Editor's evaluation: Development and evaluation of a live birth prediction model for evaluating human blastocysts from a retrospective study
Article Figures and data Abstract Editor's evaluation eLife digest Introduction Methods Results Discussion Data availability References Decision letter Author response Article and author information Metrics Abstract Background: In infertility treatment, blastocyst morphological grading is commonly used in clinical practice for blastocyst evaluation and selection, but has shown limited predictive power on live birth outcomes of blastocysts. To improve live birth prediction, a number of artificial intelligence (AI) models have been established. Most existing AI models for blastocyst evaluation only used images for live birth prediction, and the area under the receiver operating characteristic (ROC) curve (AUC) achieved by these models has plateaued at ~0.65. Methods: This study proposed a multimodal blastocyst evaluation method using both blastocyst images and patient couple’s clinical features (e.g., maternal age, hormone profiles, endometrium thickness, and semen quality) to predict live birth outcomes of human blastocysts. To utilize the multimodal data, we developed a new AI model consisting of a convolutional neural network (CNN) to process blastocyst images and a multilayer perceptron to process patient couple’s clinical features. The data set used in this study consists of 17,580 blastocysts with known live birth outcomes, blastocyst images, and patient couple’s clinical features. Results: This study achieved an AUC of 0.77 for live birth prediction, which significantly outperforms related works in the literature. Sixteen out of 103 clinical features were identified to be predictors of live birth outcomes and helped improve live birth prediction. Among these features, maternal age, the day of blastocyst transfer, antral follicle count, retrieved oocyte number, and endometrium thickness measured before transfer are the top five features contributing to live birth prediction. Heatmaps showed that the CNN in the AI model mainly focuses on image regions of inner cell mass and trophectoderm (TE) for live birth prediction, and the contribution of TE-related features was greater in the CNN trained with the inclusion of patient couple's clinical features compared with the CNN trained with blastocyst images alone. Conclusions: The results suggest that the inclusion of patient couple’s clinical features along with blastocyst images increases live birth prediction accuracy. Funding: Natural Sciences and Engineering Research Council of Canada and the Canada Research Chairs Program. Editor's evaluation This article provides important findings that have practical implications for reproductive medicine and would be of interest to IVF specialists. Based on the compelling strength of evidence, the study demonstrates significant results in improving the predictive value of the live birth model based on blastocyst evaluation and clinical features. https://doi.org/10.7554/eLife.83662.sa0 Decision letter Reviews on Sciety eLife's review process eLife digest More than 50 million couples worldwide experience infertility. The most common treatment is in vitro fertilization (IVF). Fertility specialists collect eggs and sperm from the prospective parents. They combine the egg and sperm in a laboratory and allow the fertilized eggs to develop for five days into a multi-celled blastocyst. Then, the specialists select the healthiest blastocysts and return them to the patient's uterus. Since 1978, more than 8 million children have been conceived through IVF. Yet, only about 30% of IVF attempts result in a successful birth. As a result, fertility patients often undergo multiple rounds of IVF, which can be expensive and emotionally draining. Several factors determine IVF success, one of which is the health of the blastocysts selected for transfer to the uterus. Specialists select the blastocysts using several criteria. But these human assessments are subjective and inconsistent in predicting which ones are most likely to result in a successful birth. Recent studies suggest artificial intelligence technology may help select blastocysts. Liu et al. show that using artificial intelligence to assess blastocysts and fertility patient characteristics leads to more accurate predictions about which blastocysts are likely to result in a successful birth. In the experiments, the researchers trained an artificial intelligence computer program using pictures of 17,580 blastocysts with known birth outcomes and the parents' clinical characteristics. The model identified 16 parental factors associated with birth outcomes. The top 5 most predictive parental factors were maternal age, the day of blastocyst transfer to the uterus, how many eggs were present in the ovaries, the number of eggs retrieved and the thickness of the uterus lining. The program achieved the highest prediction of healthy births so far, compared to success rates listed in other studies. Artificial intelligence-aided blastocyte selection using patient and blastocyst characteristics may improve IVF success rates and reduce the number of treatment cycles patient couples undergo. Before specialists can use artificial intelligence in their clinics, they must conduct confirmatory clinical studies that enroll patient couples to compare conventional methods and artificial intelligence. Introduction Infertility is a global health issue, affecting more than 50 million couples worldwide (Mascarenhas et al., 2012). Since the birth of the first in vitro fertilization (IVF) child in 1978, over 8 million children were born with IVF treatment (Adamson et al., 2018). Among the various factors contributing to IVF outcomes, the quality of the blastocyst (day 5 embryo) selected for transfer is critical for the success of IVF treatment. Manual grading of blastocyst development stage, inner cell mass (ICM), and trophectoderm (TE) remains the most common method for blastocyst evaluation. While the blastocyst morphological grading is widely used in clinical practice, morphological grades of the development stage, ICM, and TE have shown limited predictive power on clinical outcomes (Seli et al., 2011; Reignier et al., 2019; Bartolacci et al., 2021; Ueno et al., 2021; Xiong et al., 2022). It is desired to identify features for accurate prediction of clinical outcomes of blastocysts. To achieve this goal, the convolutional neural network (CNN) is expected to play a critical role. CNN is able to automatically detect discriminative features from images and has been the state-of-the-art method in various fields in medical imaging, such as lung cancer prediction (Ardila et al., 2019), breast cancer prediction (McKinney et al., 2020), and diabetic retinopathy screening (Bora et al., 2021). To apply CNN to predict the clinical outcome of a blastocyst, images of blastocysts with a known clinical outcome (e.g., pregnancy and live birth) are collected for the CNN model development. The area under the receiver operating characteristic (ROC) curve (AUC) is the most commonly used metric to evaluate and compare machine learning models on predicting clinical outcomes of blastocysts (Kragh and Karstoft, 2021a). The AUCs reported in the literature using CNN to predict clinical outcomes from blastocyst images range from 0.64 to 0.71 for pregnancy prediction (VerMilyea et al., 2020; Kragh et al., 2021b; Berntsen et al., 2022; Enatsu et al., 2022; Loewke et al., 2022), and are around 0.65 for live birth prediction (Miyagi et al., 2019; Nagaya and Ukita, 2021). Besides using blastocyst images, attempts have also been made to use time-lapse videos for live birth prediction. These videos contain the entire development process from days 0 to 5–7. However, results in the literature show that CNN using static images achieved similar or slightly better accuracies in predicting clinical outcomes than using time-lapse videos, for instance, AUC=0.68–0.71 (VerMilyea et al., 2020; Enatsu et al., 2022; Loewke et al., 2022) versus 0.64–0.67 (Kragh et al., 2021b; Berntsen et al., 2022) for pregnancy prediction, and 0.66 (Miyagi et al., 2019) versus 0.65 (Nagaya and Ukita, 2021) for live birth prediction. A potential reason is that the redundant frames of images in time-lapse videos may work as noise causing the model to overfit and thus leading to a lower prediction accuracy (Zhu et al., 2018; Wu et al., 2021; Tao et al., 2022). Therefore, we opted to use static blastocyst images in this study. Different from using blastocyst images alone to predict clinical outcomes, Miyagi et al., 2020 proposed to use blastocyst images together with maternal clinical features including maternal age, AMH, and BMI and reported an AUC of 0.74, the highest accuracy in literature (Miyagi et al., 2020). However, two questions remain elusive. First, the contribution of blastocyst images and the additional contribution of maternal clinical features to live birth prediction are unknown. Second, endometrium status-related features, such as endometrium thickness and pattern, are also critical factors impacting live birth outcomes (Ng et al., 2007; Bu et al., 2016; Mahutte et al., 2022), but were not considered. In this study, we quantified the effect of blastocyst images and the combined effect of both blastocyst images and patient couple’s clinical features on live birth prediction. The live birth prediction model using only blastocyst images achieved an AUC of 0.67, which was significantly outperformed by the AUC of 0.77 achieved by the model using both blastocyst images and patient couple’s clinical features (p value<0.0001). Additionally, when endometrium status-related features (e.g., endometrium thickness and pattern) were excluded, the AUC of the model using both blastocyst images and patient couple’s clinical features significantly decreased to 0.74 (p value<0.0001), indicating that the inclusion of endometrium status-related clinical features helps improve live birth prediction accuracy. Sixteen patient couple’s clinical features were identified to be most related to live birth outcomes of blastocysts, among which maternal age, the day of blastocyst transfer, antral follicle count (AFC), retrieved oocyte number, and endometrium thickness measured before transfer are the top five features contributing to live birth prediction. Additionally, the CNN heatmaps showed that the CNN mainly focused on ICM and TE for live birth prediction, and the contribution of TE-related features was greater in the CNN trained with the inclusion of patient couple’s clinical features compared with the CNN trained with blastocyst images alone. Methods Data set collection We used retrospectively collected data to develop the live birth prediction model. Transferred blastocysts with known live birth outcomes for patients who underwent frozen embryo transfer cycles from 2016 to 2020, at the Reproductive and Genetic Hospital of CITIC-Xiangya, were reviewed for inclusion in the data set. Informed consent was not necessary because this study used retrospective and fully de-identified data, no medical intervention was performed on the subject, and no biological samples from the patient were collected. This study was approved by the Ethics Committee of the Reproductive and Genetic Hospital of CITIC-Xiangya (approval number: LL-SC-2021-008). Blastocyst images were captured before transfer using a standard optical light microscope mounted with a camera. Two grayscale images were captured for each blastocyst, one focusing on ICM and the other focusing on TE. Blastocysts were cropped from the original images which have a resolution of 1024×768 and were consistently padded to 500×500 to facilitate model training. Patient couple’s clinical features consist of 103 features including maternal age and BMI, the day of blastocyst transfer, infertility diagnosis and treatment history of patient couples, ovarian stimulation protocols, maternal hormone profiles, and ultrasound results measured during the ovarian stimulation process and before transfer, and paternal semen diagnosis results (see Supplementary file 1 for a complete list). Based on p value analysis and logistic regression (LR)-based sequential forward feature selection (Solorio-Fernández et al., 2020; Raschka, 2018), 16 clinical features that are most relevant to live birth prediction were identified and used for training the machine learning model (see Figure 3). Feature selection reduces the input feature dimensions by removing redundant features and features with limited predictive power, thus improving the model generalization capability (see Figure 3—figure supplement 1). The LR-based feature selection was used due to its computing efficiency, we also presented the result of multilayer perceptron (MLP)-based feature selection in Figure 3—figure supplement 1 and Figure 3—figure supplement 2. A total of 28,118 blastocysts with known live birth outcomes were reviewed, among which 17,580 blastocysts with two blastocyst images and all the 16 clinical features available were included in the data set. Model architecture Figure 1 shows the architecture of the live birth prediction model based on multimodal blastocyst evaluation. It consists of a CNN to process blastocyst images and an MLP to process patient couple’s clinical features. Features from the CNN and the MLP are fused; thus, the model can be trained to simultaneously take into account both blastocyst images and patient couple’s clinical features for live birth prediction. The last fully connected layer in the CNN and the last fully connected layer in the MLP each output a decision-level feature, which has two variables used to classify the blastocyst into the positive or negative live birth outcome category. The adding operation fuses decision-level features from the CNN and the MLP, and the result of addition is taken as the final output of the overall live birth prediction model. Figure 1 Download asset Open asset Architecture of the live birth prediction model based on multimodal blastocyst evaluation. CNN, convolutional neural network; FC, fully connected layer; MLP, multilayer perceptron. Figure 1—source data 1 Source code to reproduce the model. https://cdn.elifesciences.org/articles/83662/elife-83662-fig1-data1-v2.zip Download elife-83662-fig1-data1-v2.zip Model implementation and training The proposed live birth prediction model used EfficientNetV2-S as the backbone CNN. EfficientNetV2-S is the baseline model in the EfficientNetV2 family, which is a new family of CNN models that provide higher accuracy and training speed than conventional models (Tan and Le, 2021). In our work, the output dimension of the final fully connected layer in EfficientNetV2-S was set to be two, representing the positive and negative live birth outcome of a blastocyst, respectively. The model was implemented using PyTorch 1.10.1 (Paszke et al., 2019). Each of the 17,580 blastocysts had two images taken at different focal planes, one focused more on TE cells and the other on ICM. Furthermore, for each blastocyst, live birth outcomes and all the 16 patient couple’s clinical features were available. The blastocysts were randomly split into 80%:10%:10% to construct the training, validation, and testing data sets. The stratified random sampling approach was used to ensure that all split data sets have the same distribution of minority and majority classes. Since the ratio of blastocysts with a positive live birth outcome in the data set is 0.368, to mitigate the model’s prediction bias toward the majority category (i.e., the negative live birth outcome), the weighted sampling approach, which can help rebalance the class distributions when sampling from an imbalanced data set (Feng et al., 2021), was employed for training the model. In the weighted sampling approach, the probability of each item to be selected is determined by its weight, and the weight of each item is assigned by inverse class frequencies. In this way, the weighted sampling approach rebalances the class distributions by oversampling the minority class and under-sampling the majority class. We also verified the approach of using weighted cross-entropy loss, which assigns greater weights to the loss caused by the prediction error of minority classes. Both approaches helped mitigate the prediction bias toward the majority class, and the results showed that the weighted sampling approach outperformed the weighted cross-entropy loss method. Model performance is subject to training hyperparameters (e.g., optimizer, learning rate, batch size, and number of layers). Hence, an automatic hyperparameter-tuning tool is used, Facebook Ax (version 0.2.2, https://github.com/facebook/Ax), to search for the optimal hyperparameters for model training. The selected hyperparameters for training the model include a batch size of 16, an SGD optimizer with a learning rate of 0.008, and a momentum of 0.39, and three hidden layers in the MLP. A dropout layer follows each hidden layer in the MLP to prevent overfitting. The number of nodes in each hidden layer is 6836, 5657, and 468, respectively. The dropout rate in each dropout layer is 0.01, 0.07, and 0.67, respectively. The model was trained with four RTX A6000 GPUs. It took about 30 hr to search for the optimal hyperparameters and about an hour to train the model using the optimal hyperparameters. Statistical analysis Statistical tests were calculated to compare clinical features between blastocysts with the positive live birth outcome and blastocysts with the negative live birth outcome. Chi-squared test was used for categorical features, t test was used for numerical features. Chi-squared test and t test were performed using Python (version 3.6). ROC curves were compared by the DeLong test implemented in MedCalc software (version 20). All statistical tests were two-tailed and considered significant if p value≤0.05. Results The inclusion of patient couple’s clinical features increased AUC for live birth prediction To quantify the individual effect of blastocyst images and the combined effect of patient couple’s clinical features, we built and compared models that (1) used only blastocyst images for live birth prediction, and (2) used both blastocyst images and patient couple’s clinical features for prediction. In addition, to quantify the specific effect of endometrium status-related features (i.e., endometrium thickness before transfer, endometrium thickness on HCG day, and endometrium pattern B (yes/no) on HCG day) on live birth prediction, a third model trained using blastocyst images and patient couple’s clinical features where endometrium status-related features were excluded, was also built and compared. Figure 2 shows the ROC curves and AUCs of the three models for predicting live birth outcomes of 1758 blastocysts (i.e., 10% of 17,580) in the test data set. Using only blastocyst images for live birth prediction gave an AUC of 0.67, with a 95% confidence interval (CI) of 0.65–0.70. Using blastocyst images and patient couple’s clinical features (endometrium status-related features excluded) significantly increased the AUC to 0.74 (95% CI: 0.72–0.76, p value<0.0001). Using both blastocyst images and patient couple’s clinical features (endometrium status-related features included) achieved a prediction AUC of 0.77 (95% CI: 0.75–0.79), which is significantly higher than using only blastocyst images for prediction (p value<0.0001) and than using blastocyst images and patient couple’s clinical features where endometrium status-related features were excluded (p value=0.007). Figure 2 Download asset Open asset Receiver operating characteristic (ROC) analysis. ROC curves of the model using only blastocyst images, the model using blastocyst images and patient couple’s clinical features where EM-status related features were excluded, and the model using blastocyst images and patient couple’s clinical features where EM-status related features were included to predict live birth outcomes of 1758 blastocysts in the test data set. AUC, area under the ROC curve; EM, endometrium; EM status-related features, endometrium thickness before transfer, endometrium thickness on HCG day, endometrium pattern B (yes/no) on HCG day. Figure 2—source data 1 Code and data used to generate the ROC curves. https://cdn.elifesciences.org/articles/83662/elife-83662-fig2-data1-v2.zip Download elife-83662-fig2-data1-v2.zip Ranking the predictive power of patient couple’s clinical features We then investigated the predictive power of each patient couple’s clinical feature in predicting live birth outcome. Figure 3 shows the 16 features that were identified to be most related to the live birth outcomes of the blastocysts. These features were ranked according to their AUCs for individually predicting the live birth outcomes of blastocysts using univariable LR. The AUC for each feature was reported as the mean AUC over a tenfold cross-validation process. Figure 3 with 2 supplements see all Download asset Open asset Ranking the predictive power of patient couple’s clinical features. The 16 patient couple’s clinical features that were identified to be most related to the live birth outcomes of the blastocysts ranked by the AUC for individually predicting live birth outcome. AUC, area under the curve. Figure 3—source data 1 Code and data used to generate the AUC ranking chart. https://cdn.elifesciences.org/articles/83662/elife-83662-fig3-data1-v2.zip Download elife-83662-fig3-data1-v2.zip CNN heatmaps In blastocyst images, what does CNN focus on to predict the live birth outcome of a blastocyst? Is there a difference in what the CNN focuses on between the model trained without and with the inclusion of patient couple’s clinical features? To answer these questions, we used the class activation mapping method to generate heatmaps. Figure 4 shows blastocyst images, corresponding heatmaps of the CNN trained without including patient couple’s clinical features, and corresponding heatmaps of the CNN trained with including patient couple’s clinical features. Figure 4 with 1 supplement see all Download asset Open asset CNN heatmaps analysis. Heatmaps of the CNN trained without and with patient couple’s clinical features. Column (A): original blastocyst images. Column (B): corresponding heatmaps of the CNN trained without including patient couple’s clinical features. Column (C): corresponding heatmaps of the CNN trained with the inclusion of patient couple’s clinical features. CNN, convolutional neural network. Figure 4—source data 1 Code and data used to generate the heatmaps shown in Figure 4. https://cdn.elifesciences.org/articles/83662/elife-83662-fig4-data1-v2.zip Download elife-83662-fig4-data1-v2.zip The blastocyst images were cropped and padded to a consistent size to facilitate model training. The padding value was calculated as the mean pixel value of blastocyst images in the data set. Heatmaps were by et al., 2020). that the CNN blastocyst images as the one focusing on ICM and the other focusing on TE. The blastocyst images shown in Figure 4 are focused on ICM, and Figure supplement 1 shows the blastocyst images. As can be in Figure when trained using only blastocyst images, the CNN mainly focuses on ICM and TE for predicting live birth outcomes. training with both blastocyst images and patient couple’s clinical features, TE-related features more to live birth prediction compared with training with blastocyst images Discussion In this study, the individual effect of blastocyst images and the combined effect of patient couple’s clinical features for live birth prediction were quantified by the AUC of the model using only blastocyst images and the model using both blastocyst images and patient couple’s clinical features. AUC of was achieved with blastocyst images only using both blastocyst images and patient couple’s clinical features to a significantly higher AUC of 0.77 in live birth prediction. endometrium status-related features were excluded from patient couple’s clinical features, the AUC of the live birth prediction model significantly decreased (p from 0.77 to 0.74, indicating the of endometrium status-related features in live birth prediction. Sixteen patient couple’s clinical features were identified to be most related to live birth outcomes of blastocysts, among which maternal age, the day of blastocyst transfer, retrieved oocyte number, and endometrium thickness before transfer are the top five features contributing to live birth prediction. This study was based on a multimodal data set collected for blastocyst evaluation. The data set 17,580 blastocysts with known live birth outcomes, blastocyst images, and 16 patient couple’s clinical features. As shown in Figure 16 patient couple’s clinical features include maternal characteristics and hormone measured and on HCG day and and before transfer endometrium status-related features (endometrium thickness on HCG day and before transfer, endometrium pattern on HCG features related to retrieved oocyte the day of blastocyst number of ovarian stimulation and paternal features ratio of A sperm semen the data set by Miyagi et al., 2020 not contain endometrium status-related features and hormone (e.g., and are IVF data sets over of clinical features and live birth outcomes and 2011; et al., 2016; et al., there are no blastocyst images in these data and thus, these data sets be used for models to evaluate blastocysts from their images. To the multimodal data our proposed model was to two including a CNN and an MLP to the model to simultaneously images and numerical clinical features for blastocyst evaluation. The and multimodal data set and the proposed model in the highest AUC value of 0.77 reported thus for predicting live birth outcomes of blastocysts. They also to quantify the predictive power of each feature in predicting the live birth outcomes of blastocysts. The blastocyst grading in and remains the most common method used by to evaluate blastocyst quality the morphological grades of blastocyst development stage, ICM and TE have limited predictive power on live birth outcomes (e.g., for live birth prediction reported by Reignier et al., 2019; Bartolacci et al., 2021; Xiong et al., 2022). Since CNN a state-of-the-art method for many attempts have been made to apply the CNN to blastocyst evaluation for predicting clinical outcomes (e.g., et al., 2020; Kragh et al., 2021b; Berntsen et al., 2022; Enatsu et al., 2022; Loewke et al., 2022; Miyagi et al., 2019; Nagaya and Ukita, 2021). Among the AUC reported in the literature using blastocyst images only were around 0.65 for live birth prediction (Miyagi et al., 2019; Nagaya and Ukita, 2021). we achieved an AUC of (see Figure with the AUC of reported in the literature using grades for live birth prediction, these results that CNN can achieve a higher prediction accuracy. As shown in Figure the CNN mainly focuses on ICM and TE. Different from the TE on the number of TE cells and the of TE cells as a the CNN to focus on specific TE the heatmaps more Miyagi et al., 2020 used both blastocyst images and maternal clinical features AMH, and to predict live birth outcomes and achieved an AUC of 0.74 (Miyagi et al., 2020). The additional contribution from the three maternal clinical features was not no AUC was reported by using blastocyst images alone. Furthermore, their in pregnancy and live endometrium status-related features were not considered in their Therefore, our study used a data set and compared the AUC of live birth prediction using only blastocyst images versus using both blastocyst images and patient couple’s clinical features. We also quantified the of endometrium status-related features in with blastocyst images to improve live birth prediction. Furthermore, we that hormone such as and features related to oocyte such as and number of and the ratio of A sperm representing semen quality are able to work with blastocyst images to improve the live birth prediction accuracy. that in this study, only the total was and or was not available for clinical feature analysis (see Supplementary file 1). This may potential bias in the of as a of live birth. of this study, by the heatmaps of the CNN trained without and with the inclusion of patient couple’s clinical features, is that the weights of TE-related features increased (see Figure A potential reason may be that TE a
Read more