Artificial Intelligence in IVF: A Need

Abstract

Predicting the outcome of in-vitro fertilization (IVF) treatment is an extremely semantic issue in reproductive medicine. Discrepancies in results among reproductive centres still exist making the construction of new systems capable to foresee the desired outcome a necessity. As such, artificial neural networks (ANNs) represent a combination of a learning, self-adapting, and predicting machine. In this review hypothesis paper we summarize the past efforts of the ANNs systems to predict IVF outcomes. This will be considered together with other statistical models, such as the ensemble techniques, Classification And Regression Tree (CART) and regression analysis techniques, discriminant analysis, and case based reasoning systems. We also summarize the various inputs that have been employed as parameters in these studies to predict the IVF outcome. Finally, we report our attempt to construct a new ANN architecture based on the Learning Vector Quantizer promising good generalization: a system filled by a complete data set of our IVF unit, formulated parameters most commonly used in similar studies, trained by a network expert, and evaluated in terms of predictive power.

Keywords:

• artificial intelligence
• classification models
• in vitro fertilization
• IVF
• neural networks

Introduction

Predicting the outcome of in vitro fertilization (IVF) treatment is an extremely semantic issue in reproductive medicine. As discrepancies in results among reproductive centres still exist and the literature continues to be filled with new methods aiming to foresee the desired outcome, there is a clear need for constructing systems to assist the human mind. Numerous methods have been proposed since 1986 trying to perform this prediction [Bancsi et al. 2004; Hunault et al. 2002; Jurisica et al. 1998; van Weert et al. 2008]. Most of them were based on well-known statistical models and a decade later newer technologies, such as artificial neural networks (ANNs) [Kaufmann et al. 1997; Milewski et al. 2009; Uyar et al. 2009; Wald et al. 2005; Yi et al. 1998]. The literature relating ANNs' capabilities and IVF prediction is rather limited.

Clinically an ANN represents a combination of a learning, self-adapting, and predicting machine. The initial objective of using an artificial network in medicine is to learn a relationship which is represented by a set of historical data. The final aim is to discover the diagnosis associated with a particular set of symptoms and/or additional data or, equally importantly, to pre- and fore- see treatment outcomes. There are certain problematic situations like an exhausting edge that highlight the limitations of the human mind capabilities. Such situations are mainly related to prediction issues and one of the usual examples is IVF outcomes. Although ANNs are hugely popular in current research on medical decision support systems [Ecke et al. 2010; Karakitsos et al. 2002; 2005; Lin et al. 2009; Markopoulos et al. 1997; Naik et al. 2008; Pantazopoulos et al. 1998], their use in IVF as prediction models is restricted to only a few studies [Kaufmann et al. 1997; Milewski et al. 2009; Uyar et al. 2009; Wald et al. 2005; Yi et al. 1998].

For the latter, this has been attributed to the experience required to train an ANN, the existing danger of overtraining the system and the trapping to local minima, together with the instability of some neural network models as predictive tools. Thus, small changes in the training data set may produce very different models and – consequently- performance when applied to new data yielding totally different results. This instability leads the generalization performance of some ANN architectures for a particular task to vary considerably, being dependent on pre-chosen data used, which carry the extra bias of their retrospective nature. In an effort to reduce the pitfalls of a single system, ANN ensemble techniques are beginning to be adopted. One of the most common include elaboration of bagging Breinman [1996], Freund and Schapire boosting [1999], or Wolpert stacking [1996]. All have produced dramatic improvements in generalization performance [Cunningham et al. 2000].

In this paper, we first summarize past efforts made of the available statistical and ANN systems to predict IVF outcomes. Second, based on the results of these studies, we describe our current effort to construct a new ANN architecture promising good generalization using a system filled by a complete data set of our IVF unit, formulated parameters most commonly used in similar studies, trained by a network expert with predictive power evaluated.

Existing Techniques

Introduction to Artificial Neural Networks

ANNs are mathematical systems which approach the functionality of small neural clusters in a very fundamental manner. Their anatomical and functional features, as well as their applications in medicine have been described previously [Cochand-Priollet et al. 2006; Cunningham et al. 2000; Karakitsos et al. 2002; 2005; Pantazopoulos et al. 1998; Papik et al. 1998].

McCulloch and Pitts [1943] were the first to show that an artificial network similar to a biological network could be constructed, by using interconnected neurons, mathematics, and algorithms. In 1949, Hebb introduced the concept of Hebbian or associative learning, while in 1962, Rosenblatt showed the way of training a network by changing neurons synaptic strengths, each time the network yielded a wrong answer. Despite the inefficiency of the system, pointed by Minsky and Papert [1969], Werbos [1990] introduced the idea to propagate the errors back to the starting layers of the system [Haykin 1994]. In addition, Hopfield [1982] showed that an asynchronous network can develop a behaviour similar to that of the human brain and, using ‘least energy’, can develop solutions to problems requiring minimization of the network output [Sima and Orponen 2001]. Figure 1 shows the classic construction of an ANN that includes a combination of interconnected neurons. Each sums the weighted inputs and passes the result from a non linear function to produce the output. Several neurons are connected to each other, such that the output of a single neuron constitutes the input for another or others. A typical structure of the popular feed forward neural network (NN) architecture is presented in Figure 2.

Figure 1.  The neuron model. For the artificial neuron, the information flows from left to right as follows: data input, multiplication of each input with a weight, summation of the weighted inputs, passing from non-linear function, and neuron output.

Figure 2.  Multilayer neural networks. The multilayer neural network is based on layers composed from neurons: the input layer accepts the inputs, the output layer produces the neural network response. Between the input and the output layer there are one or more hidden layers. In most neural network models the flow of information moves from one layer to the next.

One of the most popular models of ANNs, was proposed by Kohonen [1988]: the Learning Vector Quantization (LVQ) which intelligently partitions the feature space and the self organizing map (the so-called SOM model). The latter creates either one dimensional or multidimensional maps with neighbour clusters according to the similarities of the entered data. A very important feature of this ANN is that it does not require knowledge of the categories where the data belong during the training phase. A typical cycle for the creation of a useful ANN system involves the following steps: data collection, data pre-processing, ANN model selection, ANN parameter selection, ANN training, evaluation, data re-processing, further evaluation, and application in a production environment.

In terms of function and value, the construction of an ANN leads to the development of algorithms in order to provide a solution to a certain question/problem. It differs compared to conventional computational systems by its learning aptitude through training. It self-adapts to the exercise and cyclically changes its structural characteristics based on external or internal information flowing through the network during the learning phase. The result is the initial production of its own solving algorithm. The role of the latter is to minimize the total error of the system. In addition, the ANN can predict cases that have never been presented to the system before. Its flexibility makes it suitable to solve classification problems based on qualitative and quantitative differences [Papik et al. 1998].

ANNs' disadvantages include the empirical nature of the model development, the instability of some as predictors, the long experience required for tuning others, and the lack of understanding interactions in their hidden layers. Additionally, the decision borders set by clinicians leading to an automatic decision by the network, the limited abilities to identify possible causal relationships, the way, the time, and the type of parameters to be trained, and the pitfalls during training requires consideration. Of note, not all ANN models present instability issues, so that proper selection is mandatory before targeting the solution of a specific problem.

Introduction to the Learning Vector Quantizer

The Learning Vector Quantizer (LVQ) is a supervised neural network. Currently there are available four variations of the LVQ algorithm: LVQ1, LVQ2.1, LVQ3, and the Optimized LVQ1 “OLVQ1” [Kohonen et al. 1992; Kohonen et al. 1995]. In contrast to unsupervised classifiers, such as the self organizing map (SOM) [Kohonen 1988], where class is not required measurements of each case are required to be available together with the class that they belong (in our study the class is ‘IVF outcome’ and the measurements are the quantified expressions of numerous IVF related indices). During training, the LVQ classifier creates partitions of the feature space; each partition is characterized by a vector in its center, called codebook vector, while the class that the vector belongs characterizes the class of the assigned partition.

In order to train and evaluate the classifier, the available data are divided into 2 sets, called training and test sets, respectively. During the training phase the classifier learns a specific data set and the codebook vectors are moved in the feature space according to the learning algorithm. At first, the codebook vectors are selected from the available data, but as the training algorithm is applied, several passes of all the training vectors are performed. Usually 50 to 200 passes of all data are required for the training of the classifier. However, the exact number of passes and the required codebook vectors, together with the applied learning algorithm [Kohonen et al. 1992; 1995] are determined during training, according to classifier performance on the training set.

During the test phase the performance of the trained NN is evaluated; each unknown case (represented by a vector of IVF related measurements) from the test set is presented to the network. The class of these unknown cases is determined to be similar to the class of the partition that the vector resides. This is achieved by identifying the codebook vector which is nearest to the unknown vector and by assigning the unknown case to a class equal to the class of the nearest codebook vector.

Applications of ANNs to IVF prediction

Kauffmann et al. [1997] applied a neural network which had 4 inputs, 1 hidden layer with 4 nodes and 1 output. These authors used the variables of age, number of eggs recovered and fertilized, number of embryos transferred, and whether there was embryo freezing or not. These features have been identified as significant prognostic factors of success. The range of the network performance over the 8 data randomizations was correct in 54.9 – 79.3% of the pregnant and in 40.7 – 56.5% of the non-pregnant population. Overall the neural network managed to achieve an accuracy of 59%, being more successful in predicting success than failure. The quite low percentage probably suggested the insufficiency of the input information of characterizing the outcome, further indicating the absence of important predictor variables from the initial data set. According to the authors, these results highlighted the difficulties of forecasting biological events in these circumstances, indicating that there are other factors such as the inherent developmental potential of the oocytes, influence of treatment regiments, differences in medical and laboratory techniques, and intercycle variations which may have an important but as yet unquantifiable influence on the outcome.

Wald et al. [2005] have compared computational models for the prediction of IVF/ICSI (intra-cytoplasmic sperm injection) outcomes with surgically retrieved spermatozoa. Their dataset was comprised of 113 exemplars, derived from patients who underwent IVF/ICSI. As input features they used maternal age, sperm retrieval technique, type of spermatozoa, type of male factor, and output intrauterine pregnancy, being modeled by using linear and quadratic discriminant function analysis, logistic regression, and neural computation. A 4-hidden node neural network was found to have the highest accuracy, with a test set receiver operator characteristic curve area of 0.783. In addition, reverse regression of the network showed maternal age to be the most significant feature in predicting pregnancy, followed by sperm type.

Uyar et al. [2009] have employed the Support Vector Machine method for the prediction of IVF outcomes. An original IVF dataset for classification of embryos according to their implantation potential, including both categorical and continuous feature values, was used. They emphasized the importance of their methodology through the transformation of their categorical variables into numeric values with a pre-processing stage significantly affecting the performance of the classification. The proposed technique significantly improved the performance of IVF implantation prediction in terms of area under ROC curve, as compared to the common binary encoding and expert judgment-based transformation methods (0.712 +/- 0.032 vs. 0.676 +/- 0.033 and 0.696 +/- 0.024, respectively).

In comparison, Morales et al. [2008] proposed a decision support system based on supervised classification by Bayesian classifiers for the selection of the most promising embryos to transfer. The application of these classifiers permitted more accurate embryo selection as compared to conventional procedures that fully relied on the expertise and experience of the embryologists. Their input variables consisted of embryo morphology characteristics and patients' clinical data.

Ensemble techniques

Ensemble techniques were invented to boost the generalization capabilities and improve the stability of ANNs and their accuracy to rely on the diversity of its constitutive components. The concept is, combining the outputs of several individual predictors can logically improve the performance of a single generic output. The dynamics of such systems have been extensively reported [Carney and Cunningham 1999; Cunningham and Walsh 2002; Jiang et al. 2002; Lovell et al. 1997; Opitz and Shavlik 1996; Sharkey 1999; Shu and Burn 2004]. Cunningham et al. [2000] showed how to improve accuracy of a prediction model for use in IVF. They used bagged ensemble solutions by aggregating the output of several predictors to produce a single prediction. They managed to minimize the disagreement amongst the ensembles and to improve the overall generalization performance.

CART and regression analysis techniques

Other statistical tools generate simple graphical representations of predictive models, aiming to predict a specific outcome for an individual patient [Eastham et al. 2002]. Statistical techniques tend to be more widely used for IVF predication as compared to neural.

Dessolle et al. [2010] developed a regression analysis model to predict blastocyst transfer cancellation. The model was built from 562 consecutive first IVF cycles. It used multivariable logistic regression (MLR) analysis to test the association between the presence of cultured blastocysts on day 5 (primary outcome) and patient and cycle characteristics. The parameters introduced into the model were: the female's age and body mass index (BMI), the male's age, type of fertilization (ICSI or IVF), the total number of embryos cultured, the number of embryos cultured on day 3, and the ratio between the number of top embryos on day 3 and the total number of cultured embryos on day 3 (quality ratio); follicle-stimulating hormone (FSH) levels and type of ovarian hyperstimulation were used as well. When performance was assessed on a test set the model showed sensitivity, specificity, and positive and negative predictive values of 76.5%, 70.7%, 65.6%, and 80.5%, respectively.

Frattarrelli and Gerber [2006] through a retrospective study used a sample of 117 women undergoing IVF to build their model. By performing multiple linear regression analysis they showed that serum androgen levels do not influence IVF pregnancy outcomes, despite their correlation with IVF stimulation parameters during the woman's menstrual cycle.

Using the same statistical method, Kwee et al. [2007] evaluated various input features in order to predict the outcome in a sample of 110 women undergoing IVF. Authors found that the antral follicle count (AFC), the clomiphene citrate challenge test, and the basal FSH were the best predictors for poor response, while the basal ovarian volume, the exogenous FSH ovarian reserve test, the inhibin B, and the AFC were good predictors for the ovarian reserve.

In a retrospective study, Repping et al [2002] employed a large database composed of 892 couples with 1,569 IVF cycles. They evaluated the data by stepwise logistic regression analysis and constructed two models, one before the IVF cycle and one at the time of oocyte recovery. The male's age, the number of IVF cycles, tubal, ovulation and cervical factors together with the number of oocytes recovered have been identified as important predictors. In addition, the introduction of the prewash total motile sperm count (TMC) parameter showed to increase the predictive power of both models (ROC analysis area under curve = 0.75 and area under curve = 0.80); adversely, the addition of the postwash TMC parameter did not manage to increase the systems' performance.

Srouji et al. [2005] evaluated by logistic regression analysis 3,175 IVF cycles; they concluded that both oestradiol (E₂) and FSH levels at day 6 and maternal age constituted reliable predictors for ongoing IVF cycles. Verhagen et al. [2008] performed a metaanalysis based on 11 articles aiming to evaluate the predictive capacity of multivariate models in ovarian reserve testing. Most studies were related to the prediction of the poor ovarian response, while only 1 was related to the occurrence of pregnancy [Creus et al. 2000]. The sensitivity of the models in terms of prediction of poor ovarian response varied between 39 and 97%, while the specificity was between 50 and 96%. Authors concluded that the accuracy of multivariate models was equal to that of the AFC.

Based on 1,675 IVF treatment cycles and a logistic regression model Ottosen et al. [2007] concluded that embryo quality, patient age, and basal FSH were important predictive factors; however, the system performance was not satisfactory, demonstrating similar power to others', requiring the addition of more detailed patient and embryo data. Saith et al. [1998] investigated 53 features on a base of 200 IVF patients using tree classifiers. Their results indicated that among the plethora of features studied, only 4 had a significant prediction value: the embryo grade, the cell number, the follicle size, and the follicular fluid volume. Among the various rule sets that were constructed, a single rule set was capable to achieve a ‘take baby home’ prediction rate of 67% with an overall accuracy equal to 77%. Other factors investigated to affect IVF success rates, serving as input parameters to predictive models were: physician's ability to perform the embryo transfer, patient's stress level, anti-Mullerian hormone (AMH), FSH and luteinizing hormone (LH) at day 3, prolactin, E_2, and progesterone concentrations and the AFC, maternal age, cause of infertility, the number of embryos transferred, and their average morphology score [Angelini et al. 2006].

Other techniques

There are other reports of other statistical techniques, such as discriminant analysis [Gnoth et al. 2008; Muasher et al. 1988] and case based reasoning systems [Jurisica et al. 1998] being applied to this area. For example, Muasher et al. [1988] performed a study on 80 consecutive patients undergoing the same stimulation protocol for IVF. Paired and canonical discriminant analyses were used analyzing FSH and LH on day 3 of the cycle before treatment. Seven statistically significant different groups of patients were identified. The authors concluded that basal serum gonadotropin levels can identify different populations of patients behaving differently in terms of E₂ response, oocytes obtained and transferred, pregnancy rates, and outcomes. Using the same statistical model, Gnoth et al. [2008] evaluated and compared the predictive value of age, FSH, inhibin B and AMH for ovarian response, and clinical pregnancy rates in IVF cycles using a population of 316 patients. This study concluded that AMH was a reliable predictor of ovarian response, but other parameters, such as AFC are required to predict IVF outcome.

Jirisica et al. [1998] developed a case-based reasoning system (TA3_IVF) whose knowledge base was populated with a number of past cases, used to explore and discover relationships among data, thereby achieving a form of knowledge mining. The system was proven capable to assist in knowledge visualization, interactive data exploration and discovery, and to suggest possible modifications to IVF treatment plans, leading to the improvement of the overall success rates.

Hypothesis and Proposed Architecture for an ANN System

We propose to create an ANN model based on the LVQ classifier. In order to obtain a suitable data set, it is planned to retrospectively collect the features and the data of 300 cases, treated with IVF in our unit. This initial number of cases will be reprocessed i.e., clearing of any inconsistent entries, or entries with missing data (incomplete entries). In addition, the number may be increased during either the learning phase of the network or if a non-satisfactory prediction result is obtained. Positive and negative predictive values over a threshold of 75% are considered as satisfactory results.

In order to define the most ‘accurate’ input parameters –we will select at least 10 characteristics, based on the features used in other studies. Table 1 summarizes the primary inputs that have been employed during various efforts to predict the IVF outcome. These characteristics are presented independent of the prediction approach used in each study and the prediction outcomes. The selection criteria include the production of accurate classification results (irrelevant of the classification approach). In order to reduce the number of input parameters, data that will be highly correlated will be removed.

Table 1. Summary of features used for the prediction of IVF result.

Treatment Characteristics	Embryo Characteristics	Female Characteristics	Male Characteristics
Type of controlled ovarian hyperstimulation (i.e., long and short agonist, antagonist protocol)FSH dosageDays of stimulationBasal E2 (pg/mL)Peak E2 level (pg/mL)Basal LH (mIU/mL)Basal FSH (mIU/mL)Basal AFCInhibin BActivinAMHCA-125Serum Vascular endothelial growth factor (VEGF)Challenge tests with clomiphene citrateUltrasonographic markers such as the AFC, total ovarian volume and blood flow3D ultrasonographic endometrial volume assessment on the day of hCG administrationEndometrial thickness 3D ultrasonography and power Doppler angiographyFollicular fluid volume at egg recoveryFollicle size at egg recoveryNumber of eggs inseminatedSelection of euploid eggsNumber of normal eggs fertilized	Mean grade of embryos replacedMorphological scoring of human pronuclear zygotesMorphology score of the two best embryos available for transferCumulative embryo scoreNumber of embryos cleaved (number of embryos that have been divided)Early cleavage morphologyEarly cleavage timeEquality of blastomeresNumber of embryosFragmentation rateRegularity of cells4-cell embryos at day 2Appearance of cytoplasm (darkness)FreezingPrewash total motile sperm countPostwash total motile sperm countNucleus characteristicsVisibility of nuclei in the cellsPrewash total motile sperm count at egg recoveryPostwash total motile sperm count at egg recoveryTotal number of embryos culturedNumber of embryos cultured on day 3The ratio between the number of top embryos on day 3 and the total number of cultured embryos on day 3 (quality ratio)	Female ageInfertility factorBody Mass IndexMean ovarian volume (cm^3)TraumaPrevious live birthsPrevious abortionsDuration of infertilityDiagnosis of infertilityPrevious IVF cyclesSmoking   Transfer DataType of catheterTransfer dayPhysician performing embryo transferEmbryo transfer techniqueTransfer with or without ultrasound guidanceDifficulty of transferType of fertilization (ICSI or IVF)Luteal support regimen	Male ageFamily historyPersonal historySperm characteristicsSperm motilitySperm function testsStrict criteria for sperm morphologyChromatin packaging assessmentSperm DNA integrity assessmentHypoosmotic swelling testHemizona assayZona pellucida –induced acrosome reaction

For example, patient's characteristics, such as female age, BMI, AMH and FSH in the early follicular phase, history and results of previous clomiphene stimulated or IVF cycles, and oocytes and embryo characteristics, such as number of retrieved oocytes, early cleavage morphology, fragmentation rates, number and equality of blastomeres at day 2 and 3 and progression to blastocyst stage will be included. The final outcome parameter of the study will be the identification of two options: successful pregnancy or not.

A separation of the available data into two sets will follow data pre-processing. A training set will be used for ANN formation and a test set will be used to evaluate the performance of the system.

The proposed system architecture will be based on the LVQ NN. We chose this system due to its ability to create clusters formed by the data into the feature space. As above, each cluster will be identified by a single codebook vector belonging to a specific category (in our case the IVF outcome). The codebook vectors will be created during the training phase using the data of the training set. The decision of the LVQ classifier on unknown data will be possible using clusters, the related codebook vectors, and the data during training. It will be justified by the following procedure. Initially, an unknown entry (i.e., an IVF case with unknown outcome) will be presented to the LVQ NN, then the nearest codebook vector will be identified. The unknown case will be expected to have an outcome similar to the outcome of the nearest codebook vector and will be assigned to a class equal to that of the particular vector. Finally, the cases of the training set which are nearest to the vector, presenting the greatest similarity to the unknown one will be used for the justification of the decision. One of the important characteristic of this system and important reason for its selection, is the fast convergence, as compared to other systems, such as the back propagation family algorithms. Equally important, the system resists getting trapped in local minima caused by the data themselves, so that stability problems are not expected. During training, the adjustable characteristics of the LVQ algorithm, such as the initial number of codebook vectors, the learning rate, and the number of data iterations [Kohonen 1988] will be identified to enable the best training set performance.

The next step will be the evaluation of the performance using the test set or the complete set. If our results are satisfactory, the system could be possibly placed in a production environment for routine use. Otherwise, the steps from the beginning up to the evaluation stage will be revised. The stability of the system will be evaluated by splitting the complete data set into a training and a test set for several times.

Declaration of Interest: The authors report no conflicts of interest. The authors alone are responsible for the content and writing of the paper.

Abbreviations

IVF:	=	in vitro fertilization
ANNs:	=	artificial neural networks
CART:	=	Classification And Regression Tree
LVQ:	=	Learning Vector Quantization
SOM:	=	self organizing map
ICSI	=	intra-cytoplasmic sperm injection
MLR:	=	multivariable logistic regression
FSH:	=	follicle-stimulating hormone
AFC:	=	antral follicle count
TMC:	=	total motile sperm count
AMH:	=	anti-Mullerian hormone
LH:	=	luteinizing hormone.