{"paper_id":"0e0f7172-4f6a-4ee8-86fe-b4c731dbf1db","body_text":"In vitro  fertilization (IVF) is considered a popular technique used in\nassisted reproductive technology (ART) to promote the achievement of childbirth in the\npopulation of infertile individuals. Numerous aspects of IVF treatments have changed over\ntime. Substantial research has been conducted to improve IVF results by taking into\nconsideration its influencing factors; however, there is still a lack of knowledge about the\npredictors of IVF outcomes while the overall pregnancy rates have only reached approximately\n30% ( 1 ,  2 ).\nMany factors have been known to affect IVF outcomes\nincluding age, sperm quality, fertilization rate, embryo\nquality, frequency of transferred embryos, and endometrial thickness ( 3 ,  4 ). Determining influencing factors, could\npotentially influence the likelihood for a successful IVF\ntreatment; this would enable clinicians and physicians to\nmake better decisions in order to apply IVF based on patients’ characteristics ( 5 ). Patients who failed treatments\nmight experience adverse psychological problems such\nas depression and anxiety ( 6 ). Therefore, it is essential\nto assess factors associated with the outcome after IVF\nand determine the influencing factors. In order to reduce\npsychological and other negative outcomes after IVF, patients could evaluate the likelihood of successful IVF\nbased on their characteristics.\nThus, machine learning approaches have been designed\nto assess the relationship of an outcome and its effective\nvariables; The use of a hybrid intelligence method for\nknowledge exploring of a clinical database on IVF ( 7 ), an\nordered mechanism in comparison with naïve Bayes (NB)\nclassifier to estimate the odds of success after IVF ( 8 ),\nrandom forest (RF) and adaptive boosting in classifying\nthe state of ART ( 9 ), and logistic regression (LR) to predict implantation after blastocyst transfer ( 10 ) are some\nexamples of application of this approach on IVF data.\nHere, we used a clinical database that included each couple’s characteristics and available data on oocytes, sperm\nand embryos, as well as the cycle outcomes to classify the\nIVF outcome (successful/unsuccessful delivery) by NB, RF,\nsupport vector machine (SVM), extreme gradient boosting\n(XGBoost), linear discriminant analysis (LDA), and LR.\n\nWe conducted this historical cohort study in a referral infertility centre located in Tehran, Iran. Data from 6071 cycles performed during March 21, 2011 to March 20, 2014\nwere analysed. We included only those women for whom\nclinical pregnancy was confirmed observing an intrauterine gestational sac. The collected demographic and clinical variables comprised women’s ages, source of infertility\n(female factor, male factor, combined male-female factor\ninfertility, unexplained), infertility type (primary, secondary), body mass index (BMI), infertility duration (years),\nnumber of previous abortions, polycystic ovary syndrome\n(PCOS), number of previous IVF attempts, total number\nof retrieved oocyte, number of injected oocytes, number\nof embryos, number of transferred embryos, spermogram,\nfertilization rate after intracytoplasmic sperm injection\n(ICSI), number of two-pronuclear embryos to number of\nmetaphase II (MII) oocytes (2PN/MII ratio), and data on\nembryo quality (number of compact, blastocysts, grade\nA, grade AB, early blastocysts, A compact, and AB compact), as well as the day of the embryo transfer (ET).\nThe descriptive characteristics of the data are shown\nusing mean (standard error) and frequency (percentage)\nfor continuous and categorical variables, respectively.\nWe used the independent samples t test after checking\nthe normality of data distribution to compare the mean\nof the variables across the categories of the response. The\nchi-squaretest was used to assess the independence of categorical variables with the outcome.\nA principle component approach was utilized to reduce\nthe dimension of multiple independent variables into\nsmaller components. To do so, the variables that included\nthe numbers of compact, blastocyst, grade A, grade AB,\nearly blastocyst, A compact and AB compact, and the day of the ET were entered in the principle component analysis. The best number of components is decided according\nto the highest determined variance of the variables so that\nthe majority of variability in the independent variables is\navailable in the result antcomponents.\nFor the classification approaches, we randomly divided\nthe data into two sets of train (70%) and test (30%). The\ntrain set was used to fit the model and the validation of the\nresults was checked by the test set. In order to classify the\nstatus of delivery (successful/unsuccessful), we compared\nthe results from the following six techniques: LR, SVM,\nXGBoost, RF, NB, and LDA. Sensitivity (SE), specificity (SP), positive predictive value (PPV), negative predictive value (NPV), accuracy (ACC), area under the curve\n(AUC) and 95% confidence interval were used to assess\nthe performance of the models. In order to find more reliable results, we repeated each technique 500 times. The\nmean ACC measures are presented.\nThe statistical programing R software version 3.2.3\n(http://www.R-project.org) packages that included RF,\nNB, e1071, XGBoost, and MASS were used for data\nanalysis. The type one error was assumed as 0.05.\nThe Ethics Committee of Royan Institute (approval\nnumber: IR.ACECR.ROYAN.REC.1395.62), Tehran,\nIran approved this study. The information used in this\nstudy was obtained from the data routinely registered in\nthe patients’ medical records.\n\nAmong the assessed cycles, 4930 (81.2%) cycles resulted in successful deliveries. In the\nanalysis, 23 variables were assessed and eight variables were summarized into four\ncomponents using the principle component analysis. Finally, the association of IVF outcome\nand the 19 variables were evaluated. Table 1 lists the mean or frequency of the variables\nfor both successful and unsuccessful deliveries. The unadjusted results are shown using the\nt test and chi-square test for continuous and categorical variables, respectively. The\nduration of infertility for those who delivered successfully was 0.40 years less than those\nwith unsuccessful deliveries (t-score: 2.75, P=0.006). The mean number of previous IVF\ncycles was higher for cases without successful deliveries (t-score: 2.46, P=0.014). The\nnumber of injected oocytes among cases with successful deliveries was higher than those with\nunsuccessful deliveries (t-score:-1.99, P=0.046). Cases with successful deliveries were\nsignificantly 1.35 years younger (t-score: 8.78, P<0.001) and had 0.60\nkg/m 2  lower BMI (tscore: 4.67, P<0.001). We noted that patients with\nPCOS had more successful deliveries (chi-square: 6.83, degree of freedom [df]: 1, P=0.009).\nMale factor (chi-square: 18.25, df: 5, P=0.003), frequency of previous abortions(chi-square:\n19.62, df: 2, P<0.001), and primary type of infertility (chisquare: 5.02, df: 1,\nP=0.025) were associated with a higher probability of successful delivery. Table 1 provides\nadditional details of the patients’ characteristics.\nPatients’ characteristics in the successful and unsuccessful delivery groups\nC1; Number of compact and blastocysts, C2; Number of grade A and grade AB, C3; Number of early\nblastocysts, A compact and the day of ET, and C4; Number of AB compact, SD; Standard\ndeviation, df; Degree of freedom, IVF;  In vitro  fertilisation, BMI;\nBody mass index, PCOS; Polycystic ovary syndrome, ET; Embryo transfer, and ICSI;\nIntracytoplasmic sperm injection.\nThe principle component analysis reduced eight embryo\nfactors (number of compact, blastocyst, grade A, grade AB,\nearly blastocyst, A compact, AB compact, and day of ET)\nto four components. The components were: C1 (number\nof compact and blastocysts); C2 (number of grade A and\ngrade AB); C3 (number of early blastocysts, A compact and\nthe day of ET), and C4 (number of AB compact).\nThe six classification methods, including NB, RF, LDA\nand LR, were applied. Table 2 shows a comparison of\ntheir ACC measures. Except for LR, other classification\nmethods resulted in almost the same and high SE and\nPPV (SE>0.80, PPV>0.99). In contrast, the SP and NPV\nof LR was higher than the other approaches (SP=0.50,\nNPV=0.27). The total accuracy of the results showed\nthat LR (ACC=0.64) had the worst performance where\nas RF (ACC=0.81) had the best performance among\nthe six applied approaches. Moreover, the AUC for RF\n(AUC=60; 0.55–0.64), LDA (AUC=0.57; 0.51–0.63), LR\n(AUC=0.55; 0.49–0.61), and NB (AUC=0.53; 0.47–0.58)\nconfirmed as lightly higher accuracy for RF compared to\nthe other methods.\nA comparison of the six applied classification techniques using the accuracy measures\nSVM; Support vector machine, XGBoost; Extreme gradient boosting, LDA; Linear discriminant analysis, LR; Logistic regression, RF; Random forest, NB; Naïve Bayes, SE; Sensitivity,\nSP; Specificity, PPV; Positive predictive value, NPV; Negative predictive value, ACC; Accuracy, and AUC; area under the curve.\nThe importance of variables that affect successful delivery according to the RF approach, which\nhad the best performance among the classification approaches. C1; Number of compact and\nblastocysts, C2; Number of grade A and grade AB, C3; Number of early blastocysts, A\ncompact and the day of ET, and C4; Number of AB compact, RF; Random forest, PCOS;\nPolycystic ovary syndrome, IVF;  In vitro  fertilisation, BMI; Body mass\nindex, and ACC; Accuracy.\nFigure 1shows the importance of variables that affected\nsuccessful delivery using the RF method. The total number of embryos, number of injected oocytes, cause of infertility, women’s age, and PCOS were the affecting predictors for a successful delivery that had a higher amount\nof importance in comparison to the other variables.\n\nThe aim of this study was to compare classical regression\nbased methods with machine learning methods. We\ncompared these techniques in an attempt to gain a better\nunderstanding and prediction of IVF outcomes. The\napplication of these methods in IVF data is supposed to\nimprove efficiency by estimating the chance of success.\nGeneralizable and reliable prediction methods can help\nfulfil this purpose.\nIn the statistical analysis of our paper, six data mining\nprocedures were fitted and compared to investigate\nsuccessful delivery. Based on the ACC tools, the RF best\nfitted the data. There are several possible explanations for\nthe better performance of the RF method in this study.\nFirst, it might be explained by the fact that modern\nmodelling methods such as RF tend to be “data hungry”\nand are believed to perform better with a higher eventsper-variable ratio than classical methods ( 11 ). The larger\nnumber of continuous variables than categorical variables\nin this study could be another possible explanation for\nthe better RF performance. It may also be due to the\nfact that tree based methods like RF account for variable\ninteractions, while regression based methods like LR do\nnot ( 12 ). The goodness of fit for determining a machine\nlearning approach is a function of rates in levels of the\noutcome. Therefore, it does not seem to be quite rational\nto focus only on the total accuracy ( 13 ). A few other\nresearch indicate inconsistent performance of various\nclassification algorithms with respect to the small/\nhigh prevalence of the outcome ( 14 ,  15 ). The manner\nunder which the predictor variables influence the result\nis essential for deciding the correct form of method.\nTherefore, discrepancies could be reported in performing\nclassification techniques in various data areas.\nSeveral studies have assessed machine learningbased prediction models in different outcomes during\nART (e.g., embryo implantation, ongoing pregnancy,\nclinical pregnancy, pregnancy) ( 16 ). Uyar et al. ( 17 )\nfound that higher accuracy rates might be obtained by\nusing morphological variables of individual embryos\nutilizing NB method for implantation prediction.\nHafiz et al. ( 9 ) demonstrated that RF performed better\nthan SVM, recursive partitioning (RPART), adaptive\nboosting (Adaboost), and nearest neighbour in predicting\nimplantation outcomes of IVF and ICSI. The dataset in\ntheir study was highly unbalanced as the number of those\nwith negative implantation was more than the positive.\nThey explained the poor performance of SVM with the\nunbalanced nature of medical datasets, in particular,\nthe one used in their study. In another study, Hassan et al. ( 18 ) compared a series of classifiers (SVM, RF,\nmultilayer perceptron neural network [MLP], decision\ntree, classification and regression trees [CART] and\nartificial neural networks [ANN]) to predict pregnancy\noutcomes for IVF treatment. They reported that SVM and\nRF performed almost the same and both were better than\nthe other classifiers in terms of prediction ACC and AUC.\nThe results demonstrated that selection of a set of features\nfor each method significantly improved the prediction\nACC of pregnancy success.\nThe result of this study showed that the total number\nof embryos obtained in each cycle was associated with\nsuccessful live birth. This finding supported those reported\nby Bartmann et al. ( 19 ), who used an artificial intelligence\nsystem to calculate pregnancy chance by taking into\nconsideration the patients’ clinical and laboratory\ninformation. They showed that the number of embryos\nobtained was the best discriminant variable for pregnancy\nprediction; according to the artificial intelligent system\ndeveloped in this study, women with more embryos tended\nto have greater chances for pregnancy. It was reported that\nthe total number of embryos might be a surrogate marker\nfor hormonal factors that act via uterine receptivity ( 20 ).\nIn the current study, the number of injected oocytes was\nanother important variable that predicted IVF outcome. In\na historical cohort study on 996 infertile women, modified\nPoisson regression analysis demonstrated that females\nwho attained clinical pregnancy had a significantly\ngreater number of injected oocytes compared with those\nwho failed to achieve pregnancy ( 21 ). In another study,\nthe number of injected oocytes was positively associated\nwith the number of grade A embryos and could be a\ndeterminant of a successful ART ( 22 ). Zorn et al. ( 23 )\nconducted a study of influencing gender characteristics of\nICSI outcome in azoospermic and aspermic patients. They\nobserved a positive association between the frequency of\ninjected oocytes with reaching the blastocyst stage and\nlive birth.\nCause of infertility is another important variable that\naffects IVF outcome, which has been confirmed by the\nresults of numerous similar studies ( 24 ). Nelson and\nLawlor ( 25 ) predicted live birth and weight at birth among\ninfants born from IVF. They observed that male cause\nof infertility was linked to lower chances of successful\npregnancy in patients who did not receive ICSI. Factors\nassociated with failed treatment were evaluated by\nBhattacharya et al. ( 26 ); the results showed that the risk of\npoor fertilization was more common among patients with\ntubal disease, male factor, and endometriosis. Moreover,\nthey noted that the risk of non-live births among those\nwith tubal disease and male factor was higher than those\nwith unexplained infertility. It has been demonstrated\nthat cause of infertility plays a role in determining poor\nintermediate outcomes. Elizur et al. ( 27 ) investigated the\npredictive factors for IVF treatment pregnancy results;\nthey observed that delivery rate among those with male\nfactor was significantly higher than other aetiologies.\nA woman’s age was another significant factor for\nachieving a successful pregnancy. It has been widely\ndebated that with increasing of female age, the IVF\noutcomes become increasingly worse. Among infertile\ncases, the Society for ART (SART) stated that 47% of\nETs among women younger than 35 years of age resulted\nin successful delivery. The proportion was 38% for ages\n35–37, 28% for ages 38–40, 16% for ages 41–42 and 6%\nfor older than 42 years of age ( 28 ). Nazemian et al. ( 29 )\ninvestigated the impact of age on IVF outcome. They\nreported that cases younger than 25 years of age have\nlower fertilization rates as well as a decreased frequency\nof high quality embryos. In their study, clinical pregnancy\nand implantation rates were similar to those who were\n30-35years of age. In another research, Yan et al. ( 30 )\nevaluated the mechanism by which maternal age affects\nthe outcomes of IVF cases. Patients older than 40 years\nhad a disadvantaged IVF outcome and increased numbers\nof miscarriages.\nThe current work shows that PCOS is a potential\ninfluencing factor for live birth. Beydoun et al. ( 31 ) have\nreported that PCOS has a distinct effect on the early stages\nof pregnancy among women who undergo IVF/ICSI, but\nnot on the later stages. Ryan et al. ( 32 ), in a study of a\nlarge number of infertile women, showed that women\nwith PCOS had increased odds for childbirth. Moreover,\nPCOS significantly confounded the relationship between\nthe duration of ovarian stimulation and treatment success.\nEarlier findings also showed a greater number of oocytes\nwere retrieved in PCOS women compared to women\nwithout PCOS ( 31 ), and greater number of follicles>16\nmm and MII oocytes in PCOS women compared to\nwomen with subfertile male partners and those with\nunexplained infertility ( 33 ). These results imply a higher\namount of ovarian capacity in PCOS women and the\ncompensatory impact of this capacity ( 34 ,  35 ). However,\nthe results from a large number of studies mentioned that\nthe role of PCOS in ART success mainly depended on\nobesity, insulin resistance, and other metabolic syndrome\nfeatures ( 36 ,  37 ).\nThis study had several limitations. First, this research\nwas carried out in one infertility clinic and this limits the\ngeneraliz ability of our findings. Second, other predictors\nsuch as basal FSH and somegenetic features are potential\nfactors that were not recorded by the Centre ( 38 ,  39 ).\nThird, the distribution of IVF outcome was not balanced\n(unbalanced dataset). Fourth, the AUCs were relatively\nsmall and the performance of the models was compared\nusing accuracy tools in conjunction with the AUC.\nRF performance has less dependence on parameter\nvalues than other machine learning methods. However,\nin future investigations it might be possible to achieve\nmore improvements in this method by using optimization\nprocedures to simultaneously tune the RF parameters or\nuse RF based on conditional inference trees to address the\nproblem of variable selection ( 40 ).\nResults obtained from machine learning could help to determine the risk factors and their impact in real world\nsettings. It could also help to predict the personalized\nchance of an ART outcome before the treatment procedure.\nThis would assist clinicians decide whether it is worth to\nstartan ART procedure and would also provide infertile\ncouples information about the chances for success.\n\nThis study sought to classify IVF successful delivery\nbased on six machine learning approaches: SVM,\nXGBoost, LDA, LR, RF and NB by using couples’\ncharacteristics and available data on oocytes, sperm, and\nembryos. This study indicated that successful delivery\nafter ART is strongly dependent on various characteristics\nof the patients, which included total number of embryos,\nnumber of injected oocytes, cause of infertility, age\nof women, and PCOS. Our results indicated that the\nRF approach could be a better choice to classify ART\noutcome among other classification methods. These\nresults could assist clinicians to have a better prediction\nand management of ART treatment and advise patients\naccordingly.","source_license":"CC-BY-4.0","license_restricted":false}