https://orcid.org/0000-0003-0771-7306
Figures
Abstract
Introduction
IVF is a widely-used assisted reproductive technology with a consistent success rate of around 30%, and improving this rate is crucial due to emotional, financial, and health-related implications for infertile couples. This study aimed to develop a model for predicting IVF outcome by comparing five machine-learning techniques.
Method
The research approached five prominent machine learning algorithms, including Random Forest, Artificial Neural Network (ANN), Support Vector Machine (SVM), Recursive Partitioning and Regression Trees (RPART), and AdaBoost, in the context of IVF success prediction. The study also incorporated GA as a feature selection method to enhance the predictive models’ robustness.
Results
Findings demonstrate that AdaBoost, particularly when combined with GA feature selection, achieved the highest accuracy rate of 89.8%. Using GA, Random Forest also demonstrated strong performance, achieving an accuracy rate of 87.4%. Genetic Algorithm significantly improved the performance of all classifiers, emphasizing the importance of feature selection. Ten crucial features, including female age, AMH, endometrial thickness, sperm count, and various indicators of oocyte and embryo quality, were identified as key determinants of IVF success.
Conclusion
These findings underscore the potential of machine learning and feature selection techniques to assist IVF clinicians in providing more accurate predictions, enabling tailored treatment plans for each patient. Future research and validation can further enhance the practicality and reliability of these predictive models in clinical IVF practice.
Citation: Dehghan S, Rabiei R, Choobineh H, Maghooli K, Nazari M, Vahidi-Asl M (2024) Comparative study of machine learning approaches integrated with genetic algorithm for IVF success prediction. PLoS ONE 19(10): e0310829. https://doi.org/10.1371/journal.pone.0310829
Editor: Sanaz Alaeejahromi, Shiraz University of Medical Sciences, ISLAMIC REPUBLIC OF IRAN
Received: January 17, 2024; Accepted: September 8, 2024; Published: October 11, 2024
Copyright: © 2024 Dehghan et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: As mentioned in the method section of the manuscript, medical records of the couples undergoing IVF cycles at Helal-e-Iran Hospital were used to create the dataset. Although all data were de-identified, we are unable to publicly share these data due to ownership constraints by a third-party organization, the hospital board of trustees. Despite this limitation, we remain committed to collaborate with researchers and students who seek access to this dataset for academic and research purposes. Upon request, we will provide the necessary permissions to make the data available. Dr. Hamid Choobineh is the institutional representative, who is listed as the corresponding author and would be able to facilitate data access. Also, any request for data access could be sent to Ms Moideyan at Royesh Clinic (Moideyannahid71@gmail.com).
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
1. Background
1.1 Introduction
Confronted with the increasing prevalence of infertility, this decade has faced a sharp rise in the application of assisted reproductive technologies globally [1, 2]. In vitro fertilization (IVF), the most common and widely used procedure in ART, is a process for breeding an embryo in vitro, which, if successful, leads to pregnancy.
Since the birth of the first IVF baby in 1978 [3], efforts have been made to improve the success of IVF. However, the success rate has remained roughly constant at about 30% [4, 5] which, along with severe side effects [6, 7] and a financial burden [8, 9], make parenthood a long road for infertile couples. IVF failure can also bring emotional distress such as anxiety and stress, which may affect the quality of life and even lead to marriage failure [10, 11]. This makes "How likely can my IVF be successful?" become the most important question for infertile couples seeking treatment. Typically, to answer this question, the IVF clinicians need to consider all the demographic and clinical variables related to both female and male. With a variety of variables and their complex relations, providing an accurate estimation of success chance is challenging. Therefore, accurate models are required to predict IVF success appropriately [12, 13].
Machine learning (ML), as a subset of artificial intelligence can predict clinical outcome, by developing prediction models, based on these contributing factors. Robust prediction models can allow IVF clinicians to estimate IVF success outcome more accurately.
It is crucial to recognize that the robustness of any prediction model mainly depends on two critical factors: the choice of machine learning algorithm and the selection of the most contributing and informative features [14, 15]. Feature selection plays a pivotal role in enhancing model performance by identifying the most significant features. However, it is still a challenge to integrate the right set of features with the right ML algorithm [16, 17].
While various studies have applied machine learning techniques to develop IVF prediction models, many have relied on filter methods for feature selection [18–22], which often overlook complex interactions among variables and fail to capture the intricate relationships inherent in IVF data [23, 24]. In this study, we introduce the Genetic Algorithm as a robust wrapper method to explore the entire solution space, dynamically identifying an optimal subset of features that contribute to IVF success prediction. This approach is more flexible and effective than traditional filter methods, as it accounts for complex interactions among features.
We systematically compare the performance of five well-known machine learning algorithms—Random Forest (RF), Artificial Neural Network (ANN), Support Vector Machine (SVM), Recursive Partitioning and Regression Trees (RPART), and AdaBoost—in predicting IVF outcomes. This comparison is enhanced by the application of GA-based feature selection. By combining GA for feature selection with the aforementioned machine learning techniques, we tried to develop a robust prediction model for IVF success, offering potential improvements over existing models that rely on traditional feature selection methods.
The remainder of this paper is organized as follows: In section 1.2, we provide an overview of the previous studies, in Section 2, we describe the dataset used in this study, including the features and preprocessing steps. Section 3 describes the methodology, including the machine learning techniques applied, the Genetic Algorithm for feature selection, and the experimental setup. Section 4 presents the results of our experiments, comparing the performance of different models with and without GA-based feature selection. In Section 5, we discuss the implications of our findings, including the strengths and limitations of our approach. Finally, Section 6 concludes the paper with a summary of our contributions and suggestions for future research.
1.2 Related works
For the prediction of IVF outcome, the majority of the prediction models; have applied filter methods for feature extraction algorithms, but few studies have applied wrapper methods to explore their impact on the model performance. Table 1 illustrates these studies, a basis for comparison and discussion.
Tian et al. [25] aimed to predict fertilization failure in ART treatments using Bayesian network (BN) modeling. Analyzing 106,640 IVF/ICSI cycles from a Chinese reproductive health center, the study incorporated 24 predictors, including female, male, and treatment-related variables. BN modeling yielded a predictive model with 91.7% accuracy. Results showed strong calibration, with ROC AUCs of 0.779 (TFF vs. control) and 0.807 (TFF vs. LRF). Limitations include reliance on single-center data and absence of detailed IVF laboratory parameters, necessitating further validation and refinement.
Yang et al. [18] analyzed the risk factors affecting clinical pregnancy outcomes in patients undergoing in vitro fertilization embryo transfer (IVF-ET) and constructed a predictive model based on these factors. Data from 369 women undergoing IVF-ET were analyzed in this nested case-control study. Univariate and multivariate logistic regression analyses identified potential predictors, while a random forest model was validated using ten-fold cross-validation. Results highlighted age, BMI, cycle number, hematocrit, LH, progesterone, endometrial thickness, and FSH as predictors associated with clinical pregnancy outcome. Limitations included a small sample size and lack of external validation, emphasizing the need for larger-scale studies to validate these findings.
Wen et al. [26] developed an AI model for predicting pregnancy outcomes and multiple pregnancy risks using data from 1507 fresh embryo transfer cycles, encompassing 20 features,. Six machine learning algorithms were applied, with XGBoost demonstrating superior performance. The pregnancy prediction model achieved an accuracy of 0.716, and an AUC of 0.787. Limitations such as sample size constraints and exclusion of certain treatment modalities warrant further investigation.
Amini et al. [27] aimed to categorize successful IVF deliveries based on couples’ characteristics and available reproductive data using various classification methods. Conducted at a Tehran infertility center, the study collected data from 6071 IVF cycles spanning three years. Six machine learning approaches were employed, with Random Forest (RF) showing the highest accuracy (ACC = 0.81) in predicting successful delivery. Despite limitations such as single-center data and unrecorded predictors, the study highlights the importance of predictive modeling in optimizing IVF outcomes.
Vogiatzi et al. [28] developed an artificial neural network (ANN) for predicting live birth outcomes in ART using 257 infertile couples’ data from 426 IVF/ICSI cycles. Identified 12 significant parameters for ANN construction, achieving 74.8% accuracy. Limitations included the need for a more diverse cycle and infertility factor inclusion. Encouraged external validation and multicenter collaboration for enhanced applicability.
Qiu et al. [29] in their study analyzed data from 7188 women undergoing initial IVF at a Chinese medical center (2014–2018), developing machine learning models with pre-treatment variables. XGBoost demonstrated the best performance, achieving an AUC of 0.73. Limitations included the single-center nature of the study, exclusion of previously treated couples, and absence of family genetic history and lifestyle factors in the dataset.
2. Materials and methods
2.1 Dataset description
Medical records of couples undergoing IVF cycles at Royesh clinics, Helal-e-Iran Hospital located in Tehran, Iran, were reviewed for inclusion in this study. For each couple, only the fresh cycle of ovarian stimulation was considered. As a result, donor oocytes or embryos, frozen oocytes/embryos, and PGD/PGS cycles were excluded. A total of 812 patients met the inclusion criteria. This study was approved by the Ethics Committee of Shahid Beheshti University of Medical Sciences (IR.SBMU. RETECH.REC.1400.695). All data were de-identified and used with unique patient identifier codes. Relevant data, including demographics, medical/reproductive history of both partners, baseline information, test results, clinical diagnosis, and the treatment procedure, were extracted and recorded in a database. Fig 1 presents an image of the dataset.
All variables were extracted from textbooks, papers, and guidelines. Then an expert panel consisting eight infertility specialists was conducted. During this process, experts’ opinions toward the initial list of variables were collected based on a Likert-type scale checklist. There was an open question about whether they could add any other contributing variables that were not in the variable list. 26 variables were specified, one for treatment outcome. Table 2 shows the variables and their characteristics. The primary outcome was defined as clinical pregnancy as a positive β-HCG test result after the treatment cycle. Outcomes were obtained through a review of medical records.
2.2 Data pre-processing
Data pre-processing is an essential part of data mining, and its goal is to get the data ready for the crucial learning model stage. The goal is to improve data quality and make it easier to understand the rules generated by the models, reducing the number of variables.
First, the collected data is divided into two groups: the target variable, which was the treatment outcome, and the predictor variables, as were the remaining ones. As the dataset includes 26 variables, 25 of them are considered predictors and one is the target class.
Among the variables, “smoking” was excluded since its values were constant in more than 90% of cases. Moreover, variables including “type of infertility” and “PCOs” were also excluded while their correlation was about zero. Finally, 22 variables remained as predictors to perform classification.
None of the variables had missing values for more than 50% of the dataset. For variables with missing values less than 50%, techniques for estimating and replacing missing values were employed. The average imputation and mode imputation methods were used to replace the missing values of numerical and categorical variables. Table 2 illustrates the distribution of the normal ranges for these features. In addition, four variables: FSH, LH, AMH, AFC, and vitamin D3, were divided into three categories (high, low, and normal) to convert discrete features into nominal values. Table 3 presents the range of these variables.
2.3 Proposed feature selection model
Feature selection was performed exclusively using a Genetic Algorithm (GA). The GA was chosen for its ability to effectively explore the entire feature space, identifying the subset of features that contributed most significantly to classifier performance. Unlike filter methods, GA evaluates combinations of features by iteratively evolving a population of candidate solutions, thereby capturing complex interactions among features that may not be apparent through simpler methods. It aims to improve the performance and efficiency of data mining algorithms by reducing dimensionality, removing irrelevant or redundant features, and enhancing interpretability.
In this study, the feature selection process involves two stages. In the first stage, GA is employed as a feature reduction mechanism to identify discriminative features and remove redundant data. In the second stage, the best subset of features obtained by GA is used as input for different data mining techniques.
GA works with a population and generates improved solutions iteratively. Until satisfactory results are obtained, GA creates successive populations of potential solutions represented chromosomes. In the evaluation process, a fitness function assesses the quality of each solution. The crossover and mutation functions are two important operators that significantly affect the fitness value. Chromosomes for reproduction are selected based on their fitness value, with a higher probability of selection for those with greater fitness. The fittest chromosomes are more likely to be chosen for the recombination pool using either the roulette wheel or tournament selection methods. Mutation involves the random updating of genes. Crossover is a genetic operation that combines different characteristics from pairs of subsets to create a new subset. The offspring replaces the previous population using either the elitism or variety replacement strategy, resulting in a new population for the next generation. To improve performance, selected features based on genetic algorithms are used as input for classifiers.
To model the fitness function, three criteria need to be considered: the accuracy of the model, the number of selected features, and the cost. A chromosome will have a satisfactory fitness value if it meets the acceptable classification accuracy rate, selects only significant and informative features, and reduces costs. Chromosomes with higher fitness values are more likely to be used in the next generation, as they align with user specifications. To achieve accurate feature selection using GA, the following steps should be followed. Fig 2 illustrates the application of the GA algorithm to feature selection.
- Data scaling: It offers two benefits. First, it allows controlling attributes within a smaller numeric range rather than a larger one. Second, it helps avoid numerical calculation issues.
- Conversion: The process of converting genotype to phenotype involves transforming each feature chromosome.
- Subset selection: Feature subset refers to a selected group of features chosen for analysis or modeling.
- Fitness evaluation: It is the assessment of how well a particular solution or individual performs to the desired outcome or objective.
- Termination criteria: It determines when the process should be stopped. If the criteria are met, the process ends; otherwise, it continues with the next generation.
- Genetic operation: It is a step in which genetic operations are used to search for a better solution within the GA algorithm.
2.4 Model training & evaluation
In this study, ANN, RPART, RF, and SVM were trained to predict pregnancy outcomes after IVF. Each of these techniques was chosen for its robustness and effectiveness in handling complex datasets. In addition, we applied an Adaboost constructed on the four mentioned classifiers and compared their performance to select the most robust prediction model.
- Artificial Neural Network (ANN)
ANN consists of layers of neurons, which are interconnected by modifiable weights, represented by links between layers. The size of the input and output layers is determined by the number of variables in input and output data, respectively. In this study, we used 200 neurons with 100 epochs to train the model. - Recursive Partitioning and Regression Trees (RPART)
RPART is a recursive partitioning algorithm used for classification and regression in decision trees. The algorithm uses a binary tree structure to represent the decision rules and is built by recursively splitting the data into smaller subsets until a stopping criterion is met. The Gini index was used as the impurity measure and the maximum depth was determined as five. - Random Forest (RF)
The random forest method is a technique that uses many decision trees to solve classification and regression. The trees are built independently by randomly selecting vectors and have the same distribution in the forest. The error rate converges to a limit as the number of trees in the forest increases. We built the model using 1000 tress with maximum depth of three. - Support Vector Machine (SVM)
As a method for classification and regression, SVM uses a kernel that first maps the input space into a higher-dimensional feature. Then, a hyperplane is constructed in the transformed space to classify the dataset. we used a linear kernel with c = 1 in this study. - AdaBoost
AdaBoost, also known as Adaptive Boosting, is a machine learning technique used as an ensemble method. This meta-algorithm combines multiple weak classifiers and assigns equal initial weights to each sample. After each training round, the weights of the samples were adjusted based on their classification errors. The weight of misclassified samples is increased to give them more importance in subsequent iterations. Through this iterative process, k weak learners are obtained. Finally, a weighted combination is performed to obtain a strong learner.
The working process of Adaboost is depicted in Fig 3.
During the data preparation phase, a certain number of models are selected. The first-choice model is created, and priority is given to incorrectly classified records in this initial model. Only these records were used as input for the next model. This process continues until multiple base models are determined. It’s important to note that redundant records are allowed with all the supporting methods. Algorithm 1 illustrates how the first model is created and identifies the errors made.
Algorithm 1: Pseudocode of Adaboost
Given (x1, y1),…,(xn, yn); x1 ∈ y1, ∑y = {−1,1}
Initialize D1 (i) = 1/n
For t = 1 … T:
1. Train the weak classifier using Di
2. Get weak hypothesis ht: X→{−1,1}error
∑kht (xi) ≠ yiDt(xi)
3. Choose
4. Update
Dt+1(i) = Dt(i)/Zt
5. Output H(x) = sign (∑Tt=1atHt(x))
Records that are classified incorrectly serve as input to the subsequent model. This iterative process repeats until a predefined condition is met. As can be seen, there are ’n’ numbers of models created by using the errors from the previous model. This is how the supporting process works. Models 1, 2, 3…, N represent individual classification models that can be considered as supporting models. Various supporting models operate based on the same principle.
As the dataset was highly imbalanced (189 samples in class “yes” and 623 samples in class “no”), oversampling was used. Oversampling helps mitigate this problem by generating synthetic examples for the minority class, thereby increasing its presence in the dataset. We utilized SMOTE (Synthetic Minority Over-sampling Technique) for this purpose. SMOTE generates synthetic samples by interpolating between neighboring instances of the minority class and creating new instances along the line segments. The resulting balanced dataset provides a better representation of all classes, allowing data mining methods to learn from more diverse examples and potentially improve their performance in accurately predicting the minority class. We used the Python programming language with the help of Scikit-Learn tools. Balancing was performed only on the training set. To ensure that the synthetic data generated by SMOTE is consistent with the original data, we conducted the Kolmogorov-Smirnov (KS) test to compare the distributions of the original minority class data and the synthetic samples generated by the SMOTE algorithm. By calculating the empirical Cumulative Distribution Functions (CDFs) for each feature in both datasets, we obtained the KS statistic, which measures the maximum difference between the CDFs.
Model validation was performed by random data allocation into training and test sets and by k-fold cross-validation with 10 iterations. Four measures: accuracy, precision, recall and, F-measure were used to evaluate the performance of the prediction models. In addition, we included the AUC (area under the ROC curve) metric, commonly used in medical data mining tasks. These metrics can be calculated using the following equations:
3. Results
In this study, the performance of the five Machine learning techniques in terms of performance metrics was compared to each other using the proposed feature-selection method (GA). To achieve this, a dataset of 812 records with 22 variables was used. We utilized SMOTE to deal with the imbalanced data. The comparative study of obtained accuracy with and without the application of SMOTE on imbalanced datasets is provided in Table 4. As can be seen, SMOTE increased the classification accuracy of all techniques.
Table 5 summarizes the result of the KS test. As can be seen, high p-values from the KS test were obtained for all features that did not reject the null hypothesis. This suggests that there is no significant difference between the distributions. Finally, it can be concluded that the synthetic data is consistent with the original data and SMOTE effectively captured the characteristics of the original minority class, ensuring its adequacy for model training.
Also, Table 6 compares the performance of the five classifiers with and without the genetic algorithm as a feature selection method. For each classifier, the same training was performed but with different model parameters. 10-fold cross-validation was applied to evaluate the performance of the models on the balanced dataset.
The results indicate that Adaboost outperforms other classifiers regardless of whether the GA feature selection method is implemented or not. Its accuracy score was 87.3% without using the GA feature selection method, while it is further raised to 89.8% when GA is employed. Following Adaboost, RF surpassed other classifiers with an accuracy of 85.8% and 87.4% with and without the GA feature selection method, respectively. Overall, the results show a relatively significant difference in the performance of all classifiers when feature selection with GA was applied. When GA was used as the feature selection method, all classifiers achieved higher values for each performance measure.
The predictive abilities of the classifiers were further analyzed with receiver operating characteristic (ROC) curves. As shown in Fig 4, the five classifiers showed little differences, but Adaboost had the highest AUC of 0.910. Again, RF surpassed the other classifiers, with a slight difference from Adaboost, with an AUC score of 0.903.
By employing each classifier as an evaluation function, the Genetic Algorithm seeks to identify the optimal feature combination for each classifier. Fig 5 lists this subset of features for each classifier. The significance of each feature, as illustrated in Fig 5, reflects its frequency of selection across these classifiers.
A score of 5 on the figure indicates that the feature was selected by all five classifiers in their final feature subsets. Female Age was identified as a critical feature by all five classifiers, achieving the highest possible score of 5. Similarly, features such as AMH, endometrial thickness, number of oocytes, FSH, and number of embryo were selected by the majority of classifiers, with scores ranging from 4 to 5. In contrast, features like Vitamin D, Previous Pregnancy, and Cause of Infertility received lower scores, suggesting they were less frequently included in the feature subsets generated by the GA-based approach.
Also, ten features were identified as the most common ones: female age, AMH, Endometrial thickness, follicle size, number of oocytes, quality of retrieved oocytes (MI), quality of retrieved oocytes (MII), sperm count, sperm morphology, and quality of embryo.
Table 7 provides a detailed overview of features in the subset of a given classifier. Notably, RPART exhibits the highest number of 20 features in its subset, while SVM maintains the lowest number of 10 features.
4. Discussion
This study attempted to utilize artificial intelligence in the IVF treatment practice through proposing a model to predict pregnancy outcome for IVF patients. For this, five well-known classifiers namely, ANN, RPART, RF, SVM and, AdaBoost were compared to select the most robust one, using the genetic algorithm as the feature selection method. A dataset of 812 records with 22 variables was used in this study.
The results demonstrate that SMOTE consistently improves classifier accuracy across all models. SMOTE is beneficial to data mining as it effectively addresses class imbalance issues by oversampling the minority class through the generation of synthetic instances. By creating new samples rather than simply duplicating existing ones, SMOTE helps mitigate overfitting, preserve the underlying data structure, and improve model generalization on unseen data [30]. These characteristics make SMOTE a valuable tool for enhancing model performance and accuracy in machine learning tasks, as it provides a more balanced dataset that can be easily integrated with various algorithms. Overall, SMOTE plays a crucial role in improving the robustness and reliability of machine learning models when dealing with imbalanced datasets [31, 32].
Our findings are in contrast with the results of a study, Hafiz et al. [33] did in 2017. They compared AdaBoost with four classifiers (SVM, RPART, RF and ANN) in which RF surpassed the others with an accuracy of 83.96 and AUC of 93.74. Contrary to our results, AdaBoost gained a relatively weak performance with an accuracy of 66.99 and an AUC of 47.52. This discrepancy could be due to the difference in model design between the two studies. The difference in the number of dataset records and the features used in the two studies can affect the model performance. Moreover, AdaBoost’s capacity to provide weights to each classifier’s predictions adaptively, improves the overall model’s robustness and efficacy. In this study, we constructed AdaBoost using the four other classifiers, i.e., SVM, RPART, RF, and ANN. This could also explain the significant difference between the performance of AdaBoost in the two studies.
Although little research has been done to compare Adaboost with other classifiers to develop an IVF outcome prediction model, but in many studies, RF has been compared to determine the highest performance. Their findings show that, in most of the cases, RF has surpassed other classifiers compared with [13, 22, 27, 34, 35]. This superiority of RF, AdaBoost and other similar methods may be due to their structure and how they predict. Known as ensemble learning methods, they are based on the idea of combining several different prediction models to create a global composite model that generates reliable and accurate estimates or predictions [36]. Ensemble learning methods seek to overcome overfitting by combining several ‘‘weak” models to form a diverse and accurate model. Theoretical and experimental evidence have shown that ensemble models achieve much better prediction performance than single models [37].
Another part of the findings depicts that applying Genetic algorithm as a feature selection method resulted in a relatively significant enhancement of the performance of all classifiers for each performance measure. Feature selection has a crucial role in eliminating irrelevant or redundant features, which can affect the model’s performance in predicting the IVF outcome accurately. Feature selection helps reduce the dimensionality of the dataset by selecting a subset of relevant features [38]. Genetic Algorithm, as an optimization approach based on the concept of biological evolution, has recently been more developed compared with different feature selection algorithms, and its advantages have been verified in various medical disciplines [39–41]. For the IVF prediction model, Guh et al. pointed out that GA performed better in attribute selection than other compared methods [42].
The results from Fig 5 demonstrate the robustness of GA in identifying key features for predicting IVF success. A score of 5 in the graph signifies that all five classifiers found a particular feature important when GA was employed for feature selection. ’Female Age’ emerged as the most consistently selected feature, aligning with clinical knowledge about its significant impact on fertility outcomes [43, 44]. This finding validates the GA’s ability to prioritize features crucial for prediction across multiple machine learning models [45]. Other important features, such as AMH, endometrial thickness, number of oocytes, and embryo quality, also achieved high selection scores, indicating their relevance across various classifiers. These features are well-known indicators of ovarian reserve, embryo viability, and overall reproductive potential, further supporting their inclusion in the models. On the other hand, features like Vitamin D, Previous Pregnancy, and Cause of Infertility were less frequently selected, which might suggest that their predictive value is either limited or highly context-dependent. This variation in selection across classifiers highlights the GA’s flexibility in evaluating complex interactions between features, which may be overlooked by simpler, filter-based methods.
While every classifier applied a different set of features in the prediction model, ten features were used in common by all of the five classifiers. Women’s age, AMH, Endometrial thickness, sperm count, sperm morphology, follicle size, retrieved oocytes, quality of retrieved oocytes (MI), quality of retrieved oocytes (MII) and quality of embryo are found to be the most important features affecting the IVF outcome. These features have been indicated to have a significant contribution to the prediction of IVF outcome in previous studies [27, 29, 46–48]. Selecting an appropriate feature subset is quite vital, as it can affect the prediction model notably. Curchoe et al. pointed out that, no precise set of features has been identified as the most important predictors in successful infertility treatment [49]. Although, the main objective of using feature selection methods is to determine the most suitable subset of significant predictors.
5. Strengths and limitation
This study highlights the ability of the capability of the predicting modelling using the Adaboost as an ensemble approach, combining the predictive performance of the four classifiers (ANN, RPART, RF, and SVM). This approach produces an ensemble model with reasonably high predicted accuracy for successful IVF by combining the mentioned classifiers. Furthermore, the research shows a significant boost in predicting accuracy by combining machine learning techniques with the Genetic Algorithm for feature selection. This novel approach improves the predictability of IVF success, giving significant insights for professionals and patients alike.
As the first limitation for this study, the findings are based on data from a single medical facility in Tehran, Iran. While the size of the data is reasonable, its applicability to different demographics and healthcare settings may be limited.
Finally, the prediction models created in this work should be externally validated in other studies to check the reliability and generalizability.
6. Conclusion
In this study, we presented a comparative study of machine learning methods integrated with Genetic Algorithm (GA) for predicting the success of In vitro fertilization (IVF) procedures. The study explores the performance of five machine learning algorithms: Artificial Neural Network (ANN), Recursive Partitioning and Regression Trees (RPART), Random Forest (RF), Support Vector Machine (SVM), and AdaBoost, using GA for feature selection.
The standout performer was Adaboost, achieving an accuracy of 89.8% with GA. Random Forest also showed promise, reaching 87.4% accuracy with GA. These results emphasize the significance of feature selection in improving model performance. Overall, the study highlights the effectiveness of ensemble learning methods like AdaBoost and Random Forest in predicting IVF outcomes.
Furthermore, the application of Genetic Algorithm as a feature selection method significantly improves the performance of all classifiers, emphasizing the importance of selecting relevant features for accurate IVF outcome prediction. Ten key features, including female age, AMH, endometrial thickness, sperm count, sperm morphology, follicle size, retrieved oocytes, quality of oocyte and embryo quality, are identified as critical factors influencing IVF success.
These models provide a promising tool for IVF practitioners, allowing for more exact treatment planning. We try to provide a better insight into the IVF success for infertile couples by applying AI and ML, which may potentially decrease emotional and financial difficulties.
References
- 1. De Geyter C. Assisted reproductive technology: Impact on society and need for surveillance. Best Practice & Research Clinical Endocrinology & Metabolism. 2019;33(1):3–8. pmid:30799230
- 2. Bulletti FM, Berrettini M, Sciorio R, Bulletti C. Artificial intelligence algorithms for optimizing assisted reproductive technology programs: A systematic review. GTM. 2023;2(2).
- 3.
Taylor HS, Pal L, Sell E. Speroff’s clinical gynecologic endocrinology and infertility: Lippincott Williams & Wilkins; 2019.
- 4. Fauser BCJM. Towards the global coverage of a unified registry of IVF outcomes. Reproductive BioMedicine Online. 2019;38(2):133–7. pmid:30593441
- 5. Reproduction TEI-MCftESoH, Embryology, Wyns C, De Geyter C, Calhaz-Jorge C, Kupka MS, et al. ART in Europe, 2017: results generated from European registries by ESHRE†. Human Reproduction Open. 2021;2021(3). pmid:34377841
- 6. Nagata C, Yang L, Yamamoto-Hanada K, Mezawa H, Ayabe T, Ishizuka K, et al. Complications and adverse outcomes in pregnancy and childbirth among women who conceived by assisted reproductive technologies: a nationwide birth cohort study of Japan environment and children’s study. BMC Pregnancy and Childbirth. 2019;19(1):77. pmid:30786865
- 7.
Sullivan-Pyke CS, Senapati S, Mainigi MA, Barnhart KT, editors. In vitro fertilization and adverse obstetric and perinatal outcomes. Seminars in perinatology; 2017: Elsevier.
- 8. Inhorn MC, Patrizio P. Infertility around the globe: new thinking on gender, reproductive technologies and global movements in the 21st century. Human Reproduction Update. 2015;21(4):411–26.
- 9. Thompson C. IVF global histories, USA: between Rock and a marketplace. Reproductive Biomedicine & Society Online. 2016;2:128–35.
- 10. Gdańska P, Drozdowicz-Jastrzębska E, Grzechocińska B, Radziwon-Zaleska M, Węgrzyn P, Wielgoś M. Anxiety and depression in women undergoing infertility treatment. Ginekologia Polska. 2017;88(2):109–12. pmid:28326521
- 11. Karaca N, Karabulut A, Ozkan S, Aktun H, Orengul F, Yilmaz R, et al. Effect of IVF failure on quality of life and emotional status in infertile couples. European Journal of Obstetrics & Gynecology and Reproductive Biology. 2016;206:158–63.
- 12. Chen Z, Zhang D, Zhen J, Sun Z, Yu Q, Yin Y. Predicting cumulative live birth rate for patients undergoing in vitro fertilization (IVF)/intracytoplasmic sperm injection (ICSI) for tubal and male infertility: a machine learning approach using XGBoost. Chinese Medical Journal. 2022;135(08):997–9. pmid:35730375
- 13. Wang C-W, Kuo C-Y, Chen C-H, Hsieh Y-H, Su EC-Y. Predicting clinical pregnancy using clinical features and machine learning algorithms in in vitro fertilization. PLOS ONE. 2022;17(6):e0267554. pmid:35675328
- 14. Dessì N, Pes B. Similarity of feature selection methods: An empirical study across data intensive classification tasks. Expert Systems with Applications. 2015;42(10):4632–42.
- 15. de la Iglesia B. Evolutionary computation for feature selection in classification problems. WIREs Data Mining and Knowledge Discovery. 2013;3(6):381–407.
- 16. Swain J, VerMilyea MT, Meseguer M, Ezcurra D, Letterie G, Sánchez P, et al. AI in the treatment of fertility: key considerations. Journal of Assisted Reproduction and Genetics. 2020. pmid:32989510
- 17. Fernandez EI, Ferreira AS, Cecílio MHM, Chéles DS, de Souza RCM, Nogueira MFG, et al. Artificial intelligence in the IVF laboratory: overview through the application of different types of algorithms for the classification of reproductive data. Journal of Assisted Reproduction and Genetics. 2020;37(10):2359–76. pmid:32654105
- 18. Yang H, Liu F, Ma Y, Di M. Clinical pregnancy outcomes prediction in vitro fertilization women based on random forest prediction model: A nested case-control study. Medicine (United States). 2022;101(49):E32232.
- 19. Liu H, Zhang ZR, Gu YF, Dai CS, Shan GQ, Song HC, et al. Development and evaluation of a live birth prediction model for evaluating human blastocysts from a retrospective study. Elife. 2023;12. pmid:36810139
- 20. Mehrjerd A, Rezaei H, Eslami S, Ratna MB, Ghaebi NK. Internal validation and comparison of predictive models to determine success rate of infertility treatments: a retrospective study of 2485 cycles. Scientific Reports. 2022;12(1). pmid:35508641
- 21. Chen Z, Zhang D, Zhen J, Sun Z, Yu Q. Predicting cumulative live birth rate for patients undergoing in vitro fertilization (IVF)/intracytoplasmic sperm injection (ICSI) for tubal and male infertility: a machine learning approach using XGBoost. Chin Med J (Engl). 2022;135(8):997–9. pmid:35730375
- 22. Goyal A, Kuchana M, Ayyagari KPR. Machine learning predicts live-birth occurrence before in-vitro fertilization treatment. Scientific Reports. 2020;10(1). pmid:33262383
- 23. Yland J, Messerlian C, Mínguez-Alarcón L, Ford JB, Hauser R, Williams PL. Methodological approaches to analyzing IVF data with multiple cycles. Human Reproduction. 2018;34(3):549–57.
- 24. Maity A, Williams PL, Ryan L, Missmer SA, Coull BA, Hauser R. Analysis of in vitro fertilization data with multiple outcomes using discrete time-to-event analysis. Stat Med. 2014;33(10):1738–49. pmid:24317880
- 25. Tian T, Kong F, Yang R, Long XY, Chen LX, Li M, et al. A Bayesian network model for prediction of low or failed fertilization in assisted reproductive technology based on a large clinical real-world data. Reproductive Biology and Endocrinology. 2023;21(1). pmid:36703171
- 26. Wen JY, Liu CF, Chung MT, Tsai YC. Artificial intelligence model to predict pregnancy and multiple pregnancy risk following in vitro fertilization-embryo transfer (IVF-ET). Taiwanese Journal of Obstetrics & Gynecology. 2022;61(5):837–46.
- 27. Amini P, Ramezanali F, Parchehbaf-Kashani M, Maroufizadeh S, Omani-Samani R, Ghaheri A. Factors associated with in vitro fertilization live birth outcome: A comparison of different classification methods. Int J Fertil Steril. 2021;15(2):128–34. pmid:33687166
- 28. Vogiatzi P, Pouliakis A, Siristatidis C. An artificial neural network for the prediction of assisted reproduction outcome. Journal of Assisted Reproduction and Genetics. 2019;36(7):1441–8. pmid:31218565
- 29. Qiu J, Li P, Dong M, Xin X, Tan J. Personalized prediction of live birth prior to the first in vitro fertilization treatment: A machine learning method. Journal of Translational Medicine. 2019;17(1).
- 30. Fernández A, Garcia S, Herrera F, Chawla NV. SMOTE for learning from imbalanced data: progress and challenges, marking the 15-year anniversary. Journal of artificial intelligence research. 2018;61:863–905.
- 31. Sahlaoui H, Alaoui EAA, Agoujil S, Nayyar A. An empirical assessment of smote variants techniques and interpretation methods in improving the accuracy and the interpretability of student performance models. Education and Information Technologies. 2024;29(5):5447–83.
- 32. Yang Y, Mirzaei G. Performance analysis of data resampling on class imbalance and classification techniques on multi-omics data for cancer classification. PLOS ONE. 2024;19(2):e0293607. pmid:38422094
- 33. Hafiz P, Nematollahi M, Boostani R, Jahromi BN. Predicting implantation outcome of in vitro fertilization and intracytoplasmic sperm injection using data mining techniques. Int J Fertil Steril. 2017;11(3):184–90. pmid:28868840
- 34. Kozar N, Kovač V, Reljič M. Can methods of artificial intelligence aid in optimizing patient selection in patients undergoing intrauterine inseminations? Journal of Assisted Reproduction and Genetics. 2021;38(7):1665–73. pmid:34031765
- 35. Raef B, Maleki M, Ferdousi R. Computational prediction of implantation outcome after embryo transfer. Health Informatics Journal. 2020;26(3):1810–26. pmid:31826687
- 36. Adlung L, Cohen Y, Mor U, Elinav E. Machine learning in clinical decision making. Med. 2021;2(6):642–65. pmid:35590138
- 37. Pintelas P, Livieris IE. Special Issue on Ensemble Learning and Applications. algorithms. 2020;13(6):140.
- 38.
Altarabichi MG, Nowaczyk S, Pashami S, Mashhadi PS. Fast Genetic Algorithm for feature selection—A qualitative approximation approach. Proceedings of the Companion Conference on Genetic and Evolutionary Computation; Lisbon, Portugal: Association for Computing Machinery; 2023. p. 11–2.
- 39. Aalaei S, Shahraki H, Rowhanimanesh A, Eslami S. Feature selection using genetic algorithm for breast cancer diagnosis: experiment on three different datasets. Iranian journal of basic medical sciences. 2016;19(5):476. pmid:27403253
- 40.
Kanwal S, Rashid J, Nisar MW, Kim J, Hussain A, editors. An effective classification algorithm for heart disease prediction with genetic algorithm for feature selection. 2021 Mohammad Ali Jinnah University International Conference on Computing (MAJICC); 2021: IEEE.
- 41. Paul D, Su R, Romain M, Sébastien V, Pierre V, Isabelle G. Feature selection for outcome prediction in oesophageal cancer using genetic algorithm and random forest classifier. Computerized Medical Imaging and Graphics. 2017;60:42–9. pmid:28087102
- 42. Guh RS, Wu TCJ, Weng SP. Integrating genetic algorithm and decision tree learning for assistance in predicting in vitro fertilization outcomes. Expert Systems with Applications. 2011;38(4):4437–49.
- 43. Xu T, de Figueiredo Veiga A, Hammer KC, Paschalidis IC, Mahalingaiah S. Informative predictors of pregnancy after first IVF cycle using eIVF practice highway electronic health records. Scientific Reports. 2022;12(1):839. pmid:35039614
- 44. Ribeiro S, Sousa M. In Vitro Fertilisation and Intracytoplasmic Sperm Injection predictive factors: A review of the effect of female age, ovarian reserve, male age, and male factor on IVF/ICSI treatment outcomes. JBRA Assist Reprod. 2023;27(1):97–111. pmid:35916467
- 45. GS G, NAYAK S, CHOLLI N. Evaluation of Machine Learning Algorithms on the Prediction of Live Birth Occurrence. International Journal of Pharmaceutical Research (09752366). 2021;13(2).
- 46. Liang R, An J, Zheng Y, Li J, Wang Y, Jia Y, et al. predicting and improving the probability of live birth for women undergoing frozen-thawed embryo transfer: a data-driven estimation and simulation model. Computer Methods and Programs in Biomedicine. 2021;198. pmid:33049450
- 47. Milewska AJ, Jankowska D, Cwalina U, Citko D, Wiȩsak T, Acacio B, et al. Significance of discriminant analysis in prediction of pregnancy in IVF treatment. Studies in Logic, Grammar and Rhetoric. 2015;43(56):7–20.
- 48. Mohammadi M, Kavousi A, Madani T, Amini P, Ghaheri A. Joint modeling of in vitro fertilization outcomes among a population of Iranian infertile couples: a historical cohort study. Int J Fertil S teril. 2023;17(4):306–11.
- 49. Curchoe CL, Bormann CL. Artificial intelligence and machine learning for human reproduction and embryology presented at ASRM and ESHRE 2018. Journal of Assisted Reproduction and Genetics. 2019;36(4):591–600. pmid:30690654