Semen quality prediction method and system based on random forest

Abstract

The invention discloses a sperm quality prediction method and system based on a random forest, and the method comprises the steps: obtaining metabolite data samples of different sperms, carrying out the preprocessing of the metabolite data samples, and obtaining the preprocessed metabolite data samples; analyzing the preprocessed metabolite data sample based on a preset multi-dimensional screening standard, and determining difference characteristics related to semen quality abnormity; a preset feature screening algorithm is adopted to screen and evaluate the difference features, and a core feature subset meeting the semen quality prediction requirement is determined according to a preset optimization index; based on the core feature subset, training a pre-constructed semen quality prediction model based on a random forest algorithm; and performing quality prediction on metabolite data of to-be-detected semen by using the trained semen quality prediction model to obtain a semen quality prediction result. According to the invention, the accuracy of semen quality prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a semen quality prediction method and system based on random forest. Background Art

[0002] In the field of semen quality prediction, existing technical solutions mainly rely on traditional statistical methods and machine learning models. Traditional statistical methods, such as linear regression and logistic regression, establish prediction equations to evaluate semen quality by analyzing parameters such as concentration, vitality, and morphology of semen samples. These methods assume that the relationship between variables is linear and have strict requirements on the distribution of data, resulting in limited prediction accuracy in practical applications. With the improvement of computing power and the increase in data volume, machine learning models have gradually replaced traditional statistical methods and become the mainstream technical means. Machine learning models (such as support vector machines, gradient boosting machines, etc.) can handle nonlinear relationships and have low requirements on data distribution, which makes them perform well in processing complex semen quality data. However, although machine learning models have shown certain advantages in semen quality prediction, existing technical solutions still have the following problems: 1. Feature selection is not accurate enough: In semen quality prediction, metabolite data is usually high-dimensional and contains a large number of features (such as metabolite concentration, vitality, etc.). Existing technical solutions often rely on simple statistical methods or empirical rules in feature selection, resulting in the selected features may not be accurate enough. For example, some features that contribute significantly to the prediction of semen quality may be ignored, while some irrelevant or redundant features are retained. This inaccurate feature selection will directly affect the predictive performance of the model, resulting in unstable prediction results. 2. Insufficient model generalization ability: Existing machine learning models may perform well on training data, but often show poor generalization ability when faced with new and unseen clinical data. This is because existing technical solutions fail to fully consider the diversity and complexity of clinical data during the model training process. For example, semen samples from different patients may be affected by a variety of factors (such as living habits, genetic factors, etc.), and existing models are difficult to adapt to this diversity, resulting in poor prediction results in practical applications.

[0003] Therefore, the existing semen quality prediction algorithms have problems with inaccurate feature selection and insufficient type generalization ability, which affects the accuracy of semen quality prediction. Summary of the invention

[0004] The embodiments of the present invention provide a semen quality prediction method and system based on random forest, which can improve the accuracy of semen quality prediction.

[0005] An embodiment of the present invention provides a method for predicting semen quality based on random forest, comprising the following steps: Obtaining metabolite data samples of different semens, and preprocessing the metabolite data samples to obtain preprocessed metabolite data samples; the metabolite data samples are obtained by pre-detecting and analyzing different semens; The pre-processed metabolite data samples were analyzed based on the preset multi-dimensional screening criteria to determine the differential features associated with abnormal semen quality; The preset feature screening algorithm is used to screen and evaluate the differential features, and the core feature subset that meets the requirements for semen quality prediction is determined based on the preset optimization indicators; Based on the core feature subset, a pre-built semen quality prediction model based on a random forest algorithm is trained; The trained semen quality prediction model is used to predict the quality of the metabolite data of the semen to be tested, and the semen quality prediction result is obtained.

[0006] As an improvement of the above scheme, the pre-processed metabolite data samples are analyzed based on the preset multi-dimensional screening criteria to determine the difference characteristics related to abnormal semen quality, including: The OPLS-DA method was used to screen metabolites with variable importance projection values ​​greater than 1; The Wilcoxon signed-rank test was used to screen metabolites with an FDR-corrected P value less than 0.05; Fold difference analysis was used to screen metabolites with fold differences greater than 2 or less than 0.5; The metabolites that met the above three screening criteria were regarded as significant differential metabolites, and the significant differential metabolites were regarded as differential features associated with abnormal semen quality.

[0007] As an improvement of the above scheme, the preset feature screening algorithm is used to screen and evaluate the differential features, and the core feature subset that meets the semen quality prediction requirements is determined according to the preset optimization index, including: Using the Lasso regression method, by introducing the L1 regularization term in the loss function, the regression coefficient is dynamically adjusted so that the coefficients of irrelevant difference features shrink to zero, retaining the difference features that have predictive power for the response variable; Using the support vector machine recursive feature elimination method, the support vector machine classifier is trained based on the current differential feature set, the importance of the differential features is calculated, the differential features with the smallest weight are removed, and the above steps are repeated until the remaining differential features reach the predetermined number, thereby obtaining a core feature subset that meets the requirements for semen quality prediction.

[0008] As an improvement of the above scheme, the Lasso regression method is used to dynamically adjust the regression coefficient by introducing the L1 regularization term in the loss function, so that the coefficients of irrelevant difference features shrink to zero, and retain the difference features that have predictive power for the response variable, including: Construct a loss function that minimizes the prediction error and the L1 regularization term: , is the sample size; is the number of features; is the intercept; is the regression coefficient of the jth feature; is the value of the jth feature in the i-th sample; is the response variable of the i-th sample; is a regularization hyperparameter that controls The intensity of the penalty; among them, minimizing the prediction error is used to measure the quality of model fitting, and the L1 regularization term shrinks the coefficients of unimportant difference features to zero by weighted penalty on the absolute value of the regression coefficient; During the training process, the regression coefficients are dynamically adjusted so that the coefficients of irrelevant differential features shrink to zero, while the coefficients of relevant features maintain a larger absolute value. Ultimately, only differential features that have predictive power for the response variable are retained, and the optimal regularization parameter is selected through cross-validation to achieve a balance between model sparsity and predictive power.

[0009] As an improvement of the above scheme, the support vector machine recursive feature elimination method is used to train the support vector machine classifier based on the current difference feature set, calculate the importance of the difference feature, remove the difference feature with the smallest weight, and repeat the above steps until the remaining difference features reach a predetermined number, thereby obtaining a core feature subset that meets the requirements for semen quality prediction, including: Input the difference feature set and the response variable, wherein the difference feature set includes the significantly different metabolites screened by OPLS-DA, Wilcoxon signed rank test and difference fold analysis; the response variable is the classification label of the semen quality prediction model for semen quality prediction; Using a support vector machine algorithm, a classifier is trained based on the current differential feature set and the response variable to construct an optimal classification decision surface, wherein the classification decision surface is used to distinguish different categories of semen quality samples; The weight vector of each difference feature is calculated by the support vector machine model. The weight vector represents the contribution of each difference feature to the classification decision surface. The larger the absolute value of the weight, the greater the contribution of the feature to the classification. According to the weight vector, the importance of each difference feature is evaluated, and the importance is measured by the absolute value of the weight vector, and the difference feature with the smallest weight is removed; The difference feature set is updated, and the process of training the support vector machine classifier, calculating feature importance, and removing the difference feature with the smallest weight is repeated until the remaining difference features reach a predetermined number; The final core feature subset is output, which includes the differential features that contribute most to the prediction of semen quality.

[0010] As an improvement of the above solution, the pre-built semen quality prediction model based on the random forest algorithm is trained based on the core feature subset, including: Constructing a new training set and test set from the core feature subset; The constructed new training set is used to train the random forest model as a semen quality prediction model; during the training process, samples are randomly sampled from the training set with replacement to construct multiple different training subsets to increase sample diversity; and feature random sampling is used to randomly select some features for training each time a decision tree is constructed to reduce model complexity; at the same time, the performance of the model is optimized by adjusting specific hyperparameters of the model; specific hyperparameters include the number of decision trees and the maximum depth of the random forest model; The sample data in the test set is input into the trained random forest model. The random forest model predicts the semen quality of each test sample based on the learned correlation pattern between the core features and semen quality, and outputs the corresponding prediction results. The 10-fold cross validation method was used to evaluate the performance of the random forest model on the training set, and to evaluate the sensitivity, specificity, and AUC value of the random forest model.

[0011] Another embodiment of the present invention provides a semen quality prediction system based on random forest, including: An acquisition module is used to acquire metabolite data samples of different semens and pre-process the metabolite data samples to obtain pre-processed metabolite data samples; the metabolite data samples are obtained by pre-detecting and analyzing different semens; An analysis module is used to analyze the pre-processed metabolite data samples based on preset multi-dimensional screening criteria to determine the differential features associated with abnormal semen quality; A screening module is used to screen and evaluate differential features using a preset feature screening algorithm, and determine a core feature subset that meets the requirements for semen quality prediction based on preset optimization indicators; A training module, used for training a pre-built semen quality prediction model based on a random forest algorithm based on the core feature subset; The prediction module is used to use the trained semen quality prediction model to perform quality prediction on the metabolite data of the semen to be tested to obtain the semen quality prediction result.

[0012] As an improvement of the above solution, the analysis module is specifically used for: The OPLS-DA method was used to screen metabolites with variable importance projection values ​​greater than 1; The Wilcoxon signed-rank test was used to screen metabolites with an FDR-corrected P value less than 0.05; Fold difference analysis was used to screen metabolites with fold differences greater than 2 or less than 0.5; The metabolites that met the above three screening criteria were regarded as significant differential metabolites, and the significant differential metabolites were regarded as differential features associated with abnormal semen quality.

[0013] As an improvement of the above solution, the screening module is specifically used for: Using the Lasso regression method, by introducing the L1 regularization term in the loss function, the regression coefficient is dynamically adjusted so that the coefficients of irrelevant difference features shrink to zero, retaining the difference features that have predictive power for the response variable; Using the support vector machine recursive feature elimination method, the support vector machine classifier is trained based on the current differential feature set, the importance of the differential features is calculated, the differential features with the smallest weight are removed, and the above steps are repeated until the remaining differential features reach the predetermined number, thereby obtaining a core feature subset that meets the requirements for semen quality prediction.

[0014] As an improvement of the above solution, the training module is specifically used for: Constructing a new training set and test set from the core feature subset; The constructed new training set is used to train the random forest model as a semen quality prediction model; during the training process, samples are randomly sampled from the training set with replacement to construct multiple different training subsets to increase sample diversity; and feature random sampling is used to randomly select some features for training each time a decision tree is constructed to reduce model complexity; at the same time, the performance of the model is optimized by adjusting specific hyperparameters of the model; specific hyperparameters include the number of decision trees and the maximum depth of the random forest model; The sample data in the test set is input into the trained random forest model. The random forest model predicts the semen quality of each test sample based on the learned correlation pattern between the core features and semen quality, and outputs the corresponding prediction results. The 10-fold cross validation method was used to evaluate the performance of the random forest model on the training set, and to evaluate the sensitivity, specificity, and AUC value of the random forest model.

[0015] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: First, the metabolite data samples of different semen obtained by detection and analysis are obtained and preprocessed to provide reliable basic data. Then, the preprocessed data is deeply analyzed according to the preset multi-dimensional screening criteria, and the difference features related to abnormal semen quality are accurately located, providing a key basis for subsequent accurate prediction. Then, the preset feature screening algorithm is used to screen and evaluate these difference features, and the core feature subset is determined according to the optimization index, which effectively reduces the data dimension and highlights the key information. Based on this core feature subset, the semen quality prediction model of the pre-constructed random forest algorithm is trained, and the generalization ability and efficiency of the random forest algorithm are utilized and the advantages of random forest processing high-dimensional data, evaluating feature importance, and resisting overfitting are fully utilized. Finally, the trained model is used to predict the semen metabolite data to be tested, and the semen quality prediction result is obtained. Therefore, the embodiment of the present invention can more accurately screen out key features, effectively process complex data relationships, and construct a semen quality prediction model with high prediction accuracy and good generalization ability through the combination of multi-dimensional screening criteria, efficient feature screening algorithms and random forest models with strong generalization ability, thereby effectively improving the accuracy of semen quality prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a method for predicting semen quality based on random forests provided by one embodiment of the present invention; Figure 2 is a comparison diagram of differential metabolites between a semen quality abnormality case group and a control group provided by an embodiment of the present invention; Figure 3 is a graph of the analysis results of Lasso regression provided by an embodiment of the present invention; Figure 4 It is a diagram of analysis results of support vector machine recursive feature elimination SVM-RFE provided by one embodiment of the present invention; Figure 5 It is a prediction effect diagram of four machine learning models provided by an embodiment of the present invention; Figure 6 It is a structural schematic diagram of a semen quality prediction system based on random forest provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0017] 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.

[0018] See also Figure 1, is a flow chart of a method for predicting semen quality based on random forests provided by an embodiment of the present invention. The method for predicting semen quality based on random forests comprises the following steps: S10, obtaining metabolite data samples of different semens, and preprocessing the metabolite data samples to obtain preprocessed metabolite data samples; the metabolite data samples are obtained by pre-detecting and analyzing different semens; S11, analyzing the pre-processed metabolite data samples based on the preset multi-dimensional screening criteria to determine the differential features associated with abnormal semen quality; S12, using a preset feature screening algorithm to screen and evaluate the differential features, and determining a core feature subset that meets the semen quality prediction requirements based on a preset optimization index; S13, training a pre-built semen quality prediction model based on a random forest algorithm based on the core feature subset; S14, using the trained semen quality prediction model to perform quality prediction on the metabolite data of the semen to be tested, to obtain a semen quality prediction result.

[0019] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: First, the metabolite data samples of different semen obtained by detection and analysis are obtained and preprocessed to provide reliable basic data. Then, the preprocessed data is deeply analyzed according to the preset multi-dimensional screening criteria, and the difference features related to abnormal semen quality are accurately located, providing a key basis for subsequent accurate prediction. Then, the preset feature screening algorithm is used to screen and evaluate these difference features, and the core feature subset is determined according to the optimization index, which effectively reduces the data dimension and highlights the key information. Based on this core feature subset, the semen quality prediction model of the pre-constructed random forest algorithm is trained, and the generalization ability and efficiency of the random forest algorithm are utilized and the advantages of random forest processing high-dimensional data, evaluating feature importance, and resisting overfitting are fully utilized. Finally, the trained model is used to predict the semen metabolite data to be tested, and the semen quality prediction result is obtained. Therefore, the embodiment of the present invention can more accurately screen out key features, effectively process complex data relationships, and construct a semen quality prediction model with high prediction accuracy and good generalization ability through the combination of multi-dimensional screening criteria, efficient feature screening algorithms and random forest models with strong generalization ability, thereby effectively improving the accuracy of semen quality prediction.

[0020] Understandably, semen quality prediction models have broad prospects in clinical applications, especially in assisted reproductive technology. By using prediction models, doctors can more accurately assess the fertility potential of patients, optimize treatment plans, and thus improve the success rate. For example, during in vitro fertilization (IVF), semen quality prediction models can help doctors select the embryos most likely to succeed for transplantation, reducing the risk and cost of multiple attempts. In addition, semen quality prediction models can also be used to monitor patients' hormone levels and follicular development, and adjust treatment plans in a timely manner to ensure the best fertility results. Prediction models also play an important role in the early diagnosis of male infertility. By analyzing data on men's semen parameters, genetic factors, and lifestyle habits, prediction models can identify potential infertility risks and help doctors intervene early. For example, some prediction models can predict the risk of sperm DNA damage, which is crucial for assessing male fertility and formulating treatment strategies. Early diagnosis can not only improve treatment outcomes, but also reduce unnecessary consumption of medical resources and improve patients' quality of life. The formulation of personalized treatment plans is another important area of ​​clinical application of semen quality prediction models. By integrating multi-dimensional information such as the patient's genetic information, clinical data, and lifestyle, the prediction model can tailor the most appropriate treatment plan for each patient. For example, in cancer treatment, the prediction model can recommend the most effective drug combination and treatment strategy based on the patient's genetic mutation and the molecular characteristics of the tumor. This personalized treatment plan can not only improve the treatment effect, but also reduce side effects and improve the patient's quality of life. Based on the matched oligospermia and asthenospermia case-control population, through the screening and analysis of non-targeted metabolites in seminal plasma, combined with the results of a questionnaire survey on factors affecting semen quality, a machine learning model will be used to construct a prediction model for male semen quality, providing more powerful support for clinical practice and male reproductive health management.

[0021] As an example, the semen sample detection analysis and pretreatment process is as follows: First, the semen sample was centrifuged at 15,000 rpm at 4°C for 10 minutes, and methanol was added to separate the seminal plasma for protein precipitation, followed by overnight incubation at -20°C. The next day, the sample was centrifuged at 14,000g at 4°C for 20 minutes, the supernatant was collected and vacuum centrifuged to dryness, and the sample was re-dissolved with methanol / water (1:9), vortexed and centrifuged again to obtain the supernatant for analysis. Quality control (QC) samples were prepared by mixing 5µL of each sample, and blank samples were treated with ultrapure water. Liquid chromatography analysis used ultrapure water and methanol with 0.1% formic acid as the mobile phase for gradient elution, and the running time for each sample was 25 minutes. The qualitative identification analysis of metabolites used a multi-level qualitative system, including the OSI / SMMS system developed by the Chinese Academy of Sciences, the widely used network database, and the computational mass spectrometry structure prediction library and other tools for identification and analysis. A total of 259 metabolites were identified, as shown in Table 1. A series of algorithms were used to standardize and process data during data analysis, including global parameter setting, data normalization (sample and feature direction), 80% rule feature deletion (retain the feature when the proportion of non-zero values ​​in any category of samples is greater than or equal to 80%, otherwise delete it), peak table data filtering based on QC sample RSD values, and missing value filling (using variable minimum processing). Before machine learning modeling, all metabolomics data were standardized. The standardization method ensures that the mean of each feature is zero and the standard deviation is one, thereby eliminating the influence of different feature scales and ensuring that each feature contributes evenly to model training.

[0022] Table 1 259 metabolites Metabolites VIP value <![CDATA[P FDR Value]]> FC value (±)13-HDoHE 1.24 0.00 0.86 (1S,2R,4aR,8aR)-1-Acetoxy-7-isopropylidene-1,4a-dimethyl-6-oxodecahydro-2-naphthalenyl 2,3-dimethyl-2-oxiranecarboxylate 0.95 0.01 1.18 (1S,Z)-3-((2R,E)-7-hydroxy-2,5-dimethyloct-4-en-1-ylidene)-1-methyloctahydro-2H-quinolizin-1-ol 1.87 0.00 0.78 (2E,6E,12E)-19-(2-amino-2-oxoethyl)-9,11-dihydroxy-8-methoxy-10,12,14-trimethyl-15-oxohenicosa-2,6,12-trienedioicacid 0.94 0.00 0.77 (2S)-3-(1H-indol-3-yl)-2-[[2-(2-oxo-4-propylchromen-7-yl)oxyacetyl]amino]propanoicacid 0.95 0.01 1.18 (3S,8aS)-3-[(1R)-1-Hydroxyethyl]hexahydropyrrolo[1,2-a]pyrazine-1,4-dione 0.79 0.01 0.84 (4a'R,5'S,6'R,6a'S,7'R,10a'R,10b'R)-6'-Acetoxy-4a',6a',10b'-trimethyl-5-oxododecahydro-2'H-dispiro[furan-3,3'-benzo[f]chromene-7',2''-oxiran]-5'-yl 2-methylpropanoate 1.19 0.00 1.15 (6R)-vitamin D2 6,19-sulfurdioxide adduct / (6R)-ergocalciferol 6,19-sulfurdioxide adduct 0.75 0.02 1.16 (R)-4'-Deoxyindenestrol 1.05 0.01 1.43 1-[3-(4-methyl-2-oxo-6,7,8,9-tetrahydro-[1]benzofuro[3,2-g]chromen-3-yl)propanoyl]-4-phenylpiperidine-4-carboxylicacid 0.41 0.00 1.35 11-deoxy-16,16-dimethyl-PGE2 0.40 0.04 1.18 11Z-6s-trans-fixed bicyclicretinal 0.86 0.00 1.69 16(R)-HETE 0.92 0.01 1.18 17R,18S-Epoxy-5Z,8Z,11Z,14Z-icosatetraenoicacid 0.55 0.03 1.17 17-α-Methyltestosterone 0.95 0.02 4.06 1-Methoxy-2-hydroxyanthracene 0.93 0.00 0.76 2-(3,5-dihydroxyphenyl)-5,7-dihydroxy-6,8-bis(3-methylbut-2-enyl)-2,3-dihydrochromen-4-one 0.70 0.03 1.31 2-(4-Ethoxyphenyl)-5,6,7,8-tetramethoxy-4H-1-benzopyran-4-one 0.68 0.04 0.84 20-HETE 0.96 0.01 1.17 2-Amino-3,7-dideoxy-D-threo-hept-6-ulosonicacid 0.87 0.04 0.84 2-butyl-6-heptylpiperidine 0.61 0.04 1.23 3,5-Dihydroxy-1,4-naphthoquinone 0.11 0.01 1.17 3,6-Dimethoxyestra-1,3,5(10),6,8-pentaene-17β-carboxylic acidmethyl ester 0.65 0.04 1.22 3,7-Diethyl-11-methyltrideca-1,3,6,10-tetraene 0.18 0.00 1.22 3,8a-Dimethyl-5-methylene-2-oxo-2,4,4a,5,6,7,8,8a,9,9a-decahydronaphtho[2,3-b]furan-8-yl acetate 0.07 0.05 1.21 3-Hydroxyisovalerylcarnitine 1.49 0.00 0.83 3-Methyluridine 1.23 0.00 1.40 3β-(3-methyl-butanoyloxy)-villanovane-13α,17-diol 0.80 0.00 1.34 4-[[2-(3,4,7-trimethyl-2-oxochromen-5-yl)oxypropanoylamino]methyl]cyclohexane-1-carboxylic acid 0.65 0.01 1.34 4-hydroxy-3-[(E)-7-hydroxy-3,7-dimethyl-4-oxooct-5-enyl]-5-(3-methylbut-2-enyl)benzoicacid 0.85 0.00 1.23 5a-Androstan-17β-ol-3-one 0.97 0.01 1.20 5-Hydroxytryptophol 1.09 0.00 1.51 5-Phosphooxy-L-lysine 0.84 0.03 1.27 6,7-Dimethoxy-1-[2-(2,3,4-trimethoxy-phenyl)-vinyl]-3,4-dihydro-isoquinoline 0.71 0.03 1.44 6-[3-[(3,4-dimethoxyphenyl)methyl]-4-methoxy-2-(methoxymethyl)butyl]-4-methoxy-1,3-benzodioxole 1.36 0.00 0.82 7,11-Methano-2H,4H-oxireno[1,8]cycloocta[1,2-f][2]benzopyran-8-acetic acid, 4-(3-furanyl)dodecahydro-alpha,11-dihydroxy-4a,7,9,9-tetramethyl-10-(2-methyl-1-oxopropoxy)-2,13-dioxo-,methyl ester,(alphaR,4R,4aR,7S,8S,10S,11S,11aR,12aS)- 0.70 0.03 1.20 9-Phenanthrol 0.67 0.04 0.72 9α-Fluoro-11β,16α,17α,21-tetrahydroxypregn-4-ene-3,20-dione 0.28 0.03 1.16 Acetaminophen 1.33 0.00 1.48 Acetylsulfamethoxazole 0.59 0.03 0.81 Acetyl-T2 Toxin 0.14 0.00 0.87 Adrenosterone 0.74 0.04 1.31 Alibendol 0.81 0.01 1.22 Aminopentol 0.42 0.00 1.36 Artocarpin 1.09 0.00 0.86 Asp Glu 1.22 0.00 1.66 Asp Trp 1.46 0.00 1.61 Asp-glu 1.27 0.00 1.33 Benzoic acid, 2,4,5-trimethoxy- 1.51 0.00 0.80 Benzoic acid, 4-hydroxy-, (1aS,2aR,3S,5R,6S,7aR)-3-(acetyloxy)decahydro-5-hydroxy-2a,7a-dimethyl-5-(1-methylethyl)azuleno[5,6-b]oxiren-6-ylester 0.43 0.02 1.18 beta-D-Glucopyranose,O-6-deoxy-alpha-L-mannopyranosyl-(1->4)-O-beta-D-glucopyranosyl-(1->6)-1-O-[(2alpha,3beta,5xi,6beta,9xi)-2,3,6,23-tetrahydroxy-28-oxoolean-12-en-28-yl]- 0.44 0.00 0.86 Bifenazate 0.50 0.00 1.22 Brefeldin A 0.75 0.01 1.16 Butein 1.16 0.00 0.75 Calycanthidine 0.78 0.01 1.23 Carnosol 0.65 0.04 0.83 CARβPENTANE 0.63 0.04 0.85 Cetyltrimethylammonium bromide 0.79 0.00 1.16 Chlorophorin 0.64 0.01 1.17 Chol-11-EnicAcid 0.49 0.00 2.20 Cinmethylin 0.51 0.02 1.15 cis-5-Tetradecenoylcarnitine 1.12 0.00 0.79 cis-6,7-epoxy-9Z-heptadecene 0.20 0.02 1.95 Cisapride 0.18 0.04 1.16 cis-Hinokiresinol 0.53 0.01 5.65 Clemastinefumarate14976-57-9 0.82 0.01 1.17 CYCLOPENTOLATEHYDROCHLORIDE 1.59 0.00 0.52 D-(+)-Octopine 0.66 0.00 0.62 Denatoniumbenzoate3734-33-6 1.32 0.00 0.69 DGDG 19:0 0.18 0.04 1.26 Dienestrol 1.02 0.00 14.97 Difluprednate 0.32 0.01 1.31 Dihomomethionine 0.55 0.00 1.35 Dihydroabieticacid 1.01 0.01 1.24 DiphenhydramineHCl (Benadryl) 1.10 0.00 1.38 Exemestane 0.70 0.02 1.16 Ferrioxamine H 1.09 0.00 1.19 Furanogermenone 1.06 0.00 1.25 Glaucarubin 0.68 0.01 1.20 Gln Leu 1.14 0.00 1.20 Glu Glu 0.84 0.02 1.41 Glu Met 0.51 0.04 1.17 Glu Trp 0.92 0.03 1.22 Glu Tyr 1.73 0.00 1.36 Glutaminylthreonine 0.93 0.00 0.49 Glu-Val-Phe 1.41 0.00 2.08 Gly Phe 0.51 0.00 1.31 Glycidylstearate 0.89 0.05 1.16 Guazatine 0.80 0.00 1.19 Heterodendrin 1.01 0.01 0.85 Hirsuteine 1.05 0.00 0.86 Hydroxyprolyl-Proline 1.62 0.00 1.39 Ichangin 0.75 0.00 1.20 Iminostilbene 0.53 0.02 1.59 Kinetin 0.93 0.00 1.20 L-165,04179558-09-1 1.37 0.00 1.98 L-Carnitine 1.78 0.00 0.75 Leu-Gly-Pro 0.75 0.00 0.80 Linoleoylethanolamide 0.92 0.00 0.85 L-Palmitoylcarnitine 0.55 0.00 0.84 Lysyl-Isoleucine 1.27 0.00 1.54 METHACHOLINE 2.02 0.00 0.69 MG 17:0 0.61 0.00 1.25 MG(0:0 / 22:6(4Z,7Z,10Z,13Z,16Z,19Z) / 0:0) 1.49 0.00 0.65 ML-236C 0.57 0.03 1.21 N-(3-morpholin-4-ylpropyl)-3-(2,3,5,9-tetramethyl-7-oxofuro[3,2-g]chromen-6-yl)propanamide 1.00 0.00 0.79 N,2,3-Trimethyl-2-(1-methylethyl)butanamide 0.54 0.02 1.18 N,N-Dimethyldodecylamine N-oxide 0.90 0.00 1.39 N1,N12-Diacetylspermine 0.62 0.01 1.51 N2-Acetyl-L-lysine 1.18 0.00 1.44 N-3-oxo-hexadec-11(Z)-enoyl-L-Homoserinelactone 0.44 0.03 1.20 N-Acetyl-L-histidine 1.20 0.00 3.55 N-Acetylputrescine 0.71 0.02 1.26 NAGly 0.13 0.00 0.75 Nigakilactone N 0.71 0.03 1.42 N-Oleyl-Leucine 1.50 0.00 0.81 N-γ-L-Glutamyl-L-phenylalanine 0.67 0.00 1.27 OleoylEthanolamide-d2 1.05 0.00 0.84 Oleoylcarnitine 0.60 0.00 0.81 Olopatadine 1.27 0.01 1.27 Oxymetazolinehydrochloride 1.16 0.01 0.85 Paclobutrazol 0.44 0.00 1.17 Paliperidone 0.97 0.00 1.19 Palmitoylethanolamide 1.80 0.00 0.83 Phenylgalactoside 1.37 0.00 1.62 Pivmecillinamhydrochloride32887-03-9 1.02 0.00 1.64 Prednisone 0.67 0.00 1.24 Pregabalin (JAN / USAN / INN) 1.50 0.00 0.74 Pristimerin 1.13 0.00 1.27 Promecarb 1.08 0.01 2.14 Pteroside A 0.78 0.01 1.58 Quinidine 0.35 0.04 1.43 R-4-benzyl-3-((R)-3-hydroxy-2,2-dimethyloct-7-ynoyl)-5,5-dimethyloxazolidin-2-one 0.33 0.05 1.49 Retapamulin(USAN / INN) 0.75 0.00 1.46 Retinal 1.10 0.01 1.19 Rifamycin W 0.27 0.01 0.81 Ser Lys 1.07 0.01 1.23 SKF-95282dimaleate salt 1.41 0.00 1.45 Sphinganine 1.02 0.00 0.84 Stearidonic acid 0.81 0.04 1.26 Steryl sulfate 0.79 0.01 1.16 Succinylmonocholine 1.23 0.00 1.50 Sulprostone(USAN / INN) 0.94 0.01 1.32 Taurohyocholate 0.45 0.04 1.17 Tebuthiuron 1.35 0.00 1.44 Tepraloxydim 0.48 0.01 1.17 Tetrahomomethionine 0.80 0.02 1.26 TG003 1.05 0.00 0.83 Threoninyl-Proline 0.68 0.01 0.79 trans-1,2-Diphenylcyclobutane 0.94 0.01 1.68 trans-Hexadec-2-enoyl carnitine 0.53 0.00 0.84 Trp His 0.17 0.04 1.18 Trp Thr 0.62 0.04 1.27 Tyr Glu 0.83 0.00 1.19 Val Lys 0.15 0.03 1.18 Vanillylamine 0.87 0.03 1.19 Vernoflexin 0.79 0.02 1.22 Volemitol 2.39 0.00 0.30 Yohimbinic acidmonohydrate522-87-2 1.41 0.00 0.59 (+)-threo-2-Amino-3,4-dihydroxybutanoic acid 1.13 0.00 0.82 (±)-Jasmonicacid 0.73 0.02 1.16 (1R)-Hydroxy-(2R)-glutathionyl-1,2-dihydronaphthalene 1.30 0.00 0.83 (2R,2'R,4a'S,6'S,8a'S)-4,6'-Dihydroxy-2',5',5',8a'-tetramethyl-6-oxo-3,3',4',4a',5',6,6',7,7',8,8',8a'-dodecahydro-2'H-spiro[furo[2,3-e]isoindole-2,1'-naphthalen]-7'-yl acetate 0.49 0.00 1.34 11-Oxo-androsteroneglucuronide 1.67 0.00 0.40 16-Glucuronide-estriol 1.10 0.00 1.64 17(S)-HpDoHE 0.36 0.00 0.78 1a,1b-dihomoPGF2α 0.01 0.04 1.17 2,6-DiaminooimelicAcid 1.43 0.00 0.86 20-ethyl-PGE2 0.77 0.01 1.36 20-hydroxylipoxinA4 0.38 0.00 1.15 2'-Deoxyinosine 0.72 0.02 2.85 2-Hydroxyquinoline 0.66 0.04 1.22 2-Methoxy-3-(4-methoxyphenyl)propanoic acid 0.94 0.00 0.86 3-Cyano-L-alanine 1.53 0.00 1.41 3-Dehydroshikimate 0.69 0.05 1.38 3-Nitrobenzoicacid 0.77 0.04 1.31 3-Oxoadipate 0.77 0.02 1.17 4-Methylene-2-oxoglutarate 0.70 0.01 5.20 5-(6-hydroxy-6-methyloctyl)furan-2(5H)-one 0.69 0.02 0.83 5,10-Methylenetetrahydrofolate 1.49 0.00 1.38 5,6-Dihydrouracil 1.18 0.00 1.22 5-Hydroxypentanoate 2.14 0.00 0.76 7(S),17(S)-dihydroxy-8(E),10(Z),13(Z),15(E),19(Z)-DocosapentaenoicAcid 0.79 0.04 1.24 8-Hexanoylneosolaniol 0.59 0.00 0.85 Aciclovir 0.91 0.01 1.25 Aspartyl-Isoleucine 1.61 0.00 1.52 Aspartyl-Leucine 1.31 0.00 1.24 Aspartyl-Proline 0.70 0.00 0.85 Aspartyl-Tryptophan 1.37 0.00 1.43 Aspartyl-Valine 1.34 0.00 1.59 Asp-Glu 0.77 0.00 1.25 Asp-phe 1.47 0.00 2.45 bonactin 0.80 0.03 1.21 Butylidenephthalide 0.46 0.00 1.77 Caffeic acid 1.45 0.00 0.74 Chamazulene 1.05 0.02 1.32 Ciprostene 0.75 0.00 1.48 D-Galactono-1,4-lactone 0.54 0.04 1.20 diethyl 2,6-dimethyl-4-oxo-4h-pyran-3,5-dicarboxylate 1.95 0.00 1.54 Dihydrocoumarin 1.40 0.00 0.67 DIMBOA 0.37 0.04 1.20 Dl-P-Hydroxyphenyllactic acid 1.57 0.00 0.78 D-Norvaline 0.26 0.04 0.59 ETF 0.51 0.01 1.82 Ethopabate 1.00 0.00 1.32 Flacourtin 0.21 0.02 0.83 Genistein 0.65 0.00 1.47 Gluconate 0.71 0.03 0.84 Glutamate, γ-methyl ester 1.42 0.00 1.80 Glycerophosphoric acid 1.35 0.00 1.32 Glycyl-Phenylalanine 1.04 0.01 1.36 Hovenitin I 0.87 0.01 1.16 Hypoglycin 1.06 0.00 1.33 Hypotaurine 1.01 0.00 1.38 Imiquimod (JAN / USAN / INN) 0.86 0.01 1.33 Kanzonol P 0.18 0.05 1.23 L-(-)-3-Phenyllacticacid 2.22 0.00 0.64 L-Argininephosphate 0.99 0.00 1.40 Leucyl-Alanine 0.61 0.00 1.16 Leucyl-Tryptophan 0.47 0.01 1.18 Leukotriene B4 0.26 0.03 1.24 Leukotriene C4 0.55 0.00 1.25 LPC 22:6 1.25 0.00 0.46 LPS 22:6 0.49 0.00 0.83 Lys-Leu-OH 1.27 0.00 1.60 lysoPC 16:0 0.31 0.00 1.15 L-β-aspartyl-L-glutamic acid 1.42 0.00 1.65 Magnolol 0.75 0.01 11.33 m-Cresol 1.29 0.00 0.84 Methacycline 1.34 0.00 2.15 Methyl (3R,4R,5R)-3,4-dihydroxy-5-({2-O-[(2E)-3-phenyl-2-propenoyl]-β-D-glucopyranosyl}oxy)-1-cyclohexene-1-carboxylate 0.46 0.00 0.85 Methyldopaanhydrous 0.81 0.03 1.16 Methylmalonate 0.82 0.00 1.87 Morphine 0.72 0.00 1.30 N-(1-Deoxy-1-fructosyl)tyrosine 0.84 0.03 1.30 N(γ)-Acetyldiaminobutyrate 0.54 0.02 1.58 N-Acetyl-Dl-glutamic acid 1.13 0.01 1.20 N-Acetyl-D-phenylalanine 1.72 0.00 0.86 N-Acetyl-L-leucine 1.06 0.02 1.16 N-Acetyl-L-methionine 1.31 0.00 1.38 Naringin 0.58 0.00 0.87 Neuraminic acid 0.85 0.00 0.83 N-Methylschinifoline 1.26 0.00 1.24 o-Cresol 1.22 0.00 0.82 O-Desmethylvenlafaxine glucuronide 1.99 0.00 0.45 Oleoyl glycine 0.51 0.04 1.25 Phe-Phe 0.68 0.03 1.26 SM d34:2 1.04 0.01 0.87 Succinic acid 0.84 0.00 2.19 Succinicanhydride 0.49 0.03 2.06 Threoninyl-Phenylalanine 0.74 0.02 1.21 Tiglylglycine 0.79 0.01 0.87 trans-Cinnamate 1.85 0.00 0.71 trans-p-Hydroxycinnamicacid 1.33 0.00 0.66 Triamcinolone16,21-diacetate 0.89 0.00 1.60 Tryptophyl-Aspartate 0.56 0.01 1.83 Valyl-Aspartate 0.86 0.00 1.35 Zedoarol 0.31 0.01 0.73 α-CYANO-3-HYDROXYCINNAMICACID 1.40 0.00 1.39 As an improvement of the above embodiment, the pre-processed metabolite data samples are analyzed based on the preset multi-dimensional screening criteria to determine the difference characteristics related to abnormal semen quality, including: The OPLS-DA method was used to screen metabolites with variable importance projection values ​​greater than 1; The Wilcoxon signed-rank test was used to screen metabolites with an FDR-corrected P value less than 0.05; Fold difference analysis was used to screen metabolites with fold differences greater than 2 or less than 0.5; The metabolites that met the above three screening criteria were regarded as significant differential metabolites, and the significant differential metabolites were regarded as differential features associated with abnormal semen quality.

[0023] In this embodiment, the pre-processed metabolite data samples are deeply mined by a variety of analytical methods to comprehensively and accurately determine the differential features related to abnormal semen quality. First, the OPLS-DA method is used to decompose the data matrix into multiple parts, calculate the variable importance projection (VIP), and screen out metabolites with a VIP value greater than 1. This method can effectively extract the predicted part related to the response variable and enhance the interpretability of the model, thereby preliminarily identifying metabolites that have an important impact on semen quality. Then, the Wilcoxon signed rank test is used. This non-parametric test method compares the distribution differences of the two groups of samples. Under the premise of not relying on the assumption of normal distribution of the data, it screens out metabolites with a P value less than 0.05 after FDR correction, which can effectively avoid erroneous screening caused by inconsistent data distribution assumptions, and further explore potential metabolites related to semen quality. Finally, the difference multiple analysis is used to calculate the ratio of metabolite expression levels between the two groups of samples, and the difference multiple is greater than 2 or less than 0.5 as the standard for screening, and the metabolites with significant changes are determined from the perspective of the degree of change. Based on the screening results of these three methods, metabolites that meet the three screening criteria at the same time are determined as significant differential metabolites, which are used as differential features closely related to abnormal semen quality, providing a key feature data basis for the subsequent construction of an accurate semen quality prediction model. Therefore, this embodiment combines three different principles of OPLS-DA, Wilcoxon signed rank test and differential multiple analysis to screen differential features. OPLS-DA analyzes metabolite data from the perspective of variable importance, Wilcoxon signed rank test from the perspective of distribution differences, and differential multiple analysis from the perspective of expression level changes, avoiding the limitations of a single method. In terms of technical effects, the accuracy and reliability of the correlation between the screened differential features and abnormal semen quality are improved, and it can focus more accurately on the metabolite features that really affect semen quality, providing high-quality input data for subsequent model construction, thereby improving the performance of the entire semen quality prediction model, so that it can more effectively assist doctors in evaluating patients' fertility potential, optimizing treatment plans, etc. in scenarios such as clinical practice.

[0024] For example, in the clinical assisted reproductive technology scenario, for male patients who come for semen testing, after obtaining their semen metabolite data samples and preprocessing through this technical step, the above three methods are used to screen differential features. For example, when analyzing the semen metabolite data of a patient, the OPLS-DA method may find that certain metabolites have a high importance for the prediction of semen quality (VIP>1), and these metabolites may be involved in key physiological processes such as sperm energy metabolism or cell structure composition; the Wilcoxon signed rank test can identify metabolites with significant differences in distribution (Pfdr<0.05) between abnormal and normal semen quality samples, which may involve metabolic pathways related to reproductive hormone regulation or inflammatory response; differential fold analysis highlights metabolites with significant changes in expression levels (FC>2 or FC<0.5) between the two groups of samples, such as some metabolites closely related to sperm motility or morphology maintenance. The differential characteristics determined by these screening results are used to construct a semen quality prediction model. Doctors can use this to more accurately judge the patient's semen quality and predict his fertility potential, thereby providing an important basis for formulating personalized treatment plans. For example, for patients with poor semen quality predictions, intervention measures can be taken in advance, such as improving lifestyle habits, providing nutritional supplements or targeted treatments.

[0025] For ease of understanding, this embodiment is described in detail below: In order to identify the differential metabolites between the semen quality abnormality case group and the control group, this example combined three methods: OPLS-DA, Wilcoxon signed rank test and difference fold analysis. The screening criteria are as follows: VIP>1: selected according to the variable importance value of the OPLS-DA model; Pfdr<0.05: P value selected based on FDR adjustment for multiple hypothesis testing; FC>2 or FC<0.5: The fold change is greater than 2 or less than 0.5.

[0026] Based on the above criteria, the final significantly different metabolites were the intersection of the three criteria. Ten significantly different metabolites were screened out from 259 seminal plasma metabolites ( Figure 2 , Table 2, Table 3). These metabolites will serve as core variables for subsequent analysis.

[0027] 2.1 OPLS-DA method screening Orthogonal Projections to Latent Structures Discriminant Analysis (OPLS-DA) is a supervised learning method based on multivariate statistics for classification and explanation of inter-group differences. It is an extension of partial least squares discriminant analysis (PLS-DA) that enhances the interpretability of the model by separating the predictive part of variable X (related to the response variable Y) from the orthogonal part (not related to Y).

[0028] 2.1.1 Core algorithm: Decompose the data matrix X into three parts: prediction part, orthogonal part and residual: , : The prediction score matrix associated with Y; : predicted loading matrix; : an orthogonal fractional matrix independent of Y; : orthogonal loading matrix; : residual matrix.

[0029] 2.1.2 Utilization Information removal The orthogonal variation part that is not related to YYY in the matrix is ​​obtained by removing the orthogonal components: , 2.1.3 Based on the prediction matrix Calculate the response variable The predicted value of , used for classification and explanation of differences between groups. Variable Importance in Projection (VIP) is used to assess the importance of each variable to the model's predictive ability.

[0030] VIP value formula: , : total number of variables; : Principal component number; : No. The scores of the principal components; : No. The variable in The weights on the principal components; : No. The sum of squares of the response variable explained by the principal components; : The sum of squares of the total response variables. The screening criterion is VIP>1, indicating that the variable contributes significantly to the classification.

[0031] 2.2 Wilcoxon signed-rank test WilcoxonRankSumTest, also known as the Mann-WhitneyU test, is a nonparametric test method used to compare the distribution differences of two independent groups of samples. Unlike the t-test, the WilcoxonRankSumTest does not assume that the data comes from a normal distribution, so it is particularly useful when the sample size is small and the data distribution is uncertain. FDR correction is used to control the false positive rate of multiple hypothesis testing, and the screening criterion is Pfdr<0.05.

[0032] 2.2.1 Hypothesis: Null hypothesis (H0): The two groups of samples come from the same distribution.

[0033] Alternative hypothesis (H1): The two groups of samples come from different distributions.

[0034] 2.2.2 Data merging and sorting: After merging the two sets of sample data, sort all the data and assign ranks from small to large. If there are identical values ​​in the data (i.e. repeated ranks), use the average rank for processing.

[0035] 2.2.3 Calculate the rank sum: Assign a rank to each data point , and calculate the rank sum of the two groups of samples respectively: , Select the rank of the smaller sample group: .

[0036] 2.2.4 Calculation of the test statistic For small samples, the U statistic is used to calculate the test statistic: ; and : The size of the two sample groups; : The rank sum of the smaller sample; Use the precomputed Wilcoxon U test table to find the p-value corresponding to U.

[0037] For large samples, use the normal distribution approximation value: ; (Expectation of rank sum): ; (Standard deviation of the rank sum): ; According to the calculated Value, look up the standard normal distribution table to obtain value.

[0038] 2.3. Fold Difference Analysis Fold Change (FC) is a statistical method used to compare the difference of a variable (such as metabolite expression, gene expression, etc.) between two groups of samples (such as experimental group and control group). The core idea is to calculate the ratio of the expression level of each variable in the two groups of samples to measure the degree of change of the variable.

[0039] Basic formula: ;X 实验组 : The value of the individual variable of the experimental group; X 对照组 : The value of the individual variable of the control group; for the convenience of symmetric analysis and visualization, the logarithm of FC is usually taken: ; Standards: log2(FC)>1: indicates significant upregulation (FC>2); log2(FC)<-1: indicates significant downregulation (FC<0.5).

[0040] For example, see Figure 2 The final screened differential metabolites are shown in Table 2: Table 2. Differential metabolites in the screened semen quality abnormality case group and control group

[0041] As an improvement of the above embodiment, the preset feature screening algorithm is used to screen and evaluate the difference features, and the core feature subset that meets the semen quality prediction requirements is determined according to the preset optimization index, including: Using the Lasso regression method, by introducing the L1 regularization term in the loss function, the regression coefficient is dynamically adjusted so that the coefficients of irrelevant difference features shrink to zero, retaining the difference features that have predictive power for the response variable; Using the support vector machine recursive feature elimination method, the support vector machine classifier is trained based on the current differential feature set, the importance of the differential features is calculated, the differential features with the smallest weight are removed, and the above steps are repeated until the remaining differential features reach the predetermined number, thereby obtaining a core feature subset that meets the requirements for semen quality prediction.

[0042] In the present embodiment, the determined differential features are further processed by two complementary feature screening algorithms to obtain the core feature subset with the most predictive value. First, the Lasso regression method is adopted, and its core lies in constructing a loss function including prediction error minimization and L1 regularization term. During the training process, the L1 regularization term performs a weighted penalty on the absolute value of the regression coefficient according to the regularization hyperparameter set, so that the coefficient of the irrelevant differential features gradually approaches zero, and finally shrinks to zero, while the differential feature coefficient with high correlation to the response variable and predictive ability is retained, so as to achieve the purpose of screening out the key part from many differential features. Then the support vector machine recursive feature elimination method is used, and the support vector machine classifier is trained based on the current differential feature set, and the contribution weight of each differential feature to the classification decision surface is calculated by the classifier to measure the importance of the differential feature, and then the differential feature with the smallest weight is removed, and the process of training, calculation and removal is repeated until the remaining differential features reach a predetermined number. Through the synergy of these two algorithms, the differential features are comprehensively evaluated and screened, and the core feature subset that meets the requirements of semen quality prediction is finally determined, so as to provide core data support for building an accurate and efficient semen quality prediction model. Therefore, this embodiment comprehensively uses two different feature screening algorithms of Lasso regression and support vector machine recursive feature elimination. Lasso regression performs feature selection in high-dimensional data based on the principle of regularization, and has the advantages of automatic screening and model simplification; support vector machine recursive feature elimination uses the powerful classification ability of support vector machine to recursively screen according to feature importance. The combination of the two overcomes the limitations that may exist in a single algorithm. In terms of technical effects, it can effectively remove redundant and irrelevant difference features, reduce data dimensions, highlight features that have important contributions to semen quality prediction, improve the representativeness and predictive ability of core feature subsets, and thus improve the accuracy and efficiency of the entire semen quality prediction model. In clinical diagnosis and research, semen quality can be predicted more accurately, and doctors can be assisted to make more scientific decisions, such as early diagnosis of male infertility and providing a basis for personalized treatment.

[0043] For example, in the male infertility research scenario, for a large amount of semen sample data collected, after the differential features are determined through preliminary processing, this technical step is used for feature screening. For example, when studying the differential features of semen metabolites in a group of patients with oligospermia and asthenospermia, Lasso regression begins to intervene. By continuously adjusting the regularization parameters, the coefficients corresponding to those metabolite differential features that do not contribute much to the prediction of semen quality (such as some small molecule metabolite features with weak correlation in complex metabolic networks) are gradually shrunk to zero, retaining features that have an important impact on semen quality, such as metabolite differential features related to key enzymes of sperm energy metabolism. Subsequently, the support vector machine recursive feature elimination method comes into play, and the support vector machine classifier is constructed with these initially screened differential features, and the importance weight of each feature is calculated. Assume that in this process, some metabolite differential features related to inflammatory response but with relatively small weights in the overall semen quality prediction are found, and they are gradually removed. After multiple iterations, the core feature subset is finally determined, such as specific metabolite features closely related to sperm production and vitality. Based on the semen quality prediction model constructed based on this core feature subset, researchers can more accurately analyze the mechanism of semen quality abnormalities in this group of patients, and doctors can also more accurately predict the semen quality status of other similar patients.

[0044] In the above embodiment, specifically, the Lasso regression method is used to dynamically adjust the regression coefficient by introducing an L1 regularization term in the loss function so that the coefficients of irrelevant difference features shrink to zero and retain the difference features that have predictive power for the response variable, including: Construct a loss function that minimizes the prediction error and the L1 regularization term: , is the sample size; is the number of features; is the intercept; is the regression coefficient of the jth feature; is the value of the jth feature in the i-th sample; is the response variable of the i-th sample; is a regularization hyperparameter that controls The intensity of the penalty; among them, minimizing the prediction error is used to measure the quality of model fitting, and the L1 regularization term shrinks the coefficients of unimportant difference features to zero by weighted penalty on the absolute value of the regression coefficient; During the training process, the regression coefficients are dynamically adjusted so that the coefficients of irrelevant differential features shrink to zero, while the coefficients of relevant features maintain a larger absolute value. Ultimately, only differential features that have predictive power for the response variable are retained, and the optimal regularization parameter is selected through cross-validation to achieve a balance between model sparsity and predictive power.

[0045] For ease of understanding, this embodiment is described in detail below: Lasso regression (Least Absolute Shrinkage and Selection Operator) is a linear regression method that introduces Regularization term, to achieve the purpose of feature selection and model simplification. It is particularly effective in high-dimensional data analysis because Regularization shrinks some regression coefficients to 0, thereby automatically selecting the most predictive features.

[0046] 3.1.1 Loss Function Definition The loss function of Lasso regression consists of two parts: Minimize the prediction error (residual sum of squares): a measure of how well the model fits.

[0047] Add to Regularization term: Limits model complexity by penalizing the regression coefficients.

[0048] The loss function formula is: , in: : sample size; : number of features; :intercept; : The regression coefficient of the jth feature; : The value of the jth feature in the i-th sample; : The response variable of the i-th sample; : Regularization hyperparameter, control The intensity of the punishment.

[0049] 3.1.2 Core Mechanism of Feature Selection 3.1.2.1 The role of regularization Regularization makes the coefficients of unimportant features approach zero or even exactly zero by weighting the absolute value of the regression coefficient. When it is larger, the penalty effect is stronger, a large number of features will be eliminated, and the model will be sparser. When , Lasso regression degenerates into ordinary linear regression.

[0050] 3.1.2.2 Key Features The regularization constraint of Lasso enables feature selection to be completed by solving a constrained optimization problem: , in is a Determined constant. Through geometric interpretation, The regularized constraint is a diamond-shaped constraint region whose sharp corners tend to make the regression coefficients equal to zero.

[0051] 3.1.2.3 Screening process During the training process, Lasso regression dynamically adjusts the regression coefficients so that the coefficients of irrelevant features shrink to zero and the coefficients of relevant features maintain a larger absolute value, ultimately retaining only the response variable. Features with predictive power.

[0052] 3.1.3 Optimization Algorithm of Lasso Regression The optimal regularization parameter is selected through cross-validation (CV) to achieve a balance between model sparsity and predictive ability. The analysis effect of the optimized Lasso regression can be found in Figure 3 .

[0053] In the above embodiment, specifically, the support vector machine recursive feature elimination method is used to train a support vector machine classifier based on the current difference feature set, calculate the importance of the difference feature, remove the difference feature with the smallest weight, and repeat the above steps until the remaining difference features reach a predetermined number, thereby obtaining a core feature subset that meets the requirements for semen quality prediction, including: Input the difference feature set and the response variable, wherein the difference feature set includes the significantly different metabolites screened by OPLS-DA, Wilcoxon signed rank test and difference fold analysis; the response variable is the classification label of the semen quality prediction model for semen quality prediction; Using a support vector machine algorithm, a classifier is trained based on the current differential feature set and the response variable to construct an optimal classification decision surface, wherein the classification decision surface is used to distinguish different categories of semen quality samples; The weight vector of each difference feature is calculated by the support vector machine model. The weight vector represents the contribution of each difference feature to the classification decision surface. The larger the absolute value of the weight, the greater the contribution of the feature to the classification. According to the weight vector, the importance of each difference feature is evaluated, and the importance is measured by the absolute value of the weight vector, and the difference feature with the smallest weight is removed; The set of differential features is updated, and the process of training the support vector machine classifier, calculating feature importance, and removing the differential features with the smallest weight is repeated until the remaining differential features reach a predetermined number; The final core feature subset is output, which includes the differential features that contribute most to the prediction of semen quality.

[0054] For ease of understanding, this embodiment is described in detail below: The core idea of ​​SVM-RFE is to use the classification ability of support vector machines to recursively remove the least important features according to the contribution of the features to the classification decision surface, thereby achieving feature selection. Support vector machine (SVM) constructs an optimal hyperplane to distinguish data of different categories. The influence of features on the position of the hyperplane is measured by their corresponding weight vectors. Features with large weights (large absolute values) contribute more to classification; features with small weights (small absolute values) contribute less to classification.

[0055] Recursive feature elimination (RFE) trains the model based on the current feature set, calculates the importance (weight) of the features, removes the features with the smallest weight, and repeats the above steps until the target number of features is reached, thereby achieving feature screening.

[0056] 3.2.1 Initial feature set: input feature matrix and the response variable ,in is the number of samples, is the initial feature number.

[0057] 3.2.2 Training SVM model: Use the current feature set to train a support vector machine classifier and obtain the weight vector For linear kernel SVM, the weight vector directly represents the importance of the feature. For nonlinear kernel SVM, the feature importance can be calculated by explicitly mapping the input features.

[0058] 3.2.3 Calculate feature importance: Feature importance is defined by the following formula: ; It is The weight of a feature.

[0059] 3.2.4 Recursive feature elimination: remove features with minimum Update feature collection , and retrain the SVM.

[0060] 3.4.5 Repeat steps 3.2.2-3.2.4 until the remaining features reach the predetermined number.

[0061] 3.4.6 Output the final feature set: Sort the features from high to low in importance and select the final target features.

[0062] The analysis results of the SVM-RFE algorithm can be found in Figure 4 .

[0063] Finally, the top 10 features selected in both methods constitute the core feature subset of this embodiment (as shown in Table 3). The feature selection process uses the regression coefficient and AUC value as optimization indicators. By screening non-zero features and comparing the AUC values ​​of different feature subsets, the two models are combined to select the 9 most predictive feature subsets, see Table 3, thereby removing the variables that contribute less to the model and ensuring the simplicity and efficiency of the model.

[0064] Table 3. Feature selection and optimization results

[0065] In summary, participating Figure 3 and Figure 4 , by comprehensively using two different feature screening algorithms: Lasso regression and support vector machine recursive feature elimination. Lasso regression performs feature selection in high-dimensional data based on the principle of regularization, and has the advantages of automatic screening and model simplification; support vector machine recursive feature elimination uses the powerful classification ability of the support vector machine to recursively screen according to feature importance. The combination of the two overcomes the limitations that may exist in a single algorithm. In terms of technical effects, it can effectively remove redundant and irrelevant difference features, reduce data dimensions, highlight features that have important contributions to semen quality prediction, improve the representativeness and predictive ability of the core feature subset, and thus improve the accuracy and efficiency of the entire semen quality prediction model. In clinical diagnosis and research, it can more accurately predict semen quality and assist doctors in making more scientific decisions, such as early diagnosis of male infertility and providing a basis for personalized treatment.

[0066] As an improvement of the above embodiment, the pre-built semen quality prediction model based on the random forest algorithm is trained based on the core feature subset, including: Constructing a new training set and test set from the core feature subset; The constructed new training set is used to train the random forest model as a semen quality prediction model; during the training process, samples are randomly sampled from the training set with replacement to construct multiple different training subsets to increase sample diversity; and feature random sampling is used to randomly select some features for training each time a decision tree is constructed to reduce model complexity; at the same time, the performance of the model is optimized by adjusting specific hyperparameters of the model; specific hyperparameters include the number of decision trees and the maximum depth of the random forest model; The sample data in the test set is input into the trained random forest model. The random forest model predicts the semen quality of each test sample based on the learned correlation pattern between the core features and semen quality, and outputs the corresponding prediction results. The 10-fold cross validation method was used to evaluate the performance of the random forest model on the training set, and to evaluate the sensitivity, specificity, and AUC value of the random forest model.

[0067] In this embodiment, firstly, a new training set and test set are constructed from the core feature subset to ensure that only key features highly related to semen quality are included in the data set, providing an accurate data basis for model training. Then, the random forest model is trained using the new training set. During the training process, the random sampling of samples generates multiple different training subsets by extracting samples with replacement, so that the model can be exposed to a variety of sample distributions and enhance the adaptability to different sample situations; the random sampling of features randomly selects some features each time a decision tree is constructed to avoid the model from over-relying on specific features, effectively reducing the complexity of the model. At the same time, the model performance is optimized by adjusting specific hyperparameters such as the number of decision trees and the maximum depth, so that the model is better able to learn the correlation pattern between core features and semen quality. Afterwards, the test set samples are input into the trained model for prediction and the results are output. Finally, a 10-fold cross validation is used to comprehensively evaluate the performance of the model on the training set, calculate indicators such as sensitivity, specificity and AUC values, measure the model performance from different angles, and ensure the validity and reliability of the model. Therefore, this embodiment trains the random forest model by combining random sampling of samples, random sampling of features and hyperparameter optimization. Random sampling of samples and features increases the robustness and generalization ability of the model and avoids overfitting, while hyperparameter optimization makes the model performance reach a better state. In terms of technical effects, the accuracy of the model's prediction of semen quality is improved, and it can more accurately identify the status of semen quality, which can assist doctors in making more reasonable decisions in clinical applications. For example, in assisted reproductive technology, it helps to select more suitable embryos for transplantation and improve the success rate; in the diagnosis of male infertility, it can detect potential problems earlier, intervene in treatment in time, reduce unnecessary waste of medical resources, and improve the patient's reproductive health management level.

[0068] For example, in the daily work scenarios of reproductive medicine centers, for the patient group who comes for semen examination, after the core feature subset is obtained through preliminary processing, the steps of this technology are followed. For example, when constructing training sets and test sets, relevant data is selected from the semen data of many patients based on core features. When training the random forest model, random sampling of samples enables the model to learn under different combinations of patient samples. For example, different proportions of samples of young healthy men and older infertile men are extracted at a certain time, so that the model can adapt to the characteristics of samples of different ages and fertility conditions; random sampling of features allows the model to focus on different core feature combinations when constructing decision trees. Suppose that features related to sperm motility and morphology are randomly selected for training at a certain time. The model is optimized by adjusting hyperparameters such as the number of decision trees to 100 and the maximum depth to 8. When testing semen samples of new patients, the model can accurately predict their semen quality. For example, for a patient with fertility problems, the model predicts that there may be problems with his semen quality based on the core characteristics of his semen. The doctor can then conduct further examinations and develop personalized treatment plans based on this, such as suggesting improved lifestyle habits or targeted treatments to increase the patient's chances of fertility. At the same time, a 10-fold cross-validation is used to ensure that the model also has good predictive performance for other similar patients, thus ensuring the reliability of the model in practical applications.

[0069] For ease of understanding, this embodiment is described in detail below: In order to ensure the generalization ability of the model, the research data is divided into 70% as a training set and 30% as a test set. The training set is used to build and tune the model, and the test set is used to evaluate the performance of the model on unseen data. The 10-fold cross-validation technique is used to further evaluate the model performance of the training set. Through multiple rounds of training and validation, the stability of the model can be ensured and the performance of the model can be further optimized.

[0070] This example constructs a variety of common machine learning models and compares their performance. These models include: Random Forest (RF), XGBoost, NeuralNetwork, Lightweight Gradient Boosting Machine (LightGBM). By training on cohort study data, the prediction effects of different models were compared and the best performing model was selected.

[0071] 5. Model Performance Evaluation During model evaluation, multiple performance metrics were used, including: Sensitivity: reflects the model's ability to identify positive samples; Specificity: reflects the model's ability to identify negative samples; AUC (Area Under Curve): Measures the overall classification ability of the model, especially in the case of class imbalance.

[0072] According to the above evaluation indicators, the optimal model combination is determined. As shown in Table 4, among the four commonly used machine learning models, the random forest (RF) model performs better. Figure 5 As shown in the figure, among the four commonly used machine learning models, the random forest (RF) model performed best, with an AUC value of 0.823, a sensitivity of 0.821, and a specificity of 0.679.

[0073] Table 4. Prediction results of four machine learning models

[0074] Random Forest is an algorithm based on ensemble learning. It improves the prediction performance and robustness of the model by building multiple decision trees. Its core idea is to reduce the variance and overfitting risk of a single decision tree through random sampling of samples (Bootstrap Sampling) and random sampling of features (Feature Sampling), while using ensemble prediction to obtain higher classification or regression accuracy. The Random Forest model has the following significant features: Efficient processing of high-dimensional data: Randomly selected feature subsets reduce the computational complexity, making it suitable for high-dimensional data.

[0075] Feature importance evaluation: The model calculates the importance of each feature by its contribution to the split purity, providing a basis for subsequent feature screening.

[0076] Robustness and resistance to overfitting: Enhance the generalization ability of the model through sub-sampling and feature randomness.

[0077] The application process of the random forest model after Lasso regression and support vector machine recursive feature elimination (SVM-RFE) screening features: The filtered feature subset Applied to build a more efficient random forest model, including the following steps: 1. Data preparation Only the selected feature subset is retained in the original dataset , construct a new training set and test set .

[0078] 2. Random Forest Model Training and Prediction Train a random forest model using the filtered data: , The training process is still based on random sampling of samples and features, and optimizes model performance through specific hyperparameters (such as the number of decision trees and maximum depth).

[0079] For the test set samples Make predictions: .

[0080] 3. Model Validation and Performance Evaluation The model performance was evaluated using 10-fold cross validation and compared with the original model to verify the effectiveness of the selected features. Performance indicators include sensitivity, specificity, AUC, etc. AUC is calculated by constructing the ROC curve: .

[0081] See also Figure 6 , is a schematic diagram of a semen quality prediction system based on random forests according to an embodiment of the present invention. The semen quality prediction system based on random forests comprises: The acquisition module 10 is used to acquire metabolite data samples of different semens and pre-process the metabolite data samples to obtain pre-processed metabolite data samples; the metabolite data samples are obtained by pre-detecting and analyzing different semens; An analysis module 11 is used to analyze the pre-processed metabolite data samples based on a preset multi-dimensional screening standard to determine the difference characteristics related to abnormal semen quality; A screening module 12 is used to screen and evaluate the differential features using a preset feature screening algorithm, and determine a core feature subset that meets the semen quality prediction requirements based on a preset optimization index; A training module 13 is used to train a pre-built semen quality prediction model based on a random forest algorithm based on the core feature subset; The prediction module 14 is used to use the trained semen quality prediction model to perform quality prediction on the metabolite data of the semen to be tested, and obtain a semen quality prediction result.

[0082] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: First, the metabolite data samples of different semen obtained by detection and analysis are obtained and preprocessed to provide reliable basic data. Then, the preprocessed data is deeply analyzed according to the preset multi-dimensional screening criteria, and the difference features related to abnormal semen quality are accurately located, providing a key basis for subsequent accurate prediction. Then, the preset feature screening algorithm is used to screen and evaluate these difference features, and the core feature subset is determined according to the optimization index, which effectively reduces the data dimension and highlights the key information. Based on this core feature subset, the semen quality prediction model of the pre-constructed random forest algorithm is trained, and the generalization ability and efficiency of the random forest algorithm are utilized and the advantages of random forest processing high-dimensional data, evaluating feature importance, and resisting overfitting are fully utilized. Finally, the trained model is used to predict the semen metabolite data to be tested, and the semen quality prediction result is obtained. Therefore, the embodiment of the present invention can more accurately screen out key features, effectively process complex data relationships, and construct a semen quality prediction model with high prediction accuracy and good generalization ability through the combination of multi-dimensional screening criteria, efficient feature screening algorithms and random forest models with strong generalization ability, thereby effectively improving the accuracy of semen quality prediction.

[0083] As an improvement of the above embodiment, the analysis module is specifically used for: The OPLS-DA method was used to screen metabolites with variable importance projection values ​​greater than 1; The Wilcoxon signed-rank test was used to screen metabolites with an FDR-corrected P value less than 0.05; Fold difference analysis was used to screen metabolites with fold differences greater than 2 or less than 0.5; The metabolites that met the above three screening criteria were regarded as significant differential metabolites, and the significant differential metabolites were regarded as differential features associated with abnormal semen quality.

[0084] As an improvement of the above embodiment, the screening module is specifically used for: Using the Lasso regression method, by introducing the L1 regularization term in the loss function, the regression coefficient is dynamically adjusted so that the coefficients of irrelevant difference features shrink to zero, retaining the difference features that have predictive power for the response variable; Using the support vector machine recursive feature elimination method, the support vector machine classifier is trained based on the current differential feature set, the importance of the differential features is calculated, the differential features with the smallest weight are removed, and the above steps are repeated until the remaining differential features reach the predetermined number, thereby obtaining a core feature subset that meets the requirements for semen quality prediction.

[0085] As an improvement of the above embodiment, the training module is specifically used for: Constructing a new training set and test set from the core feature subset; The constructed new training set is used to train the random forest model as a semen quality prediction model; during the training process, samples are randomly sampled from the training set with replacement to construct multiple different training subsets to increase sample diversity; and feature random sampling is used to randomly select some features for training each time a decision tree is constructed to reduce model complexity; at the same time, the performance of the model is optimized by adjusting specific hyperparameters of the model; specific hyperparameters include the number of decision trees and the maximum depth of the random forest model; The sample data in the test set is input into the trained random forest model. The random forest model predicts the semen quality of each test sample based on the learned correlation pattern between the core features and semen quality, and outputs the corresponding prediction results. The 10-fold cross validation method was used to evaluate the performance of the random forest model on the training set, and to evaluate the sensitivity, specificity, and AUC value of the random forest model.

[0086] It can be understood that the various embodiments of the semen quality prediction system based on random forest can correspond to the contents of the relevant embodiments of the semen quality prediction method based on random forest mentioned above, which will not be elaborated here.

[0087] Wherein, if the module / unit integrated in the semen quality prediction system based on random forest is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media does not include electrical carrier signals and telecommunication signals.

[0088] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.

[0089] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A semen quality prediction method based on random forest, characterized in that: The following steps are involved: Obtaining metabolite data samples of different semens, and preprocessing the metabolite data samples to obtain preprocessed metabolite data samples; the metabolite data samples are obtained by pre-detecting and analyzing different semens; The pre-processed metabolite data samples were analyzed based on the preset multi-dimensional screening criteria to determine the differential features associated with abnormal semen quality; The preset feature screening algorithm is used to screen and evaluate the differential features, and the core feature subset that meets the requirements for semen quality prediction is determined based on the preset optimization indicators; Based on the core feature subset, a pre-built semen quality prediction model based on a random forest algorithm is trained; The trained semen quality prediction model is used to predict the quality of the metabolite data of the semen to be tested, and the semen quality prediction result is obtained.

2. The semen quality prediction method based on random forest as claimed in claim 1, characterized in that: The pre-processed metabolite data samples are analyzed based on the preset multi-dimensional screening criteria to determine the difference characteristics related to abnormal semen quality, including: The OPLS-DA method was used to screen metabolites with variable importance projection values ​​greater than 1; The Wilcoxon signed-rank test was used to screen metabolites with an FDR-corrected P value less than 0.05; Fold difference analysis was used to screen metabolites with fold differences greater than 2 or less than 0.5; The metabolites that met the above three screening criteria were regarded as significant differential metabolites, and the significant differential metabolites were regarded as differential features associated with abnormal semen quality.

3. The semen quality prediction method based on random forest as claimed in claim 2, characterized in that: The method uses a preset feature screening algorithm to screen and evaluate the differential features, and determines a core feature subset that meets the semen quality prediction requirements based on preset optimization indicators, including: Using the Lasso regression method, by introducing the L1 regularization term in the loss function, the regression coefficient is dynamically adjusted so that the coefficients of irrelevant difference features shrink to zero, retaining the difference features that have predictive power for the response variable; Using the support vector machine recursive feature elimination method, the support vector machine classifier is trained based on the current differential feature set, the importance of the differential features is calculated, the differential features with the smallest weight are removed, and the above steps are repeated until the remaining differential features reach the predetermined number, thereby obtaining a core feature subset that meets the requirements for semen quality prediction.

4. The method for predicting semen quality based on random forest as claimed in claim 3, characterized in that: The Lasso regression method introduces an L1 regularization term into the loss function to dynamically adjust the regression coefficient so that the coefficients of irrelevant difference features shrink to zero, retaining the difference features that have predictive power for the response variable, including: Construct a loss function that minimizes the prediction error and the L1 regularization term: Construct a loss function that minimizes the prediction error and the L1 regularization term: , is the sample size; is the number of features; is the intercept; is the regression coefficient of the jth feature; is the value of the jth feature in the i-th sample; is the response variable of the i-th sample; is a regularization hyperparameter that controls The intensity of the penalty; among them, minimizing the prediction error is used to measure the quality of model fitting, and the L1 regularization term shrinks the coefficients of unimportant difference features to zero by weighted penalty on the absolute value of the regression coefficient; During the training process, the regression coefficients are dynamically adjusted so that the coefficients of irrelevant differential features shrink to zero, while the coefficients of relevant features maintain a larger absolute value. Ultimately, only differential features that have predictive power for the response variable are retained, and the optimal regularization parameter is selected through cross-validation to achieve a balance between model sparsity and predictive power.

5. The method for predicting semen quality based on random forest as claimed in claim 4, characterized in that: The support vector machine recursive feature elimination method is used to train a support vector machine classifier based on the current difference feature set, calculate the importance of the difference feature, remove the difference feature with the smallest weight, and repeat the above steps until the remaining difference features reach a predetermined number, thereby obtaining a core feature subset that meets the requirements for semen quality prediction, including: Input the difference feature set and the response variable, wherein the difference feature set includes the significantly different metabolites screened by OPLS-DA, Wilcoxon signed rank test and difference fold analysis; the response variable is the classification label of the semen quality prediction model for semen quality prediction; Using a support vector machine algorithm, a classifier is trained based on the current differential feature set and the response variable to construct an optimal classification decision surface, wherein the classification decision surface is used to distinguish different categories of semen quality samples; The weight vector of each difference feature is calculated by the support vector machine model. The weight vector represents the contribution of each difference feature to the classification decision surface. The larger the absolute value of the weight, the greater the contribution of the feature to the classification. According to the weight vector, the importance of each difference feature is evaluated, and the importance is measured by the absolute value of the weight vector, and the difference feature with the smallest weight is removed; The difference feature set is updated, and the process of training the support vector machine classifier, calculating feature importance, and removing the difference feature with the smallest weight is repeated until the remaining difference features reach a predetermined number; The final core feature subset is output, which includes the differential features that contribute most to the prediction of semen quality.

6. The method for predicting semen quality based on random forests according to claim 1, characterized in that: The method of training a pre-built semen quality prediction model based on a random forest algorithm based on the core feature subset comprises: Constructing a new training set and test set from the core feature subset; The constructed new training set is used to train the random forest model as a semen quality prediction model; during the training process, samples are randomly sampled from the training set with replacement to construct multiple different training subsets to increase sample diversity; and feature random sampling is used to randomly select some features for training each time a decision tree is constructed to reduce model complexity; at the same time, the performance of the model is optimized by adjusting specific hyperparameters of the model; specific hyperparameters include the number of decision trees and the maximum depth of the random forest model; The sample data in the test set is input into the trained random forest model. The random forest model predicts the semen quality of each test sample based on the learned correlation pattern between the core features and semen quality, and outputs the corresponding prediction results. The 10-fold cross validation method was used to evaluate the performance of the random forest model on the training set, and to evaluate the sensitivity, specificity, and AUC value of the random forest model.

7. A semen quality prediction system based on random forest, characterized in that: include: An acquisition module, used for acquiring metabolite data samples of different semens, and preprocessing the metabolite data samples to obtain preprocessed metabolite data samples; The metabolite data samples are obtained by testing and analyzing different semen in advance; An analysis module is used to analyze the pre-processed metabolite data samples based on preset multi-dimensional screening criteria to determine the differential features associated with abnormal semen quality; A screening module is used to screen and evaluate differential features using a preset feature screening algorithm, and determine a core feature subset that meets the requirements for semen quality prediction based on preset optimization indicators; A training module, used for training a pre-built semen quality prediction model based on a random forest algorithm based on the core feature subset; The prediction module is used to use the trained semen quality prediction model to perform quality prediction on the metabolite data of the semen to be tested to obtain the semen quality prediction result.

8. The semen quality prediction system based on random forest according to claim 7, characterized in that: The analysis module is specifically used for: The OPLS-DA method was used to screen metabolites with variable importance projection values ​​greater than 1; The Wilcoxon signed-rank test was used to screen metabolites with an FDR-corrected P value less than 0.05; Fold difference analysis was used to screen metabolites with fold differences greater than 2 or less than 0.5; The metabolites that met the above three screening criteria were regarded as significant differential metabolites, and the significant differential metabolites were regarded as differential features associated with abnormal semen quality.

9. The semen quality prediction system based on random forest according to claim 8, characterized in that: The screening module is specifically used for: Using the Lasso regression method, by introducing the L1 regularization term in the loss function, the regression coefficient is dynamically adjusted so that the coefficients of irrelevant difference features shrink to zero, retaining the difference features that have predictive power for the response variable; Using the support vector machine recursive feature elimination method, the support vector machine classifier is trained based on the current differential feature set, the importance of the differential features is calculated, the differential features with the smallest weight are removed, and the above steps are repeated until the remaining differential features reach the predetermined number, thereby obtaining a core feature subset that meets the requirements for semen quality prediction.

10. The semen quality prediction system based on random forest according to claim 9, characterized in that: The training module is specifically used for: Constructing a new training set and test set from the core feature subset; The constructed new training set is used to train the random forest model as a semen quality prediction model; during the training process, samples are randomly sampled from the training set with replacement to construct multiple different training subsets to increase sample diversity; and feature random sampling is used to randomly select some features for training each time a decision tree is constructed to reduce model complexity; at the same time, the performance of the model is optimized by adjusting specific hyperparameters of the model; specific hyperparameters include the number of decision trees and the maximum depth of the random forest model; The sample data in the test set is input into the trained random forest model. The random forest model predicts the semen quality of each test sample based on the learned correlation pattern between the core features and semen quality, and outputs the corresponding prediction results. The 10-fold cross validation method was used to evaluate the performance of the random forest model on the training set, and to evaluate the sensitivity, specificity, and AUC value of the random forest model.

Images