|Year : 2018 | Volume
| Issue : 4 | Page : 224-229
Predictive modeling in reproductive medicine
Jing Lin1, Xiao-Xi Sun2
1 Shanghai Ji Ai Genetics and IVF Institute, Obstetrics and Gynecology Hospital of Fudan University, Shanghai 200011, China
2 Shanghai Ji Ai Genetics and IVF Institute, Obstetrics and Gynecology Hospital of Fudan University; Laboratory of Female Reproductive Endocrine-Related Diseases, Shanghai 200011, China
|Date of Submission||28-Aug-2018|
|Date of Web Publication||11-Jan-2019|
Shanghai Ji Ai Genetics and IVF Institute, Obstetrics and Gynecology Hospital of Fudan University, 588 Fangxie Road, Shanghai 200011
Source of Support: None, Conflict of Interest: None
The accurate prediction of fertility outcomes is an extremely interesting and challenging task in reproductive medicine. Efforts in this area focus on classic statistical models and newer technologies, including machine learning. The modeling process has three steps, namely, data preparation, model selection and data fitting, and model validation. Here, we present a review of studies on these methods of fertility prediction. Various databases were searched using relevant keywords. Original studies with full-text available on this topic were included for review. Earlier studies explored prediction models for spontaneous pregnancy prognosis, reproductive outcomes after intrauterine insemination and in vitro fertilization, and implantation potential based on embryo morphology and morphokinetic data. Future directions for predictive modeling in reproductive medicine include solving problems presented by big data, identifying novel informative features, balancing predictive power and result interpretability, and validating models with gold-standard study designs.
Keywords: Machine Learning; Prediction Model; Predictive Modeling; Regression Model; Reproductive Medicine
|How to cite this article:|
Lin J, Sun XX. Predictive modeling in reproductive medicine. Reprod Dev Med 2018;2:224-9
| Introduction|| |
The accurate prediction of fertility outcome is an extremely interesting and challenging task in reproductive medicine. Clinical gut feelings or experience-based predictions have only slight-to-fair reproducibility, and consequently poor predictive quality. As there is an obvious need for eliminating the interphysician discrepancies, numerous alternative objective methods have been discovered to assist in human decision-making. Efforts in this area focus on classic statistical models and newer technologies, including machine learning. Here, we present a review of studies on the classical statistics and machine learning methods used in fertility prediction.
| Development of Prediction Models|| |
Methods and algorithms
In line with the well-known no-free-lunch theorem, which describes the connection between methodologies and learning tasks, there is no single model that works universally and optimally for all situations. It is difficult to foresee which model will be better at making predictions based on a given dataset or to explain model performances once prediction results are available. Therefore, there is no reason to set limits on prediction methods in the pursuit of good predictive power.
Classical statistics often focus on odds ratios provided by logistic regression and hazard ratios provided by Cox proportional hazards regression. Machine learning methods have taken the data analysis world by storm, offering a wider range of options for solving the same problems and even the information-rich tasks such as embryo images or time-lapse data. The k-nearest neighbor algorithm is an instance-based learning method, which assigns class based on nearest-neighbor decision rule. Decision trees map features to outcomes with a splitting procedure which recursively splits the source set at each node or branch point into disjointed subsets based on the value of a particular feature. The support vector machine (SVM) categorizes the input data by applying the statistics of support vectors and constructing a hyperplane in much higher dimensional space. In addition, ensemble modeling of various learning algorithms has been explored to deliver a superior prediction.
Deep learning is a subfield of machine learning based on deep neural networks, a concept inspired by the structure and function of biological neural networks. An artificial neural network (ANN) presents a system of stacked layers with interconnected processors or nodes, which can compose increasingly complex features in each successive layer. The raw information is fed into the input layer, proceeds through the hidden layers via a system of weighted connections, and then output values of these transformed features are produced in the output layer to predict outcomes. In clinical settings, the input layer may represent medical data, the output layer may represent prognosis subclasses, and the multiple hidden layers may represent feature detectors for capturing higher-order correlations.
Three steps for predictive modeling
The modeling process has three basic steps [Figure 1] used iteratively until a desired model is constructed. First of all, we need accurate, actionable, and accessible data to extract informative features for accurate classification, which is the key ingredient of any successful model. Full consideration should be given to common data-related problems such as data cleansing, variable preparation, and handling of missing values and outliers. For complex learning tasks, more flexibility is introduced by some higher-order combinations or transformation of features for the identification of the best fitting variables.
|Figure 1: Three basic steps for predictive modeling. The model-building steps are used iteratively until a desired model has been constructed. Step 1: Data preparation, Step 2: Model selection and data fitting, Step 3: Model validation. ROC: Receiver operating characteristic.|
Click here to view
In the model selection and data-fitting step, prior knowledge, assumptions, and plots of data are used to select the form of the model. An appropriate data-fitting model is constructed using selected methods and the best candidate variables. The developed model should perform well against the original dataset.
Finally, validation measures are used to assess how well a model will generalize to new unseen data with regard to the ability of the model to distinguish samples with different labels (discrimination) and the agreement between predicted and observed probabilities (calibration). For regression-based frameworks, common validation measures involve mean squared error and coefficients of correlation. For classification-based frameworks, the area under the receiver operating characteristic (ROC) curve (area under the ROC curve [AUC], also known as c-statistic) is used as a measure of goodness of fit. The confusion matrices can also be used to visualize model performance, especially on data with more than two labels.
The model-building process is repeated if the validation step identifies problems with the current model and uses information from the validation step for model tuning.
| Predictive Modeling in Reproductive Medicine|| |
A systematic search was performed in July 2018 on different databases, including PubMed, Web of Science, Chinese National Knowledge Infrastructure, and Wanfang Data. For the search, various combination of keywords, such as prediction, regression, machine learning, artificial intelligence, data-mining, assisted reproduction, pregnancy, embryo, in vitro fertilization (IVF), and intrauterine insemination (IUI), were used. Original studies analyzing prediction models in reproductive medicine were included for review. The search results were merged, duplicates were removed, and the articles were then screened by title and abstract. For further evaluation, full texts of the filtered articles were obtained. A manual search of reference lists of the primary studies and reviews was also performed. The commonly studied models were selected, and the results were qualitatively presented.
Prediction of natural pregnancy
Many earlier studies have explored prediction models for spontaneous pregnancy prognosis in untreated subfertile couples. In 1994, Eimers et al. employed a Cox regression model to estimate the likelihood of treatment-independent pregnancies leading to live births within 1 year in 996 subfertile couples from a Dutch university fertility center. The relevant prognostic factors included age of the female, duration of subfertility, type of subfertility (primary or secondary), subfertility history in the male's family, postcoital test (PCT) result, nonprogressive PCT result, and percentage of motile sperm cells. The calibration and discriminative ability of the predictions from Eimers model were externally validated in 1,061 couples from Canadian university fertility centers in 2002.
In 1995, the Collins model was developed with the proportional hazards analysis based on data derived from 873 untreated couples from 11 Canadian university hospitals. It included a series of predictors, namely female age, duration of subfertility, pregnancy history, endometriosis, tubal defects, and male defects. This model had a moderate discriminative power as shown by the ROC analysis (AUC = 0.59).
The Snick model was developed in 1997 to estimate the likelihood of live births in 726 subfertile couples from a Dutch primary care fertility center. It identified the duration of subfertility, tubal defects, ovulation defects, and PCT result as significant and independent predictors of reproductive outcomes in primary care settings.
Hunault et al. went even further by creating two integrated models based on the three previously published frameworks to predict the probability of spontaneous pregnancies leading to live births. Referral status of the couple, female age, duration of subfertility, primary or secondary female subfertility, PCT result, and percentage of motile sperm cells were considered to possess independent prognostic information in the first model. The second model included the same predictors except the PCT result. The discrimination and calibration of the models were assessed in an external validation study on 302 couples from two Dutch university hospitals in 2005. In 2007, a prospective validation study was conducted for evaluate the general applicability of these adapted models for calculating spontaneous ongoing pregnancies (www.freya.nl/probability.php) in 3,021 couples from 38 hospitals in the Netherlands. The prediction models had strong calibrations and relatively poor discriminative capacities (c-statistic = 0.59).
Models such as these can be used to correctly stratify couples into clinically useful risk strata and assist fertility specialists in the tailoring of personalized management. If the chances of conceiving without treatment are favorable, watchful waiting may be a viable option. In contrast, if the chances are low, the couple should be advised to undergo treatment.
Prediction of pregnancy after intrauterine insemination
IUI is an assisted conception technique that places a processed motile semen sample directly into a woman's uterus. Similar tools have been created for the prediction of IUI outcomes. In 1996, a retrospective analysis of 260 IUI cycles employed logistic regression to predict pregnancy based on clinical and laboratory variables. Duration of infertility, follicle number, endometrial thickness, and progressive motility were included in this model.
To overcome the methodological shortcoming of small sample size, Steures et al. developed a logistic regression model in 2004 to predict ongoing pregnancy after IUI in 3,371 couples from four Dutch fertility centers. This model had a satisfactory agreement between the predicted and observed rates of ongoing pregnancy (calibration) in all subgroups of couples with pregnancy probabilities; however, it had moderate discrimination (AUC = 0.59). The Steures model was further evaluated in a prospective validation study of 1,079 couples from seven Dutch hospitals. The calibration was <1.5% in all the four prognostic groups, and the AUC was 0.56.
As the main outcome parameter in these studies was clinical pregnancy, which is not as ideal a measure of reproductive success as live birth, a cross-sectional study in 2008 performed a logistic regression analysis to predict live births. The prognostic factors identified in this model were duration of infertility, number of treatment cycles, and number of dominant follicles before administration of human chorionic gonadotropin.
Considering the important inter-cycle correlations between repeat cycles from the same individual, generalized estimating equations were used to create a logistic regression model for predicting pregnancy after ovulation induction. The AUC values were 0.6427, 0.7389, and 0.6942 for clinical pregnancy, multiple pregnancy, and spontaneous abortion, respectively, when incorporating all significant or nearly significant variables.
The predictive models can be used to distinguish between couples who are likely to benefit from IUI and those who are not. However, there is so far no standard approach for modeling IUI outcomes.
Prediction of pregnancy after in vitro fertilization and intracytoplasmic sperm injection
IVF and intracytoplasmic sperm injection (ICSI) are two essential techniques for assisted reproduction. Satisfactory precision in the prediction of IVF/ICSI implantation or pregnancy success is important for infertile couples. In the historical development of predictive models for use in clinical practice, some fundamental challenges have been encountered. To date, most of the models are limited by the retrospective nature of their design, small sample size, nonstandardized outcome measures, and lack of external validation, all of which have led to extensive revisions and refinements. The Templeton model is a noteworthy example of a model that offers successful prediction of assisted reproduction outcome. It employs logistic regression analysis to predict the live birth rate per cycle based on the clinical characteristics of 36,961 cycles in the UK between 1991 and 1994. The following five variables had predictive values: female age, pregnancy history, infertility duration, infertility cause, and previous IVF attempts. The model was well calibrated, as confirmed by the Hosmer–Lemeshow test (P = 0.73). External validation was performed for 1,253 couples who underwent 2,756 IVF cycles in the Netherlands from 1991 to 1999. Arvis et al. modified the Templeton model by adding three more predictors (smoking habit, body mass index, and follicle-stimulating hormone), resulting in higher discrimination (c-statistic = 0.75).
Nelson et al. analyzed 144,018 cycles and developed a novel model to calculate the likelihood of live births per IVF attempt (http://www.IVFpredict.com), which included the same predictors as the established Templeton model and the following additional predictors: number of treatment cycles, use of donor oocytes, ICSI, and type of hormonal preparation. The new model showed improved calibration and discrimination compared to the Templeton model.
In 2011, the association between the number of oocytes and live births across all female age groups was studied by a logistic regression model based on 400,135 IVF cycles. The goodness-of-fit for this model was assessed using the Hosmer–Lemeshow test (P = 0.86), and the discriminative capacity of the model was modest (c-statistic = 0.65) in the derivation cohort. In the validation cohort, the Hosmer–Lemeshow test showed a significant P value (P = 0.04) and the c-statistic was 0.66.
Most studies on patient cycle-specific characteristics and their association with implantation efficiency are based on treatments with multiple-embryo transfers in a single cycle, and the traceability of the exact fate of each transferred embryo is difficult. Vaegter et al. applied the prospective data of 8,451 IVF/ICSI single-embryo transfers to generalized estimating equation regression models to estimate the odds of live birth. Seven out of 100 collected variables were defined as significant independent factors and are as follows: female age, infertility cause, treatment history, Ovarian Sensitivity Index, endometrial thickness, embryo score, and female height. The predictive ability of the model was checked using the Hosmer–Lemeshow test (P = 0.90) and the c-statistic (c-statistic = 0.67) in the training set. Comparable results were obtained in an internal test set.
Numerous machine learning techniques have been proposed for embryo-based implantation predictions that are expected to be applied in clinical practice. Their main strength is their ability to identify highly nonlinear associations for evaluating copious amounts of data compared to other statistical methods. Nelson et al. used the boosted tree method to predict live birth probabilities based on clinical variables and anti-Müllerian hormone from 2,124 IVF cycles between 2006 and 2010. Its performance was temporarily assessed against an independent validation set comprising 1,121 cycles from 2011 to 2012. Uyar et al. showed that the Naïve Bayes classifier was better at predicting in IVF/ICSI cases than the k-nearest neighbor, SVM, decision tree, multilayer perceptron, and radial basis function network based on 2,453 embryos transferred after ICSI. After analysis of 486 patients, Hafiz et al. discovered that the random forest classifiers and recursive partitioning with AUC values of 0.84 and 0.82, respectively, provided superior performance over 1-nearest neighbor, SVM, and adaptive boosting.
Neural networks offer powerful collective computation techniques for pattern recognition. In 1997, Kaufmann et al. constructed a neural network using female age, number of eggs recovered, number of embryos transferred, and frozen embryo transfer as variables and achieved an overall accuracy of 59%. Wald et al. found that a 4-hidden node neural network outperformed the other comparable methods (logistic regression and discriminant function analyses) with a test set AUC of 0.783. An ANN-based information technology platform was capable of recognizing a wider range of associations in an automated and efficient manner to improve assisted reproduction outcome and was suggested in routine practice for the treatment of infertility. The continuous creation of increasingly better tools for modeling IVF/ICSI outcomes with numerical estimates is essential for the promotion of individualized health care in assisted reproductive technology (ART) centers.
Embryo selection has routinely relied on developmental and morphological criteria used in various morphology scoring and classification systems. Computer-assisted technology offers a promising approach to overcome large intra- and inter-observer variations and provides standardized, objective, and accurate prediction of embryo implantation potential.
In 2004, a new method for embryo assessment leading to births was supplemented with image processing and a Total Recognition by Adaptive Classification Experiments (TRACE) pattern recognition algorithm. The automatic machine program extracted a set of feature variables from the training images to form rules for classifying embryos (birth or no birth) and used the test set for internal validation. The recognition of the TRACE algorithm was further compared with that of experts in a blind dataset comprising 103 embryos transferred to 35 patients and demonstrated the superiority of the algorithmic judgment over human experts.
As embryo morphology was not the only factor influencing implantation, Morales et al. brought together the clinical characteristics of patients and a reduced subset of embryo morphological attributes to develop an intelligent decision support system on the basis of supervised classification using Bayesian classifiers. Novel prediction models were developed using an ensemble of neural networks with local binary patterns descriptors and multivariate adaptive regression splines, allowing a selection of more promising embryo batches to be transferred to mothers-to-be.
Time-lapse parameters contain morphokinetic information useful for the analysis of embryo quality and implantation success. In 2011, Meseguer et al. proposed a multivariable hierarchical model based on 247 transferred embryos obtained after ICSI, which contained three independently significant predictors: time to 5-cell embryo (t5), time of the second cell cycle (cc2 = t3–t2), and time between division to four cells and subsequent division to three cells (s2 = t4–t3). The Meseguer model was acknowledged to have a lower accuracy in an unselected population than that in the original publication.
Milewski et al. constructed a logistic regression model for predicting implantation potential with the following parameters: t2, t3, t4, t5, cc2, s2, levels of fragmentation, and female age. Principal component analysis, an advanced data-mining method, was applied in the model-building process with the assumption that morphokinetic parameters from the same embryo were not independent events. The quality of the created model was evaluated by ROC analysis on an internal validation data set (AUC = 0.70). Later, an interesting predictive model merging principal component analysis and ANN was presented. The AUC values obtained by this approach were 0.75 and 0.71 in the derivation and validation sets, respectively.
Embryo morphokinetics provide new markers linked to good-quality embryos cultured in the IVF laboratory. The application of data-mining methods, from bench to bedside, enables the creation of more accurate and useful models to reduce the subjectivity of human decisions in assisted reproduction. IVF centers can use inspirations from the established models to build their own center-specific models to cater to their specific needs.
| Future Directions|| |
Relevant literature has been, and will continue to be, filled with simple or sophisticated models predicting reproductive outcomes for the future of personalized medicine. Skills and expertise in machine learning are still rare in the medical field. There are some challenges that are yet to be addressed to ensure that machine learning algorithms are accepted by clinicians and adopted in clinical practice.
Machine learning can be utilized to solve problems presented by big data. Models derived from clinical variables, genomic data, embryo images, morphokinetic parameters, and real-world data have the potential to overcome the difficulties presented by any of the data types. Given the high-dimensional and limited sample-size data, several feature reduction and cross-validation methods have been explored in an attempt to find an optimal prediction model that does not underfit or overfit the data.
Machine learning technology does not merely approximate what a trained physician can do with high accuracy but also takes an unbiased approach to identify unexpected informative variables that are not readily obvious to clinicians. Novel informative features are essential clues in the construction of an improved model. There is little-to-no value in simply using established predictors in a more innovative algorithm packaging. The discovery of robust descriptors in the machine learning process allows for superior prediction.
Predictive modeling is a science as well as an art. The methods of handling and presenting data to arrive at logical inferences and valid conclusions can be considered art in the sense that the models require careful and expert explanation. Methods such as neural networks produce improved predictive power but cannot produce transparent and interpretable rules compared to methods such as logistic regression analysis and decision trees. The hard-to-explain estimates make it difficult for frontline clinicians to understand a model, follow it, and communicate its results.
Our health system is opposed to completely trusting a model and acting on its results. The validation of prediction models using gold-standard databases, such as randomized controlled trials and prospective cohort studies, is highly important for their establishment as decision support tools adopted into the clinical practice in the aspect of evidence-based medicine. Predictive modeling in reproductive medicine is a discipline where human and machine complement each other. There is a clear need for an open and transparent platform for code sharing and feedback collection.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
| References|| |
Deo RC. Machine learning in medicine. Circulation 2015;132:1920-30. doi: 10.1161/CIRCULATIONAHA.115.001593.
Gómez D, Rojas A. An empirical overview of the no free lunch theorem and its effect on real-world machine learning classification. Neural Comput 2016;28:216-28. doi: 10.1162/NECO_a_00793.
Zhang Z. Introduction to machine learning: K-nearest neighbors. Ann Transl Med 2016;4:218. doi: 10.21037/atm.2016.03.37.
Hammann F, Drewe J. Decision tree models for data mining in hit discovery. Expert Opin Drug Discov 2012;7:341-52. doi: 10.1517/17460441.2012.668182.
Heikamp K, Bajorath J. Support vector machines for drug discovery. Expert Opin Drug Discov 2014;9:93-104. doi: 10.1517/17460441.2014.866943.
Siristatidis C, Pouliakis A, Chrelias C, Kassanos D. Artificial intelligence in IVF: A need. Syst Biol Reprod Med 2011;57:179-85. doi: 10.3109/19396368.2011.558607.
van Loendersloot L, Repping S, Bossuyt PM, van der Veen F, van Wely M. Prediction models in in vitro
fertilization; where are we? A mini review. J Adv Res 2014;5:295-301. doi: 10.1016/j.jare.2013.05.002.
Eimers JM, te Velde ER, Gerritse R, Vogelzang ET, Looman CW, Habbema JD. The prediction of the chance to conceive in subfertile couples. Fertil Steril 1994;61:44-52. doi: 10.1016/S0015-0282(16)56451-6.
Hunault CC, Eijkemans MJ, te Velde ER, Collins JA, Habbema JD. Validation of a model predicting spontaneous pregnancy among subfertile untreated couples. Fertil Steril 2002;78:500-6. doi: 10.1016/S0015-0282(02)03244-2.
Collins JA, Burrows EA, Wilan AR. The prognosis for live birth among untreated infertile couples. Fertil Steril 1995;64:22-8. doi: 10.1016/S0015-0282(16)57650-X.
Snick HK, Snick TS, Evers JL, Collins JA. The spontaneous pregnancy prognosis in untreated subfertile couples: The Walcheren primary care study. Hum Reprod 1997;12:1582-8. doi: 10.1093/humrep/12.7.1582.
Hunault CC, Habbema JD, Eijkemans MJ, Collins JA, Evers JL, te Velde ER. Two new prediction rules for spontaneous pregnancy leading to live birth among subfertile couples, based on the synthesis of three previous models. Hum Reprod 2004;19:2019-26. doi: 10.1093/humrep/deh365.
Hunault CC, Laven JS, van Rooij IA, Eijkemans MJ, te Velde ER, Habbema JD. Prospective validation of two models predicting pregnancy leading to live birth among untreated subfertile couples. Hum Reprod 2005;20:1636-41. doi: 10.1093/humrep/deh821.
van der Steeg JW, Steures P, Eijkemans MJ, Habbema JD, Hompes PG, Broekmans FJ, et al.
Pregnancy is predictable: A large-scale prospective external validation of the prediction of spontaneous pregnancy in subfertile couples. Hum Reprod 2007;22:536-42. doi: 10.1093/humrep/del378.
Tomlinson MJ, Amissah-Arthur JB, Thompson KA, Kasraie JL, Bentick B. Prognostic indicators for intrauterine insemination (IUI): Statistical model for IUI success. Hum Reprod 1996;11:1892-6. doi: 10.1093/oxfordjournals.humrep.a019513.
Steures P, van der Steeg JW, Mol BW, Eijkemans MJ, van der Veen F, Habbema JD, et al.
Prediction of an ongoing pregnancy after intrauterine insemination. Fertil Steril 2004;82:45-51. doi: 10.1016/j.fertnstert.2003.12.028.
Custers IM, Steures P, van der Steeg JW, van Dessel TJ, Bernardus RE, Bourdrez P, et al.
External validation of a prediction model for an ongoing pregnancy after intrauterine insemination. Fertil Steril 2007;88:425-31. doi: 10.1016/j.fertnstert.2006.12.007.
Erdem A, Erdem M, Atmaca S, Korucuoglu U, Karabacak O. Factors affecting live birth rate in intrauterine insemination cycles with recombinant gonadotrophin stimulation. Reprod Biomed Online 2008;17:199-206. doi: 10.1016/S1472-6483(10)60195-2.
Goldman RH, Batsis M, Petrozza JC, Souter I. Patient-specific predictions of outcome after gonadotropin ovulation induction/intrauterine insemination. Fertil Steril 2014;101:1649-55.e1-2. doi: 10.1016/j.fertnstert.2014.02.028.
Templeton A, Morris JK, Parslow W. Factors that affect outcome of in vitro
fertilisation treatment. Lancet 1996;348:1402-6. doi: 10.1016/S0140-6736(96)05291-9.
Smeenk JM, Stolwijk AM, Kremer JA, Braat DD. External validation of the Templeton model for predicting success after IVF. Hum Reprod 2000;15:1065-8. doi: 10.1093/humrep/15.5.1065.
Arvis P, Lehert P, Guivarc'h-Levêque A. Simple adaptations to the Templeton model for IVF outcome prediction make it current and clinically useful. Hum Reprod 2012;27:2971-8. doi: 10.1093/humrep/des283.
Nelson SM, Lawlor DA. Predicting live birth, preterm delivery, and low birth weight in infants born from in vitro
fertilisation: A prospective study of 144,018 treatment cycles. PLoS Med 2011;8:e1000386. doi: 10.1371/journal.pmed.1000386.
Sunkara SK, Rittenberg V, Raine-Fenning N, Bhattacharya S, Zamora J, Coomarasamy A. Association between the number of eggs and live birth in IVF treatment: An analysis of 400 135 treatment cycles. Hum Reprod 2011;26:1768-74. doi: 10.1093/humrep/der106.
Vaegter KK, Lakic TG, Olovsson M, Berglund L, Brodin T, Holte J. Which factors are most predictive for live birth after in vitro
fertilization and intracytoplasmic sperm injection (IVF/ICSI) treatments? Analysis of 100 prospectively recorded variables in 8,400 IVF/ICSI single-embryo transfers. Fertil Steril 2017;107:641-8.e2. doi: 10.1016/j.fertnstert.2016.12.005.
Nelson SM, Fleming R, Gaudoin M, Choi B, Santo-Domingo K, Yao M. Antimüllerian hormone levels and antral follicle count as prognostic indicators in a personalized prediction model of live birth. Fertil Steril 2015;104:325-32. doi: 10.1016/j.fertnstert.2015.04.032.
Uyar A, Bener A, Ciray HN. Predictive modeling of implantation outcome in an in vitro
fertilization setting: An application of machine learning methods. Med Decis Making 2015;35:714-25. doi: 10.1177/0272989X14535984.
Hafiz P, Nematollahi M, Boostani R, Namavar Jahromi B. Predicting implantation outcome of in vitro
fertilization and intracytoplasmic sperm injection using data mining techniques. Int J Fertil Steril 2017;11:184-90. doi: 10.22074/ijfs.2017.4882.
Kaufmann SJ, Eastaugh JL, Snowden S, Smye SW, Sharma V. The application of neural networks in predicting the outcome of in vitro
fertilization. Hum Reprod 1997;12:1454-7. doi: 10.1093/humrep/12.7.1454.
Wald M, Sparks A, Sandlow J, Van-Voorhis B, Syrop CH, Niederberger CS. Computational models for prediction of IVF/ICSI outcomes with surgically retrieved spermatozoa. Reprod Biomed Online 2005;11:325-31. doi: 10.1016/S1472-6483(10)60840-1.
Siristatidis C, Vogiatzi P, Pouliakis A, Trivella M, Papantoniou N, Bettocchi S, et al
. Predicting IVF outcome: A proposed web-based system using artificial intelligence.In Vivo
Patrizi G, Manna C, Moscatelli C, Nieddu L. Pattern recognition methods in human-assisted reproduction. Int Trans Oper Res 2004;11:365-79. doi: 10.1111/j.1475-3995.2004.00464.x.
Manna C, Patrizi G, Rahman A, Sallam H. Experimental results on the recognition of embryos in human assisted reproduction. Reprod Biomed Online 2004;8:460-9. doi: 10.1016/S1472-6483(10)60931-5.
Morales DA, Bengoetxea E, Larrañaga P. Selection of human embryos for transfer by Bayesian classifiers. Comput Biol Med 2008;38:1177-86. doi: 10.1016/j.compbiomed.2008.09.002.
Manna C, Nanni L, Lumini A, Pappalardo S. Artificial intelligence techniques for embryo and oocyte classification. Reprod Biomed Online 2013;26:42-9. doi: 10.1016/j.rbmo.2012.09.015.
Chen F, De Neubourg D, Debrock S, Peeraer K, D'Hooghe T, Spiessens C. Selecting the embryo with the highest implantation potential using a data mining based prediction model. Reprod Biol Endocrinol 2016;14:10. doi: 10.1186/s12958-016-0145-1.
Meseguer M, Herrero J, Tejera A, Hilligsøe KM, Ramsing NB, Remohí J. The use of morphokinetics as a predictor of embryo implantation. Hum Reprod 2011;26:2658-71. doi: 10.1093/humrep/der256.
Fréour T, Le Fleuter N, Lammers J, Splingart C, Reignier A, Barrière P. External validation of a time-lapse prediction model. Fertil Steril 2015;103:917-22. doi: 10.1016/j.fertnstert.2014.12.111.
Milewski R, Milewska AJ, Kuczyńska A, Stankiewicz B, Kuczyński W. Do morphokinetic data sets inform pregnancy potential? J Assist Reprod Genet 2016;33:357-65. doi: 10.1007/s10815-016-0649-9.
Milewski R, Kuczyńska A, Stankiewicz B, Kuczyński W. How much information about embryo implantation potential is included in morphokinetic data? A prediction model based on artificial neural networks and principal component analysis. Adv Med Sci 2017;62:202-6. doi: 10.1016/j.advms.2017.02.001.
Patel MJ, Khalaf A, Aizenstein HJ. Studying depression using imaging and machine learning methods. Neuroimage Clin 2016;10:115-23. doi: 10.1016/j.nicl.2015.11.003.
Bologna G, Hayashi Y. Characterization of symbolic rules embedded in deep DIMLP networks: A challenge to transparency of deep learning. J Artif Intell Soft 2017;7:265-86. doi: 10.1515/jaiscr-2017-0019.
Leushuis E, van der Steeg JW, Steures P, Bossuyt PM, Eijkemans MJ, van der Veen F, et al.
Prediction models in reproductive medicine: A critical appraisal. Hum Reprod Update 2009;15:537-52. doi: 10.1093/humupd/dmp013.