The algorithm will see you now: artificial intelligence in the prediction of pregnancy – ESHRE

A web-based cohort study suggests that, if machine learning algorithms are provided with a sufficiently wide range of predictive data, they can be induced to analyse epidemiologic data and predict the probability of conception with a discrimination accuracy which exceeds earlier studies.

One focus for AI research has been in predicting the chance of pregnancy - with varying success. A study last year found an AI-based model outperformed clinicians in assessing embryo viability, while a poster from last years annual meeting of preliminary research into predicting embryo ploidy showed that the algorithm tended to classify embryos as aneuploid.(1,2)

Adding to this evidence base, a new large prospective study has now found that algorithms are able to forecast the probability of conception among couples trying to get pregnant if given a wide range of data on predictors of fecundability (defined as the per-cycle probability of conception).(3) Based on a study participation cohort of more than 4000 women, results showed an overall discrimination performance of around 70% for six different supervised machine-learning algorithms in distinguishing between women who were likely to conceive and those who were not.

It was an outcome which, the authors say, exceeds results from predictive models in previous studies and demonstrates that such models can be created with reasonable discrimination using self-reported data. They add that this is in the absence of more detailed medical information such as laboratory or imaging tests.

Earlier work in this area has focused primarily on identifying individual risk factors for infertility. Several predictive models have been developed in sub-fertile populations but with limited power and using little or no data on lifestyle, environmental and sociodemographic factors. In contrast, a total of 163 predictors of fecundability were considered in this new study to anticipate the cumulative likelihood of pregnancy over six and 12 menstrual cycles.

The data were based on 4133 women from the ongoing Pregnancy Study Online (PRESTO), a web-based preconception cohort study which is analysing the impact of environmental and behavioural factors on fertility and pregnancy. Participants in the study were aged 2144 years, from the US or Canada, were not using fertility treatment, reported no more than one menstrual cycle of pregnancy attempt at study entry, and were actively trying to conceive at enrolment (20132019).

The female patients completed extensive questionnaires at enrolment (eg, marital status, reproductive and diet history, male partner characteristics, etc). Some of this information (eg, menstrual cycle dates) was updated via follow-up questionnaires completed bimonthly for 12 months, or until conception/cessation of pregnancy attempts or study withdrawal.

Next, the data were used to develop models to predict the probability of pregnancy. These were based on three time periods: pregnancy in fewer than 12 menstrual cycles (model I, n = 3195); pregnancy within six menstrual cycles (model II, n = 3476); and the average probability of pregnancy per menstrual cycle (model III, n = 4133). Additional models were also developed for women (n = 1957) who had never been pregnant but had no history of infertility: pregnancy in fewer than 12 menstrual cycles (model IV); pregnancy within six menstrual cycles (model V); and predicting fecundability (model VI). Six different supervised machine learning algorithms were then applied to each model to establish how each algorithm performed.

Results showed 86% of women in model I became pregnant and 69% in model II within the timeframes. For all six algorithms, the AUC (for prediction accuracy) was as follows: model I 68-70% (SD: 0.8%-1.9%); model II 65-66% (SD: 1.9%-2.6%); model III (63%); model IV 69.5% (SD: 1.4%); model V 65.6% (SD: 2.9); and model VI 60.2% concordant index.

Female age, female BMI and history of infertility were the predictors inversely associated with pregnancy in all models. The predictors positively associated with pregnancy in the first three models were having previously breastfed an infant and using multivitamins or folic acid supplements. Among the nulligravid women, the most important predictors were female age, female BMI, male BMI, use of a fertility app, attempt time at study entry and perceived stress.

The authors conclude that the findings are especially relevant for couples planning a pregnancy and for clinicians caring for women coming off contraception to have a baby. However, they add that the models do need to be validated in external populations before they can become a counselling tool.

1. VerMilyea M, Hall J, Diakiw S, at al. Development of an artificial intelligence-based assessment model for prediction of embryo viability using static images captured by optical light microscopy during IVF. Human doi: 10.1093/humrep/deaa0132. Aparicio Ruiz B, Bori L, Paya E, et al. Applying artificial intelligence for ploidy prediction: The concentration of IL-6 in spent culture medium, blastocyst morphological grade and embryo morphokinetics as variables under consideration. Human Reprod 2021; doi.org/10.1093/humrep/deab127.0663. Yland J, Wang T, Zad Z, et al. Predictive models of pregnancy based on data from a preconception cohort study. Human Reprod 2022; 1-13; doi.org/10.1093/humrep/deab280

Read more:

The algorithm will see you now: artificial intelligence in the prediction of pregnancy - ESHRE