Abstract
Purpose
To develop and validate a predictive model for live birth (LB) outcomes in single euploid frozen embryo transfers (seFET) based on patient’s characteristics and embryo parameters.
Methods
A retrospective cohort study was performed including 1979 seFET performed between March 2017 and December 2023. Prediction models were built using logistic regression (LR), random forest classifier (RFC), support vector machines (SVM), and a gradient booster (XGBoost). Considered variables associated with LB outcomes were blastocyst expansion, blastocyst inner cell mass (ICM) and TE quality, day (D) of TE biopsy (D5, D6, and D7), female age and body mass index (BMI), distance from the uterine fundus at embryo transfer, endometrial preparation as natural cycles (NC) or hormonal replacement therapy (HRT), and endometrial thickness. Model performance was evaluated using area under the precision-recall curve and calibration metrics.
Results
Variables that were negatively associated with LB rate were BMI (OR = 0.79 [0.64–0.96], P = 0.020 for overweight and OR = 0.76 [0.60–0.95], P = 0.015 for obese class I/II), ICM grade B (OR = 0.72 [0.57–0.90], P = 0.005) or C (OR = 0.21 [0.15–0.30], P < 0.001), TE grade C (OR = 0.32 [0.24–0.43], P < 0.001), and blastocyst biopsied on D6 (OR = 0.66 [0.55–0.80], P < 0.001 or D7 (OR = 0.19[0.09–0.37], P < 0.001). The LR model was the best in terms of overall classification performance (C-statistics: 0.626 ± 0.018 vs. 0.606 ± 0.018, 0.581 ± 0.018, 0.601 ± 0.017, LR vs. RFC, XGBoost, and SVM, respectively, P < 0.001). A prediction model of LB outcome was developed and is free to access: https://artfertilityclinics.shinyapps.io/ABLE/.
Conclusion
LR demonstrated a stable validation performance and superior LB prediction, aiding as a predictive tool for patient counselling and assessing success in seFET cycles.
Supplementary Information
The online version contains supplementary material available at 10.1007/s10815-025-03524-3.
Keywords: FET cycles, Prediction model, Live birth, Euploidy, Day of biopsy, Machine learning models
Introduction
Blastocyst selection for seFET cycles represents the ultimate step for couples seeking a positive clinical outcome. It is essential to prioritize embryos with the highest implantation potential within each clinical context, and to inform patients about their predictive pregnancy chances [1]. Blastocyst selection is mainly based on morphological characteristics, and when embryos of the same morphological grade are available to be transferred, a lack of consensus on embryo selection and prioritization do exist [2, 3].
In seFET cycles, it is fundamental to inform and counsel patients regarding their potential success based on the morphological grade of the embryo selected from a cohort (if exists); however, the inter-individual variability in embryo assessment from a cohort cannot be disregarded. This results in a misleading strategy with a lack of discriminating power on which embryo should be prioritized to transfer [3]. Nevertheless, embryo morphology per se is not the only factor that should be considered when selecting an embryo in seFET cycles, since many variables can positively or negatively influence embryo implantation potential [4].
Recently, data have emerged integrating clinical outcomes into prediction models for optimizing embryo selection [5, 6]. Previous studies based on machine learning models in IVF have focused mainly on selecting the embryo with the highest morphology grade, predicting blastocyst formation, anticipating aneuploidy risks, and implantation aiming to improve IVF outcomes [7–9]. Although few studies have discerned patients’ factors in combination with embryo morphology parameters and the effect on implantation outcomes in seFET cycles [10–13], robust predictive models of LB outcomes to forecast patients about their likelihood clinical chances a priori are still lacking.
In an effort to increase clinical outcomes in seFET cycles, developing an accurate individualized embryo selection approach when a pool of embryos is available in seFET cycles is paramount. This is crucial to provide a clear understanding of the likelihood of success and, therefore, managing expectations. Predictive models typically incorporate several factors such as female age and/or BMI, causes of infertility, or infertility duration [14, 15]. Albeit, the performance of predictive models can vary significantly, with some showing a tendency to overfit data training or perform poorly in validation cohorts [16]. Despite these efforts, a more robust and well-calibrated model that can reliably predict outcomes across diverse patient populations is needed. This study aimed to develop and validate a predictive model for LB outcomes to estimate the likelihood of success for patient’s counselling. Therefore, the effectiveness of different predictive models incorporating key IVF parameters were evaluated to identify the most reliable approach for this purpose.
Materials and methods
Ethical approval
Approval for this study was obtained from the local Research Ethics Committee (REFA102). All couples consented to use their cycle information for retrospective data analysis.
Study design
All patients undergoing autologous seFET cycles between March 2017 and December 2023 were considered for the study. All blastocysts with a quality ≥ BL3 CC (Gardner) that underwent TE biopsy on D5, D6, or D7 analyzed by next-generation sequencing (NGS) for PGT-A and were utilized in a subsequent seFET cycle in a NC or HRT cycle were considered for analysis.
The aim of the study was to develop and validate the most accurate predictive model for LB outcomes in seFET cycles based on blastocyst’s morphology, day of TE biopsy, and patient’s characteristics. Female age and male age, female BMI, anti-Mullerian hormone (AMH), infertility type (primary and secondary), duration of infertility (years), endometrial preparation (EP) protocols (NC vs HRT cycles), distance from the uterus fundus after embryo transfer (mm), blastocyst’s grade of expansion, ICM and TE quality, and day of TE biopsy were all variables considered with potential significance based on the literature already published and clinical plausibility. Therefore, different prediction models were built using multiple approaches including LR, RFC, SVM, and a XGBoost to compare which one have superior scores predicting LB outcomes. The predictive model that performed best was subsequently subjected to further performance testing with a final static fourfold cross-validation.
Blastocyst grading
Blastocyst grade of expansion and ICM and TE quality were assessed immediately before TE biopsy and graded accordingly to a modified Gardner scoring [17]. Blastocysts were evaluated based on the hours post insemination (hpi) on D5 (120 ± 4 hpi) of culture and if quality was deemed suboptimal and expansion or number of cells present in the TE was insufficient to perform TE biopsy accordingly to the routine clinical practice (< BL3 CC), blastocysts were left in culture to be re-evaluated on D6 (144 ± 4 hpi). The final day of blastocyst assessment to perform TE biopsy was on D7 (168 ± 4 hpi) and blastocysts graded < BL3 CC were not eligible for biopsy and were subsequently discarded. Embryo grading was performed by senior embryologists by pairs to reduce the inter-variability. Additionally, internal quality control is performed monthly according to the standard operation procedures (SOP) to ensure embryo grading is assessed accurately across all embryologists.
Embryo biopsy and tubing procedure
All blastocysts with sufficient quality (≥ BL3 CC) available on D5, D6, or D7 were subjected to TE biopsy for PGT-A analysis. A biopsy pipette (Origio, CooperSurgical) possessing an internal diameter of 30 µm was utilized to aspirate three to ten TE cells. Immediately before performing TE biopsy, a drilling was performed on the zona pellucida with a laser and the TE cells were taken and gently loosened through the application of a 2.2-ms intensity laser pulse, followed by a completed detachment using a mechanical flicking technique. TE biopsies were washed, placed in 0.2-mL Eppendorf PCR tubes containing 2.5 µL phosphate-buffered saline, and stored at − 20 °C until further processing for PGT-A analysis.
Preimplantation genetic testing for aneuploidies of TE samples
The genetic laboratory used the Ion ReproseqTM PGS kit for PGT-A, which is based on next-generation sequencing (NGS) [Thermo Fisher Scientific, USA] to analyze the biopsied TE cells. The test was validated for 24-chromosome aneuploidy screening on the human genome. The genetic material of the biopsied cells was extracted and whole genome was amplified and sequenced using the Ion Chef™ and Ion GeneStudio S5 Prime System (Thermo Fisher Scientific, Inc., MA, USA). Data analysis was performed with Ion Reporter software, which aligns the reads using the human genome build (hg19) (Thermo Fisher Scientific, USA). The workflow of the PGT-A analysis was based on a transition penalty [TP] of − 3 (default high sensitivity) and a baseline tile size of 2 Mbp for consistently calling sub-chromosomal copy number of variation (CNV) events. According to the internal validation conducted by the genetic laboratory, the Ion ReproSeq™ PGS Kit demonstrates a high sensitivity of over 99% for detecting whole chromosome and sub-chromosome aneuploidies with a resolution of 10 Mb. This high sensitivity ensures that the test can reliably detect aneuploid embryos, thereby minimizing the risk of false negatives. The specificity of the Ion ReproSeq™ PGS Kit for detecting these aneuploidies is 99%. This high specificity ensures that the test can accurately identify euploid embryos, reducing the occurrence of false positives. Mosaic embryos were not considered for this study (neither low/high mosaic embryos).
Type of EP protocols and embryo transfer (ET) procedure
The EP protocol (NC or HRT cycles) was individually chosen according to physician’s discretion and patients’ characteristics. In NC, supplementation of 100 mg of progesterone commenced in the evening of the ovulation day, followed by three times daily until pregnancy test.
In HRT cycles, patients were supplemented with 4 mg estradiol until day 2 and increased to 6 mg until endometrial lining with a trilaminar appearance was seen [18]. Thereafter, daily progesterone with starting dose of 100 mg was administered vaginally (day 0), which was increased on day 1 to three times daily and continued until pregnancy test.
Embryo selection for transfer was based on blastocysts morphological characteristics and day of TE biopsy accordingly with the pool of embryos available from each patient. All seFET cycles were performed after five full days of P4 administration was given regardless of the day of blastocyst biopsy (D5, D6, or D7) and all ETs were performed under abdominal ultrasound guidance using a soft pass catheter (GUARDIATM AccessET® Catheter, Cook® Medical, USA).
Data collection and statistical analysis
Patient’s characteristics are presented by median and inter-quartile (IQR) ranges for continuous variables and frequencies and percentages for categorical variables. To analyze the association of patient and embryo characteristics with LB outcomes, regression analyses were performed using multilevel logistics with random intercepts for transfers of the same couple. The variables considered were female and male age at time of oocyte retrieval, female BMI defined as normal (18.5–24.9 kg/m2), overweight (25–29.9 kg/m2), and obese class I/II (> 30 kg/m2), AMH, infertility type (primary and secondary), duration of infertility (years), EP protocol (NC vs HRT cycles), distance from the uterus fundus after embryo transfer (mm), blastocyst grade of expansion (BL3-BL6), ICM and TE quality (A, B, or C), and day of TE biopsy (D5, D6, or D7). These metrics included the most popular patient and embryo level variables that are commonly obtained in IVF treatments and those which are often associated with a successful outcome.
Prediction models were built using multiple approaches, including LR, RFC, SVM, and XGBoost [19, 20]. Sample size estimation showed 1150 cycles would be enough to create a 10-parameter logistic regression model with a Snell-Cox R2 of 0.075, shrinkage of 0.9, and event prevalence of 50% [21]. It was deemed a 60%:40% splitting of the patients would consistently allow the adequate number of cycles in each repeated training set. Model training and validation were done with near-exhaustive repeated cross-validation (60:40), which was repeated 2000 times. Dataset split was made based on random allocation of patients rather than cycles so that the model was trained and tested on different patients. Tuning parameters were selected to achieve stable overall performance among each paired training and validation set, preventing overfitting (low learning rate, limiting number of trees, nodes, and tree depth) and between repeated samplings. Considered performance metrics at large were C-statistics, C-statistics shrinkage between training and validation, calibration intercept, and calibration slope. All models were compared against each other, and the best-performing one was subjected to further performance testing with a final static fourfold cross-validation. Models were built with an aim to create the most parsimonious model without sacrificing discrimination or calibration performance. All statistically significant variables in the regression analysis were considered initially and then were removed 1 by 1 to reach the most parsimonious model that does not sacrifice performance. Assessed metrics were area under the precision-recall curve (AUPRC) and area under the receiver operating curve (AUROC). An online application was deployed for external use and further validation. Using calibration plot as guidance, probabilistic estimates were categorized to different prognostic groups and outcomes were compared between the prognostic categories. All analyses were performed with R for Statistical Computing Software. Group comparisons were performed using the Mann–Whitney U or chi-squared test. For multiple group comparisons, the Kruskal–Wallis test was used with Bonferroni correction.
Results
Patient characteristics
A total of 1979 seFET cycles of 1459 couples were included in the analysis. The characteristics of couples included in the analysis are shown in Table 1. Those patients who underwent only one seFET cycle were significantly older (P < 0.041) and had lower AMH levels (median: 2.2 vs. 2.5 ng/mL, P = 0.001). No significant differences were observed in terms of BMI, infertility type, or duration of infertility (P = 0.561, 0.430, and 0.418, respectively) (Table 1).
Table 1.
General characteristics of patients who underwent single and multiple cycles of seFET
| Variable | Single seFET | Multiple seFET | P value |
|---|---|---|---|
| Female age | |||
| < 35 | 568 (52.4) | 201 (53.7) | 0.041 |
| 35–37 | 174 (16.0) | 76 (20.3) | |
| 38–40 | 196 (18.1) | 64 (17.1) | |
| > 40 | 147 (13.5) | 33 (8.8) | |
| BMI (kg/m2) | |||
| Normal weight | 419 (38.6) | 146 (39.0) | 0.561 |
| Overweight | 388 (35.8) | 142 (38.0) | |
| Obese class I or II | 278 (25.6) | 86 (23.0) | |
| AMH (ng/mL) | 2.2 (1.1 to 3.8) | 2.5 (1.5 to 4.5) | 0.001 |
| Infertility type | |||
| Primary | 486 (44.8) | 177 (47.3) | 0.430 |
| Secondary | 599 (55.2) | 197 (52.7) | |
| Duration of infertility (years) | 2.0 (1.0 to 4.0) | 2.0 (1.0 to 4.0) | 0.418 |
Values are expressed as total numbers (percentage) for categorical variables or median (range) for continuous variables
Univariable and multivariable logistic regression analyses
After performing univariable mixed-effect logistic regression analyses, results showed that BMI categories, EP protocols, blastocyst expansion, ICM quality, TE quality, day of biopsy, and endometrial thickness were all significantly associated with LB outcome (Table 2). Based on multivariable analysis, except for blastocysts expansion, all these variables remained significant to predict LB outcomes; therefore, they were used in the predictive models (Table 2). In the final model, the parameters that were ranked most important to least important for achieving a LB outcome were BMI, ICM quality, TE quality, day of TE biopsy, and type of EP protocol.
Table 2.
Mixed logistic regression analysis including factors associated with live birth in seFET cycles
| Variables | Live birth no | Live birth yes | OR (95% CI, P value) | aOR (95% CI, P value) |
|---|---|---|---|---|
| Female age category (years old) | ||||
| < 35 | 539 (50.1) | 536 (49.9) | Reference | - |
| 35–37 | 173 (48.9) | 181 (51.1) | 1.05 (0.83–1.34, P = 0.679) | - |
| 38–40 | 159 (48.0) | 172 (52.0) | 1.09 (0.85–1.39, P = 0.503) | - |
| > 40 | 106 (48.4) | 113 (51.6) | 1.07 (0.80–1.43, P = 0.639) | - |
| BMI (kg/m2) | ||||
| Normal weight | 349 (45.4) | 419 (54.6) | Reference | Reference |
| Overweight | 367 (51.5) | 346 (48.5) | 0.79 (0.64–0.96, P = 0.020) | 0.76 (0.60–0.95, P = 0.017) |
| Obese class I or II | 261 (52.4) | 237 (47.6) | 0.76 (0.60–0.95, P = 0.015) | 0.75 (0.58–0.96, P = 0.025) |
| AMH (ng/mL) | 3.3 (3.1) | 3.2 (3.1) | 0.99 (0.96–1.02, P = 0.530) | - |
| Male age | 37.1 (7.0) | 37.5 (7.0) | 1.01 (1.00–1.02, P = 0.194) | - |
| Infertility type | ||||
| Primary | 451 (48.8) | 473 (51.2) | Reference | - |
| Secondary | 526 (49.9) | 529 (50.1) | 0.96 (0.80–1.14, P = 0.642) | - |
| Duration of infertility | 3.1 (2.7) | 3.1 (2.9) | 1.00 (0.97–1.03, P = 0.976) | - |
| Cycle preparation | ||||
| HRT | 598 (52.5) | 541 (47.5) | Reference | Reference |
| NC | 379 (45.1) | 461 (54.9) | 1.34 (1.12–1.61, P = 0.001) | 1.33 (1.08–1.62, P = 0.006) |
| BL-ICM quality | ||||
| A | 147 (39.1) | 229 (60.9) | Reference | Reference |
| B | 635 (47.3) | 708 (52.7) | 0.72 (0.57–0.90, P = 0.005) | 0.83 (0.64–1.08, P = 0.167) |
| C | 195 (75.0) | 65 (25.0) | 0.21 (0.15–0.30, P < 0.001) | 0.34 (0.22–0.53, P < 0.001) |
| BL-TE quality | ||||
| A | 161 (41.5) | 227 (58.5) | Reference | Reference |
| B | 577 (46.4) | 667 (53.6) | 0.82 (0.65–1.03, P = 0.091) | 0.86 (0.66–1.12, P = 0.260) |
| C | 239 (68.9) | 108 (31.1) | 0.32 (0.24–0.43, P < 0.001) | 0.56 (0.38–0.83, P = 0.004) |
| BL-Expansion | ||||
| BL3 | 163 (58.4) | 116 (41.6) | Reference | Reference |
| BL4 | 349 (48.0) | 378 (52.0) | 1.52 (1.15–2.01, P = 0.003) | 1.16 (0.85–1.59, P = 0.351) |
| BL5 | 447 (47.2) | 501 (52.8) | 1.57 (1.20–2.07, P = 0.001) | 1.21 (0.89–1.65, P = 0.227) |
| BL6 | 18 (72.0) | 7 (28.0) | 0.55 (0.21–1.30, P = 0.191) | 0.43 (0.16–1.14, P = 0.089) |
| Day of biopsy | ||||
| 5 | 478 (44.1) | 606 (55.9) | Reference | Reference |
| 6 | 458 (54.3) | 386 (45.7) | 0.66 (0.55–0.80, P < 0.001) | 0.76 (0.62–0.94, P = 0.013) |
| 7 | 41 (80.4) | 10 (19.6) | 0.19 (0.09–0.37, P < 0.001) | 0.33 (0.15–0.72, P = 0.005) |
| Endometrial thickness, mm | 7.8 (1.5) | 8.0 (1.7) | 1.08 (1.02–1.15, P = 0.005) | 1.10 (1.03–1.17, P = 0.004) |
| Distance from fundus, mm | 7.1 (2.7) | 6.9 (2.5) | 0.97 (0.94–1.00, P = 0.074) | 0.97 (0.93–1.00, P = 0.079) |
Values are expressed as total numbers (percentage) for categorical variables or median (range) for continuous variables. Mixed-effects model using random intercepts for the repeated cycle of the same patient. OR odds ratio, aOR adjusted odds ratio, BMI body mass index, AMH anti-Mullerian hormone, HRT hormone replacement therapy, NC natural cycle, ICM inner cell mass, TE trophectoderm, BL blastocyst
Different predictive models to assess live birth outcome
In 2000 repeated validation samples, the LR model was the best in terms of overall classification performance (C-statistics: 0.626 ± 0.018 vs. 0.606 ± 0.018, 0.581 ± 0.018, and 0.601 ± 0.017, for LR vs. RFC, SVM, and XGBoost, respectively, P < 0.001) (Fig. 1, Supplemental Table 1). Moreover, RFC, SVM, and XGBoost models substantially overfitted to the training set as expected and were marked by significantly higher C-statistic shrinkage compared to the LR model (P < 0.001, Fig. 1, Supplemental Table 1). Additionally, the LR model showed the best calibration performance in the validation samples (P < 0.001, Fig. 1, Supplemental Table 1). Deemed superior, the final model fitted with LR and was subjected to static fourfold cross-validation. In the fourfold cross-validation, the LR model had a mean AUPRC of 0.62 and a mean AUROC of 0.63 (Fig. 2). The LR model was well-calibrated in all cross-validation samples (Supplemental Fig. 1). Based on the distribution and clumping of predicted probabilities on LB outcome, four prognosis categories according to the predictive probability were created (poor, fair, good, and top) (Table 3).
Fig. 1.
Comparison of predictive models to assess live birth. Performance comparison of different tabular classifiers as predictive models to assess live birth outcome in terms of C-statistics, shrinkage, calibration intercept, and slope. Solid vertical lines indicate the optimal value for that metric
Fig. 2.
Precision-recall and receiver operating characteristics curves of the final model in fourfold cross-validation simples. The mean area under the precision-recall curve was 0.62. The mean area under the ROC curve was 0.63
Table 3.
Population and transfer characteristics of prognostic groups predicted by the logistic regression model
| Prognostic groups based on the model | |||||
|---|---|---|---|---|---|
| Variables | Poor | Fair | Good | Top | P value |
| Live birth | |||||
| No | 203 (75.2) | 253 (54.3) | 473 (43.8) | 48 (29.6) | < 0.001 |
| Yes | 67 (24.8) | 213 (45.7) | 608 (56.2) | 114 (70.4) | |
| Female age categories (years old) | |||||
| < 35 | 131 (48.5) | 240 (51.5) | 608 (56.2) | 96 (59.3) | 0.034 |
| 35–37 | 45 (16.7) | 90 (19.3) | 189 (17.5) | 30 (18.5) | |
| 38–40 | 57 (21.1) | 71 (15.2) | 178 (16.5) | 25 (15.4) | |
| > 40 | 37 (13.7) | 65 (13.9) | 106 (9.8) | 11 (6.8) | |
| BMI categories (kg/m2) | |||||
| Normal weight | 98 (36.3) | 103 (22.1) | 469 (43.4) | 98 (60.5) | < 0.001 |
| Overweight | 99 (36.7) | 150 (32.2) | 411 (38.0) | 53 (32.7) | |
| Obese class I or II | 73 (27.0) | 213 (45.7) | 201 (18.6) | 11 (6.8) | |
| AMH (ng/mL) | 2.0 (1.1 to 3.6) | 2.3 (1.3 to 4.1) | 2.6 (1.3 to 4.4) | 2.2 (1.4 to 3.7) | 0.011 |
| Male age | 38.0 (34.0 to 43.0) | 37.0 (32.0 to 42.0) | 37.0 (32.0 to 41.0) | 36.0 (33.0 to 40.8) | 0.014 |
| Infertility type | |||||
| Primary | 112 (41.5) | 211 (45.3) | 529 (48.9) | 72 (44.4) | 0.120 |
| Secondary | 158 (58.5) | 255 (54.7) | 552 (51.1) | 90 (55.6) | |
| Duration of infertility | 3.0 (1.0 to 4.0) | 3.0 (1.0 to 5.0) | 2.0 (1.0 to 4.0) | 2.0 (1.0 to 3.0) | < 0.001 |
| Cycle preparation | |||||
| HRT | 179 (66.3) | 359 (77.0) | 588 (54.4) | 13 (8.0) | < 0.001 |
| NC | 91 (33.7) | 107 (23.0) | 493 (45.6) | 149 (92.0) | |
| BL-ICM quality | |||||
| A | 1 (0.4) | 20 (4.3) | 251 (23.2) | 104 (64.2) | < 0.001 |
| B | 39 (14.4) | 416 (89.3) | 830 (76.8) | 58 (35.8) | |
| C | 230 (85.2) | 30 (6.4) | 0 (0.0) | 0 (0.0) | |
| BL-TE quality | |||||
| A | 2 (0.7) | 42 (9.0) | 280 (25.9) | 64 (39.5) | < 0.001 |
| B | 60 (22.2) | 302 (64.8) | 784 (72.5) | 98 (60.5) | |
| C | 208 (77.0) | 122 (26.2) | 17 (1.6) | 0 (0.0) | |
| BL-Expansion | |||||
| BL3 | 79 (29.3) | 69 (14.8) | 115 (10.6) | 16 (9.9) | < 0.001 |
| BL4 | 86 (31.9) | 154 (33.0) | 422 (39.0) | 65 (40.1) | |
| BL5 | 104 (38.5) | 231 (49.6) | 533 (49.3) | 80 (49.4) | |
| BL6 | 1 (0.4) | 12 (2.6) | 11 (1.0) | 1 (0.6) | |
| Day of biopsy | |||||
| 5 | 43 (15.9) | 171 (36.7) | 726 (67.2) | 144 (88.9) | < 0.001 |
| 6 | 181 (67.0) | 290 (62.2) | 355 (32.8) | 18 (11.1) | |
| 7 | 46 (17.0) | 5 (1.1) | 0 (0.0) | 0 (0.0) | |
| Endometrial thickness, mm | 7.6 (6.8 to 8.6) | 7.3 (6.7 to 8.1) | 7.8 (7.1 to 8.8) | 9.6 (7.9 to 10.8) | < 0.001 |
| Distance from fundus, mm | 6.0 (5.0 to 8.0) | 6.0 (5.0 to 8.0) | 6.0 (5.0 to 8.0) | 6.0 (5.0 to 7.0) | 0.505 |
Values are expressed as total numbers (percentage) for categorical variables or median (range) for continuous variables. BMI body mass index, AMH anti-Mullerian hormone, HRT hormone replacement therapy, NC natural cycle, ICM inner cell mass, TE trophectoderm, BL blastocyst
Predictive model based on logistic regression model and prognosis categories
The LB outcomes in poor, fair, good, and top categories were 24.8%, 45.7%, 56.2%, and 70.4%, respectively (Table 3). Women in good and top prognosis categories were younger (P = 0.034), leaner in terms of BMI (P < 0.001), and had shorter infertility duration and ETs were performed in a NC more often than HRT cycles (P < 0.001) (Table 3). The ICM and TE quality of the transferred embryos was higher and had earlier TE biopsy days (D5) in women included in the good and top prognosis categories (P < 0.001 for both, Table 3). While endometrial thickness was similar in the first three prognosis groups (median: 7.6 mm, 7.3 mm, and 7.8 mm, poor, fair, and good groups, respectively), women included in the top prognosis category had significantly thicker endometrium compared to the rest of the groups (median: 9.6 mm, P < 0.001). Based on this analysis performed, a prediction model of LB outcome was developed and deployed an easy to access web application named ABLE (Assessing Birth Likelihood for Euploid Blastocysts) and it is free to access and use and for external validation at https://artfertilityclinics.shinyapps.io/ABLE/.
Discussion
Embryo selection represents the ultimate critical step in IVF for a successful outcome. Many factors can be negatively associated with seFET outcomes regardless of whether good-quality blastocysts are being transferred. Transparent communication about the selected embryos’ potential to be transferred based on machine learning models would strengthen patient involvement, build trust, and foster supportive fertility care experiences. Nevertheless, existing prediction models based on clinical outcomes that incorporate solely euploid embryos while integrating patient and cycle characteristics are still lacking. Therefore, this work aimed to develop and validate a predictive model for LB outcomes in seFET cycles integrating blastocyst morphology, day of biopsy, and patient characteristics, which can provide a more accurate clinical prediction outcome. Results from this study demonstrated that when these multiple variables were analyzed altogether, the LR model emerged as the most effective predictor in this clinical data setting. The model presented in this study can serve as a valuable tool for assessing embryo implantation potential and subsequently LB outcomes in seFET cycles.
Embryo evaluation and selection for seFET cycles are mainly based on static embryo morphology, time-lapse morpho kinetics, and/or euploidy outcomes [22, 23]. Nevertheless, predicting implantation is still a challenging process. Embryo selection based solely on morphological assessment results in higher inter- and intra-operator variability and inconsistency [24]. Therefore, algorithm guided by artificial intelligence is a promising tool to be integrated into embryo selection and pregnancy prediction outcomes [25–28].
While many studies have focused only on embryo quality and biopsy day for predicting implantation potentials [1, 29–31], the present study combined multiple variables to predict the LB likelihood considering the pool of euploid embryos available for transfer. The model’s categorization provides a more personalized estimate of each patient’s chances of success. Taken as example the extreme cases, patients in the “top” category, who had factors such as a thicker endometrium, higher-quality embryos, and undergo natural cycle EP, are more likely to have a chance for live birth rates above 70%. In contrast, patients in the “poor” category had much lower chances of success. By breaking down complex data into easy-to-understand intuitive categories with clearly defined success rates, reproductive specialists can set more realistic expectations to patients and support informed decision-making.
Furthermore, the developed prediction model can be easily accessible online and be used as a tool when selecting an embryo to transfer and to inform patients regarding their clinical chances based on the embryo selection made. The present model confirmed that BMI, blastocysts quality, day of TE biopsy, type EP protocol, and endometrial thickness are highly associated with a positive clinical outcome as already been published [18, 32–35]. Yet, none of these variables has been previously analyzed altogether, as presented in the current study. A recent review [36] on clinical prediction models based on national data from the USA and the UK has shown that female age, the duration of infertility, female BMI, and AMH are important pre-treatment predictors, while the number of eggs collected adds important value in IVF prediction models. However, none of these studies included seFET cycles, making this study a valuable contribution to the existing literature. Female age was not included as a predictor factor in the model developed by this study since it did not show any association with LB outcome in the multivariate analysis, reinforcing the findings reported by Lawrenz et al. [37] where it is shown that LB rate is similar when euploid blastocysts are transferred across all female ages from ≤ 35 to > 40 years old. Moreover, the integration of euploidy with the day of TE biopsy and blastocysts quality confirms the importance and originality of the present study. It is important to emphasize that while BMI was a statistically significant negative predictor of live birth, its clinical impact appears modest compared to key embryo–related factors such as ICM and TE grades or the timing (day) of biopsy. However, BMI may contribute to cumulative risk when combined with other suboptimal factors. Our model’s online calculator is designed to help clinicians contextualize this effect by illustrating how BMI interacts with other patient and embryo characteristics to influence the overall likelihood of success. The developed application available online (https://artfertilityclinics.shinyapps.io/ABLE/) can be a useful tool for clinicians to discuss the possibilities of success of the treatment when a seFET cycle is planned, and by embryologist to select the best embryo available from a pool of euploid blastocysts (if exists) at the time of embryo warming.
According to the existing literature, various predictive models for LB outcomes are available [38–40]. Most of these models incorporate similar variables to those implemented to develop the online app ABLE presented in this study, with the notable exception of euploid embryos, which enhances the robustness of our predictive model. Previous studies cited have typically relied on a single predictive method; nevertheless, the present study undertakes a comprehensive comparison of multiple predictive models and has demonstrated that the LR model outperformed other techniques, such as RFC, SVM, and XGBoost, for predicting LB outcomes. It is worth noting that this should not be construed as the definitive best model overall. Depending on the clinical use case, most of these models are expected to perform similarly. By including multiple algorithms, we aimed to illustrate that performance differences can be subtle when models are appropriately applied. This adds valuable information to the current body of knowledge, as few studies have systematically compared these methodologies in the context of embryo transfer outcomes [10, 41, 42].
Furthermore, the present study is unique in providing not only a comparison of different models but also an online application that allows for real-time predictions of LB outcomes. By integrating the most relevant clinical and embryological variables for seFET cycles, the app serves as a practical resource for clinicians. The accessibility of this online tool facilitates external validation, provided that similar clinical parameters are considered when selecting a single euploid embryo for transfer in a deferred cycle. This promotes broader applicability and helps establish the robustness of the model in diverse clinical settings. The prediction model can be intentionally used for patient counselling and planning with the purpose of managing expectations and informing probabilities of success with the embryo/s already available for a seFET cycle.
Limitations of the present study include that the developed prediction model should be validated with an external cohort in different geographical locations. Moreover, the overall strength of the logistic regression model, which is the most predictive but still has a modest predictive value (AUC 0.63), needs to be acknowledged. Therefore, it is important to recognize that predicting LB outcomes after seFET cycles is still very challenging, and additional work investigating beyond the most considered variables is needed. Extrapolation of these results to other clinical practices or population with great accurate predictions in patients undergoing different treatments than the one analyzed still needs to be elucidated. Future work could involve external validation in collaboration with other centers to further strengthen the model’s applicability across diverse settings.
In conclusion, this study remarks the potential of a model developed that predicts LB outcomes based on patients and embryos characteristics in seFET cycles which can be accessible online. In addition, this study highlights the importance of comparing different prediction models for LB outcomes and which variables should be taken into consideration in seFET cycles. To our knowledge, this is the first online model developed which can aid a real benefit for patient counselling regarding their likelihood of LB, offering a clearer assessment of success for patients undergoing seFET cycles.
Supplementary Information
Below is the link to the electronic supplementary material.
Author contribution
AA, DN, and EK conceived and designed the research, curated the data, and wrote the original draft preparation. EK performed the data analysis and built and validated the model. DN reviewed and edited the final version of the manuscript. AA, IE, AB, LM, and HF contributed to the methodology. All authors read and approved the manuscript.
Data availability
Data regarding any of the subjects in the study has not been previously published unless specified. Data will be made available to the editors of the journal for review or query upon request.
Declarations
Ethical approval
Approval for this study was obtained from the local Research Ethics Committee (REFA102). All couples consented to use their cycle information for retrospective data analysis.
Conflict of interest
The authors declare no competing interests.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Irani M, Reichman D, Robles A, Melnick A, Davis O, Zaninovic N, Xu K, Rosenwaks Z. Morphologic grading of euploid blastocysts influences implantation and ongoing pregnancy rates. Fertil Steril. 2017;107(3):664–70. 10.1016/j.fertnstert.2016.11.012. [DOI] [PubMed] [Google Scholar]
- 2.Tiitinen A. Single embryo transfer: why and how to identify the embryo with the best developmental potential. Best Pract Res Clin Endocrinol Metab. 2019;33(1):77–88. 10.1016/j.beem.2019.04.001. [DOI] [PubMed] [Google Scholar]
- 3.Widra EA. Where next for improved embryo selection? Fertil Steril. 2023;120(2):217. 10.1016/j.fertnstert.2023.06.024. [DOI] [PubMed] [Google Scholar]
- 4.Cimadomo D, Rienzi L, Conforti A, Forman E, Canosa S, Innocenti F, Poli M, Hynes J, Gemmell L, Vaiarelli A, Alviggi C, Ubaldi FM, Capalbo A. Opening the black box: why do euploid blastocysts fail to implant? A systematic review and meta-analysis. Hum Reprod Update. 2023;29(5):570–633. 10.1093/humupd/dmad010. [DOI] [PubMed] [Google Scholar]
- 5.Charnpinyo N, Suthicharoenpanich K, Onthuam K, Engphaiboon S, Chaichaowarat R, Suebthawinkul C, Siricharoen P. Embryo selection for IVF using machine learning techniques based on light microscopic images of embryo and additional factors. Annu Int Conf IEEE Eng Med Biol Soc. 2023;2023:1–4. 10.1109/EMBC40787.2023.10340767. [DOI] [PubMed] [Google Scholar]
- 6.Morales DA, Bengoetxea E, Larrañaga P. Selection of human embryos for transfer by Bayesian classifiers. Comput Biol Med. 2008;38(11–12):1177–86. 10.1016/j.compbiomed.2008.09.002. [DOI] [PubMed] [Google Scholar]
- 7.van Loendersloot LL, van Wely M, Limpens J, Bossuyt PM, Repping S, van der Veen F. Predictive factors in in vitro fertilization (IVF): a systematic review and meta-analysis. Hum Reprod Update. 2010;16(6):577–89. 10.1093/humupd/dmq015. [DOI] [PubMed] [Google Scholar]
- 8.Afnan M, Afnan MAM, Liu Y, Savulescu J, Mishra A, Conitzer V, Rudin C. Data solidarity for machine learning for embryo selection: a call for the creation of an open access repository of embryo data. Reprod Biomed Online. 2022;45(1):10–3. 10.1016/j.rbmo.2022.03.015. [DOI] [PubMed] [Google Scholar]
- 9.Corani G, Magli C, Giusti A, Gianaroli L, Gambardella LM. A Bayesian network model for predicting pregnancy after in vitro fertilization. Comput Biol Med. 2013;43(11):1783–92. 10.1016/j.compbiomed.2013.07.035. [DOI] [PubMed] [Google Scholar]
- 10.Bamford T, Smith R, Easter C, Dhillon-Smith R, Barrie A, Montgomery S, Campbell A, Coomarasamy A. Association between a morphokinetic ploidy prediction model risk score and miscarriage and live birth: a multicentre cohort study. Fertil Steril. 2023;120(4):834–43. 10.1016/j.fertnstert.2023.06.006. [DOI] [PubMed] [Google Scholar]
- 11.Motato Y, de los Santos MJ, Escriba MJ, Ruiz BA, Remohí J, Meseguer M. Morphokinetic analysis and embryonic prediction for blastocyst formation through an integrated time-lapse system. Fertil Steril. 2016;105(2):376-84.e9. 10.1016/j.fertnstert.2015.11.001. [DOI] [PubMed] [Google Scholar]
- 12.Abdala A, Elkhatib I, Bayram A, Arnanz A, El-Damen A, Melado L, Lawrenz B, Fatemi HM, De Munck N. Day 5 vs day 6 single euploid blastocyst frozen embryo transfers: which variables do have an impact on the clinical pregnancy rates? J Assist Reprod Genet. 2022;39(2):379–88. 10.1007/s10815-021-02380-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Melado Vidales L, Lawrenz B, Vitorino RL, Patel R, Ruiz FJ, Marques LM, Bayram A, Elkhatib I, Fatemi H. Clinical and laboratory parameters associated with cycle outcomes in patients undergoing euploid frozen blastocyst transfer. Reprod Biomed Online. 2023;46(6):917–25. 10.1016/j.rbmo.2023.02.014. [DOI] [PubMed] [Google Scholar]
- 14.Bardet L, Excoffier JB, Salaun-Penquer N, Ortala M, Pasquier M, Mathieu d’Argent E, Massin N. Comparison of predictive models for cumulative live birth rate after treatment with ART. Reprod Biomed Online. 2022;45(2):246–55. 10.1016/j.rbmo.2022.03.020. [DOI] [PubMed] [Google Scholar]
- 15.Shingshetty L, Cameron NJ, Mclernon DJ, Bhattacharya S. Predictors of success after in vitro fertilization. Fertil Steril. 2024;121(5):742–51. 10.1016/j.fertnstert.2024.03.003. [DOI] [PubMed] [Google Scholar]
- 16.Ratna MB, Bhattacharya S, McLernon DJ. External validation of models for predicting cumulative live birth over multiple complete cycles of IVF treatment. Hum Reprod. 2023;38(10):1998–2010. 10.1093/humrep/dead165. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Gardner DK, Schoolcraft WB. In vitro culture of human blastocyst. In: Jansen R, Mortimer D, editors. Towards reproductive certainty: infertility and genetics beyond 1999. Parthenon Press; 1999. p. 378–88. [Google Scholar]
- 18.Ata B, Liñán A, Kalafat E, Ruíz F, Melado L, Bayram A, Elkhatib I, Lawrenz B, Fatemi HM. Effect of the endometrial thickness on the live birth rate: insights from 959 single euploid frozen embryo transfers without a cutoff for thickness. Fertil Steril. 2023;120(1):91–8. 10.1016/j.fertnstert.2023.02.035. [DOI] [PubMed] [Google Scholar]
- 19.Chen T, Guestrin C (2016) XGBoost: a scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '16). Association for Computing Machinery, USA, 785–794. 10.1145/2939672.2939785.
- 20.Kowadlo G, Mittelberg Y, Ghomlaghi M, Stiglitz DK, Kishore K, Guha R, Nazareth J, Weinberg L. Development and validation of ‘Patient Optimizer’ (POP) algorithms for predicting surgical risk with machine learning. BMC Med Inform Decis Mak. 2024;24(1):70. 10.1186/s12911-024-02463-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Riley RD, Snell KI, Ensor J, Burke DL, Harrell FE Jr, Moons KG, Collins GS. Minimum sample size for developing a multivariable prediction model: PART II - binary and time-to-event outcomes. Stat Med. 2019;38(7):1276–96. 10.1002/sim.7992. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.ALPHA Scientists in Reproductive Medicine; ESHRE Special Interest Group Embryology. Istanbul consensus workshop on embryo assessment: proceedings of an expert meeting. Reprod Biomed Online. 2011;22(6):632–46. 10.1016/j.rbmo.2011.02.001. [DOI] [PubMed] [Google Scholar]
- 23.Zaninovic N, Irani M, Meseguer M. Assessment of embryo morphology and developmental dynamics by time-lapse microscopy: is there a relation to implantation and ploidy? Fertil Steril. 2017;108(5):722–9. 10.1016/j.fertnstert.2017.10.002. [DOI] [PubMed] [Google Scholar]
- 24.Storr A, Venetis CA, Cooke S, Kilani S, Ledger W. Inter-observer and intra-observer agreement between embryologists during selection of a single Day 5 embryo for transfer: a multicenter study. Hum Reprod. 2017;32(2):307–14. 10.1093/humrep/dew330. [DOI] [PubMed] [Google Scholar]
- 25.Berntsen J, Rimestad J, Lassen JT, Tran D, Kragh MF. Robust and generalizable embryo selection based on artificial intelligence and time-lapse image sequences. PLoS ONE. 2022;17(2):e0262661. 10.1371/journal.pone.0262661. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Kragh MF, Karstoft H. Embryo selection with artificial intelligence: how to evaluate and compare methods? J Assist Reprod Genet. 2021;38(7):1675–89. 10.1007/s10815-021-02254-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Duval A, Nogueira D, Dissler N, Maskani Filali M, Delestro Matos F, Chansel-Debordeaux L, Ferrer-Buitrago M, Ferrer E, Antequera V, Ruiz-Jorro M, Papaxanthos A, Ouchchane H, Keppi B, Prima PY, Regnier-Vigouroux G, Trebesses L, Geoffroy-Siraudin C, Zaragoza S, Scalici E, Sanguinet P, Cassagnard N, Ozanon C, De La Fuente A, Gómez E, Gervoise Boyer M, Boyer P, Ricciarelli E, Pollet-Villard X, Boussommier-Calleja A. A hybrid artificial intelligence model leverages multi-centric clinical data to improve fetal heart rate pregnancy prediction across time-lapse systems. Hum Reprod. 2023;38(4):596–608. 10.1093/humrep/dead023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Salih M, Austin C, Warty RR, Tiktin C, Rolnik DL, Momeni M, Rezatofighi H, Reddy S, Smith V, Vollenhoven B, Horta F. Embryo selection through artificial intelligence versus embryologists: a systematic review. Hum Reprod Open. 2023;2023(3):hoad031. 10.1093/hropen/hoad031. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Ganer Herman H, Volodarsky-Perel A, Ton Nu TN, Machado-Gedeon A, Cui Y, Shaul J, Dahan MH. Does embryo quality at transfer affect obstetric outcomes and placental findings? Hum Reprod. 2023;38(5):853–9. 10.1093/humrep/dead045. [DOI] [PubMed] [Google Scholar]
- 30.Gordon CE, Lanes A, Thomas A, Racowsky C. Day of trophectoderm biopsy and embryo quality are associated with outcomes following euploid embryo transfer. J Assist Reprod Genet. 2022;39(11):2539–46. 10.1007/s10815-022-02613-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Liu X, Lou H, Zhang J, Du M, Du Y, Wu S, Guan Y, Liu J. Clinical outcome analysis of frozen-thawed embryo transfer on Day 7. Front Endocrinol (Lausanne). 2022;9(13):1082597. 10.3389/fendo.2022.1082597. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Sermondade N, Huberlant S, Bourhis-Lefebvre V, Arbo E, Gallot V, Colombani M, Fréour T. Female obesity is negatively associated with live birth rate following IVF: a systematic review and meta-analysis. Hum Reprod Update. 2019;25(4):439–51. 10.1093/humupd/dmz011. [DOI] [PubMed] [Google Scholar]
- 33.Zou H, Kemper JM, Hammond ER, Xu F, Liu G, Xue L, Bai X, Liao H, Xue S, Zhao S, Xia L, Scott J, Chapple V, Afnan M, Morbeck DE, Mol BWJ, Liu Y, Wang R. Blastocyst quality and reproductive and perinatal outcomes: a multinational multicentre observational study. Hum Reprod. 2023;38(12):2391–9. 10.1093/humrep/dead212. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Hammond ER, Cree LM, Morbeck DE. Should extended blastocyst culture include day 7? Hum Reprod. 2018;33(6):991–7. 10.1093/humrep/dey091. [DOI] [PubMed] [Google Scholar]
- 35.Xu J, Zhang S, Jin L, Mao Y, Shi J, Huang R, Han X, Liang X, Zhang C. The effects of endometrial thickness on pregnancy outcomes of fresh IVF/ICSI embryo transfer cycles: an analysis of over 40,000 cycles among five reproductive centers in China. Front Endocrinol (Lausanne). 2022;24(12):788706. 10.3389/fendo.2021.788706. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.McLernon DJ, Bhattacharya S. Quality of clinical prediction models in in vitro fertilisation: which covariates are really important to predict cumulative live birth and which models are best? Best Pract Res Clin Obstet Gynaecol. 2023;86:102309. 10.1016/j.bpobgyn.2022.102309. [DOI] [PubMed] [Google Scholar]
- 37.Lawrenz B, Kalafat E, Ata B, Gallego RD, Melado L, Bayram A, Elkhatib I, Fatemi H. Ongoing pregnancy rates in single euploid frozen embryo transfers remain unaffected by female age: a retrospective study. Reprod Biomed Online. 2024;49(2):104074. 10.1016/j.rbmo.2024.104074. [DOI] [PubMed] [Google Scholar]
- 38.Gao H, Liu DE, Li Y, Wu X, Tan H. Early prediction of live birth for assisted reproductive technology patients: a convenient and practical prediction model. Sci Rep. 2021;11(1):331. 10.1038/s41598-020-79308-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Liu X, Chen Z, Ji Y. Construction of the machine learning-based live birth prediction models for the first in vitro fertilization pregnant women. BMC Pregnancy Childbirth. 2023;23(1):476. 10.1186/s12884-023-05775-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Awadalla MS, Bendikson KA, Ho JR, McGinnis LK, Ahmady A. A validated model for predicting live birth after embryo transfer. Sci Rep. 2021;11(1):10800. 10.1038/s41598-021-90254-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41.te Velde ER, Nieboer D, Lintsen AM, Braat DD, Eijkemans MJ, Habbema JD, Vergouwe Y. Comparison of two models predicting IVF success; the effect of time trends on model performance. Hum Reprod. 2014;29(1):57–64. 10.1093/humrep/det393. [DOI] [PubMed] [Google Scholar]
- 42.Ratna MB, Bhattacharya S, Abdulrahim B, McLernon DJ. A systematic review of the quality of clinical prediction models in in vitro fertilisation. Hum Reprod. 2020;35(1):100–16. 10.1093/humrep/dez258. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data regarding any of the subjects in the study has not been previously published unless specified. Data will be made available to the editors of the journal for review or query upon request.