Method for predicting fertilization disorder occurrence risk and optimizing in-vitro fertilization scheme

Abstract

The invention provides a method for predicting fertilization disorder occurrence risks and optimizing an in-vitro fertilization scheme. Firstly, clinical routine examination and test indexes are screened; and constructing a prediction model based on a classic ordered Logistic regression model and a Bayesian network model. And screening main indexes in combination with expert evaluation to construct a prediction model. Secondly, constructing and optimizing a prediction model by using a machine learning algorithm; the model comprises an equalization model and a non-equalization model. Finally, a prediction system is constructed according to the optimal prediction model and used for selecting an in-vitro fertilization scheme. The system can predict the probability of complete non-fertilization and low fertilization rate under the condition of using different in-vitro fertilization technologies according to basic information of a patient, and helps reproductive clinicians to formulate an optimal individualized scheme.

Description

Technical Field

[0001] The present invention belongs to the field of human assisted reproduction technology, and specifically relates to a method for predicting the risk of fertilization disorders and optimizing in vitro fertilization programs. Background Art

[0002] In the assisted reproductive technology treatment of infertility, fertilization disorders include complete infertilization and low fertilization rate. If fertilization disorders occur, there may be no embryos to be transferred, which may lead to the cancellation of the cycle, increasing the risk of re-treatment of patients and the corresponding economic burden, which has become one of the important constraints affecting the effectiveness and safety of assisted reproductive technology. There are many known or unknown factors affecting fertilization disorders, and among the series of clinical information and laboratory indicators commonly used in clinical practice, no specific indicators with high predictive value for fertilization disorders have been found. Therefore, in the current clinical diagnosis and treatment process, there is a lack of a reliable method that can accurately predict fertilization disorders.

[0003] In the past, researchers usually used single or individual indicators (such as age, semen parameters, anti-Mullerian hormone (AMH), sperm DNA fragmentation index, etc.) to predict fertilization disorders, but due to the many factors that affect fertilization disorders, it is difficult to achieve good prediction results. In recent years, researchers have begun to try to use multiple clinical information and laboratory indicators for prediction, such as: using the classic ordered logistic regression model and the multi-factor prediction model constructed by the Nomogram model. Although the model's prediction ability has been improved to a certain extent, the effect is still not obvious. Summary of the invention

[0004] In order to solve the above problems existing in the prior art, the method for predicting the risk of fertilization disorder of the present invention specifically comprises the following steps:

[0005] Step 1: Screening of routine clinical examination indicators;

[0006] The classical ordered logistic regression model and the Bayesian network model were used to construct a fertilization disorder prediction model; key predictive indicators with significant predictive ability for fertilization disorders were selected from the above two models and evaluated; the main indicators that can be promoted and used in the clinical diagnosis and treatment process were screened out based on the selection criteria of routine use, rapid acquisition and low cost;

[0007] Step 2: Use machine learning algorithms to build and optimize a fertilization disorder prediction model;

[0008] According to the indicators screened in step 1, a 10-fold*10-fold cross-validation method was used to evaluate and screen multiple models in reproductive clinical big data to select the model with the best accuracy;

[0009] Step 3: constructing a prediction system according to the prediction model determined in step 2, and selecting an in vitro fertilization plan based on the prediction results obtained by the system;

[0010] According to the fertilization disorder prediction model determined in step 2, a system for selecting the best in vitro fertilization technology is developed. The system is used to make predictions based on the basic information of both infertile patients. It can predict the probability of complete infertilization and low fertilization rate under the conditions of using different in vitro fertilization technologies, thereby assisting reproductive clinicians in formulating and making decisions on individualized best plans for patients.

[0011] Furthermore, the main indicators described in step 1 include:

[0012] Female medical history information: age, body mass index, pregnancy number, parity, adverse pregnancy and childbirth history, cause of infertility, type of infertility, and previous history of in vitro fertilization failure;

[0013] Male medical history information: male age, body mass index and semen parameters;

[0014] Treatment plan and examination results of the current cycle: number of bilateral antral follicles, ovulation induction plan and number of eggs retrieved.

[0015] Among them, the causes of infertility include fallopian tube factors, uterine abnormalities, ovarian abnormalities, history of low ovarian reserve, endometriosis and genetic factors; the types of infertility include primary and secondary; semen parameters include normal, oligoasthenospermia, severe oligoasthenospermia, azoospermia and teratospermia; ovulation induction plans include stimulation cycle, micro-stimulation cycle and natural cycle.

[0016] Furthermore, the 10*10-fold cross-validation method in step 2 is as follows: in each cross-validation, the data set is randomly divided into 10 equal parts, 9 of which are used as training data sets to train the model, and the remaining 1 is used as a validation data set to validate the model, and the same cross-validation process is repeated 10 times in total; each validation will result in a corresponding accuracy rate, and the accuracy rate of 10*10 model validations is used as an evaluation indicator of the model prediction accuracy. Finally, among the multiple models constructed, the model accuracy is used for screening to determine the best model.

[0017] Furthermore, the machine learning algorithm in step two includes: feature selection, data preprocessing, sample balancing, data set partitioning and data analysis.

[0018] Furthermore, the data analysis in step 2 yields the following:

[0019] Model accuracy on the test set: Each model is evaluated on an independent test dataset and the accuracy of the model is reported;

[0020] Confusion Matrix: For each model, a confusion matrix is ​​provided, which shows the correspondence between the model's predictions and the actual results;

[0021] ROC curve: For each model, a ROC curve is plotted, which is used to evaluate the classification performance of the model and provides a measure of the model performance.

[0022] Furthermore, based on the method described in step 2, two prediction models, balanced and non-balanced, were constructed respectively.

[0023] Among them, the non-balanced prediction model is a prediction model based on the original sample population. The predicted probability calculated by the model represents the population probability, that is, the actual probability of occurrence. When used, the predicted probability is compared with the proportions of normal fertilization, low fertilization and complete infertilization in the general population. This prediction value is easy for professional doctors to understand and interpret.

[0024] Among them, the balanced prediction model is a prediction model based on sample balanced processing. The predicted probability calculated by the model represents the individual probability, that is, the predicted probability of the sample data after balanced processing. This prediction value is convenient for patients to understand in a popular way.

[0025] Beneficial effects of the present invention:

[0026] 1. The prediction indicators used in this technical solution are all routine clinical examination and test indicators. All indicators are mandatory items for all reproductive clinical institutions, and do not include any special examination items or high examination fees, which is convenient for promotion in more medical institutions.

[0027] 2. Compared with previous technical solutions, this technical solution further improves the prediction accuracy and AUC of fertilization disorders, especially between different subgroups, both of which are >90%, and the sensitivity and specificity are both >80%.

[0028] 3. The technical solution includes a system for selecting the best in vitro fertilization technology, which can predict the probability of fertilization disorders (including different subgroups) under the conditions of using different in vitro fertilization technologies (IVF or ICSI) by directly inputting some basic information of both infertile patients, thereby assisting reproductive clinicians in formulating and making decisions on the best individualized plan for each patient.

[0029] 4. This technical solution provides two types of prediction probabilities. One is the population probability prediction result based on the original sample population. This prediction value is easy for professional doctors to understand and interpret. The other is the individual probability prediction result based on sample equalization (the baseline probabilities of normal fertilization, low fertilization, and complete infertilization are all standardized to 1 / 3). This prediction value is easy for patients to understand. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A flowchart of the method for predicting the risk of fertilization disorders and optimizing in vitro fertilization programs according to the present invention;

[0031] Figure 2 The machine learning model of fertilization barriers constructed based on non-balanced raw data is shown;

[0032] Figure 3 A machine learning model of fertilization barriers constructed based on balanced data is shown;

[0033] Figure 4 Example 1 of a model system for predicting fertilization failure;

[0034] Figure 5 Example 2 of a model system for predicting fertilization failure. DETAILED DESCRIPTION

[0035] The present invention will be further described below in conjunction with specific embodiments. It should be understood that the embodiments are only used to illustrate the present invention and are not intended to limit the scope of protection of the present invention. In addition, it should be understood that after reading the contents disclosed in the present invention, those skilled in the art may make various changes or modifications to the present invention, and these equivalent forms also fall within the scope of protection defined by the present invention.

[0036] The method for predicting the risk of fertilization disorder of the present invention specifically comprises the following steps:

[0037] Step 1: Screen routine clinical examination and test indicators.

[0038] The classical ordered logistic regression model and Bayesian network model were used to construct a fertilization failure prediction model.

[0039] In the logistic regression model, a prediction model containing only clinical indicators and a prediction model containing laboratory biochemical indicators were constructed respectively. Among them, laboratory biochemical indicators refer to hormone levels such as FSH (gonadotropin), E2 (estradiol), P (progesterone) and AMH (anti-Mullerian hormone). The results showed that the prediction model containing laboratory biochemical indicators was better than the prediction model containing only clinical indicators. Among them, primary infertility, history of IVF failure, ovulation induction method, number of oocytes retrieved <5, IVF-ET fertilization method (compared with ICSI) and AMH level <1.1ng / ml were predictors of LFR (low fertilization) and TFF (total fertilization failure) risks. The prediction accuracy of the model was 91.23%. ROC analysis of the control group and TFF showed an AUC of 0.742 (95%CI: 0.710-0.774), a sensitivity of 65.3%, and a specificity of 74.2%.

[0040] In the Bayesian network model, a directed acyclic graph for predicting fertilization disorders was constructed, and it was found that fertilization mode, infertility type, and number of eggs directly affected fertilization disorders, and age and genetic factors were the initial factors affecting other variables. The prediction accuracy of this model was 91.3%. The AUC of the control group and TFF was 0.779 (95% CI: 0.766-0.791), the sensitivity was 71.2%, and the specificity was 70.1%. The AUC of the control group and LFR was 0.619 (95% CI: 0.605-0.634), the sensitivity was 64.9%, and the specificity was 52.6%.

[0041] Key predictive indicators with significant predictive power for fertilization disorders were selected from the above two models, and reproductive clinical experts were organized to score these indicators one by one. Based on the selection criteria of routine use, rapid acquisition, and low cost, the main indicators that can be promoted and used in the current actual clinical diagnosis and treatment process were screened out. The indicators included in the prediction model include:

[0042] (1) Medical history of the woman: age, body mass index, number of pregnancies, parity, adverse pregnancy and childbirth history, causes of infertility (fallopian tube factors, uterine abnormalities, ovarian abnormalities, history of low ovarian reserve, endometriosis, genetic factors), type of infertility (primary / secondary), and history of previous in vitro fertilization failure (yes / no);

[0043] (ii) Male partner’s medical history information: male partner’s age, body mass index, and semen parameters (normal, oligoasthenospermia, severe oligoasthenospermia, azoospermia, teratospermia);

[0044] (III) Treatment plan and examination results of the current cycle: number of bilateral antral follicles, ovulation induction plan (stimulation cycle, micro-stimulation cycle, natural cycle) and number of eggs retrieved.

[0045] Step 2: Use machine learning algorithms to build and optimize the fertilization disorder prediction model.

[0046] According to the indicators screened in step 1, a machine learning algorithm was used to construct a fertilization disorder prediction model in 100,000 reproductive clinical big data. The 10*10-fold cross-validation method was used to evaluate and screen multiple models to select the model with the best accuracy. Specifically: in each cross-validation, the data set was randomly divided into 10 parts, 9 of which were used as training data sets to train the model in turn, and the remaining 1 was used as a validation data set to validate the model. The same cross-validation process was repeated 10 times in total. Each validation will result in a corresponding accuracy rate. The accuracy rate of 10*10 (i.e. 100) model validations is used as an evaluation indicator for the model prediction accuracy. Finally, among the multiple models constructed, the model accuracy is used for screening to determine the best model (i.e. the model with the highest accuracy).

[0047] The above machine learning algorithms include:

[0048] 1) Feature selection: Based on data features or prior knowledge, delete some features that do not have characteristic properties or are not obviously helpful for result judgment.

[0049] 2) Data preprocessing: detect and process outliers, fill in missing values, convert discrete variables into dummy variables, and standardize and normalize continuous variables.

[0050] 3) Sample balance: Using oversampling, based on the original data ratio, the number of samples in the fertilization disorder group was increased, so that the number of samples in the fertilization disorder group and the control group was basically balanced.

[0051] 4) Dataset division: The training set, validation set, and test set are randomly divided in a ratio of 7:1:2. The training set is used to train the model, that is, to determine the parameters such as the model weight and bias. The validation set is used to select the model and hyperparameters. The test set is used to evaluate the final model after training.

[0052] 5) Data analysis.

[0053] a) In data analysis, multiple machine learning algorithms and ensemble models were used, including support vector machines, random forests, K nearest neighbors, XGBoost, CatBoost, and neural networks. Each model was trained on the training set, and the optimal parameters were selected on the validation set.

[0054] b) The final report includes the following:

[0055] Model accuracy on test set: Each model is evaluated on an independent test dataset and the accuracy of the model is reported.

[0056] Confusion Matrix: For each model, a confusion matrix is ​​provided that shows the correspondence between the model's predictions and the actual results.

[0057] ROC curve: For each model, a ROC curve is plotted, which is used to evaluate the classification performance of the model and provides a measure of the model performance.

[0058] Based on the above method, two prediction models, balanced and non-balanced, were constructed:

[0059] (1) Prediction model based on the original sample population: The predicted probability calculated by this model represents the population probability, that is, the true probability of occurrence. When used, the predicted probability is compared with the proportion of normal fertilization, low fertilization and complete infertilization in the general population (91.05%, 4.07% and 4.88% respectively). This prediction value is easy for professional doctors to understand and interpret. However, the ROC of this model did not reach the ideal state.

[0060] (2) Prediction model based on sample balancing: The predicted probability calculated by this model represents the individual probability, that is, the predicted probability after the sample data has been balanced. For each patient, the baseline probability of normal fertilization, low fertilization, and no fertilization is standardized to 1 / 3, and the predicted probability is compared with 1 / 3. This prediction value is easy for patients to understand.

[0061] Figure 2 The process of constructing a machine learning model of fertilization failure based on unbalanced raw data is demonstrated. By considering the actual fertilization failure and the original proportion of the control group sample size, multiple machine learning methods were used to construct a fertilization failure prediction model. Figure 2 In the left figure, different machine learning models for fertilization failure prediction are listed and their accuracy is shown. After sorting by accuracy, the top-ranked model (Catboost) is selected and Figure 2 The receiver operating curve (ROC) and area under the curve (AUC) of the model are plotted in the right figure. As can be seen from the figure, the AUC of the model in distinguishing TFF from the control group is 0.755, and in distinguishing LFR from TFF, the AUC is 0.775.

[0062] Figure 3 The process of building a fertilization disorder machine learning model based on balanced data is presented. Before building the model, we processed the data according to the structure of the original data so that the sample size of the fertilization disorder group and the control group was balanced. Then, we used multiple machine learning methods to predict the model. Figure 3 The left graph shows the accuracy of different machine learning models. Sorting by accuracy, we find that the WeightedEnsemble_L2 model is the best model. Figure 3 The right figure shows the receiver operating curve (ROC) and area under the curve (AUC) results of the model. When distinguishing between TFF and the control group, the AUC of the model is 0.957, and when distinguishing between LFR and TFF, the AUC is 0.977.

[0063] Step three: construct a prediction system based on the prediction model determined in step two, and select an in vitro fertilization plan based on the prediction results obtained by the system.

[0064] According to the fertilization obstacle prediction model determined in step 2, a system for selecting the best solution for in vitro fertilization technology is developed. The system is used to make predictions based on the basic information of both infertile patients. It can predict the probability of complete infertilization and low fertilization rate under the conditions of using different in vitro fertilization technologies (i.e., conventional in vitro fertilization technology [IVF] or intracytoplasmic sperm injection technology [ICSI]). This helps reproductive clinicians to formulate and make decisions on individualized best solutions for patients.

[0065] The specific steps of system development are as follows:

[0066] 1) Front-end framework

[0067] The front-end design adopts the html+css+javascript application framework, which mainly includes two pages: data collection and result display.

[0068] In the data collection page, a form is provided in the form of a form, covering three parts: serial number, indicator name and indicator value. Each indicator value is a selection type input box, corresponding to the possible value range of each indicator, which requires the patient to manually select. The value of each indicator value input box is set to be consistent with the displayed string to facilitate subsequent development.

[0069] On the result display page, the predicted outcomes of the two fertilization methods, ICSI and IVF-ET, and the probabilities of normal, low fertilization, and complete fertilization failure are listed in a table. In addition, the probability of the three results is intuitively displayed in the form of a pie chart through echarts rendering.

[0070] 2) Backend Framework

[0071] This application adopts a front-end and back-end separation architecture. The back-end design adopts the Flask architecture. After entering data on the front-end, the main process includes:

[0072] The data is transmitted to the backend via POST;

[0073] The backend reads the data separately from the received form and combines the data with different fertilization methods to form a pandas.DataFrame. The application processes the different fertilization methods and combines them with the input data for subsequent data analysis and processing.

[0074] Retrieve the trained model and perform classification training for different situations. The results include the final predicted outcome and the probability of the three outcomes.

[0075] The application returns the results of different situations to the front-end result display page in the form of a list. The final prediction results are generated based on the classification results of the model and returned to the front-end together with the probabilities of each result for users to view.

[0076] The following are two real cases of applying the best IVF selection system.

[0077] Case 1:

[0078] The basic information of the patients is as follows: the female is 25 years old, with a body mass index of 21.5kg / m2, 0 pregnancies and 0 miscarriages, and 3 miscarriages, and is diagnosed with endometriosis. The male is 28 years old, with a body mass index of 22.3kg / m2, and the semen test results are normal. The female has 10 antral follicles on both sides, and the ovulation induction regimen is a stimulation cycle. The number of eggs obtained after the injection of hCG on the trigger day is 21.

[0079] like Figure 4 As shown, according to the above information input into the prediction system, the system will automatically calculate the probability of complete infertilization and low fertilization rate under the conditions of using different in vitro fertilization technologies (IVF or ICSI) based on the input information.

[0080] Depend on Figure 4 It can be seen that, using the prediction model based on the original sample population, the probability of normal fertilization after ICSI and IVF for this infertile couple is 97.06% and 90.56%, respectively, the probability of low fertilization is only 2.32% and 5.83%, and the probability of complete fertilization failure is only 0.62% and 3.61%. It can be seen that if the infertile couple implements conventional IVF technology, the probability of normal fertilization is close to the normal fertilization rate in the whole population (91.05%), and the risk of fertilization disorders (low fertilization and complete fertilization failure) is less than 10%. Therefore, it is recommended to choose conventional IVF technology with relatively low cost and less invasive operation for this patient. Based on the prediction model after sample balancing, the probability of normal fertilization after ICSI and IVF is 97.59% and 92.97%, respectively, which is more than 1 / 3 of the probability at the individual level. Therefore, conventional IVF technology is also recommended.

[0081] Case 2:

[0082] The basic information of the patients is as follows: the female is 40 years old, with a body mass index of 20kg / m2, 1 pregnancy and 0 births, secondary infertility, diagnosed with low ovarian reserve, and no previous history of ART. The male is 44 years old, with a body mass index of 18.3kg / m2, and the semen examination results show severe oligospermia and asthenospermia. The female has 5 antral follicles on both sides, and the ovulation induction regimen is a stimulation cycle. The number of eggs obtained after the injection of hCG on the trigger day is 3.

[0083] like Figure 5As shown, the above information is entered into the prediction system, and the system will automatically calculate the probability of complete infertilization and low fertilization rate under the conditions of using different in vitro fertilization technologies (IVF or ICSI) based on the entered information.

[0084] Depend on Figure 5 It can be seen that using the prediction model based on the original sample population, the probability of normal fertilization after ICSI for this infertile couple is close to 100%, while the probability of complete fertilization failure after IVF-ET is as high as 50%. Using the prediction model based on sample balancing, the probability of normal fertilization after ICSI is 72.9%, which is significantly higher than IVF-ET. Therefore, considering the above two prediction results, ICSI is recommended.

[0085] In summary, this in vitro fertilization technology optimal plan selection system has the advantages of individualized prediction, technology comparison, and sample balanced processing, providing reproductive clinicians with a reliable tool for formulating optimal plans and making decisions, thereby improving the fertilization success rate and individualized treatment effects.

[0086] The above is an explanation of the embodiments of the present invention. However, the present invention is not limited to the above embodiments. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the risk of fertilization disorders and optimizing in vitro fertilization programs, characterized in that The following steps are involved: Step 1: Screening of routine clinical examination and laboratory indicators; The classical ordered logistic regression model and the Bayesian network model were used to construct a fertilization disorder prediction model; key predictive indicators with significant predictive ability for fertilization disorders were selected from the above models and evaluated; the main indicators that can be promoted and used in the clinical diagnosis and treatment process were screened out based on the selection criteria of routine use, rapid acquisition and low cost; Step 2: Use machine learning algorithms to build and optimize a fertilization disorder prediction model; According to the indicators screened in step 1, a 10-fold*10-fold cross-validation method was used to evaluate and screen multiple models in reproductive clinical big data to select the model with the best accuracy; Step 3: constructing a prediction system according to the prediction model determined in step 2, and selecting an in vitro fertilization plan based on the prediction results obtained by the system; According to the fertilization disorder prediction model determined in step 2, a system for selecting the best in vitro fertilization technology is developed. The system is used to make predictions based on the basic information of both infertile patients. It can predict the probability of complete infertilization and low fertilization rate under the conditions of using different in vitro fertilization technologies, thereby assisting reproductive clinicians in formulating and making decisions on individualized best plans for patients.

2. The method according to claim 1, characterized in that The key indicators described in step one include: Female medical history information: age, body mass index, pregnancy number, parity, adverse pregnancy and childbirth history, cause of infertility, type of infertility, and previous history of in vitro fertilization failure; Male medical history information: male age, body mass index and semen parameters; Treatment plan and examination results of the current cycle: number of bilateral antral follicles, ovulation induction plan and number of eggs retrieved.

3. The method according to claim 2, characterized in that in, The causes of infertility include fallopian tube factors, uterine abnormalities, ovarian abnormalities, history of low ovarian reserve, endometriosis and genetic factors; the types of infertility include primary and secondary; semen parameters include normal, oligoasthenospermia, severe oligoasthenospermia, azoospermia and teratospermia; ovulation induction plans include stimulation cycle, micro-stimulation cycle and natural cycle.

4. The method according to claim 2 or 3, characterized in that: The specific method of 10*10-fold cross-validation in step 2 is as follows: in each cross-validation, the data set is randomly divided into 10 equal parts, 9 of which are used as training data sets to train the model, and the remaining 1 is used as a validation data set to validate the model. The same cross-validation process is repeated 10 times in total; each validation will result in a corresponding accuracy rate, and the accuracy rate of 10*10 model validations is used as an evaluation indicator of the model prediction accuracy. Finally, among the multiple models constructed, the model accuracy is used for screening to determine the best model.

5. The method according to claim 4, characterized in that The machine learning algorithms in step two include: feature selection, data preprocessing, sample balancing, data set partitioning, and data analysis.

6. The method according to claim 5, characterized in that The data analysis in step 2 yields the following: Model accuracy on the test set: Each model is evaluated on an independent test dataset and the accuracy of the model is reported; Confusion Matrix: For each model, a confusion matrix is ​​provided, which shows the correspondence between the model's predictions and the actual results; ROC curve: For each model, a ROC curve is plotted, which is used to evaluate the classification performance of the model and provides a measure of the model performance.

7. The method according to any one of claims 1 to 6, characterized in that Based on the method described in step 2, two prediction models, balanced and non-balanced, were constructed respectively.

8. The method according to claim 7, characterized in that The non-balanced prediction model is a prediction model based on the original sample population. The predicted probability calculated by the model represents the population probability, that is, the actual probability of occurrence. When used, the predicted probability is compared with the proportions of normal fertilization, low fertilization and complete infertilization in the general population. This prediction value is easy for professional doctors to understand and interpret.

9. The method according to claim 7, characterized in that: The balanced prediction model is a prediction model based on sample balanced processing. The predicted probability calculated by the model represents the individual probability, that is, the predicted probability after the sample data is balanced. This prediction value is easy for patients to understand.

Images