| Objective:At present,the IVF technology is increasingly mature and widely used in the treatment of infertility,but it has some problems,such as a low success rate of pregnancy outcome,expensive and time-consuming treatment,and difficult decisionmaking of the treatment plan,which bring great trouble to clinicians and patients.In the IVF treatment cycle,how to determine the protocol of ovarian superovulation to obtain the appropriate number of oocytes and how to choose a reasonable embryo transfer strategy,are very critical issues affecting the clinical pregnancy outcome.However,these questions should be addressed by clinical and embryonic physicians that have rich experience and knowledge,and the differences in experience and knowledge of different physicians will also have a great impact on the outcome of IVF treatment.The purpose of this study was to establish a regression model for predicting Gn starting dose,a regression model for predicting the number of oocytes retrieved,and a classification model for the pregnancy outcome with a machine learning method from the retrospective analysis of the patients undergoing IVF cycle treatment in our hospital.Based on the established regression models,the comprehensive evaluation index and method of its influencing factors are designed to realize the quantitative and accurate evaluation of the importance and trend of the influencing factors.According to the established models and evaluation method,the individualized clinical treatment plan can be formulated,the complications of IVF can be avoided or reduced,and the clinical live birth rate can be improved.At the same time,it can help patients set reasonable expectations on the IVF treatment results,simplify doctor-patient communication,and increase patient compliance.Part Ⅰ The modeling of Gn starting dose and the comprehensive evaluation of the influencing factorsMethods:A total of 1555 patients were collected from October 2019 to December 2020 in the Reproductive Center of the People’s Hospital of Wuhan University.The correlation analysis was used to explore the relationship between the female age,cause of infertility,type of infertility,duration of infertility,BMI,AFC,bFSH,E2,LH,AMH and COS treatment regimen and the Gn staring dose.The ANN model and SVR model of Gn starting dose were established by incorporating the screened highly correlated factors into machine learning.By comparing the prediction error and regression coefficient of the two models,the Gn starting dose prediction model with better modeling results was selected.A comprehensive evaluation index of the influence factor based on a machine learning model is defined to quantitatively and accurately reflect the importance and trend of each influence factor on the Gn starting dose.The data were analyzed by SPSS 24.0 software.Pearson and Spearman correlation analysis was utilized for single factor correlation analysis.P<0.05 was considered to be significant.The programming and implementation of ANN and SVR machine learning algorithms are completed in MATLAB R2021a.Results:1.Univariate analysis showed that age,duration of infertility,type of infertility,BMI,AFC,bFSH,E2,LH,AMH and COS treatment regimen were closely related to the starting dose of Gn（P<0.05）.2.The statistically significant impact factors were used as predictors,and the adaptive weight adjustment method based on the number of retrieved oocytes was designed to build the ANN model and SVR model of Gn starting dose.From the training set,test set and the overall sample,the ANN model is much better than the SVR model in terms of both prediction error（RMSE,29.87 vs.34.15;34.21 vs.37.27;31.45 vs.34.76）and precision（regression coefficient R,0.953 vs.0.947,0.942 vs.0.928 and 0.951 vs.0.943）.3.Based on the prediction model,a comprehensive evaluation of each influencing factor is conducted,with the results showing that:the ranking of continuous factors influencing the Gn starting dose was age（miv*=1）,AMH（miv*=-0.841）,AFC（miv*=-0.499）,BMI（miv*=0.439）,LH（miv*=-0.270）,bFSH（miv*=0.233）,years of infertility（miv*=0.014）,E2（miv*=-0.004）.According to the ranking,the top three influencing factors are age,AMH and AFC,then followed by BMI,LH and bFSH,and E2 has the lowest influence.Summary:1.A model capable of accurately predicting the Gn starting dose is established,the prediction accuracy is high,and the prediction error is low.2.The model can quantitatively evaluate the influencing factors of continuous variables included in the model.For the prediction of Gn starting dose,the top three influencing factors are age,AMH and AFC,followed by BMI,LH and bFSH.3.The established prediction model of Gn starting dose is of great significance to assist clinicians in customizing the initial dose of Gn and reducing the uncertainty caused by the differences between doctors’ clinical experiences.Part Ⅱ The modeling of the number of oocytes retrieved of hyperstimulation ovarian response and the comprehensive evaluation of influencing factorsMethods:From October 2019 to December 2020,a total of 1365 patients with IVF oocyte retrieval cycles were collected from the Reproductive Center of the People’s Hospital of Wuhan University.Correlation analysis was used to explore the close correlation between the number of retrieved oocytes and the female age,type of infertility,duration of infertility,cause of infertility,BMI,AFC,bFSH,E2,LH,AMH,COS treatment regimen,Gn days,Gn dose,and E2 level on HCG day.The selected significant factors were brought into machine learning,and the ANN model and SVR model of the number of retrieved oocytes were established.The RMSE and regression coefficient R of the two models were compared,and the model with better modeling results was selected as the final prediction model.Based on the comprehensive evaluation index,the influence trend and degree of each influencing factor on the number of retrieved oocytes can be quantitatively and accurately reflected.SPSS 24.0 software was used for data statistics,and Pearson and Spearman correlation analysis was used for single factor correlation analysis.P<0.05 was considered to be significant.The programming and implementation of ANN and SVR machine learning algorithms are completed in MATLAB R2021a.Results:1.Univariate analysis showed that age,infertility type,infertility factors,duration of infertility,AFC,bFSH,AMH,COS treatment regimen,Gn days,Gn dose,and E2 level on HCG day were all significant factors affecting the number of oocytes retrieved（P<0.05）.2.The ANN model and SVR model for predicting the number of retrieved oocytes were established by using the significant influencing factors as predictors.From the training set,the test set and the overall sample,the ANN model is far superior to the SVR model in terms of both prediction error（RMSE,2.59 vs.3.66;2.80 vs.3.89 and 2.63 vs.3.70）and precision（regression coefficient R,0.896 vs.0.810,0.840 vs.0.782 and 0.882 vs.0.799）.3.Based on the prediction model,the comprehensive evaluation results of the influencing factors were as follows:age（miv*=-0.354）,duration of infertility（miv*=-0.039）,AFC（miv*=1）,bFSH（miv*=-0.131）,AMH（miv*=0.314）,days of Gn（miv*=0.234）,dose of Gn（miv*=0.219）,and E2 level on the HCG day（miv*=0.951）.By ranking the influence degree from high to low,the top three factors were AFC,HCG,E2 level on the HCG day,and age;followed by AMH,days of Gn,dose of Gn,and bFSH;and the influence of infertility years on the number of retrieved oocytes was low.Summary:1.An ANN model capable of accurately predicting the number of oocytes retrieved is established,the prediction accuracy is high,the prediction error is low,and the prediction results are superior to models that are constructed based on any influencing factor alone.2.For the modeling of the number of oocytes retrieved,the first three influence factors with the largest importance were AFC,E2 level on the HCG day,and age;followed by AMH,days of Gn,dose of Gn,and bFSH.3.As an auxiliary tool,the established prediction model can customize the ovulation stimulation scheme according to the comprehensive evaluation value of each influencing factor and the prediction result of the model,which can help obtain the target number of retrieved oocytes,and provide a good basis for the subsequent IVFET process.Part Ⅲ The modeling of live birth result of the pregnancy outcomeMethods:Atotal of 1405 patients undergoing IVF-ET in the Reproductive Center of Rennin Hospital of Wuhan University from October 2019 to December 2020 were collected and analyzed by univariate and multivariate analysis,so as to find the significantly related factors to the live birth outcomes from the following factors:basic clinical characteristics（female age,BMI,infertility type,infertility years,cause infertility）,the clinical cycle characteristics（COS treatment regimen,Gn starting dose,OSI,E2 level on the HCG day,P level on the HCG day,LH level on the HCG day,endometrial thickness on the HCG day）,and embryo transfer indicators（2PN number;the number of transferable embryos;embryo transfer strategy）.The selected statistical meaningful factors were brought into machine learning as predictors to establish ANN model and SVM model for predicting the outcome of live birth,and finally,the one with better prediction results was selected as the final prediction model.SPSS 24.0 software was used for data analysis,independent sample t-test and χ2 test was used for univariate analysis,binary logistic regression analysis was used for multivariate analysis,and P<0.05 was considered to be significant.The programming and implementation of ANN and SVM machine learning algorithms are completed in MATLAB R2021a.Results:1.Results of univariate and multivariate analysis:the factors closely related to the outcome of live birth（P<0.05）were female age,OSI,treatment regimen,Gn starting dose,endometrial thickness on the day of HCG,P level on the day of HCG and embryo transfer strategy（the period and number of embryos transferred）.2.By using the above factors as predictors,the ANN model and SVM model of the IVF live birth outcome were established.Compared with ANN model,SVM model has much better results for both training set and test set:training set（AUC:0.912 vs.0.726,precision:86.41%vs.67.23%,sensitivity:75.58%vs.67.72%,accuracy:84.78%vs.75.52%,F1-score:80.63%vs.67.47%）;test set（AUC:0.854 vs.0.701,precision:77.50%vs.62.96%,sensitivity:76.86%vs.68.92%,accuracy:80.43%vs.71.04%,F1score:77.18%vs.65.81%）.Summary:I.An SVM model was established to accurately predict IVF live birth outcomes with high AUC,precision,sensitivity,accuracy,and F1 score.2.The established model is constructed based on the basic clinical characteristics of infertile patients,COS treatment,ovarian responsiveness index（OSI）,P level on the HCG day,endometrial thickness on the HCG day,embryo transfer strategy and others,covering all the essential processes in the IVF treatment of fresh transfer cycle.3.As a scientific tool for predicting the outcome of live birth,the established model can guide clinicians and embryologists to choose appropriate embryo transfer strategy and also help patients to have an objective and correct assessment of the outcome of IVF.It is of great significance to make decisions on IVF treatment,alleviate the psychological burden of patients and increase their compliance.Part Ⅳ Application and clinical validation of the proposed modelsMethods:From January 2021 to February 2021,the patients seeking for IVF-ET in the Reproductive Center of Renmin Hospital of Wuhan University were involved,and these patients now have complete COS results and pregnancy outcomes.Eighty-two patients were selected in the clinical application to verify the modeling effect of the models.The prototype software of IVF machine learning prediction/decision system（v1.0）is developed.The data calculation and processing related to the application and verification of the model are completed in MATLAB R2021a,and the code and graphical interface development of the software are completed in Visual Studio 2019 and QT 6.0,respectively.Results:1.The Gn starting dose predicted from the model can approach the actual value with good precision,with the mean absolute error of only 11.26 units.The recommended Gn starting dose can make the number of oocytes retrieved close to the optimal number of 16.2.The predicted oocyte number from the model was close to the actual oocyte number in IVF treatment,with the average error between them only 1.64.The modelbased decision of the COS treatment regime was realized,and the decision results were consistent with the actual regime used by clinicians:71.95%of the patients’ modelbased treatment regime was consistent with the actual strategy used in the clinic.3.The metrics of the prediction results of the live birth outcome model were:AUC=0.8825,accuracy=90.57%,sensitivity=75%,accuracy=74.39%,F1-score=82.05%.The model had good and valid prediction results.The optimal decision of the embryo transfer strategy is realized based on the proposed model,and compared with the decision of the clinician that purely relies on experience,the embryo transfer strategy based on the model could have better pregnancy outcome for some patients.Summary:1.The established Gn starting dose model,the number of retrieved oocytes model and the live birth model are all good summaries and reflections of the experience and knowledge of the senior clinicians.2.Based on the three models,the clinical applications are realized for the accurate recommendation of Gn starting dose,the decision-making of treatment plan,and the decision-making of embryo transfer strategy.3.The decision-making results from the models have been successfully applied to many aspects of the clinical treatment of IVF cycle,which has good scientific and clinical significance for assisting clinicians to make scientific and objective decisions,and ultimately achieving the goals of a better live birth rate,lower treatment cost and fewer side effects of IVF. |
PDF Full Text Request