indicating whether the new patient has sepsis-complicated DIC. 0 indicates that the new patient does not have sepsis-complicated DIC, and 1 indicates that the new patient does not have sepsis-complicated DIC. Other probabilities in the output range of 0 to 1 indicate the probability that the new patient has sepsis-complicated DIC. The model collects the optimal subset of features of new patients in real time and predicts whether the patient has sepsis-complicated DIC in real time, which can effectively save manpower and diagnostic costs.

[0183] Based on the medical data amplification method introduced in the above embodiments, the present application also provides a medical data amplification device. Figure 6 is a schematic diagram of the structure of the device, such as Figure 6 As shown, the medical data expansion device includes:

[0184] A historical data set acquisition module 601 is used to acquire a historical clinical indicator data set; the historical clinical indicator data set includes a plurality of historical clinical indicator data samples formed by examining a target disease; the historical clinical indicator data samples include a diagnosis result indicator and a plurality of detection indicators, and the diagnosis result indicator is used to indicate whether the target disease is diagnosed;

[0185] A similar sample determination module 602 is used to extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and determine K similar samples for the target sample based on the difference between the detection index of other historical clinical indicator data samples and the corresponding detection index of the target sample; K is a positive integer greater than 1;

[0186] The new sample generation module 603 is used to randomly generate K new clinical indicator data samples based on the detection index of the target sample, the detection index of each of the K similar samples and the diagnosis result index of the target sample; the diagnosis result index of the K new clinical indicator data samples is consistent with the diagnosis result index of the target sample.

[0187] Optionally, the new sample generation module 603 includes:

[0188] A first calculation unit is used to calculate the difference between the target sample and the kth similar sample of the K similar samples one by one for the detection index, and obtain the absolute value of the difference of each detection index; k is an integer from 1 to K;

[0189] A second calculation unit is used to generate a random number in the range of 0 to 1, and multiply the random number by the absolute value to obtain a product;

[0190] The third calculation unit is used to add the product and the corresponding detection index of the target sample, and the various added calculation results are used as multiple detection indexes in the kth new clinical indicator data sample, and the diagnosis result index of the target sample is used as the diagnosis result index of the kth new clinical indicator data sample.

[0191] Optionally, the multiple detection indicators are specifically N detection indicators, where N is an integer greater than 1; the N detection indicators include one or more initial key indicators;

[0192] The similar sample determination module 602 includes:

[0193] A weight acquisition unit, used to acquire the similarity calculation weights corresponding to the N detection indicators respectively; wherein the similarity calculation weights corresponding to the initial key indicator are higher than the similarity calculation weights corresponding to the remaining detection indicators;

[0194] A similarity calculation unit is used to extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and obtain a similarity representation value between the other historical clinical indicator data samples and the target sample based on the difference between the detection index of the other historical clinical indicator data samples and the corresponding detection index of the target sample, and the similarity calculation weights corresponding to the N detection indexes respectively;

[0195] The similar sample determination unit is used to determine the K historical clinical indicator data samples with the highest similarity represented by the similarity representation value as the K similar samples of the target sample.

[0196] Optionally, the medical data augmentation device further comprises: a data set division module, used for dividing a plurality of historical clinical indicator data samples in the historical clinical indicator data set into a first training set and a first test set according to a first preset ratio before determining K similar samples for the target sample; the target sample and the K similar samples of the target sample are both taken from the first training set;

[0197] The medical data augmentation device also includes:

[0198] A traversal module, used for traversing the first training set until each historical clinical indicator data sample in the first training set is sequentially used as a target sample, and K new clinical indicator data samples are generated multiple times;

[0199] The first construction module is used to construct a backup new sample data set based on the K new clinical indicator data samples generated previously;

[0200] A second construction module is used to randomly extract a plurality of new clinical indicator data samples from the standby new sample data set according to a proportional coefficient w, and to construct a usable data set based on the extracted plurality of new clinical indicator data samples and the first training set; w takes a value in the range of 0 to 1;

[0201] an accuracy testing module, configured to train a prediction model for a target disease using the available data set, and to test the accuracy of the prediction model using the first test set;

[0202] A proportional coefficient updating module, used to dynamically update the w in the range of 0 to 1 according to a preset step size, so as to construct available data sets corresponding to different w for multiple times and obtain accuracy rates corresponding to different w;

[0203] The optimal proportional coefficient determination module is used to use w corresponding to the highest accuracy rate as the optimal proportional coefficient corresponding to the standby new sample data set, and use the available data set constructed according to the optimal proportional coefficient as the historical clinical indicator expansion data set.

[0204] Based on the training method of the sepsis complicated with DIC prediction model introduced in the above embodiments, the present application also provides a training device for the sepsis complicated with DIC prediction model. The implementation of the device is described below in conjunction with the accompanying drawings.

[0205] Figure 7 A schematic diagram of the structure of a training device for a sepsis-complicated DIC prediction model provided in an embodiment of the present application. Figure 7 As shown in , the training device of the sepsis complicated with DIC prediction model includes:

[0206] The augmented data set acquisition module 701 is used to acquire a historical clinical indicator augmented data set and determine M feature selection methods; the historical clinical indicator augmented data set is constructed by the medical data augmentation method introduced in the previous embodiment; M is an integer greater than 1; the target disease is sepsis complicated with DIC;

[0207] A feature selection module 702 is used to perform feature selection in the historical clinical indicator amplification data set using the M feature selection methods respectively, to obtain M groups of detection indicator importance data, wherein a group of detection indicator importance data corresponds to a feature selection method, and a group of detection indicator importance data includes feature importance representation values ​​corresponding to N detection indicators respectively;

[0208] An indicator screening module 703 is used to screen indicators for each set of detection indicator importance data based on the feature selection criteria of the corresponding feature selection method to generate M feature subsets;

[0209] The data set screening module 704 is used to screen the detection index of each sample in the historical clinical index amplification data set by using the initial optimal subset and the M feature subsets to form a data set after index screening; the initial optimal subset includes the initial key index;

[0210] The model training module 705 is used to use whether the patient is diagnosed with DIC as a target indicator and train a sepsis-complicated DIC prediction model using the data set after the indicator screening.

[0211] Optionally, the data set screening module 704 includes:

[0212] A first construction unit is used to use the union of each of the M feature subsets and the initial optimal subset as an extension set corresponding to each of the M feature subsets;

[0213] A second construction unit is used to use the intersection of the M feature subsets as a minimum feature subset, and use the union of the initial optimal subset and the minimum feature subset as an extended set of the minimum feature subset;

[0214] A third construction unit is used to create a maximum feature subset based on the feature importance representation value corresponding to each detection indicator in the M groups of detection indicator importance data in the M feature subsets, and use the union of the initial optimal subset and the maximum feature subset as an extension set of the maximum feature subset;

[0215] The screening unit is used to screen the detection indicators of each sample in the historical clinical indicator amplification data set by using the initial optimal subset, the M feature subsets, the extended sets corresponding to the M feature subsets, the minimum feature subset, the extended set of the minimum feature subset, the maximum feature subset and the extended set of the maximum feature subset to form a data set after indicator screening.

[0216] Optionally, the screening unit is specifically used for:

[0217] Dividing the plurality of samples in the historical clinical indicator amplification data set into a second training set and a second test set according to a second preset ratio;

[0218] Select any one indicator set from the initial optimal subset, the M feature subsets, the extended sets corresponding to the M feature subsets, the minimum feature subset, the extended set of the minimum feature subset, the maximum feature subset, and the extended set of the maximum feature subset;

[0219] For the samples in the second training set and the second test set, multiple detection indexes are extracted according to the selected index sets, and diagnostic result indicators are extracted to form a third training set and a third test set corresponding to each selected index set.

[0220] Optionally, the model training module 705 is specifically used for:

[0221] Whether or not DIC was diagnosed was taken as the target indicator, and the third training set and the third test set corresponding to each selected indicator set were used for model training and testing respectively. The indicator set corresponding to the model with the highest test evaluation score was taken as the optimal feature subset, and the model was used as a prediction model for sepsis complicated by DIC.

[0222] Optionally, the third building block is specifically used for:

[0223] For the sth feature subset among the M feature subsets, based on the feature importance representation value corresponding to each detection indicator in the sth feature subset in the sth group of detection indicator importance data, the standardized weight of each detection indicator in the sth feature subset in the sth feature subset is calculated; s takes a value between 1 and M, and s is an integer;

[0224] The average value of the standardized weights of the same detection indicator in different feature subsets is calculated to obtain the average value of the standardized weights corresponding to each detection indicator;

[0225] According to the standardized weight average values ​​corresponding to each detection index and the preset feature selection criteria, feature screening is performed on each detection index in the M feature subsets, and the screened detection indexes are used to construct the largest feature subset.

[0226] Based on the training method and training device of the sepsis-complicated DIC prediction model introduced in the above embodiments, the present application also provides a sepsis-complicated DIC prediction device. The device uses the sepsis-complicated DIC prediction model obtained through training. Figure 8 is a schematic diagram of the prediction device. Figure 8 As shown, the predictive devices for sepsis complicated by DIC include:

[0227] The indicator data acquisition module 801 is used to acquire the indicator data of the target object according to the detection indicators in the indicator-screened data set; the indicator-screened data set is formed after screening according to the training method of the sepsis complicated with DIC prediction model introduced in the above embodiment;

[0228] The DIC prediction module 802 is used to call the sepsis complicated by DIC prediction model to predict the probability of the target object suffering from sepsis complicated by DIC based on the indicator data of the target object; the sepsis complicated by DIC prediction model is obtained after training based on the sepsis complicated by DIC prediction model training method introduced in the previous embodiment.

[0229] The present application also provides a computer-readable storage medium, which is used to store program code, and the program code is used to execute the steps of the medical data augmentation method introduced in the previous embodiment, or the program code is used to execute the steps of the training method of the sepsis complicated by DIC prediction model introduced in the previous embodiment.

[0230] The present application also provides a processor, which is used to run a computer program. When the computer program is running, the steps of the medical data augmentation method introduced in the above embodiment are executed, or when the computer program is running, the steps of the training method of the sepsis complicated by DIC prediction model introduced in the above embodiment are executed.

[0231] It should be noted that each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is merely schematic, in which the unit described as a separate component may or may not be physically separated, and the component prompted as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0232] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A medical data augmentation method, characterized in that: include: Acquire a historical clinical indicator data set; the historical clinical indicator data set includes a plurality of historical clinical indicator data samples formed by examining a target disease; the historical clinical indicator data samples include a diagnosis result indicator and a plurality of detection indicators, the diagnosis result indicator being used to indicate whether the target disease is diagnosed; Extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and determine K similar samples for the target sample based on the difference between the detection index of other historical clinical indicator data samples and the corresponding detection index of the target sample; K is a positive integer greater than 1; Based on the detection index of the target sample, the detection index of each of the K similar samples and the diagnosis result index of the target sample, K new clinical indicator data samples are randomly generated; the diagnosis result index of the K new clinical indicator data samples is consistent with the diagnosis result index of the target sample.

2. The method according to claim 1, characterized in that The step of randomly generating K new clinical indicator data samples based on the detection index of the target sample, the detection index of each of the K similar samples, and the diagnostic result index of the target sample comprises: For the kth similar sample among the K similar samples, the difference between the target sample and the kth similar sample is calculated one by one for each detection index to obtain the absolute value of the difference between each detection index; k is an integer from 1 to K; Generate a random number in the range of 0 to 1, and multiply the random number by the absolute value to obtain a product; The product is added to the corresponding detection index of the target sample, and the resulting addition results are used as multiple detection indexes in the kth new clinical indicator data sample, and the diagnosis result index of the target sample is used as the diagnosis result index of the kth new clinical indicator data sample.

3. The method according to claim 1, characterized in that The multiple detection indicators are specifically N detection indicators, where N is an integer greater than 1; the N detection indicators include one or more initial key indicators; Extracting any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and determining K similar samples for the target sample based on the difference between the detection index of other historical clinical indicator data samples and the corresponding detection index of the target sample, including: Obtaining similarity calculation weights corresponding to the N detection indicators respectively; wherein the similarity calculation weights corresponding to the initial key indicator are higher than the similarity calculation weights corresponding to the remaining detection indicators; Extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and obtain similarity representation values ​​between other historical clinical indicator data samples and the target sample based on differences between detection indicators of other historical clinical indicator data samples and detection indicators corresponding to the target sample, and similarity calculation weights corresponding to the N detection indicators respectively; The K historical clinical indicator data samples with the highest similarity represented by the similarity representation values ​​are determined as K similar samples of the target sample.

4. The method according to claim 3, characterized in that Before determining K similar samples for the target sample, the method further includes: Dividing a plurality of historical clinical indicator data samples in the historical clinical indicator data set into a first training set and a first test set according to a first preset ratio; the target sample and K similar samples of the target sample are both taken from the first training set; The method further comprises: Traversing the first training set until each historical clinical indicator data sample in the first training set is sequentially used as a target sample, and K new clinical indicator data samples are generated multiple times; Construct a backup new sample data set based on the K new clinical indicator data samples generated previously; A plurality of new clinical indicator data samples are randomly extracted from the standby new sample data set according to a proportional coefficient w, and an available data set is constructed based on the extracted plurality of new clinical indicator data samples and the first training set; w is in the range of 0 to 1; Using the available data set to train a prediction model for a target disease, and using the first test set to test the accuracy of the prediction model; Dynamically update the w in the range of 0 to 1 according to a preset step size, so as to construct available data sets corresponding to different w for multiple times and obtain the accuracy rates corresponding to different w; The w corresponding to the highest accuracy rate is used as the optimal proportional coefficient corresponding to the spare new sample data set, and the available data set constructed according to the optimal proportional coefficient is used as the historical clinical indicator expansion data set.

5. A method for training a prediction model for sepsis complicated with DIC, characterized in that: include: Obtain historical clinical indicator expansion data sets and determine M feature selection methods; The historical clinical index amplification data set is constructed by the medical data amplification method according to claim 4; M is an integer greater than 1; the target disease is sepsis complicated by DIC; Using the M feature selection methods respectively, perform feature selection in the historical clinical indicator amplification data set to obtain M groups of detection indicator importance data, wherein one group of detection indicator importance data corresponds to one feature selection method, and one group of detection indicator importance data includes feature importance representation values ​​corresponding to N detection indicators respectively; For each set of detection index importance data, the indexes are screened based on the feature selection criteria of the corresponding feature selection method to generate M feature subsets; Using the initial optimal subset and the M feature subsets, the detection index of each sample in the historical clinical index amplification data set is screened to form a data set after index screening; the initial optimal subset includes initial key indicators; Whether or not DIC was diagnosed was used as the target indicator, and the prediction model for sepsis complicated with DIC was trained using the data set after indicator screening.

6. The method according to claim 5, characterized in that The method uses the initial optimal subset and the M feature subsets to screen the detection index of each sample in the historical clinical index amplification data set to form a data set after index screening, including: Taking the union of each of the M feature subsets and the initial optimal subset as the expansion set corresponding to each of the M feature subsets; Taking the intersection of the M feature subsets as the minimum feature subset, and taking the union of the initial optimal subset and the minimum feature subset as the extended set of the minimum feature subset; Based on the feature importance representation values ​​corresponding to the respective detection indicators in the M sets of detection indicator importance data, a maximum feature subset is created, and a union of the initial optimal subset and the maximum feature subset is used as an extended set of the maximum feature subset; The detection indicators of each sample in the historical clinical indicator amplification data set are screened by using the initial optimal subset, the M feature subsets, the extended sets corresponding to the M feature subsets, the minimum feature subset, the extended set of the minimum feature subset, the maximum feature subset and the extended set of the maximum feature subset to form a data set after indicator screening.

7. The method according to claim 6, characterized in that The method uses the initial optimal subset, the M feature subsets, the extension sets corresponding to the M feature subsets, the feature minimum subset, the extension set of the feature minimum subset, the feature maximum subset, and the extension set of the feature maximum subset to screen the detection index of each sample in the historical clinical index amplification data set to form a data set after index screening, including: Dividing the plurality of samples in the historical clinical indicator amplification data set into a second training set and a second test set according to a second preset ratio; Select any one indicator set from the initial optimal subset, the M feature subsets, the extended sets corresponding to the M feature subsets, the minimum feature subset, the extended set of the minimum feature subset, the maximum feature subset, and the extended set of the maximum feature subset; For the samples in the second training set and the second test set, multiple detection indexes are extracted according to the selected index sets, and diagnostic result indicators are extracted to form a third training set and a third test set corresponding to each selected index set.

8. The method according to claim 7, characterized in that The method uses whether DIC is diagnosed as a target indicator and uses the data set after the indicator screening to train a sepsis-complicated DIC prediction model, including: Whether or not DIC was diagnosed was taken as the target indicator, and the third training set and the third test set corresponding to each selected indicator set were used for model training and testing respectively. The indicator set corresponding to the model with the highest test evaluation score was taken as the optimal feature subset, and the model was used as a prediction model for sepsis complicated by DIC.

9. The method according to claim 6, characterized in that The creating a maximum feature subset based on the feature importance representation value corresponding to each detection indicator in the M feature subsets in the M groups of detection indicator importance data includes: For the sth feature subset among the M feature subsets, based on the feature importance representation value corresponding to each detection indicator in the sth feature subset in the sth group of detection indicator importance data, the standardized weight of each detection indicator in the sth feature subset in the sth feature subset is calculated; s takes a value between 1 and M, and s is an integer; The average value of the standardized weights of the same detection indicator in different feature subsets is calculated to obtain the average value of the standardized weights corresponding to each detection indicator; According to the standardized weight average values ​​corresponding to each detection index and the preset feature selection criteria, feature screening is performed on each detection index in the M feature subsets, and the screened detection indexes are used to construct the largest feature subset.

10. A medical data augmentation device, characterized in that: include: A historical data set acquisition module, used to acquire a historical clinical indicator data set; the historical clinical indicator data set includes a plurality of historical clinical indicator data samples formed by examining a target disease; the historical clinical indicator data samples include a diagnosis result indicator and a plurality of detection indicators, the diagnosis result indicator is used to indicate whether the target disease is diagnosed; A similar sample determination module, used to extract any historical clinical indicator data sample from the historical clinical indicator data set as a target sample, and determine K similar samples for the target sample based on the difference between the detection index of other historical clinical indicator data samples and the corresponding detection index of the target sample; K is a positive integer greater than 1; A new sample generation module is used to randomly generate K new clinical indicator data samples based on the detection index of the target sample, the detection index of each of the K similar samples and the diagnosis result index of the target sample; the diagnosis result index of the K new clinical indicator data samples is consistent with the diagnosis result index of the target sample.

11. A training device for a sepsis-complicated DIC prediction model, characterized in that: include: An augmented data set acquisition module is used to acquire an augmented data set of historical clinical indicators and determine M feature selection methods; The historical clinical index amplification data set is constructed by the medical data amplification method according to claim 4; M is an integer greater than 1; the target disease is sepsis complicated by DIC; A feature selection module, used to respectively use the M feature selection methods to perform feature selection in the historical clinical indicator amplification data set to obtain M groups of detection indicator importance data, wherein a group of detection indicator importance data corresponds to a feature selection method, and a group of detection indicator importance data includes feature importance representation values ​​corresponding to N detection indicators respectively; An indicator screening module is used to screen indicators for each set of detection indicator importance data based on the feature selection criteria of the corresponding feature selection method to generate M feature subsets; A data set screening module, used to screen the detection index of each sample in the historical clinical index amplification data set using the initial optimal subset and the M feature subsets to form a data set after index screening; the initial optimal subset includes initial key indicators; The model training module is used to take whether the patient is diagnosed with DIC as the target indicator and train the sepsis-complicated DIC prediction model using the data set after indicator screening.

12. A device for predicting sepsis complicated with DIC, characterized in that: include: An indicator data acquisition module, used to acquire indicator data of a target object according to the detection indicators in the indicator-screened data set; the indicator-screened data set is formed after screening according to the training method of the sepsis complicated with DIC prediction model according to any one of claims 5-9; A DIC prediction module is used to call a sepsis complicated by DIC prediction model to predict the probability that the target object suffers from sepsis complicated by DIC based on the indicator data of the target object; the sepsis complicated by DIC prediction model is obtained after training based on the training method of the sepsis complicated by DIC prediction model according to any one of claims 5-9.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store program code, and the program code is used to execute the steps of the medical data augmentation method according to any one of claims 1 to 4, or the program code is used to execute the steps of the training method of the sepsis complicated by DIC prediction model according to any one of claims 5 to 9.

14. A processor, characterized in that: The processor is used to run a computer program, and when the computer program is running, the steps of the medical data augmentation method according to any one of claims 1 to 4 are executed, or when the computer program is running, the steps of the training method of the sepsis complicated by DIC prediction model according to any one of claims 5 to 9 are executed.

Images