Predicting the chances of live birth for couples undergoing IVF-ICSI: a novel instrument to advise patients and physicians before treatment

Introduction

Infertility is a disease of the reproductive system defined by the failure to achieve a clinical pregnancy after 12 months or more of regular unprotected sexual intercourse¹.

The prevalence of infertility ranged from 3.5% to 16.7% in developed countries². Most recent portuguese data estimate that 9.8% of couples are infertile³.

Infertility is a multifactorial disease⁴. Several treatments have been proposed, with the most innovative being IVF developed by Robert Edwards⁵ in 1978. The two most important Medically Assistant Reproduction Techniques (MAR) are IVF and its subtype ICSI⁶. According to the European Society of Human Reproduction and Embryology (ESHRE), the number of MAR cycles between 1997 and 2014 increased by 13%, reaching 776 556 cycles in Europe in 2015⁷.

Despite the increasing number of MAR treatments, IVF success is not guaranteed⁸. In healthy young couples, the probability of achieving a live birth varies between 20% and 25% per month. This likelihood may increase up to 60% with MAR techniques⁹. According to the study by Malizia et al., between 38% and 49% of couples who start MAR treatments remain childless, even after undergoing up to six MAR cycles¹⁰. In Portugal, the last report of MAR showed that the treatments success rate is around 25–30% per started cycle¹¹.

MAR treatments are expensive, time-consuming, stressful, and may lead to anxiety, depression, or marital problems^12–16. There are also complications to take into consideration such as ovarian hyperstimulation syndrome, bleeding and infection, as well as multiple or premature births¹⁷.

The success rates fluctuate between studies¹⁸. There have been various efforts to build prediction models to assist physicians and patients in predicting MAR success^18–22. To our knowledge, the first predictive model ever built in this context is from Templeton et al. in 1996. A logistic regression model to predict the probability of live birth for an individual woman was developed using the woman’s age, number of previous live birth or pregnancies not resulting in a live birth, whether these were a result of previous IVF treatment, female causes of infertility, duration of infertility and the number of previous unsuccessful IVF treatments¹⁸.

These authors found that the success of IVF decreased with female age and that women between 25 and 30 years were the most likely to have a live birth¹⁸.

Considering the evolution of technology and the inclusion of other variables, Nelson and Lawlor developed a new model in 2011, using the same mathematical techniques. They included the most prevalent causes of infertility, the source of the egg (donor or patient’s own), type of hormonal preparation used (antioestrogen, gonadotrophin, or hormone replacement therapy), whether or not ICSI was used, and the number of previous cycles (1, 2 or 3)²¹.

In 2014, Velde et al. used their cohort study to validate Templeton and Nelson and Lawlor’s model’s performance. They found that Templeton’s model underestimated success rates while Nelson’s overestimated it²³.

Other relevant models were the one developed by Marca et al., which predicts live birth in assisted reproduction based on serum AMH and woman’s age²⁰, and the McLernon et al. study²² that estimated the cumulative first live birth over a maximum of 6 complete cycles using data from 23 417 women in the UK.

Nonetheless, when Leijdekkers performed an external validation of McLernon’s model²⁴ with dutch women, he found that there was an overestimation of the results and he decided to include biomarkers such as AMH, AFC and woman’s body weight. Other studies indicate it is important to include covariables such as BMI, ethnicity and ovarian reserve¹⁹, and corroborate that the use of variables such as BMI, AFC, AMH, ethnicity and smoking status should be predictors in the model^4,25–27.

Furthermore, a machine learning approach was also proposed, including decision trees, genetic algorithms, and k-nearest neighbors classifiers^9,28–31. In 2019, in the study of Qiu et al.³², 4 predictive models based on supervised learning algorithms were created. The purpose of these models was to estimate the cumulative probability of having a live birth before the first IVF treatment starts, using pre-treatment variables including BMI and AMH. One of the models was a logistic regression and achieved an AUROC of 0.71.

The main aim of this study was to find a pre-treatment predictor for achieving a live birth before an IVF/ICSI treatment.

Methods

This was a retrospective study of 739 IVF/ICSI cycles, performed between 2012 and 2016, in Centro de Infertilidade e Reprodução Medicamente Assistida (CIRMA) at Hospital Garcia de Orta, E.P.E., Almada, Portugal.

Only couples with autologous gametes with a live birth or without any frozen embryos available were considered.

The study was approved by the Hospital’s Ethics Committee for Health of the institution. Patients were informed that, under complete anonymity, the data may be used in scientific papers for public presentation or publication³³.

The model’s dependent variable/primary outcome was classified as 1 (if, at least, one baby was born) and 0 (otherwise). This decision was based on other studies in this area^18–22.

The following variables were analyzed: woman’s and man’s age (years), duration of infertility (months), cause of infertility (categorized as tubal factor, endometriosis, disovulation, male factor, both female and male factor - depending on whether the cause of infertility underlies the woman or the man - multiple female factors, unexplained infertility or other), woman’s and man’s Body Mass Index (BMI) (kg/m²), Anti-Müllerian Hormone (AMH) (ng/mL), Antral Follicle Count (AFC), woman’s and man’s ethnicity (categorized as African, Asian, Caucasian, Gipsy, Indian and Mixture), woman’s and man’s smoking status (never, previous and present) and woman’s and man’s previous live children (yes or no) were considered. AFC was obtained by transvaginal ultrasound, and serum AMH levels were measured by blood analysis. There should be no bias associated with this study. However, possible sources of bias may arise from couples’ responses during the consultation and the entry of the data in the database.

A summary statistical evaluation was performed for each variable. A univariate analysis was first performed. All continuous variables were compared using the t-student test and the categorical one using the Chi-square test. A p-value < 0.05 was considered statistically significant³⁴.

A binary logistic regression³⁵ was developed using the primary outcome as binary (no = 0 and yes = 1). An automatic backward selection process based on the Wald statistic was used to determine the final and the best logistic regression model³⁶.

As recommended by Hosmer and Lemeshow³⁵, the interactions between some variables were included to increase the reliability of the model. As Fisher³⁷ explained, the primary outcome can depend not only on one individual baseline characteristic, but also on the relationship between them. The variables in interaction were: woman’s age and duration of infertility³⁸, BMI and oligospermia³⁹, AFC and woman’s BMI⁴⁰, and AFC and AMH²⁶.

The model’s performance was measured with the discriminatory power assessed using the Area Under the Curve (AUC) value^41,42 and the model was internally validated using the bootstrapping technique with 1000 iterations⁴³. All statistical procedures were computed using IBM SPSS Statistics 25, and it was considered an α level of 0.05.

The STROBE cross-sectional reporting guidelines were adopted in this article⁴⁴.

Results

A total of 737 cycles were evaluated. Overall 31.4% of the cycles had at least one live birth.

Baseline characteristics of couples are presented in Table 1 and Table 2. The average age for females was 34.04 years, and 36.14 years for males. Younger women and men were more likely to achieve a live birth. The same was verified for men and women with lower BMI and shorter infertility duration. However, these results were only statistically significant for woman’s age, man’s age, AFC and AMH (p < 0.05).

Table 2 shows that most women and men had never smoked. Most men and women were Caucasian and male factor was the most prevalent infertility cause. Most men and women had no previous live births (90.1% and 87.4%). Chi-square test revealed that no statistically significant differences for live birth (p > 0.05) for categorical variables.

Table 3 shows the binary logistic regression parameters computed with SPSS. It shows that from the 14 variables evaluated, only cause of infertility (male factor), man’s BMI, man’s ethnicity (mixture) and AMH were statistically significant. Apart from these variables, the interactions between duration of infertility and woman’s age, duration of infertility and man’s BMI, AFC and AMH, AFC and woman’s age, AFC and woman’s BMI, and AFC and cause of infertility (disovulation) were also statistically significant (p < 0.05). The interactions between infertility duration and woman’s age and man’s BMI, AFC and AMH, woman’s age, woman’s BMI and cause of infertility (disovulation) were also statistically significant (p-value < 0.05).

According to regression coefficients of the final model, an increase of an AMH unit raises the odds of IVF-ICSI success by 0.172 times, ceteris paribus. Similarly, the interaction between AFC and woman’s BMI decreases the same probability by 0.003 times, and interaction between AFC and woman’s age increases it by 0.004 times. Male factor is the only cause of infertility that enters individually into the final model raising the chances of success by 0.420 times.

The ROC curve test for the discriminatory ability of the final prediction model (Figure 1) had an AUC equal to 0.700 (95% CI 0.660–0.741). This model was also assessed by its performance metrics, shown in Table 4.

Figure 1. Receiver Operating Characteristic (ROC) curve of the final model.

Discussion

So far, providing an accurate prediction of the chances of achieving a live birth after FIV-ISCI treatments has not been an easy task⁴⁵. In this study, a novel prediction model was built, including almost all clinical factors reported as important in the literature.

To our knowledge, this is the first model for live birth FIV-ICSI prediction that accounts for man’s ethnicity. It also includes variables such as BMI and AFC, which were not included in many models before¹⁹. This model includes all the characteristics that have been indicated as important in the literature to predict the chances of live birth in MAR^4,18,45.

The Templeton model¹⁸ considers the existence of previous IVF cycles and has been externally validated by Nelson and Lawlor²¹. Therefore, both models can be applied before IVF is started and predict the success of IVF/ICSI treatment. During this study, it was not possible to obtain information about the number of previous cycles per couple, so this factor was not included.

This model supports earlier findings in this scientific area such as the importance of AMH and AFC to predict live birth as predicted by other studies^20,26 and corroborates indirectly the importance of woman’s age to predict the result as shown by Templeton et al.¹⁸. Nonetheless, it is crucial to refer that the increase of woman’s age, when in interaction with duration of infertility, significantly decreases the probability of a living birth, which is in accordance to Nelson’s study²¹.

Concerning the main reason for treatment, both male factor and disovulation (with AFC interaction) are the only causes that emerged in the final model. When present, each one of them raises the chances of success, which might be explained with the IVF-ICSI technique itself once the problem of sperm and ovulation anomalies are overcome. Possibly, women with ovulation problems and men with sperm anomalies have higher chances of success with IVF-ICSI technique because sperm and oocytes are medically collected and interact directly, although in an in vitro environment⁶.

Another notable finding is that an increase of man’s BMI (per si) or woman’s BMI (with AFC interaction) decreased the chances of live birth, which is in accordance with the scientific literature. Women who are overweight are known to have ovulatory problems and great risks of miscarriage, in the same way that obesity may adversely affect male reproduction by endocrine, thermal, genetic, and sexual mechanisms⁴⁶.

Concerning the man’s ethnicity, we found that if the man is of mixture ethnicity, the odds of success are around 3.8. This finding must be interpreted cautiously as the sample is composed of 92% Caucasian males, with only 2.6% of males being of mixture ethnicity. Up to now, there doesn’t seem to be any biological plausibility for this find, further researchers is required to validate this possible relationship between mixture ethnicity in men and increased success in IVF/ICSI treatments.

A limitation of our model is that it is restricted to pre-treatment stages. Thereby, the physician must explain to patients when using our model that their success probability invariably changes during the cycles process. That is, the resulting probability of our model should be considered a baseline one. Taking into account that this study was performed with data from only one medical center, one limitation of these models is that only an internal validation was made. Moreover, the large temporal spectrum of the data (2012 to 2016) could mean that treatments made with older technology may have different results than those made with more recent ones (namely, vitrification procedures or culture media).

Many studies have accounted for the chances of live birth after IVF or ICSI treatment^18–24. When compared with other studies, the most similar one is Dhillon et al.’s model¹⁹ due to the variables integrated into the model. However, in our study, there were no limitations on socioeconomic status since our data has been extracted from a public hospital with universal access⁴⁷ and so the results can be generalized to people of all socioeconomic backgrounds. This is an advantage over the Dhillon et al.’s model¹⁹, where it is estimated that 75% of couples paid for their treatment and therefore their model could not be generalized to all social classes.

Predictive ability of the prediction models in medicine has been assessed by the AUC value^41,42. In general, AUC for prediction models in reproductive medicine is rather low, ranging between 0.59 and 0.64⁴². Table 5 shows the values of models’ AUC predicting the chances of live birth for couples undergoing IVF-ICSI. Although McLernon et al.²² had the highest value of the AUROC curve, that value decreased on Leijedekkers validation to 0.62²⁴. The model developed in this study had an AUC of 0.700, which is the second-highest value and so has a comparable discriminatory ability with these previous models.

Taking into account that Coppus et al.’s systematic review concluded that prediction models in reproductive medicine would be limited to an AUC value of 0.65 due to the relatively heterogeneous group of subfertile patients⁴², this study can be considered to improve upon the current available models.

The present model predicts the specific probability of a live birth based on easily acquirable couple characteristics before starting a treatment. It might help patients to understand the limitations of an IVF/ICSI treatment in their particular case and also aid physicians to compare different treatment strategies. Furthermore, it might support institutions to predict the likelihood of repeat treatment based on the specific characteristics of each couple.

In the near future, we intend to perform an external validation of the developed model with a new dataset from CIRMA. Follow this, it is also planned to perform an external geographical validation to check if there is a geographical influence on chances of live birth for couples undergoing IVF-ICSI treatment.

Conclusions

Many couples with fertility problems ask themselves if they should undergo an IVF-ICSI treatment. Our novel model provides an estimated probability of their chances of live birth before the start of a MAR treatment. This is the first model with portuguese data and considers the important variables described in literature such as AFC, AMH and woman’s age. This tool may help physicians to shape couples’ expectations, conceding them the opportunity to plan their treatments and to prepare both emotionally and financially. Hence we are developing a user-friendly interface to help physicians in their clinical practice.

Data availability