Machine model construction system based on representation mining

Abstract

The invention discloses a machine model construction system based on representation mining, and relates to the technical field of machine learning, and the system comprises a data collection module which is used for constructing a first training set and a first test set based on a blood routine examination sheet; the model construction module is used for training the initial prediction model based on the first training set to obtain a prediction model; the model evaluation module is used for carrying out dimension contribution analysis according to the Shap graph; the data set optimization module is used for optimizing the first training set and the first test set according to the dimension contribution result to obtain a second training set and a second test set; and the model optimization module is used for training and evaluating the prediction model again according to the second training set and the second test set to obtain an optimal prediction model and an optimal prediction dimension. According to the method, through the constructed optimal prediction model and the screened optimal prediction dimension, the accuracy and efficiency of diagnosis can be improved in a complex scene in which multiple diseases coexist.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and more particularly to a system for constructing a machine model based on representation mining. Background Art

[0002] In recent years, with the development of medical technology and bioinformatics, routine blood test, as a simple, fast and economical clinical test method, has become increasingly important in disease diagnosis. Routine blood test can not only provide basic information about human blood components, such as white blood cell count (WBC), red blood cell count (RBC), hemoglobin concentration (HGB), platelet count (PLT), etc., but also indirectly reflect the body's immune status and inflammatory response level through the changing trend of indicators.

[0003] However, in actual clinical applications, especially when patients suffer from multiple diseases at the same time, the change pattern of blood routine indicators becomes extremely complicated. This is because different types of diseases (such as infectious diseases, autoimmune diseases, tumors, etc.) may cause completely different change trends in blood routine indicators, and these change trends may interfere with each other in the state of multiple diseases coexisting, forming a comprehensive change pattern that is difficult to intuitively understand and analyze.

[0004] Traditional statistical analysis methods and manual interpretation methods are difficult to effectively extract meaningful information from complex routine blood data, especially when faced with the coexistence of multiple diseases. In recent years, in response to the above problems, the use of machine learning and artificial intelligence technologies to process and analyze routine blood data has been explored to improve the accuracy of diagnosis and achieve personalized diagnosis and treatment. Machine learning algorithms can automatically learn and identify complex nonlinear relationships, thereby revealing the laws hidden behind large amounts of routine blood data. For example, algorithms such as support vector machines (SVMs), decision trees, random forests, and neural networks have been applied to the classification and prediction tasks of routine blood data, showing high diagnostic efficacy.

[0005] However, most current studies are still limited to analyzing blood routine data under the condition of a single disease, while there are few studies on the changing patterns of blood routine data under the condition of multiple diseases coexisting, and there is a lack of effective automated processing tools. Therefore, how to develop a machine learning model construction system that can effectively process blood routine data under the condition of multiple diseases coexisting has become an urgent problem to be solved. Summary of the invention

[0006] In view of this, the present invention provides a system for constructing a machine model based on representation mining, which overcomes the above-mentioned defects.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A system for constructing a machine model based on representation mining, comprising:

[0009] A data acquisition module, used to identify and extract target data based on a blood routine test sheet, and to construct a first training set and a first test set based on the target data;

[0010] A model building module, used for building an initial prediction model, and training the initial prediction model based on the first training set to obtain a prediction model;

[0011] A model evaluation module, used to generate a Shap graph according to the first test set and the prediction model, and perform dimension contribution analysis according to the Shap graph;

[0012] A data set optimization module, used for optimizing the first training set and the first test set according to the dimension contribution result to obtain a second training set and a second test set;

[0013] The model optimization module is used to retrain and evaluate the prediction model according to the second training set and the second test set; and determine whether the evaluation result meets the stopping condition. If not, the second test set is transmitted to the model evaluation module for iteration. If so, the optimal prediction model and the optimal prediction dimension are output.

[0014] Optionally, the data acquisition module further includes a preprocessing module, and the preprocessing module includes:

[0015] A digitization unit, used for converting text data in the target data into numerical data;

[0016] A label classification unit, used to label the target data after numerical processing according to a preset label category to obtain classified data;

[0017] A data tilt processing unit, used for performing a balancing process on the classified data when the difference in quantity between any two types of data exceeds a maximum standard threshold;

[0018] The standardization unit is used to perform normalization on the dimension of the classified data after the equalization processing.

[0019] Optionally, the data tilt processing unit includes:

[0020] A judging subunit, used to judge whether the quantity difference between any two types of data exceeds the maximum standard threshold, and if so, determine the sample data to be balanced;

[0021] A distance calculation subunit, used to calculate the distance between each sample point and other sample points in the sample data to be balanced, to obtain a sample distance;

[0022] A sample synthesis subunit, configured to determine k nearest neighbors of any one of the sample points according to the sample distance, and generate a synthetic sample based on any one of the nearest neighbors and any one of the sample points by using a linear interpolation method;

[0023] The data updating subunit is used to merge the synthesized sample into the sample data to be balanced, and perform iterative calculation until the quantity difference between the sample data to be balanced and another type of data is less than the maximum standard threshold.

[0024] Optionally, the model building module includes:

[0025] An architecture design unit, configured to define a hyperparameter search space and generate a plurality of candidate models based on the hyperparameter search space;

[0026] A model training unit, configured to train the plurality of candidate models using the first training set, and evaluate the trained plurality of candidate models according to predefined performance indicators;

[0027] The model output unit is used to output the prediction model according to the evaluation result.

[0028] Optionally, the model training unit includes:

[0029] A data set construction subunit, used to randomly divide the first training set into n subsets, use any of the subsets as a validation set, and n-1 of the subsets as training sets to construct n data sets;

[0030] A training evaluation subunit, used for training a plurality of candidate models using the training sets in the n data sets;

[0031] A performance evaluation subunit is used to evaluate the performance of the multiple candidate models using the validation set in the data set, and calculate the average performance index of each candidate model.

[0032] Optionally, the model evaluation module includes:

[0033] A Shap graph generating unit, configured to generate a Shap graph according to the first test set and the prediction model;

[0034] The dimension contribution analysis unit is used to determine the contribution value of each dimension based on the Shap diagram.

[0035] Optionally, the data set optimization module includes:

[0036] A dimension screening unit, used to determine the dimensions to be eliminated according to the contribution value of each dimension;

[0037] A data optimization unit is used to optimize the first training set and the first test set according to the dimension to be eliminated to obtain a second training set and a second test set.

[0038] Optionally, the model optimization module includes:

[0039] A model training unit, used for retraining and evaluating the prediction model according to the second training set and the second test set;

[0040] The model iteration optimization unit is used to determine whether the evaluation result meets the stopping condition. If not, the second test set is transmitted to the model evaluation module for iteration. If yes, the optimal prediction model and optimal prediction dimension are output.

[0041] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a system for building a machine model based on representation mining. By constructing the optimal prediction model and screening the optimal prediction dimension, the accuracy of the prediction can be improved, especially in complex scenarios where multiple diseases coexist. It can improve the accuracy and efficiency of diagnosis and provide more reliable decision support for clinicians. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention 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 embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0043] Figure 1 A schematic diagram of the system structure provided by the present invention;

[0044] Figure 2(a) is a 25-dimensional Shap graph provided by the present invention; Figure 2(b) is a 19-dimensional Shap graph provided by the present invention; Figure 2(c) is a 16-dimensional Shap graph provided by the present invention; Figure 2(d) is a 13-dimensional Shap graph provided by the present invention; Figure 2(e) is an 11-dimensional Shap graph provided by the present invention; Figure 2(f) is a 9-dimensional Shap graph provided by the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] On the one hand, this embodiment discloses a system for constructing a machine model based on representation mining, such as Figure 1 As shown, including:

[0047] A data acquisition module, used to identify and extract target data based on a blood routine test sheet, and to construct a first training set and a first test set based on the target data;

[0048] A model building module is used to build an initial prediction model and train the initial prediction model based on the first training set to obtain a prediction model;

[0049] A model evaluation module, used to generate a Shap graph according to the first test set and the prediction model, and perform dimension contribution analysis according to the Shap graph;

[0050] A data set optimization module, used for optimizing the first training set and the first test set according to the dimension contribution results to obtain a second training set and a second test set;

[0051] The model optimization module is used to retrain and evaluate the prediction model based on the second training set and the second test set; and determine whether the evaluation result meets the stopping condition. If not, the second test set is transmitted to the model evaluation module for iteration. If so, the optimal prediction model and the optimal prediction dimension are output.

[0052] In one embodiment, the data acquisition module includes an identification unit and a division unit;

[0053] An identification unit, used for identifying and extracting target data from a blood routine test sheet;

[0054] The division unit is used to integrate the target data and divide the integrated target data into a first training set and a first test set according to a set ratio; the ratio set in this embodiment is 8:2.

[0055] In one embodiment, the preprocessing module includes:

[0056] A digitization unit, used for converting text data in the target data into numerical data;

[0057] The label classification unit is used to label the target data after numerical processing according to the preset label category to obtain classified data;

[0058] A data tilt processing unit, used to balance the classified data when the difference in quantity between any two types of data exceeds a maximum standard threshold;

[0059] The standardization unit is used to perform normalization on the dimensions of the balanced categorical data.

[0060] In one embodiment, the data tilt processing unit includes:

[0061] A judgment subunit is used to judge whether the quantity difference between any two types of data exceeds the maximum standard threshold, and if so, determine the sample data to be balanced;

[0062] The distance calculation subunit is used to calculate the distance between each sample point and other sample points in the sample data to be balanced to obtain the sample distance;

[0063] A sample synthesis subunit, used to determine the k nearest neighbors of any sample point according to the sample distance, and generate a synthetic sample based on any nearest neighbor and any sample point by linear interpolation;

[0064] The data updating subunit is used to merge the synthesized samples into the sample data to be balanced and perform iterative calculation until the quantity difference between the sample data to be balanced and another type of data is less than the maximum standard threshold.

[0065] Furthermore, routine blood test data usually includes [gender, age, RBC, MCV, PDW, WBC, NEUT%, LYMPH%, EO%, BASO%, NEUT#, LYMPH#, BASO#, HGB, HCT, MCH, MCHC, R-CV, PLT, MPV, PCT, MONO#, MONO%, EO#, P-LCR, R-SD, P-LCC]. In the above dimensions, gender is usually a character, so the gender is converted into a corresponding numerical value through a digitization unit. The specific steps are: counting all text values; assigning a different integer identifier to each text; replacing the text data with the corresponding integer value; in this embodiment, the identifiers assigned to male and female are 0 and 1, respectively.

[0066] The processing steps of the label classification unit are as follows: determine the label category based on historical data or demand data, and build a classification list based on the determined label category, which includes the label type and the reference threshold of each parameter; assign corresponding labels to the target data based on the comparison relationship between each parameter value in the target data and the reference threshold of each parameter in the classification list.

[0067] The processing steps of the data tilt processing unit are: obtain the quantity difference of any two categories of data in each category, and compare it with the preset maximum threshold to determine whether it is higher than the maximum threshold. If so, use SMOTE oversampling for equalization processing. The specific steps are:

[0068] Obtain all sample points of any type of samples to be balanced, and calculate the distance between each sample point and other sample points;

[0069] Determine the k nearest neighbors (k is a predefined integer) of each sample point based on the calculated distances between the sample points, randomly select a sample from the k nearest neighbors, and use linear interpolation (i.e., randomly select a point between the original sample and the selected neighbor) to generate a synthetic sample; add the synthetic sample to the original sample;

[0070] The above process is repeated several times until the number of samples in any category to be balanced meets the preset standard (the difference in the number of data between any two categories in each category is less than the maximum threshold).

[0071] Furthermore, SMOTE oversampling not only increases the amount of sample data, but also increases data diversity, reduces the risk of overfitting, and can better generalize the performance of the model.

[0072] The processing steps of the standardization unit are: normalization processing is performed on the target data after equalization processing through the Z-score formula.

[0073] In one embodiment, the model construction module includes an architecture design unit, a model training unit, and a model output unit;

[0074] An architecture design unit, used to define a hyperparameter search space and generate multiple candidate models based on the hyperparameter search space;

[0075] A model training unit is used to train multiple candidate models using the first training set and evaluate the performance of the models according to defined performance indicators. Common indicators include accuracy, recall, F1 score, etc., and the multiple candidate models after training are evaluated depending on the task type (classification, regression, etc.);

[0076] The model output unit is used to output the prediction model according to the evaluation results.

[0077] Furthermore, all adjustable model parameters are defined in the architecture design unit. Model parameters include the number of neural network layers, the number of nodes in each layer, activation functions, regularization term coefficients, etc. In order to increase the effectiveness of the search, a reasonable range or a set of discrete values ​​can be specified for each hyperparameter. Based on the defined hyperparameter space, algorithms such as random search, grid search, Bayesian optimization, etc. are used to generate different model combinations. These algorithms can effectively explore the hyperparameter space and find the best possible configuration.

[0078] Furthermore, when building a neural network model, the model can be built based on Python's Kerastuner library. Kerastuner is an automatic optimization library based on Tensorflow models, which can find the optimal model structure within given parameters. In order to verify the stability of the model, K10 (ten-fold cross validation) is used for model validation during the training process, and the cross-validation optimal model is used as the final prediction model.

[0079] In one embodiment, the model training unit includes a data set construction subunit, a training evaluation subunit, and a performance evaluation subunit;

[0080] A data set construction subunit is used to randomly divide the first training set into n subsets, and use any subset as a validation set and n-1 subsets as training sets to construct n data sets;

[0081] A training evaluation subunit, used to train multiple candidate models using training sets in n data sets;

[0082] The performance evaluation subunit is used to evaluate the performance of multiple candidate models using the validation set in the dataset and calculate the average performance index of each candidate model.

[0083] Furthermore, in this embodiment, the data set constructs a subunit, randomly divides the training data set into 10 subsets, all of which can be used as independent validation sets, and the 9 subsets other than the validation set are used as training sets;

[0084] The training evaluation subunit uses the training set to train the candidate model and evaluates the performance of the candidate model on the validation set until each subset is used as a validation set once;

[0085] The performance evaluation subunit is used to calculate the average performance index of the candidate model after training.

[0086] In this embodiment, the final model is tested with the first test set, the confusion matrix and related indicators are calculated, the model performance is evaluated, and the current performance results are recorded. The specific indicators are shown in Table 1:

[0087] Table 1

[0088]

[0089] In one embodiment, the model evaluation module includes:

[0090] A Shap graph generating unit, used to generate a Shap graph according to the first test set and the prediction model;

[0091] The dimension contribution analysis unit is used to determine the contribution value of each dimension based on the Shap diagram.

[0092] Furthermore, in this embodiment, the Shap library and matplotlib library of Python are used to draw the Shap diagram of the test set to analyze the dimensional contribution of the model. The test data is shown in Table 2, and the Shap diagram is obtained as shown in Figure 2(a)-2(f) As shown in the figure, the Shap features are arranged from top to bottom according to the contribution. The horizontal axis represents the contribution of the dimension to the model to distinguish between brucellosis and non-brucellosis. The larger the span, the greater the contribution (the wider). The color represents the positive and negative mathematical influence score on distinguishing between brucellosis and non-brucellosis within the algorithm. Taking the center line as the reference, the clearer the color on the left and right sides, the stronger the significance of the feature in distinguishing between brucellosis and non-brucellosis.

[0093] Table 2

[0094]

[0095]

[0096]

[0097] In one embodiment, the data set optimization module includes:

[0098] A dimension screening unit, used to determine the dimensions to be eliminated according to the contribution value of each dimension;

[0099] The data optimization unit is used to optimize the first training set and the first test set according to the dimension to be eliminated to obtain the second training set and the second test set.

[0100] In one embodiment, the model optimization module includes:

[0101] A model training unit, used for retraining and evaluating the prediction model according to the second training set and the second test set;

[0102] The model iteration optimization unit is used to determine whether the evaluation result meets the stopping condition. If not, the second test set is transmitted to the model evaluation module for iteration. If so, the optimal prediction model and the optimal prediction dimension are output.

[0103] Furthermore, the dimensions with lower contribution in the Shap diagram are recorded. In this embodiment, the contribution of the dimension is judged based on two points. One is the dimension with short Shap span, and the other is the dimension with chaotic Shap color (red and blue are not clearly distinguished). Dimensions that meet one of these two dimensions are low contribution dimensions, and these dimensions are usually arranged at the end of the Shap diagram.

[0104] Remove the selected low-contribution dimensions from the training and test sets, train the model again, and evaluate the performance of the test set. If the accuracy and AUC have not decreased significantly, repeat the Shap analysis operation and continue to remove low-contribution dimensions until the accuracy and AUC of the model show a significant decrease. The model before the accuracy and AUC decrease is the optimal model, and the training dimensions of this model are the key related dimensions of the disease, that is, the blood routine characterization of the disease. The changes in blood routine caused by the disease can be understood by comparing the group means (with and without the disease) to understand the changes in the incidence of these key dimensions.

[0105] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0106] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A system for constructing a machine model based on representation mining, characterized in that: include: A data acquisition module, used to identify and extract target data based on a blood routine test sheet, and to construct a first training set and a first test set based on the target data; A model building module, used for building an initial prediction model, and training the initial prediction model based on the first training set to obtain a prediction model; A model evaluation module, used to generate a Shap graph according to the first test set and the prediction model, and perform dimension contribution analysis according to the Shap graph; A data set optimization module, used for optimizing the first training set and the first test set according to the dimension contribution result to obtain a second training set and a second test set; A model optimization module, used for retraining and evaluating the prediction model according to the second training set and the second test set; And determine whether the evaluation result meets the stopping condition. If not, transmit the second test set to the model evaluation module for iteration. If so, output the optimal prediction model and the optimal prediction dimension.

2. A system for constructing a machine model based on representation mining according to claim 1, characterized in that: The data acquisition module also includes a preprocessing module, which includes: A digitization unit, used for converting text data in the target data into numerical data; A label classification unit, used to label the target data after numerical processing according to a preset label category to obtain classified data; A data tilt processing unit, used for performing a balancing process on the classified data when the difference in quantity between any two types of data exceeds a maximum standard threshold; The standardization unit is used to perform normalization on the dimension of the classified data after the equalization processing.

3. A system for constructing a machine model based on representation mining according to claim 2, characterized in that: The data tilt processing unit comprises: A judging subunit, used to judge whether the quantity difference between any two types of data exceeds the maximum standard threshold, and if so, determine the sample data to be balanced; A distance calculation subunit, used to calculate the distance between each sample point and other sample points in the sample data to be balanced, to obtain a sample distance; A sample synthesis subunit, configured to determine k nearest neighbors of any one of the sample points according to the sample distance, and generate a synthetic sample based on any one of the nearest neighbors and any one of the sample points by using a linear interpolation method; The data updating subunit is used to merge the synthesized sample into the sample data to be balanced, and perform iterative calculation until the quantity difference between the sample data to be balanced and another type of data is less than the maximum standard threshold.

4. The system for constructing a machine model based on representation mining according to claim 1, characterized in that: The model building module includes: An architecture design unit, configured to define a hyperparameter search space and generate a plurality of candidate models based on the hyperparameter search space; A model training unit, configured to train the plurality of candidate models using the first training set, and evaluate the trained plurality of candidate models according to predefined performance indicators; The model output unit is used to output the prediction model according to the evaluation result.

5. A system for constructing a machine model based on representation mining according to claim 4, characterized in that: The model training unit comprises: A data set construction subunit, used to randomly divide the first training set into n subsets, use any of the subsets as a validation set, and n-1 of the subsets as training sets to construct n data sets; A training evaluation subunit, used for training a plurality of candidate models using the training sets in the n data sets; A performance evaluation subunit is used to evaluate the performance of the multiple candidate models using the validation set in the data set, and calculate the average performance index of each candidate model.

6. The system for constructing a machine model based on representation mining according to claim 1, characterized in that: The model evaluation module includes: A Shap graph generating unit, configured to generate a Shap graph according to the first test set and the prediction model; The dimension contribution analysis unit is used to determine the contribution value of each dimension based on the Shap diagram.

7. The system for constructing a machine model based on representation mining according to claim 6, characterized in that: The data set optimization module includes: A dimension screening unit, used to determine the dimensions to be eliminated according to the contribution value of each dimension; A data optimization unit is used to optimize the first training set and the first test set according to the dimension to be eliminated to obtain a second training set and a second test set.

8. The system for constructing a machine model based on representation mining according to claim 1, characterized in that: The model optimization module includes: A model training unit, used for retraining and evaluating the prediction model according to the second training set and the second test set; The model iteration optimization unit is used to determine whether the evaluation result meets the stopping condition. If not, the second test set is transmitted to the model evaluation module for iteration. If so, the optimal prediction model and the optimal prediction dimension are output.

Images