Machine learning in predicting infertility treatment success: A systematic literature review of techniques.

Infertility, an emotionally stressful disability with a considerable social burden, is a growing public health challenge. Infertility is defined as the failure of a couple to conceive after 1 year of regular, unprotected intercourse.[ 1 2 ] Globally, it is estimated to affect 8–12% of couples of reproductive age, which constitutes 186 million people.[ 3 ] Nowadays, more couples seek infertility treatment due to increased public awareness, convenient availability, and improvements in infertility treatments. Assisted reproductive technology (ART) is one of the major developments that has had a great impact on infertility treatment. It encompasses all techniques involving direct manipulation of oocytes outside of the body. The most common forms of ART are in vitro fertilization (IVF) and intracytoplasmic sperm injection (ICSI).[ 4 ] IVF comprises a sequence of steps beginning with ovarian stimulation and oocyte retrieval, fertilization, and embryo culture in the laboratory and embryo transfer to the uterus for implantation. ICSI is very similar to IVF, with the difference that in this method during artificial insemination, sperm is injected into the cytoplasm of the oocyte.[ 2 ] So far, more than 8 million IVF children have been delivered alive worldwide.[ 5 ] Although ART turns out to be a breakthrough in infertility treatment, it does not guarantee success. Compared to the pregnancy rate of 20% in healthy couples, the overall success rate has just been slightly higher and reaches 30–40%.[ 2 6 ] Also, there are complications associated with this treatment that could threaten both mother and child’s health and life.[ 6 7 8 9 ] Besides the low rate of success and risk of adverse effects, ART is an expensive treatment procedure. Infertile couples undergo financial burdens because of the high costs of it.[ 10 ] This, in addition to the complexity of decision-making in medicine,[ 11 12 ] makes it crucial to uptake ART as a treatment for infertile couples. Infertility clinicians use patients’ specific characteristics, together with their knowledge and clinical experience, to recommend this treatment when it is likely to succeed. Yet, due to the number of variables and complex relations among them, the prediction of infertility treatment success turns out to be a challenging task. The complexity could bias the clinician’s judgment and affect clinical outcomes.[ 13 14 ] Artificial intelligence has increasingly progressed over the past decade, and its applications have extended to various fields, including medicine. Machine learning (ML), a subset of artificial intelligence, enables systems to detect complex patterns in biomedical data and use them in line with clinical activities.[ 15 16 ] Systems based on ML techniques could train a classifier model with a large amount of medical data, which make it suitable for medical prediction.[ 16 ] In reproductive medicine, the ML techniques could help to improve the prediction of infertility treatment success resulting in making informed decisions for infertile couples. A predictive model of ART success based on ML techniques could provide a robust basis for the estimation of pregnancy chance by identifying interlinks among variables, making a realistic expectation for both clinicians and couples. Predictive models can vary depending on the number and combination of features, size, and variety of the dataset and type of ML techniques. Accordingly, they could have different performances in terms of accuracy, area under receiver operating characteristic (AUC), sensitivity, specificity, etc., This variation highlights the importance of examining these elements in the context of infertility treatment. By thoroughly examining the previous studies, this review aims to contribute to a better understanding of the predictive modeling landscape in infertility treatment. Such insights increase our understanding of the field and guide future researchers and practitioners in making informed decisions about choosing algorithms, feature selection, and dataset characteristics to develop more robust and clinically relevant prediction models. To the best of our knowledge, no systematic review has yet addressed this topic. This study aimed to identify predictive models of ART success and their determinants.

As illustrated in Figure 2 , searching databases resulted in 3655 records in total (PubMed: 802, Embase: 928, Scopus: 1281, and Web of Science: 644). A total of 1585 duplicates were removed using EndNote. The title and abstract of the remaining 2070 articles were screened, of which 106 eligible articles were identified for full-text assessment. In the next step, the authors reviewed the full text of the articles and excluded 79 articles, according to the inclusion or exclusion criteria. Studies were excluded if they did not apply a predictive model for IVF/ICSI outcome (27 records), did not use ML techniques in the predictive modeling (9 records), or did not include important study details (5 records). In addition, articles were excluded if there were conference papers, review articles, book chapters (26 records), and non-English papers (1 record) and if the full text was not available (11 records). The final set included 27 articles published from 2009 to 2023 [ Table 1 ]. Study selection and data extraction Details of studies identified RF: random forest, SVM: support vector machine, NB: naïve Bayes, MLP: multi-layer perceptron, LR: logistic regression, LDA: liner discriminant analysis, KNN: k-nearest neighbor, ANN: artificial neural network, DT: decision tree, MLR: multivariate logistic regression, RPART: recursive partitioning, RBF: radial basis network, BN: Bayesian network, LGBM: light gradient boosting machine, GNB: gradient naïve Bayes A large amount of high-quality training data is required to obtain an accurate prediction model. The first step is to collect and store data appropriately. In predicting IVF/ICSI success, retrospective data are usually extracted from the medical records of infertile couples who visit an infertility treatment center. These data are reported as participants or treatment cycles in studies. Table 2 presents the dataset characteristics and features. Two studies were excluded as they reported the number of transferred embryos instead of treatment cycles or the number of participants. Among all the studies, only one study merged the data of multiple IVF centers across the UK.[ 26 ] The analysis applied to the data showed that the median number of participants in each study was 1048 and the mean number was 12723 (standard deviation (SD) =33542). The minimum number of participants was 62,[ 29 ] and the maximum was 141160.[ 26 ] Details of datasets in reviewed studies In less than half of the studies, feature selection methods were used to identify features with greater predictive power. The goal of feature selection methods is to help create more robust models to increase predictive performance. The feature selection methods used in the studies can be divided into three main categories: expert opinions, statistical methods, and ranking and ML techniques. In the first group, features are selected mainly based on subject knowledge, expert opinion, and clinical practice guidelines (CPGs). Only four studies used this method for feature selection.[ 19 24 25 45 ] In the second group, statistical methods are used to select features. Two studies used t-tests, correlation analysis, and multivariable logistic regression.[ 27 30 ] The third group used ranking algorithms (hill-climbing algorithm, principal component analysis method, information gain, and Johnson algorithm) and ML techniques (decision tree (DT), genetic algorithm, and random forest (RF)). Seven studies used these algorithms to define the features.[ 22 29 32 38 39 40 42 ] In eleven studies, no feature selection method was employed. Furthermore, three studies did not report any strategy for feature selection. All studies except one[ 31 ] reported the features elaborately in a constructed table. The mean number of features, used in each study, was 21 (SD = 16). The minimum number of features was 6,[ 35 ] and the maximum was 78.[ 25 ] In general, 107 various features were reported in all reviewed studies [ Table 3 ]. The features used in the studies could be classified into eight main categories: demographics (3 features), medical history (19 features), laboratory findings (24 features), ovary stimulation (10 features), oocyte (15 features), sperm (13 features), and embryo (23 features). Female age was the most common feature used in all of the identified studies. Follicle-stimulating hormone (FSH) was the second most frequently used feature in 19 studies (73.07%). Other than women’s age and FSH, the cause of infertility and duration of infertility were repeated in 13 (50%) and 12 (46.15%) studies, respectively. Estradiol, endometrial thickness, and body mass index (BMI) as three common features were used in 11 studies (42.30%). Fifty-five (51.4%) features were used only once in studies. Features in reviewed studies The mean number of ML techniques employed in each study was 3 (SD = 2). Most studies (96.3%) applied a supervised approach to develop predictive models, and only one study applied an unsupervised approach to model the data. A wide range of different types and combinations of ML techniques have been used in the studies. Among all, support vector machine (SVM) was the most frequently applied technique (44.44%). Both RF and artificial neural network (ANN) were applied in 40.74% of studies. DT and naïve Bayes (NB) were applied in 37.3% and 29.62% of studies, respectively. The number of studies applying ML techniques to IVF or ICSI prediction has grown during the past decade. Figure 3 displays the trend of applied techniques in the studies. The top most-frequent ML techniques were compared together. ML publication and applied techniques by year There was considerable variety in the way authors chose to evaluate the predictive model performance. Nineteen different indicators have been used in studies to evaluate the model performance. Of the 24 studies included in the review, 20 studies (74.07%) reported AUC as their performance indicator. Accuracy (55.55%), sensitivity (40.74%), and specificity (25.92%) were also commonly reported. Seventeen percent of studies reported positive predictive value (PPV), while negative predictive value (NPV), F-measure, recall, and false alarm ratio were all reported by 8%. Other indicators were used only in one study. Fourteen studies have compared the performance of multiple ML techniques to choose the best one, based on performance indicators. In eight of these studies, RF performed the best. Table 4 presents the comparison among the six most frequent techniques in pairs. The most common comparison was between RF and SVM. RF outperformed in 83% of the cases (five of six studies). Five studies compared RF and ANN, in which RF surpassed all of them. Also, RF compared with NB in four studies and performed better in all of them. Two studies compared RF and DT, which RF outperformed. Comparing technique performance: Counts of studies where technique outperformed technique B NB and SVM were compared in two studies, while NB demonstrated superior performance in both of them. NB also outperformed ANN in two studies. One study compared SVM and ANN, while SVM showed superior performance. SVM performed better in comparison with DT in another one. Only one study compared XGBoost and RF, in which XGBoost performed better. Also, XGBoost presented superior performance rather than SVM in another study.

In this study, we aimed to gain insight into the application of ML techniques to predict infertility treatment success. To the best of our knowledge, this is the first review that systematically identifies predictive models of ART outcome success and their determinants. In this review, female age was identified as the most significant feature in studies. In the majority of reviewed studies, supervised ML approaches have been used to predict success. The most common technique used in the studies was SVM, followed by RF, ANN, DT, and simple Bayes. This systematic review provides insight into the evolving landscape of ML applications in predicting ART success and guiding future researchers and practitioners in developing more robust and clinically relevant prediction models. The synthesis of current knowledge underscores the importance of standardized methodologies and understanding the factors influencing ART outcomes for informed decision-making in infertility treatment. There are no conflicts of interest.

This review was performed to review the application of ML techniques to predict infertility treatment success. Of the 3655 records retrieved through searching four literature databases, 27 eligible studies matched the inclusion criteria. These studies mainly focused on developing ML predictive models for the ART outcome. Results showed that the number of ML publications in ART prediction has increased over the past 5 years. Since 2017, the number of publications predicting the outcome of infertility treatment has grown substantially. The interest in using ML to predict ART outcomes has increased in recent years, but still there is substantial scope for further applications. These findings are consistent with those of Dhombres et al .,[ 46 ] who observed a significant rising trend in AI contribution in obstetrics and gynecology, a wide range of AI approaches applied across all the subdomains. In this review, female age was identified as the most significant feature, as it was reported in all reviewed studies as well. This was consistent with Meena and Vijayalakshmi’s research findings, as they found female age to be the most significant factor influencing female fertility through statistical analysis and data mining techniques. However, Meena and Vijayalakshmi[ 47 ] found BMI and thyroid-stimulating hormone (TSH) levels as other influencing factors, which contradicts the findings of this review. Curchoe and Bormann[ 48 ] pointed out that not a specific set of features has been determined as the most contributing factors in successful infertility treatment till now. This could be known and elucidated through the use of artificial intelligence. The findings showed that the studies used 107 various features in predictive models. The plentifulness of factors affecting the outcome of ART indicates the usefulness of ML methods for prediction.[ 49 ] This is in line with Vaughan et al .’s[ 13 ] research findings as well, who concluded that the abundance of features related to infertility treatment is well suited for the application of ML. The results showed that eleven studies did not use the feature selection method and three others did not report whether they used any feature selection or not. The robustness of an ML model depends on the contribution of predicting features in the model. The accuracy of a predictive model mainly depends on how influential variables are incorporated. “Feature selection method” as a crucial step in data preprocessing can identify the most significant features needed to improve the performance.[ 50 51 ] Besides, overfitting is an event that leads to poor model performance and is associated with applying too many irrelevant features in the model. Overfitting occurs when a predictive model adapts to the details of the training dataset and affects the model performance on new data.[ 52 ] Using the feature selection method to identify the most significant variables could overcome overfitting. The vast majority of the studies have applied supervised ML approaches to predict success. In the supervised learning approach, an ML technique uses a labeled dataset to develop a predictive model. Then, the predictive model maps the input data to predefined groups or classes. As classes were defined before examining the data, it is referred to as supervised learning.[ 14 ] Results revealed that supervised ML approaches could be suitable for the prediction of infertility treatment success. This is in agreement with the findings of Wang et al .[ 16 ] who reported the successful application of supervised learning methods to the prediction of ART success. Although nearly all studies used supervised ML techniques, there was significant heterogeneity in the applied modeling approaches. Studies differed concerning the ML techniques and features used for predictive modeling for infertility. The most common technique reported in studies was SVM, followed by RF, ANN, DT, and NB. In this review, ML techniques were compared to performance indicators in 14 studies. According to findings, in more than half of these studies, RF outperforms other ML techniques in supervised learning tasks, while the ensemble techniques are less vulnerable to overfitting[ 53 ] and are recognized as superior to individual classifiers in prediction performance.[ 54 ] Most studies reported AUC, accuracy, sensitivity, and specificity as performance indicators. Although further metrics are required for better predictive performance, only a limited number of studies have considered PPV and NPV, despite their critical role in the performance prediction assessment.[ 55 ] There are limitations to this review, which can be considered in future studies. First, despite our efforts to perform a comprehensive search across multiple online databases, we may have missed non-English studies reporting ML predictive models. This review may likely be influenced by publication bias. Positive results are typically more likely to be published.[ 56 ] Selective outcome reporting may have affected our findings since we found no studies reporting failed predictive models. Nevertheless, in this review, an attempt was made to report and compare the performance of all the different predictive models in the studies. This review highlights the need for continued research and innovation in the application of artificial intelligence in reproductive medicine. Also, it provides a snapshot of the current landscape in this field. The combination of features, along with a thorough understanding of ML techniques, is identified as two critical elements for the development of predictive ART models. Considering this will have a major impact on this realm, as technology and data availability continue to advance.

This study was conducted according to Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to report on evidence from the studies that were included in this systematic review.[ 17 ] The questions examined in this review were as follows: What features were used in predictive modeling for infertility success? What ML techniques were applied in predictive modeling? What performance indicators were reported in reviewed studies? A systematic search was conducted in PubMed, Web of Science, Scopus, and Embase, in the time frame of 2000 to 2022, based on terms in the title, abstract, MeSH terms, and keywords. The final literature search was performed on November 18, 2022. The detailed search strategy is outlined in Figure 1 . Different terms were applied for the two main concepts of “ART” and “ML.” Box “A” contains all the terms on the “ART” concept, and box “B” contains all terms related to the concept of “ML.” MeSH terms and Emtree terms are seen as highlighted ones. The search was performed by combining these two groups using the Boolean operator “AND.” Truncation symbols, phrase searching, and other techniques were used to do a comprehensive and specific search. Search strategy (Keywords related to (a) ART & (b) ML) This review focused only on the full text of original papers published in English that addressed predictive models of ART (IVF/ICSI) outcomes based on an ML technique. Studies on the selection of embryos based on analyzing data of embryo images and predictive models based on any sort of deep learning approach were excluded. In the screening step, three authors (ShD, RR, and HM) examined the papers by their titles and abstracts and irrelevant studies were removed. In the eligibility step, the full text of the selected papers was independently evaluated by two authors (ShD and RR) according to the inclusion or exclusion criteria. The third author (HM) was available to solve disagreements by brainstorming and consensus. A qualitative and quantitative analysis of the studies was conducted. Data extraction involved collecting data in studies on dataset characteristics (number of records, number and types of features involved, and feature selection method), ML techniques (number and types of methods), and predictive model performance indicators. The quantitative analysis was conducted in Statistical Package for the Social Sciences (SPSS) version 26. The first author’s name, publication year, and the study’s location were also extracted. Disagreements were again solved by consensus involving the third author. The result of this analysis is adjusted in the form of a structured table. Two authors independently appraised the quality of the papers using the PROBAST checklist. PROBAST is a tool developed to evaluate the bias risk and applicability of prediction model studies.[ 18 ] It was created to give an organized and transparent technique for assessing the methodological quality of prediction model investigations. It has 20 signaling questions divided into four domains (participants, predictors, outcome, and analysis). This study was approved by the Ethics Committee of Shahid Beheshti University of Medical Sciences (IR.SBMU. RETECH.REC.1400.695). The authors declare that no human subjects were involved in this research.