A method for predicting pregnancy in patients with pathological ovarian aging

Abstract

The present invention relates to the technical field of pregnancy prediction, and discloses a method for predicting pregnancy in patients with pathological ovarian aging. The method comprises the following steps: collecting sample data, obtaining sample features and converting and interpolating them; using the Boruta algorithm to screen candidate predictor variables for the formed complete sample, and obtaining FSH trajectory subgroups based on latent class mixed growth model modeling; forming baseline and longitudinal training samples based on the screening variable subsets, FSH trajectory subgroups and FSH longitudinal measurements, respectively; constructing a pregnancy prediction model for patients with pathological ovarian aging based on a random forest model on two training samples, and performing prediction; then verifying and comparing the optimal model in the test set; and using the optimal model in external samples to perform pregnancy prediction and evaluate the model prediction effect. The beneficial effect of the present invention is to evaluate the clinical pregnancy probability of patients with pathological ovarian aging through a pregnancy prediction model, guide doctors to optimize assisted reproductive treatment decisions, and improve the patient's pregnancy success rate.

Description

Technical Field

[0001] The present invention relates to the technical field of pregnancy prediction, and in particular to a method for predicting pregnancy in patients with pathological ovarian aging. Background Art

[0002] At present, there is still a lack of pregnancy prediction models for women with pathological ovarian aging, especially joint prediction models based on large sample, long-term prospective cohort studies and FSH dynamic change trajectories. Establishing an effective clinical pregnancy prediction model for people with pathological ovarian aging will help patients make fertility decisions and reduce medical expenses. At the same time, it will guide clinicians to provide personalized fertility counseling for different patients and optimize assisted reproductive treatment plans, thereby improving the success rate of clinical pregnancy and helping patients realize their fertility wishes.

[0003] To solve the above problems, the present application proposes a method for predicting pregnancy in patients with pathological ovarian aging. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting pregnancy in patients with pathological ovarian aging, so as to solve the problems in the prior art proposed in the above background technology, such as low pregnancy rate in patients with pathological ovarian aging, lack of clinical pregnancy prediction model, and difficulty in timely intervention through early identification of patients with pregnancy potential.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for predicting pregnancy in patients with pathological ovarian aging, the method comprising the following steps:

[0006] Step 1: Collect sample data and obtain sample characteristics;

[0007] Step 1.1. Collect information on 545 patients with pathological ovarian aging from the Chinese Ovarian Aging CoHort (COACH). The inclusion criteria for patients with pathological ovarian aging were: (1) age at enrollment ≤ 40 years; (2) serum FSH level ≥ 15 IU / L (at least twice, with an interval of > 4 weeks). Patients who were followed up twice or more after enrollment were included in trajectory analysis and prediction model construction, and were labeled with pregnancy labels and non-pregnancy labels according to known pregnancy outcomes;

[0008] Step 1.2, select sample characteristics related to pregnancy and longitudinal FSH measurements in the sample data according to preset criteria;

[0009] Step 2: Convert and interpolate the sample features described in step 1.2. The specific conversion and interpolation process is as follows:

[0010] Step 2.1: Convert and encode the sample features to make them suitable for the input of the training model based on the machine learning algorithm;

[0011] Step 2.2: If some of the sample features have missing values, then use the multiple interpolation method to fill in the missing values ​​of different features. If there are no missing values, no filling is required. Finally, a complete sample with no missing values ​​is formed and proceed to step 3.

[0012] Step 3. Use the Boruta algorithm to screen candidate predictor variables for the complete sample formed in step 2.2. The Boruta algorithm is a feature selection algorithm based on random forests. It screens features by comparing the importance of original features with randomly generated "shadow features". The core idea of ​​the algorithm is that if a feature is statistically significantly better than a shadow feature, then the feature is considered important for the prediction task. In this patent, the data is divided into a training set and a test set with a sample ratio of 7:3, and the Boruta algorithm is applied to the training set. The final subset of key classification variables screened is entered into step 5 to build a prediction model;

[0013] Step 4: For the longitudinal FSH measurements described in step 1.2, use the latent class growth mixture model (LCGMM) to determine the different trajectory pattern subgroups to which each individual FSH measurement belongs;

[0014] Step 5: The training samples formed in Step 2, Step 3 and Step 4 are used to construct a pregnancy prediction model for patients with pathological ovarian aging based on the random forest algorithm, which includes a baseline model and a longitudinal model. The specific model construction is as follows:

[0015] Step 5.1, construct a training sample data set, which is consistent with the training set in step 3 when the Boruta algorithm is used to screen candidate predictor variables, and the total sample is divided into a training set and a test set with a sample ratio of 7:3, including a baseline model training set and a longitudinal model training set;

[0016] Step 5.2, balance the training samples in the training set;

[0017] Step 5.3: Shuffle the cleaned training samples and divide them into n sets using stratified random sampling, where n-1 sets are used as training sets to train the model and the remaining set is used as a test set to evaluate the model, where n is a natural number.

[0018] Step 5.4, set the number of decision trees contained in the random forest to X, specify the number of variables used for binary trees in the node to Y, and use the default values ​​in the randomForest package for the remaining parameters;

[0019] Step 6: Use the pregnancy prediction model for patients with pathological ovarian aging based on random forest constructed in step 5 to predict pregnancy for the test samples, and compare the prediction effects of the baseline model and the longitudinal model;

[0020] Step 7, using the random forest-based longitudinal prediction model for pregnancy in patients with pathological ovarian aging selected in step 6 to predict pregnancy in external samples and evaluate the prediction effect of the model;

[0021] Step 7.1. A total of 456 patients with pathological ovarian aging (age ≤ 40 years, FSH ≥ 15 IU / L) who visited the Reproductive Hospital Affiliated to Shandong University from January 2010 to December 2021 (excluding the COACH cohort) were collected as an external validation set;

[0022] Step 7.2, inputting the external test set sample data constructed in step 7.1 into the trained random forest-based pregnancy prediction model for patients with pathological ovarian aging described in step 5 to evaluate the performance of the external prediction ability of the longitudinal model;

[0023] Furthermore, in step 1.2, the preset criteria are screening factors related to female fertility and pregnancy, including the patient's age, body mass index (BMI), age at menarche, education level, number of pregnancies, number of births, whether there is amenorrhea, whether there are chromosomal abnormalities, FSH, luteinizing hormone (LH), estradiol (E2), testosterone (T), anti-Mullerian hormone (AMH), AFC, thyroid stimulating hormone (TSH), alanine aminotransferase (ALT), aspartate aminotransferase (AST), glutamyl transpeptidase (GGT), alkaline phosphatase (ALP), total protein (TP), albumin (ALB), globulin (GLO), total cholesterol (TC), triglycerides (TG), high-density lipoprotein (HDL), low-density lipoprotein (LDL), and also include longitudinal FSH measurements during patient follow-up.

[0024] Furthermore, in step 2.1, the sample features were transformed and encoded to make them consistent with the input of the training model, including converting the skewed distribution of FSH, TSH, ALT, LH and GGT into a normal distribution, and calculating the range of FSH during the entire follow-up period (maximum FSH measurement value - minimum FSH measurement value during the patient's follow-up period) and the change value during the follow-up period (last FSH measurement value - first FSH measurement value during the patient's follow-up period);

[0025] In step 2.2, if some of the sample characteristics have missing values, they will be deleted when the missing values ​​or null values ​​are less than the preset values. When the missing values ​​or null values ​​are more than the preset values, both continuous variables and categorical variables will be filled using multiple interpolation methods.

[0026] Furthermore, step 4 specifically includes:

[0027] Step 4.1, LCGMM can divide the group into a finite number of "trajectory groups" according to the individual development trajectory over time. The individuals in each "trajectory group" have the same or similar dynamic change trajectory pattern. At the same time, the model will give the probability that each individual belongs to a certain "trajectory group". For trajectory grouping, LCGMM does not define the number of groups in advance, but regards the number of groups as a latent variable. The optimal estimate of the number of groups is given by combining the maximum likelihood function and its model fitting rules. Therefore, LCGMM is regarded as completing the latent classification on the basis of the finite mixture model. LCGMM assumes that the growth curve of the dynamic change of FSH measurement value of the research group is heterogeneous, which may contain There are different change trajectories. Each individual belongs to and only belongs to one potential category. The definition of the latent category is a discrete random variable as follows: Potential categories The probability of The multinomial logistic model is used for calculation, as follows: in, is a covariate that affects the individual's latent class membership and is set to a null value. For Category The intercept term of is a time-independent covariate The corresponding latent class parameter is set to null value, and the individual No. Repeated measurement index It is expressed as follows: in, is a fixed effect covariate that is the same across categories, corresponding to the general fixed effect coefficient ; is the class-specific fixed effect covariate, corresponding to the class-specific fixed effect coefficient ; is the individual random effect covariate, corresponding to the individual random effect ; is an autocorrelated process; is a random error;

[0028] Step 4.2: Based on the formula proposed in 4.1, after further traversing 2-4 potential categories, select the optimal FSH trajectory model according to the following criteria:

[0029] (1) Average posterior probability ≥ 0.7;

[0030] (2) The minimum proportion of individuals in each group is ≥5.0%;

[0031] (3) The Bayesian Information Criterion (BIC) is the lowest;

[0032] (4) Prior clinical knowledge;

[0033] Step 4.3: According to the selected optimal FSH trajectory model, attribute each individual to its posterior probability The highest trajectory subgroup is assigned a label to the corresponding FSH trajectory subgroup.

[0034] Furthermore, in step 5.1, the baseline model training set is constructed based on whether menopause, baseline FSH level, baseline LH level, baseline TSH level, baseline ALT level, baseline GGT level, baseline GLO level, and baseline AFC obtained after processing in step 2 and screening in step 3. On the basis of the baseline model training set, FSH trajectory subgroups, FSH ranges, and FSH change values ​​are further included as the longitudinal model training set;

[0035] In step 5.2, the ROSE algorithm is used to balance the training samples of the training set, that is, to balance the data of the pregnancy group and the data of the non-pregnancy group, wherein the data of the pregnancy group and the data of the non-pregnancy group refer to the samples marked with the pregnancy label and the non-pregnancy label in step 1.1 respectively;

[0036] In step 5.3, the cleaned training samples are shuffled and a ten-fold cross validation is performed, that is, the training samples are divided into 10 sets by stratified random sampling, and one of the sets is used as the test set and the other 9 sets are used as the training set to train the model. Each test will produce a corresponding result, and the average value of the final result is used as the evaluation result of the model effect;

[0037] In step 5.4, the training samples of the training set are encoded as 1 and 0 with pregnancy label and non-pregnancy label respectively as supervision signals. After ten trainings through ten-fold cross validation, the trained pregnancy prediction model for patients with pathological ovarian aging is obtained. The number of decision trees with the highest model accuracy is set to 500, and the number of variables used for the binary tree in the specified node is 3.

[0038] Furthermore, step 6 specifically includes:

[0039] Step 6.1, construct a data set of test samples. This data set is consistent with the test set in the training set and the test set in the Boruta algorithm in step 3 when screening candidate predictor variables. The total sample is divided into a training set and a test set with a sample ratio of 7:3. The predictor variables included in the baseline model test set are consistent with the variables screened in step 5.1. The predictor variables included in the longitudinal model are further included in the FSH trajectory subgroups, FSH ranges and FSH change values ​​on the basis of the baseline model test set;

[0040] Step 6.2: Input the test set sample data constructed in step 6.1 into the trained random forest-based pregnancy prediction model for patients with pathological ovarian aging described in step 5 to obtain the performance of the baseline model and the longitudinal model in predictive ability and the conclusion on whether the longitudinal model is better than the baseline model.

[0041] Furthermore, in step 5.2, the ROSE algorithm is used to balance the training samples of the training set, specifically including:

[0042] Consider the training set ,in are the category labels, namely non-pregnant and pregnant, is a certain whole A realization of (from the overall a sample of ), the probability density of the population unknown, Indicates category The ROSE algorithm generates a new artificial sample through the following steps:

[0043] (1) Select with probability 1 / 2 , ;

[0044] (2) Based on probability In the training set Select sample Make ;

[0045] (3) From Sampling in is a probability distribution centered at , is the scale parameter matrix;

[0046] First, select a sample from the training set, and then generate a new sample in the neighborhood of this sample. The width of this neighborhood is determined by Decide, usually, is a unimodal and symmetric probability distribution, given a class label When , generating new samples is equivalent to The kernel density estimate is used to sample, where the kernel function is , kernel function and "window width" The choice of is a pure kernel density estimation problem;

[0047] Given a label, the conditional density is as follows: Finally, the ROSE algorithm generates artificial samples with balanced categories through conditional kernel density estimation of the independent variables corresponding to each category, thus completing the balance of training sample data.

[0048] Furthermore, in order to facilitate clinical use, the prediction method also includes step 8, wherein the random forest-based longitudinal prediction model for pregnancy in patients with pathological ovarian aging selected in step 6 is used to build a trajectory grouping prediction and pregnancy probability prediction website.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] The present invention constructs a pregnancy prediction model specifically for evaluating the clinical pregnancy probability of patients with pathological ovarian aging. Compared with the existing methods of predicting pregnancy outcomes based on 3D-US or establishing an IVF / ICSI patient ovarian response prediction model, the method of the present invention, from sample screening to external validation of the model, all steps are based on the population of patients with pathological ovarian aging, and the prospective cohort population is used as the basis for modeling, and the dynamic change trajectory and longitudinal measurement value of the FSH indicator are innovatively incorporated, so that the model can more accurately predict the patient's pregnancy probability. The present invention not only helps patients make fertility decisions and reduce medical expenses, but also guides clinicians to conduct personalized fertility consultations, optimize assisted reproductive treatment plans, improve clinical pregnancy success rates, and help patients realize their fertility wishes. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of a method for predicting pregnancy in patients with pathological ovarian aging according to the present invention;

[0052] Figure 2 A schematic diagram of a model of a method for predicting pregnancy in patients with pathological ovarian aging according to the present invention;

[0053] Figure 3 A schematic diagram of a subset of key predictive variables screened by the Boruta algorithm of a method for predicting pregnancy in patients with pathological ovarian aging according to the present invention;

[0054] Figure 4 It is a FSH longitudinal trajectory diagram fitted by the LCGMM model of a method for predicting pregnancy in patients with pathological ovarian aging according to the present invention;

[0055] Figure 5 It is an internal validation ROC curve diagram of a baseline model and a longitudinal model of pregnancy in patients with pathological ovarian aging in a method for predicting pregnancy in patients with pathological ovarian aging of the present invention;

[0056] Figure 6 It is an internal validation calibration curve diagram of a baseline model and a longitudinal model of pregnancy in patients with pathological ovarian aging in a method for predicting pregnancy in patients with pathological ovarian aging of the present invention;

[0057] Figure 7It is an external validation ROC curve diagram of a pregnancy model of a patient with pathological ovarian aging in a method for predicting pregnancy in a patient with pathological ovarian aging of the present invention;

[0058] Figure 8 It is an external validation calibration curve diagram of a pregnancy model of a patient with pathological ovarian aging in a method for predicting pregnancy in a patient with pathological ovarian aging of the present invention;

[0059] Fig. 9 A schematic diagram of the FSH trajectory of a prediction website established based on a pregnancy prediction model for patients with pathological ovarian aging according to a method for predicting pregnancy in patients with pathological ovarian aging of the present invention;

[0060] Fig.10 A schematic diagram of pregnancy probability prediction on a prediction website established based on a pregnancy prediction model for patients with pathological ovarian aging according to a method for predicting pregnancy for patients with pathological ovarian aging of the present invention. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0062] See also Figure 1-Figure 10 The present invention provides a technical solution: a method for predicting pregnancy in patients with pathological ovarian aging, the method comprising the following steps:

[0063] Step 1: Collect sample data and obtain sample characteristics;

[0064] Step 1.1. In one embodiment of the present invention, the collection of sample data may adopt the following standards: Information on 545 patients with pathological ovarian aging from the Chinese Ovarian Aging CoHort (COACH) was collected. The inclusion criteria for patients with pathological ovarian aging were: (1) age at enrollment ≤ 40 years old; (2) serum FSH level ≥ 15 IU / L (at least twice, with an interval of > 4 weeks). Patients who were followed up twice or more after enrollment were included in the trajectory analysis and prediction model construction. The follow-up method could be achieved by admission or visit to the hospital, and pregnancy labels and non-pregnancy labels were respectively marked according to known pregnancy outcomes;

[0065] Step 1.2, select sample characteristics related to pregnancy and longitudinal FSH measurements in the sample data according to preset criteria;

[0066] Step 2: Convert and interpolate the sample features described in step 1.2. The specific conversion and interpolation process is as follows:

[0067] Step 2.1: Convert and encode the sample features to make them suitable for the input of the training model based on the machine learning algorithm;

[0068] Step 2.2: If some of the sample features have missing values, then use the multiple interpolation method to fill in the missing values ​​of different features. If there are no missing values, no filling is required. Finally, a complete sample with no missing values ​​is formed and proceed to step 3.

[0069] Step 3. Use the Boruta algorithm to screen candidate predictor variables for the complete sample formed in step 2.2. The Boruta algorithm is a feature selection algorithm based on random forests. It screens features by comparing the importance of original features with randomly generated "shadow features". The core idea of ​​the algorithm is that if a feature is statistically significantly better than a shadow feature, then the feature is considered important for the prediction task. In this patent, the data is divided into a training set and a test set with a sample ratio of 7:3, and the Boruta algorithm is applied to the training set. The final subset of key classification variables screened is entered into step 5 to build a prediction model;

[0070] Step 4: For the longitudinal FSH measurements described in step 1.2, use the latent class growth mixture model (LCGMM) to determine the different trajectory pattern subgroups to which each individual FSH measurement belongs;

[0071] Step 5: The training samples formed in Step 2, Step 3 and Step 4 are used to construct a pregnancy prediction model for patients with pathological ovarian aging based on the random forest algorithm, which includes a baseline model and a longitudinal model. The specific model construction is as follows:

[0072] Step 5.1, construct a training sample data set, which is consistent with the training set in step 3 when the Boruta algorithm is used to screen candidate predictor variables, and the total sample is divided into a training set and a test set with a sample ratio of 7:3, including a baseline model training set and a longitudinal model training set;

[0073] Step 5.2, balance the training samples in the training set;

[0074] Step 5.3: Shuffle the cleaned training samples and divide them into n sets using stratified random sampling, where n-1 sets are used as training sets to train the model and the remaining set is used as a test set to evaluate the model, where n is a natural number.

[0075] Step 5.4, set the number of decision trees contained in the random forest to X, specify the number of variables used for binary trees in the node to Y, and use the default values ​​in the randomForest package for the remaining parameters;

[0076] Step 6: Use the pregnancy prediction model for patients with pathological ovarian aging based on random forest constructed in step 5 to predict pregnancy for the test samples, and compare the prediction effects of the baseline model and the longitudinal model;

[0077] Step 7, using the random forest-based longitudinal prediction model for pregnancy in patients with pathological ovarian aging selected in step 6 to predict pregnancy for external samples and evaluate the prediction effect of the model;

[0078] Step 7.1. In one embodiment of the present invention, a total of 456 patients with pathological ovarian aging (age ≤ 40 years old, FSH ≥ 15 IU / L) who visited the Reproductive Hospital Affiliated to Shandong University from January 2010 to December 2021 (excluding the COACH cohort population) were collected as an external validation set;

[0079] Step 7.2: Input the external test set sample data constructed in step 7.1 into the trained random forest-based pregnancy prediction model for patients with pathological ovarian aging described in step 5 to evaluate the performance of the external prediction ability of the longitudinal model.

[0080] After evaluating and internally validating the model, it can be considered that the longitudinal pregnancy prediction model established in this study is a model with good predictive ability. In order to facilitate clinical use, a trajectory grouping prediction and pregnancy probability prediction website was built based on the pregnancy prediction model established in this study. The website can predict the trajectory group to which the patient belongs based on 2 or more repeated measurements of FSH according to the LCGMM model fitted by the cohort data; according to the random forest model established by the cohort of this study, the covariates related to model training are input to output the probability that the individual belongs to the pregnancy outcome classification.

[0081] Preferably, in step 1.2, the preset criteria are screening factors related to female fertility and pregnancy, including the patient's age, BMI, age at menarche, education level, number of pregnancies, number of births, whether there is amenorrhea, whether there are chromosomal abnormalities, FSH, LH, E2, T, AMH, AFC, TSH, ALT, AST, GGT, ALP, TP, ALB, GLO, TC, TG, HDL, LDL, and also include longitudinal FSH measurements during patient follow-up.

[0082] In step 2.1, the sample features were transformed and encoded to make them consistent with the input of the training model, including converting the skewed distribution of FSH, TSH, ALT, LH and GGT into a normal distribution, and calculating the range of FSH during the entire follow-up period (the maximum FSH measurement value - the minimum FSH measurement value during the patient's follow-up period) and the change value during the follow-up period (the last FSH measurement value - the first FSH measurement value during the patient's follow-up period);

[0083] In step 2.2, if some of the sample characteristics have missing values, they will be deleted when the missing values ​​or null values ​​are less than the preset values. When the missing values ​​or null values ​​are more than the preset values, both continuous variables and categorical variables will be filled using multiple interpolation methods.

[0084] It should be noted that a total of 26 variables to be screened were included in step 3. After candidate predictive variables were analyzed using the Boruta algorithm, the key predictive variable subset for predicting pregnancy was menopause, baseline FSH level, baseline LH level, baseline TSH level, baseline ALT level, baseline GGT level, baseline GLO level, and baseline AFC.

[0085] Step 4 specifically includes:

[0086] Step 4.1, LCGMM can divide the group into a finite number of "trajectory groups" according to the individual development trajectory over time. The individuals in each "trajectory group" have the same or similar dynamic change trajectory pattern. At the same time, the model will give the probability that each individual belongs to a certain "trajectory group". For trajectory grouping, LCGMM does not define the number of groups in advance, but regards the number of groups as a latent variable. The optimal estimate of the number of groups is given by combining the maximum likelihood function and its model fitting rules. Therefore, LCGMM is regarded as completing the latent classification on the basis of the finite mixture model. LCGMM assumes that the growth curve of the dynamic change of FSH measurement value of the research group is heterogeneous, which may contain There are different change trajectories. Each individual belongs to and only belongs to one potential category. The definition of the latent category is a discrete random variable as follows: Potential categories The probability of The multinomial logistic model is used for calculation, as follows: in, is a covariate that affects the individual's latent class membership and is set to a null value. For Category The intercept term of is a time-independent covariate The corresponding latent class parameter is set to null value, and the individual No. Repeated measurement index It is expressed as follows: in, is a fixed effect covariate that is the same across categories, corresponding to the general fixed effect coefficient ; is the class-specific fixed effect covariate, corresponding to the class-specific fixed effect coefficient ; is the individual random effect covariate, corresponding to the individual random effect ; is an autocorrelated process; is a random error;

[0087] Step 4.2: Based on the formula proposed in 4.1, substitute the repeated FSH measurements and measurement time of each individual, set the longitudinal FSH measurement value as a linear or nonlinear function of time (the number of months between each measurement date and the baseline date), express it as time, quadratic term of time, or cubic term, and further traverse 2-4 potential categories, and select the optimal FSH trajectory model according to the following criteria:

[0088] (1) Average posterior probability ≥ 0.7;

[0089] (2) The minimum proportion of individuals in each group is ≥5.0%;

[0090] (3) The Bayesian Information Criterion (BIC) is the lowest;

[0091] (4) Prior clinical knowledge;

[0092] Step 4.3: According to the criteria described in step 4.2, the optimal FSH trajectory model is selected as three sets of quadratic term models. Then, each individual is assigned to its posterior probability The highest trajectory subgroup is assigned a label to the corresponding FSH trajectory subgroup.

[0093] In step 5.1, the baseline model training set is constructed based on whether menopause, baseline FSH level, baseline LH level, baseline TSH level, baseline ALT level, baseline GGT level, baseline GLO level, and baseline AFC obtained after processing in step 2 and screening in step 3. On the basis of the baseline model training set, FSH trajectory subgroups, FSH ranges, and FSH change values ​​are further included as the longitudinal model training set;

[0094] In step 5.2, the ROSE algorithm is used to balance the training samples of the training set, that is, to balance the data of the pregnancy group and the data of the non-pregnancy group, wherein the data of the pregnancy group and the data of the non-pregnancy group refer to the samples marked with the pregnancy label and the non-pregnancy label in step 1.1 respectively;

[0095] In step 5.3, the cleaned training samples are shuffled and a ten-fold cross validation is performed, that is, the training samples are divided into 10 sets by stratified random sampling, and one of the sets is used as the test set and the other 9 sets are used as the training set to train the model. Each test will produce a corresponding result, and the average value of the final result is used as the evaluation result of the model effect;

[0096] In step 5.4, the training samples of the training set are encoded as 1 and 0 with pregnancy label and non-pregnancy label respectively as supervision signals. After ten trainings through ten-fold cross validation, the trained pregnancy prediction model for patients with pathological ovarian aging is obtained. The number of decision trees with the highest model accuracy is set to 500, and the number of variables used for the binary tree in the specified node is 3.

[0097] Step 6 specifically includes:

[0098] Step 6.1, construct a data set of test samples. This data set is consistent with the test set in the training set and the test set in the Boruta algorithm in step 3 when screening candidate predictor variables. The total sample is divided into a training set and a test set with a sample ratio of 7:3. The predictor variables included in the baseline model test set are consistent with the variables screened in step 5.1. The predictor variables included in the longitudinal model are further included in the FSH trajectory subgroups, FSH ranges and FSH change values ​​on the basis of the baseline model test set;

[0099] Step 6.2: Input the test set sample data constructed in step 6.1 into the trained random forest-based pregnancy prediction model for pathological ovarian aging patients described in step 5, evaluate the baseline prediction model and the longitudinal prediction model through the ROC curve and calibration curve, and obtain the performance of the baseline model and the longitudinal model prediction ability and whether the longitudinal model is better than the baseline model. The AUC of the baseline prediction model for predicting patient pregnancy is 0.644, and the AUC of the longitudinal prediction model for predicting patient pregnancy is 0.727. Draw a calibration chart for calibration evaluation. The calibration chart indicates that in the baseline prediction model, its calibration curve deviates greatly from the diagonal line, and the calibration is not ideal; while in the longitudinal prediction model, its calibration curve is basically close to the ideal diagonal line. For patients with a calculated predicted probability lower than 0.625, the slope of the calibration curve is <1, and the intercept is greater than 0. The model may slightly underestimate the risk of disease; for patients with a calculated predicted risk higher than 0.625, the model may slightly overestimate the actual risk of disease. The net reclassification index (NRI) and the integrated discrimination improvement index (IDI) are used to evaluate the improvement of the model prediction ability after adding longitudinal indicators. The results showed that compared with the baseline model, after adding the longitudinal indicators, the NRI was 59.13% (15.72%~102.54%), the P value was 0.008, and the IDI was 4.45% (1.00%~8.80%), the P value was 0.045, that is, after adding the longitudinal indicators, the performance of the prediction model was significantly improved. Therefore, the present invention selects the longitudinal prediction model as the final pregnancy prediction model for patients with pathological ovarian aging.

[0100] In step 5.2, the ROSE algorithm is used to balance the training samples in the training set, specifically including:

[0101] Consider the training set ,in are the category labels, namely non-pregnant and pregnant, is a certain whole A realization of (from the overall a sample of ), the probability density of the population unknown, Indicates category The ROSE algorithm generates a new artificial sample through the following steps:

[0102] (1) Select with probability 1 / 2 , ;

[0103] (2) Based on probability In the training set Select sample Make ;

[0104] (3) From Sampling in is a probability distribution centered at , is the scale parameter matrix;

[0105] First, select a sample from the training set, and then generate a new sample in the neighborhood of this sample. The width of this neighborhood is determined by Decide, usually, is a unimodal and symmetric probability distribution, given a class label When , generating new samples is equivalent to The kernel density estimate is used to sample, where the kernel function is , kernel function and "window width" The choice of is a pure kernel density estimation problem;

[0106] Given a label, the conditional density is as follows: Finally, the ROSE algorithm generated artificial samples with balanced categories through conditional kernel density estimation of the independent variables corresponding to each category, that is, the training sample data was balanced. In step 7.2, the external test set sample data constructed in step 7.1 was input into the trained random forest-based pathological ovarian aging patient pregnancy prediction model described in step 5 to evaluate the performance of the external prediction ability of the longitudinal model. By constructing the ROC curve of the pregnancy prediction model for ovarian aging patients, the AUC of the longitudinal prediction model for predicting patient pregnancy was 0.708, and the calibration curve was basically close to the ideal diagonal. For patients with a calculated prediction probability higher than 0.5, the slope of the calibration curve was <1 and the intercept was greater than 0. The model may have slightly underestimated the risk of disease, which further proved the stability of the prediction model.

[0107] For the convenience of clinical use, preferably, the method further includes step 8, in which the longitudinal pregnancy prediction model for pathological ovarian aging patients based on random forest selected in step 6 is used to build a trajectory grouping prediction and pregnancy probability prediction website. After the model is evaluated and internally and externally verified in steps 6 and 7, it can be considered that the longitudinal pregnancy prediction model established in the method is a model with good prediction ability. For the convenience of clinical use, a trajectory grouping prediction and pregnancy probability prediction website are built according to the pregnancy prediction model established in the method. The website is fitted with the LCGMM model obtained by the cohort data. The results of the LCGMM trajectory model fitting process are shown in Table 1. The trajectory group to which the patient belongs can be predicted based on repeated measurements of FSH for 2 or more times; according to the random forest model established in the cohort of this study, the covariates related to model training are input, and the probability that the individual belongs to the pregnancy outcome classification can be output. For example, a patient is 29 years old, primary infertility, and regular menstruation. The baseline FSH level at enrollment was 18.42 IU / L. Follow-up visits and monitoring of FSH levels were conducted at 6 months, 11 months, and 19 months after enrollment, and the measured values ​​were 11.14 IU / L, 14.26 IU / L, and 13.00 IU / L, respectively. The above FSH levels and test times were entered into the website to obtain the probability of belonging to each trajectory group; further, by entering the amenorrhea status, FSH level, LH level, etc. at the time of enrollment, the probability of pregnancy of the patient was 65.40%, and corresponding clinical intervention could be given.

[0108]

[0109] Working principle:

[0110] The present invention is a pregnancy prediction model for patients with pathological ovarian aging that has been calculated and tested through a series of rigorous statistical methods. By collecting relevant hormone and biochemical indicator data of patients with pathological ovarian aging, and marking pregnancy labels and non-pregnancy labels according to whether the pregnancy outcome is, sample features related to pregnancy in the sample data are selected, and the sample features are converted and encoded, and features with missing values ​​in the sample features are filled, and then a data set is constructed with the converted, encoded and filled sample features, and the data set is divided into a training set and a test set by means of a non-repeated sampling technology, and the pregnancy baseline prediction variables are further screened by the Boruta algorithm; in the modeling process, ROSE data balance is first performed on the training samples of the training set, and then the random forest prediction model of whether the FSH longitudinal indicators (including FSH trajectory subgroups, FSH extreme values ​​during follow-up and FSH change values ​​during follow-up) are compared to determine the final pregnancy prediction model for patients with pathological ovarian aging and after ten-fold cross validation, the trained pregnancy prediction model for patients with pathological ovarian aging is obtained, and finally, the trained model is applied to pregnancy prediction for patients with pathological ovarian aging.

[0111] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.

Claims

1. A method for predicting pregnancy in patients with pathological ovarian aging, characterized in that: The method comprises the following steps: Step 1: Collect sample data and obtain sample characteristics; Step 1.1: Set the sample data collection standards. Patients who meet the standards are included in trajectory analysis and prediction model construction, and are labeled with pregnancy labels and non-pregnancy labels according to known pregnancy outcomes. Step 1.2, select sample characteristics related to pregnancy and longitudinal FSH measurements in the sample data according to preset criteria; Step 2: Convert and interpolate the sample features described in step 1.

2. The specific conversion and interpolation process is as follows: Step 2.1: Convert and encode the sample features to make them suitable for the input of the training model based on the machine learning algorithm; Step 2.2: If some of the sample features have missing values, then use the multiple interpolation method to fill in the missing values ​​of different features. If there are no missing values, no filling is required. Finally, a complete sample with no missing values ​​is formed and proceed to step 3. Step 3: Use the Boruta algorithm to screen candidate predictive variables for the complete sample formed in step 2.2; split the data into a training set and a test set with a sample ratio of 7:3, and apply the Boruta algorithm to the training set. Finally, the subset of key classification variables obtained by screening is used to build a prediction model in step 5; Step 4, for the longitudinal FSH measurements described in step 1.2, use the latent class growth mixture model LCGMM to determine the different trajectory pattern subgroups to which each individual FSH measurement value belongs; Step 5, using the training samples formed in Step 2, Step 3 and Step 4 to construct a pregnancy prediction model for patients with pathological ovarian aging based on a random forest algorithm, including a baseline model and a longitudinal model; Step 6: Use the pregnancy prediction model for patients with pathological ovarian aging based on random forest constructed in step 5 to predict pregnancy for the test samples, and compare the prediction effects of the baseline model and the longitudinal model; Step 7, using the random forest-based longitudinal prediction model for pregnancy in patients with pathological ovarian aging selected in step 6 to predict pregnancy in external samples and evaluate the prediction effect of the model; Step 4 specifically includes: Step 4.1, LCGMM assumes that the growth curve of the dynamic change of FSH measurement value in the research group is heterogeneous, containing G different change trajectories, and each individual belongs to and only belongs to one latent category. The belonging of the latent category is defined as a discrete random variable as follows: c i =g;g=1,...,G Potential Class c i The probability of ci The multinomial logistic model is used for calculation, as follows: Among them, X ci is a covariate that affects the individual's latent class membership, set to the null value, ε 0g is the intercept term of category g, ε 1g is the time-independent covariate X ci The corresponding latent class parameter is set to the null value, and the j-th repeated measurement indicator Y of individual i is ij It is expressed as follows: Among them, X L1i (t ij ) is a fixed effect covariate that is the same for all categories, corresponding to the general fixed effect coefficient β; X L2i (t ij ) is the class-specific fixed effect covariate, corresponding to the class-specific fixed effect coefficient ν g ; Z i (t ij ) is the individual random effect covariate, corresponding to the individual random effect μ ig ;ω i (t ij ) is the autocorrelation process; ε ij is a random error; Step 4.2: Based on the formula proposed in 4.1, after further traversing 2-4 potential categories, select the optimal FSH trajectory model according to the following criteria: (1) Average posterior probability ≥ 0.7; (2) the minimum proportion of individuals in each group was ≥5.0%; (3) The Bayesian Information Criterion (BIC) is the lowest; (4) prior clinical knowledge; Step 4.3: According to the selected optimal FSH trajectory model, attribute each individual to its posterior probability π ig The highest trajectory subgroup and the label of the corresponding FSH trajectory subgroup were assigned; In step 5, the specific model is constructed as follows: Step 5.1, construct a training sample data set, which is consistent with the training set in step 3 when the Boruta algorithm is used to screen candidate predictor variables, and the total sample is divided into a training set and a test set with a sample ratio of 7:3, including a baseline model training set and a longitudinal model training set; Step 5.2, balance the training samples in the training set; Step 5.3: Shuffle the cleaned training samples and divide them into n sets using stratified random sampling, of which "n-1" sets are used as training sets to train the model, and the remaining set is used as a test set to evaluate the model, where n is a natural number. Step 5.4, set the number of decision trees contained in the random forest to X, specify the number of variables used for binary trees in the node to Y, and use the default values ​​in the randomForest package for the remaining parameters; In step 5.1, the baseline model training set is constructed based on whether menopause, baseline FSH level, baseline LH level, baseline TSH level, baseline ALT level, baseline GGT level, baseline GLO level, and baseline AFC obtained after processing in step 2 and screening in step 3. On the basis of the baseline model training set, FSH trajectory subgroups, FSH ranges, and FSH change values ​​are further included as the longitudinal model training set; In step 5.2, the ROSE algorithm is used to balance the training samples of the training set, that is, to balance the data of the pregnancy group and the data of the non-pregnancy group, wherein the data of the pregnancy group and the data of the non-pregnancy group refer to the samples marked with the pregnancy label and the non-pregnancy label in step 1.1 respectively; In step 5.3, the cleaned training samples are shuffled and a ten-fold cross validation is performed; In step 5.4, the training samples of the training set are encoded as 1 and 0 with pregnancy label and non-pregnancy label respectively as supervision signals. After ten trainings through ten-fold cross validation, the trained pregnancy prediction model for patients with pathological ovarian aging is obtained. The number of decision trees with the highest model accuracy is set to 500, and the number of variables used for the binary tree in the specified node is 3.

2. The method for predicting pregnancy in patients with pathological ovarian aging according to claim 1, characterized in that: In step 1.2, the preset criteria are screening factors related to female fertility and pregnancy.

3. The method for predicting pregnancy in patients with pathological ovarian aging according to claim 1, characterized in that: In step 2.1, the sample features are transformed and encoded to make them consistent with the input of the training model, including converting the skewed distribution of FSH, TSH, ALT, LH and GGT into a normal distribution, calculating the range of FSH during the entire follow-up period, that is, the maximum FSH measurement value - the minimum FSH measurement value during the patient's follow-up period, and the change value during the follow-up period, that is, the last FSH measurement value - the first FSH measurement value during the patient's follow-up period; In step 2.2, if some of the sample characteristics have missing values, they will be deleted when the missing values ​​or null values ​​are less than the preset values. When the missing values ​​or null values ​​are more than the preset values, both continuous variables and categorical variables will be filled using multiple interpolation methods.

4. The method for predicting pregnancy in patients with pathological ovarian aging according to claim 1, characterized in that: Step 6 specifically includes: Step 6.1, construct a data set of test samples. This data set is consistent with the test set in the training set and the test set in the Boruta algorithm in step 3 when screening candidate predictor variables. The total sample is divided into a training set and a test set with a sample ratio of 7:

3. The baseline model test set contains predictor variables that are consistent with the variables screened in step 5.

1. The longitudinal model contains predictor variables that further include FSH trajectory subgroups, FSH ranges, and FSH change values ​​on the basis of the baseline model test set; Step 6.2: Input the test set sample data constructed in step 6.1 into the trained random forest-based pregnancy prediction model for patients with pathological ovarian aging described in step 5 to obtain the performance of the baseline model and the longitudinal model in predictive ability and the conclusion on whether the longitudinal model is better than the baseline model.

5. The method for predicting pregnancy in patients with pathological ovarian aging according to claim 1, characterized in that: In step 5.2, the ROSE algorithm is used to balance the training samples in the training set, specifically including: Training set T n ={(x i , y i ), i = 1, 2, …, n}, where y i ∈{y0, y1} is the class label, representing non-pregnancy and pregnancy respectively, and x i is an implementation of a certain population x ∈ R d , where x i is a sample from the population x. The probability density f(x) of the population is unknown, and n j < n represents the number of classes y j . The ROSE algorithm generates a new artificial sample through the following steps: (1) Choose y=y with probability 1 / 2 j , j = 1, 2; (2) With probability p i =1 / n j In the training set T n Select samples (x i ,y i ) makes y i =y; (3) From Sampling in is a probability distribution centered at x i , H j is the scale parameter matrix; First, select a sample from the training set, and then generate a new sample in the neighborhood of this sample. The width of this neighborhood is determined by H j Decide, usually, is a unimodal and symmetric probability distribution, given a class label Y j When generating new samples, it is equivalent to generating new samples by f(x|Y j ) is sampled using the kernel density estimate, where the kernel function is Kernel function K and "window width" H j The choice of is a pure kernel density estimation problem; Given a label, the conditional density is as follows: Finally, the ROSE algorithm generates artificial samples with balanced categories through conditional kernel density estimation of the independent variables corresponding to each category, thus completing the balance of training sample data.

6. The method for predicting pregnancy in patients with pathological ovarian aging according to claim 1, characterized in that: The prediction method also includes step 8, which is to use the random forest-based longitudinal prediction model for pregnancy in patients with pathological ovarian aging selected in step 6 to build a trajectory grouping prediction and pregnancy probability prediction website.

7. The method for predicting pregnancy in patients with pathological ovarian aging according to claim 2, characterized in that: The screening factor variables include the patient's age, body mass index (BMI), age at menarche, education level, number of pregnancies, number of births, whether there is amenorrhea, whether there is chromosomal abnormality, FSH, luteinizing hormone (LH), estradiol (E2), testosterone (T), anti-mullerian hormone (AMH), antral follicle count (AFC), thyroid stimulating hormone (TSH), alanine aminotransferase (ALT), aspartate aminotransferase (AST), glutamyl transpeptidase (GGT), alkaline phosphatase (ALP), total protein (TP), albumin (ALB), globulin (GLO), total cholesterol (TC), triglycerides (TG), high-density lipoprotein (HDL), and low-density lipoprotein (LDL). It also includes longitudinal FSH measurements during the patient's follow-up.

Images