A classification method for azoospermia surgery decision-making based on Cross-FiBiNet

Abstract

The present invention provides a method for classifying azoospermia surgery decision based on Cross‑FiBiNet. The method includes: obtaining a data set; preprocessing the data set; dividing the preprocessed data set into three different fields according to the patient's clinical hormone index, the patient's chromosome information and the patient's gene information; dividing the data set after being divided into three different fields into a training set, a validation set and a test set; building a Cross‑FiBiNet network model; inputting the training set into the Cross‑FiBiNet network model for training, inputting the validation set into the trained Cross‑FiBiNet network model for optimization, and saving the final optimized Cross‑FiBiNet network model; inputting the test set into the final optimized Cross‑FiBiNet network model to predict the patient's azoospermia surgery decision classification result, and obtaining a prediction result. The present invention effectively utilizes the patient's clinical hormone index, the patient's gene information, and the patient's chromosome information to construct an accurate azoospermia surgery plan prediction model, and realizes the prediction of the patient's azoospermia diagnosis mode.

Description

Technical Field

[0001] The present invention relates to the technical field of azoospermia surgery decision classification, and in particular to an azoospermia surgery decision classification method based on a Cross-FiBiNet network. Background Art

[0002] Azoospermia refers to a disease in which there are no sperm in male semen. According to the different causes, azoospermia can be divided into two types: primary azoospermia and secondary azoospermia. Among them, primary azoospermia or non-obstructive azoospermia (NOA) refers to problems in the male reproductive system itself, which leads to the inability to form or discharge sperm normally; while secondary azoospermia or obstructive azoospermia (OA) refers to other factors such as infection, drugs, radiation, etc. that affect sperm production or discharge. Therefore, OA patients need to undergo detailed examinations and diagnoses by doctors to determine the specific type and cause of azoospermia in order to develop corresponding treatment plans.

[0003] The patient's clinical hormone indicators, chromosome information and gene information are important sources of information for in-depth understanding of the cause of azoospermia and selection of surgical options. Among them, clinical hormone indicators such as testosterone, follicle-stimulating hormone (FSH) and luteinizing hormone levels (LH levels) can provide important clues about the patient's reproductive system function. Chromosome information includes karyotype abnormalities, such as microdeletions of azoospermia factor (AZF), which are common in azoospermic patients. Genetic information involves gene mutations related to reproductive function, which is crucial for understanding the patient's genetic basis and possible causes. However, accurate prediction of azoospermia in patients is still a challenging problem. Although clinical indicators such as clinical hormone indicators, chromosome information and gene information are important bases for doctors to make decisions on surgical options for azoospermia, clinicians generally describe chromosome and gene information in text. Doctors need to sort out and classify them based on clinical prior knowledge before they can be used in common classification methods such as machine learning. This method is highly subjective, has low accuracy and is time-consuming. With the emergence of increasingly complex computer technology and greater digital storage, artificial intelligence (AI) has been introduced to assist data analysis to identify patterns between complex variables and develop prediction algorithms. In the field of male infertility, the application of AI is mainly used for sperm recovery rate after azoospermia surgery. In the field of decision-making for azoospermia surgery, there is a lack of corresponding research and exploration. Basically, no one has done relevant research on the task of azoospermia surgery decision-making, that is, the task of azoospermia surgery decision-making itself is also a brand new technical starting point. Therefore, no researchers in related fields can provide the corresponding original data sets for azoospermia surgery decision-making.

[0004] Furthermore, it is considered that the Feature Interaction BipartiteNetwork (FiBiNet) is a neural network model commonly used for prediction tasks. FiBiNet combines the advantages of bilinear feature interaction and attention mechanism, and can efficiently capture the complex relationship between features. The FiBiNet model dynamically weights and interacts features through a bilinear interaction layer (Bilinear Interaction Layer) and an attention layer (SENet Attention Layer), thereby improving the expressive power of the model and then evaluating the importance of features; subsequently, the impact of various features on azoospermia can be explored based on network parameters, which has great clinical medical value. This is an important reason why the present invention chooses this network as the basic network.

[0005] However, the FiBiNet network is not capable of modeling complex feature relationships. Although FiBiNet enhances feature representation through bilinear interaction, it mainly focuses on feature importance assessment and capturing linear and second-order interaction relationships, and has limited modeling capabilities for high-order nonlinear relationships such as azoospermia surgery decisions. At the same time, the breadth of feature interactions of the FiBiNet network is limited. FiBiNet mainly focuses on the importance of specific features to model output, but may be insufficient in the breadth and diversity of feature interactions, and cannot well explore the correlation between hormone indicators, chromosome information, and gene information. It is necessary to introduce a feature interaction module that is more suitable for high-dimensional sparse variables.

[0006] In general, FiBiNet enhances feature representation through feature importance and bilinear interaction, but how to significantly improve the performance and accuracy of the FiBiNet classification network is particularly important. Therefore, we urgently need to build a new azoospermia surgery decision classification network, which is based on the FiBiNet network and improved, so that the improved classification network can not only improve the richness of feature representation, but also highlight important features, thereby improving the success rate of surgery and patient prognosis. Summary of the invention

[0007] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0008] To this end, the purpose of the present invention is to propose a classification method for azoospermia surgery decision-making based on Cross-FiBiNet.

[0009] In order to achieve the above-mentioned purpose, the technical solution of the present invention provides a method for classifying azoospermia surgery decision based on Cross-FiBiNet, the method comprising: step S1: acquiring a data set; wherein the data set is a tabular data containing the patient's examination information; the patient's examination information comprises: the patient's clinical hormone index, the patient's genetic information, and the patient's chromosome information; step S2: preprocessing the data set; step S3: dividing the preprocessed data set into three different fields according to the patient's clinical hormone index, the patient's chromosome information, and the patient's genetic information; wherein each field represents a feature combination; step S4: dividing the data set into three different fields according to a preset ratio; The data set after the same field is divided into a training set, a validation set and a test set; step S5: building a Cross-FiBiNet network model; wherein the Cross-FiBiNet network model includes a local information cross-processing module and a global information cross-processing module; the local information cross-processing module includes three sub-local information cross-processing modules; each sub-local information cross-processing module includes: an embedding layer, a squeeze-excitation network layer, an encoding layer, a bilinear interaction layer and a merging layer; the global information cross-processing module includes: a combination layer, a cross network layer, a deep network layer and a combined output layer; the embedding layer is used to receive the patient's clinical hormone index, the patient's chromosome information, or the patient's gene information respectively The squeezing excitation network layer is used to receive the output of the embedding layer, and extract the dependency between the low-dimensional dense features output by the embedding layer, and obtain the output features containing the importance of local features after pooling, compression and weighting; the encoding layer is used to receive the output features containing the importance of local features, and further enhance the output features containing the importance of local features, so as to help realize the more effective integration of the important feature information in the output features containing the importance of local features into the final feature representation, and enable the Cross-FiBiNet network model to capture the relationship between features. More complex interactions and dependencies, as well as the introduction of more levels and transformations into the coding layer itself, to help the Cross-FiBiNet network model converge better during the training process and reduce the risk of overfitting; the bilinear interaction layer is used to receive the data sets of the fields corresponding to the patient's clinical hormone indicators, the patient's chromosome information, or the patient's gene information, and receive the features output by the coding layer, splice the received data sets of the corresponding fields and the features output by the coding layer, calculate the bilinear interactions between each pair of fields, and output the local crossover results of the patient's clinical hormone indicators, the local crossover results of the patient's chromosome information, or the local crossover results of the patient's gene information;The merging layer is used to merge the local crossover results of the patient's clinical hormone indicators, the local crossover results of the patient's chromosome information, and the local crossover results of the patient's gene information to form a new feature representation; the combining layer is used to receive the new feature representation, and the data sets of the fields corresponding to the patient's clinical hormone indicators, the patient's chromosome information, and the patient's gene information, respectively, and use the new feature representation as a sparse feature, and use the data sets of the fields corresponding to the patient's clinical hormone indicators, the patient's chromosome information, and the patient's gene information as dense features, and then splice the sparse features and the dense features and pass them to the cross network layer and the deep network layer at the same time; the cross network layer is used to perform cross calculations on its own input features in a recursive manner; wherein the output of each layer of the cross network layer is calculated as the output of the previous layer The output and the intersection of the input features of each layer; the deep network layer is used to improve the expression ability of its own input features and to enhance the nonlinear modeling ability of the Cross-FiBiNet network model; the combined output layer is used to receive the output of the cross network layer and the output of the deep connection layer, and then generate the final prediction output through the full connection layer; step S6: input the training set into the Cross-FiBiNet network model for training, input the validation set into the trained Cross-FiBiNet network model for optimization, and save the final optimized Cross-FiBiNet network model; step S7: input the test set into the final optimized Cross-FiBiNet network model to predict the patient's azoospermia surgery decision classification result, and obtain the patient's azoospermia surgery decision classification prediction result. ;

[0010] Preferably, step S2 specifically includes: performing data cleaning on the data set to delete the examination information of patients in the data set that is irrelevant to the predicted results of the azoospermia surgery decision classification; counting the number of null values ​​of the examination indicators corresponding to the examination information of a certain patient in the data set, and when the number exceeds a preset number, deleting the examination information of the patient to implement the setting of a null value screening mechanism to screen out the examination information of the patient corresponding to the preset number of missing information; performing Word2Vec encoding on the text information of the patient's examination information; and normalizing each examination indicator in the patient's examination information.

[0011] Beneficial effects of the present invention:

[0012] The azoospermia surgery decision classification method based on Cross-FiBiNet provided by the present invention realizes the ability to improve the model feature importance evaluation through the feature importance module of FiBiNet, and the influence analysis of various features on azoospermia can be explored according to the network parameters in the future. Specifically, through the bilinear interaction module of FiBiNet, the model is efficiently interactive, and the model is efficiently interactive within the patient's clinical hormone index, the patient's genetic information, and the patient's chromosome information, and the potential information of the feature is fully mined. Further, by introducing a special coding layer and a deep learning network to enhance the modeling ability of complex feature relationships of the model, the shortcomings of FiBiNet in modeling high-order nonlinear relationships are supplemented. The introduction of the cross network layer enhances the breadth and diversity of the model feature interaction and enriches the breadth of the FiBiNet feature interaction. The improved FiBiNet network not only has a stronger complex relationship modeling ability and feature interaction breadth in azoospermia surgery decision-making, but also enhances the interpretability and computational efficiency of the model, and can effectively use the patient's clinical hormone index, the patient's genetic information, and the patient's chromosome information to construct an accurate azoospermia surgery plan prediction model to achieve the prediction of the patient's azoospermia diagnosis mode.

[0013] Additional aspects and advantages of the invention will become apparent from the following description, or may be learned by practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic flow chart of a method for classifying azoospermia surgery decision based on Cross-FiBiNet according to an embodiment of the present invention is shown;

[0015] Figure 2 A schematic diagram of the overall architecture of a Cross-FiBiNet network model constructed according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0016] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0017] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0018] Figure 1FIG. 1 is a schematic flow chart of a method for classifying azoospermia surgery decision based on Cross-FiBiNet according to an embodiment of the present invention. Figure 1 As shown, the azoospermia surgery decision classification method based on Cross-FiBiNet includes:

[0019] Step S1: Obtain a data set;

[0020] Step S2: preprocessing the data set;

[0021] Step S3: Divide the preprocessed data set into three different fields according to the patient's clinical hormone index, the patient's chromosome information, and the patient's gene information;

[0022] Step S4: Divide the data set after being divided into three different fields into a training set, a validation set, and a test set according to a preset ratio;

[0023] Step S5: Build the Cross-FiBiNet network model;

[0024] Step S6: input the training set into the Cross-FiBiNet network model for training, input the validation set into the trained Cross-FiBiNet network model for optimization, and save the final optimized Cross-FiBiNet network model;

[0025] Step S7: Input the test set into the finally optimized Cross-FiBiNet network model to predict the patient's azoospermia surgery decision classification result, and obtain the patient's azoospermia surgery decision classification prediction result.

[0026] In this embodiment, the data set is a table data containing the patient's examination information; the patient's examination information includes: the patient's clinical hormone index, the patient's gene information, and the patient's chromosome information.

[0027] In this embodiment, each field represents a feature combination.

[0028] In this embodiment, if Figure 2 As shown, the Cross-FiBiNet network model includes a local information cross-processing module and a global information cross-processing module; the local information cross-processing module includes three sub-local information cross-processing modules; each sub-local information cross-processing module includes: an embedding layer, a squeeze-excitation network layer, an encoding layer, a bilinear interaction layer, and a merging layer; the global information cross-processing module includes: a combination layer, a cross network layer, a deep network layer, and a combination output layer;

[0029] Specifically, the embedding layer is used to receive the data sets of the fields corresponding to the clinical hormone indicators of the patient, the chromosome information of the patient, or the gene information of the patient, and convert the data sets of the corresponding fields into corresponding low-dimensional dense features; the squeeze excitation network layer is used to receive the output of the embedding layer, and extract the dependency relationship between the low-dimensional dense features output by the embedding layer, and obtain the output features containing the importance of local features after pooling, compression and weighting; the encoding layer is used to receive the output features containing the importance of local features, and further enhance the output features containing the importance of local features, so as to help realize the more effective integration of the important feature information in the output features containing the importance of local features into the final feature representation, and enable the Cross-FiBiNet network model to capture more complex interactions and dependencies between features, and In order to realize that the coding layer itself introduces more layers and transformations, so as to help the Cross-FiBiNet network model converge better during the training process and reduce the risk of overfitting; a bilinear interaction layer, which is used to receive the data sets of the fields corresponding to the patient's clinical hormone indicators, the patient's chromosome information, or the patient's gene information, and receive the features output by the coding layer, splice the received data sets of the corresponding fields and the features output by the coding layer, calculate the bilinear interaction between each pair of fields, and output the local intersection results of the patient's clinical hormone indicators, the local intersection results of the patient's chromosome information, or the local intersection results of the patient's gene information; a merging layer, which is used to merge the local intersection results of the patient's clinical hormone indicators, the local intersection results of the patient's chromosome information, and the local intersection results of the patient's gene information to form a new feature representation;

[0030] Specifically, the combination layer is used to receive the new feature representation, and the data sets of the fields corresponding to the patient's clinical hormone indicators, the patient's chromosome information, and the patient's gene information, respectively, and use the new feature representation as a sparse feature, and use the data sets of the fields corresponding to the patient's clinical hormone indicators, the patient's chromosome information, and the patient's gene information as dense features, and concatenate the sparse features and the dense features and pass them to the cross network layer and the deep network layer at the same time; the cross network layer is used to cross-calculate its own input features in a recursive manner; wherein the output of each layer of the cross network layer is calculated as the intersection of the output of the previous layer and the input features of each layer; the deep network layer is used to improve the expression ability of its own input features; the deep network layer increases the number of network layers so that the Cross-FiBiNet network model can learn The invention discloses a method for generating a cross-network model comprising: a first layer for generating a cross-network model, a second layer for generating a cross-network model, and a third layer for generating a cross-network model. The method comprises: generating a cross-network model with a plurality of cross-network layers, a plurality of cross-network layers for generating a cross-network model, and a plurality of cross-network layers for generating a cross-network model. The method comprises: generating a cross-network model with a plurality of cross-network layers, a plurality of cross-network layers for generating a cross-network model, and a plurality of cross-network layers for generating a cross-network model. The method comprises: generating a cross-network model with a plurality of cross-network layers, a plurality of cross-network layers for generating a cross-network model, and a plurality of cross-network layers for generating a cross-network model. The method comprises: generating a cross-network model with a plurality of cross-network layers, a plurality of cross-network layers for generating a cross-network model, and a plurality of cross-network layers for generating a cross-network model.

[0031] In this embodiment, the purpose of selecting the FiBiNet neural network model is first described. Specifically, through the feature importance module of FiBiNet, the ability to improve the model feature importance assessment is achieved, and the impact analysis of various features on azoospermia can be explored based on network parameters. Specifically, through the bilinear interaction module of FiBiNet, efficient feature interaction of the model is achieved, and efficient feature interaction of the model is achieved within the patient's clinical hormone indicators, the patient's genetic information, and the patient's chromosome information, fully tapping the potential information of the features. Further, the purpose of improving the FiBiNet neural network model is described. Specifically, by introducing special coding layers and deep learning networks to enhance the modeling ability of complex feature relationships of the model, the shortcomings of FiBiNet in modeling high-order nonlinear relationships are supplemented. The introduction of the cross network layer improves the breadth and diversity of model feature interactions and enriches the breadth of FiBiNet feature interactions. The improved FiBiNet network not only has stronger complex relationship modeling capabilities and feature interaction breadth in azoospermia surgery decision-making, but also enhances the model's interpretability and computational efficiency. It can effectively utilize the patient's clinical hormone indicators, chromosome information, and gene information to build an accurate azoospermia surgery plan prediction model and realize the prediction of the patient's azoospermia diagnostic pattern.

[0032] In one embodiment of the present invention, step S2 specifically includes: performing data cleaning on the data set to delete the examination information of patients in the data set that is not related to the predicted results of the azoospermia surgery decision classification; counting the number of examination indicators corresponding to the examination information of a patient in the data set that are null values, and when the number exceeds a preset number, deleting the examination information of the patient to implement a null value screening mechanism to screen out the examination information of the patient corresponding to the preset number of missing information; performing Word2Vec encoding on the text information of the patient's examination information; and normalizing each examination indicator in the patient's examination information.

[0033] The technical solution of the present invention will be demonstrated below with a specific embodiment. The specific implementation steps of the Cross-FiBiNet-based azoospermia surgery decision classification method of the present invention are as follows:

[0034] 1. Obtain a data set; wherein the data set is a table data containing the patient's examination information; the patient's examination information includes: the patient's clinical hormone index, the patient's gene information, and the patient's chromosome information.

[0035] 2. Preprocess the data set; First, clean the data set to delete the patient examination information in the data set that is not related to the predicted results of the azoospermia surgery decision classification, so as to achieve the characteristics of medical clinical data and delete irrelevant clinical data (such as name, age, clinic number, and hospitalization number). Set a suitable null value screening mechanism to filter out patient data with too much missing information, and such information is unusable. Then, word2vec encode the text information in the table and normalize the numerical information such as hormone indicators.

[0036] 3. The preprocessed data set is divided into three different fields according to the patient's clinical hormone index, the patient's chromosome information, and the patient's gene information; each field represents a feature combination. The purpose of this is to convert the data into a form suitable for subsequent model processing.

[0037] 4. Build the Cross-FiBiNet network model.

[0038] First, the three field feature combinations are converted into low-dimensional dense vectors through the Embedding layer. Each field becomes a low-dimensional vector after passing through the Embedding layer. The output of the Embedding layer is a two-dimensional matrix, where each row represents the embedded representation of a field. The purpose of this step is to convert high-dimensional sparse features into low-dimensional dense features to facilitate subsequent feature interaction calculations.

[0039] Furthermore, these low-dimensional dense features are passed through the SENET Layer (squeeze excitation network layer) to extract the dependencies between the features, and after pooling, compression and weighting, an output containing the importance of local features is obtained. The encoding layer then receives the output features containing the importance of local features, and further enhances the output features containing the importance of local features, so as to help realize the more effective integration of the important feature information in the output features containing the importance of local features into the final feature representation, and enable the Cross-FiBiNet network model to capture more complex interactions and dependencies between features, and to enable the encoding layer itself to introduce more layers and transformations, so as to help the Cross-FiBiNet network model better converge during the training process and reduce the risk of overfitting.

[0040] Furthermore, the original data and the output containing the importance of local features are spliced ​​and input into the BilinearInteraction Layer to calculate the bilinear interaction between each pair of fields. For each pair of fields, the bilinear interaction is calculated, and the local intersection results of the patient's clinical hormone indicators, the local intersection results of the patient's chromosome information, or the local intersection results of the patient's gene information are output. The local intersection results of the patient's clinical hormone indicators, the local intersection results of the patient's chromosome information, and the local intersection results of the patient's gene information are then merged through the merging layer to form a new feature representation, which serves as a sparse feature for the subsequent global information cross-processing part. The purpose of this step is to capture the second-order interaction information between features to enhance the expressiveness of the model.

[0041] Furthermore, the new feature representation output by the merging layer is received through the combining layer, as well as the patient's clinical hormone index, the patient's chromosome information, and the patient's genetic information corresponding to the field data set, the new feature representation is used as a sparse feature, the patient's clinical hormone index, the patient's chromosome information, and the patient's genetic information corresponding to the field data set is used as a dense feature, the sparse feature and the dense feature are concatenated and passed to the cross network layer and the deep network layer at the same time. The cross network layer is used to cross-calculate its own input features in a recursive manner; wherein the output of each layer of the cross network layer is calculated as the intersection of the output of the previous layer and the input features of each layer; the deep network layer is used to improve the expression ability of its own input features, and to enhance the nonlinear modeling ability of the Cross-FiBiNet network model; through the multi-layer cross network, higher-order feature interaction information can be captured, and then the interactivity of the global information can be integrated and improved, and the model performance can be further improved.

[0042] Finally, the output layer receives the output of the cross network layer and the output of the deep connection layer, and then generates the final prediction output through the fully connected layer. The purpose of this step is to integrate the feature representations of each layer to form the final model output.

[0043] 5. Train and debug the network. Input the training set into the Cross-FiBiNet network model for training, and input the validation set into the trained Cross-FiBiNet network model for optimization.

[0044] Specifically, during the training process, the hyperparameters of the Cross-FiBiNe network model, such as the number of cross layers, the number of DNN hidden layer units, and the Dropout ratio, are adjusted to improve network performance. Then, in the SE layer (i.e., the squeeze-excitation network layer), different scaling ratios are tried, the effects are compared, and the scaling ratio that best suits the existing data is selected. Finally, based on the performance of the current model during training, the learning rate is adjusted using the decay strategy to improve the accuracy.

[0045] 6. Evaluate the network and save the final optimized Cross-FiBiNet network model.

[0046] The present invention uses accuracy, precision, recall and F1 score as indicators to evaluate model performance. Accuracy represents the proportion of all correct predictions (positive and negative) to the total, and the formula is:

[0047]

[0048] The precision rate is the ratio of the correct positive predictions to the total positive predictions. The formula is:

[0049]

[0050] The recall rate is the ratio of correctly predicted positives to all actual positives, and the formula is:

[0051]

[0052] As a comprehensive indicator, the F1 score is the harmonic average of precision and recall, which is a more comprehensive evaluation of a classifier. The larger the F1-score, the higher the model quality. The formula is:

[0053]

[0054] Among them, TP (True Positives) represents the number of samples correctly predicted as positive categories, TN (True Negatives) represents the number of samples correctly predicted as negative categories, FP (False Positives) represents the number of samples incorrectly predicted as positive categories, and FN (False Negatives) represents the number of samples incorrectly predicted as negative categories.

[0055] 7. Input the test set into the finally optimized Cross-FiBiNet network model to predict the patient's azoospermia surgery decision classification results, and obtain the patient's azoospermia surgery decision classification prediction results.

[0056] 8. Comparative experiment.

[0057] Based on the gold standard of clinical decision making, the trained deep learning network model is used to classify and predict the patient's surgical plan, and the prediction effect is compared with that of several machine learning models. The accuracy of various models obtained by the present invention is shown in Table 1. As shown in Table 1, the Cross-FiBiNet network model has satisfactory stability and accuracy, with a model accuracy of 0.9194, a recall value of 0.9194, a correct rate of 0.9194, and an F1 value of 0.9194.

[0058] Table 1 Four evaluation index values ​​of six models

[0059]

[0060] The comparative experiment involved several machine learning and deep learning models, including decision tree (DT), logistic regression (Logistic), K-nearest neighbors (KNN), and tabular network (TabNet). Among them, the decision tree is a supervised learning algorithm with a tree structure. It starts from the root node and gradually splits the nodes according to the value of the feature to form a decision path. It is suitable for classification and regression tasks. Logistic regression is a generalized linear model, which is mainly used for binary classification problems. It maps the input features to probability values ​​through a logical function and then classifies them according to the threshold. The K-nearest neighbor algorithm is an instance-based learning algorithm. It calculates the distance between the new data point and all the data points in the training data set, finds the nearest K neighbors, and then predicts the label of the new data point through the labels of these neighbors. TabNet is a deep learning model for tabular data. It selects and processes input features through a sequential attention mechanism, so that the model can adaptively focus on important features and improve the interpretability and performance of the model.

[0061] In summary, the present invention uses the patient's clinical hormone indicators, the patient's genetic information, and the patient's chromosome information to develop a Cross-FiBiNet azoospermia surgery decision classification method. The Cross-FiBiNet network model uses FiBiNet as the basic network, and through the integration of multiple technical means such as feature importance evaluation, bilinear interaction, depth and cross-layer interaction, it can significantly improve the performance of the model and provide more accurate support for clinical decision-making.

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A classification method for azoospermia surgery decision based on Cross-FiBiNet, characterized in that: include: Step S1: Acquire a data set; wherein the data set is a table data containing the patient's examination information; the patient's examination information includes: the patient's clinical hormone index, the patient's gene information, and the patient's chromosome information; Step S2: preprocessing the data set; Step S3: dividing the preprocessed data set into three different fields according to the patient's clinical hormone index, the patient's chromosome information, and the patient's gene information; wherein each field represents a feature combination; Step S4: Divide the data set after being divided into three different fields into a training set, a validation set, and a test set according to a preset ratio; Step S5: construct a Cross-FiBiNet network model; wherein the Cross-FiBiNet network model includes a local information cross processing module and a global information cross processing module; the local information cross processing module includes three sub-local information cross processing modules and a merging layer; each sub-local information cross processing module includes: an embedding layer, a squeeze excitation network layer, a coding layer, and a bilinear interaction layer; the global information cross processing module includes: a combination layer, a cross network layer, a deep network layer and a combination output layer; The embedding layers of the three local information cross-processing modules are used to respectively receive the data sets of the fields corresponding to the patient's clinical hormone index, the patient's chromosome information, and the patient's gene information, and convert the data sets of the corresponding fields into corresponding low-dimensional dense features; the squeeze-excitation network layer is used to receive the output of the embedding layer, and extract the dependency relationship between the low-dimensional dense features output by the embedding layer, and obtain the output features containing the importance of local features after pooling, compression and weighting; The encoding layer is used to receive the output features containing the importance of local features, and further enhance the output features containing the importance of local features, so as to help realize more effective integration of important feature information in the output features containing the importance of local features into the final feature representation, and enable the Cross-FiBiNet network model to capture more complex interactions and dependencies between features, and to enable the encoding layer itself to introduce more layers and transformations, so as to help the Cross-FiBiNet network model better converge during training and reduce the risk of overfitting; The bilinear interaction layers of the three local information crossover processing modules are used to respectively receive the data sets of the fields corresponding to the patient's clinical hormone index, the patient's chromosome information, and the patient's gene information, and receive the features output by the encoding layer, splice the received data sets of the corresponding fields and the features output by the encoding layer, calculate the bilinear interaction between each pair of fields, and output the local crossover results of the patient's clinical hormone index, the local crossover results of the patient's chromosome information, or the local crossover results of the patient's gene information; The merging layer is used to merge the local crossover results of the patient's clinical hormone index, the local crossover results of the patient's chromosome information, and the local crossover results of the patient's gene information to form a new feature representation; The combination layer is used to receive the new feature representation, and the data sets of the fields corresponding to the patient's clinical hormone index, the patient's chromosome information, and the patient's gene information, respectively, and use the new feature representation as a sparse feature, and use the data sets of the fields corresponding to the patient's clinical hormone index, the patient's chromosome information, and the patient's gene information as dense features, and concatenate the sparse features and the dense features and pass them to the cross network layer and the deep network layer at the same time; The cross network layer is used to perform cross calculation on its own input features in a recursive manner; wherein the output of each layer of the cross network layer is calculated as the intersection of the output of the previous layer and the input features of each layer; The deep network layer is used to improve the expression ability of its own input features and to enhance the nonlinear modeling ability of the Cross-FiBiNet network model; The combined output layer is used to receive the output of the cross network layer and the output of the deep network layer, and then generate the final prediction output through the fully connected layer; Step S6: inputting the training set into the Cross-FiBiNet network model for training, inputting the verification set into the trained Cross-FiBiNet network model for optimization, and saving the final optimized Cross-FiBiNet network model; Step S7: inputting the test set into the finally optimized Cross-FiBiNet network model to predict the patient's azoospermia surgery decision classification result, and obtaining the patient's azoospermia surgery decision classification prediction result.

2. The azoospermia surgery decision classification method based on Cross-FiBiNet according to claim 1, characterized in that: The step S2 specifically includes: Performing data cleaning on the data set to delete examination information of patients in the data set that is irrelevant to the predicted result of azoospermia surgery decision classification; Counting the number of times when the examination index corresponding to the examination information of a certain patient in the data set is a null value, and when the number exceeds a preset number, deleting the examination information of the patient, so as to implement a null value screening mechanism, so as to screen out the examination information of the patient corresponding to the preset number of missing information; Performing Word2Vec encoding on the text information of the patient's examination information; and Each examination index in the examination information of the patient is normalized.

Images