Predicting the success of IVF: external validation of the van Loendersloot's model

Is the predictive model for IVF success proposed by van Loendersloot et al. valid in a different geographical and cultural context?

The model discriminates well but was less accurate than in the original context where it was developed.

Several independent groups have developed models that combine different variables with the aim of estimating the chance of pregnancy with IVF but only four of them have been externally validated. One of these four, the van Loendersloot's model, deserves particular attention and further investigation for at least three reasons; (i) the reported area under the receiver operating characteristics curve (c-statistics) in the temporal validation setting was the highest reported to date (0.68), (ii) the perspective of the model is clinically wise since it includes variables obtained from previous failed cycles, if any, so it can be applied to any women entering an IVF cycle, (iii) the model lacks external validation in a geographically different center.

Retrospective cohort study of women undergoing oocyte retrieval for IVF between January 2013 and December 2013 at the infertility unit of the Fondazione Ca' Granda, Ospedale Maggiore Policlinico of Milan, Italy. Only the first oocyte retrieval cycle performed during the study period was included in the study. Women with previous IVF cycles were excluded if the last one before the study cycle was in another center. The main outcome was the cumulative live birth rate per oocytes retrieval.

Seven hundred seventy-two women were selected. Variables included in the van Loendersloot's model and the relative weights (beta) were used. The variable resulting from this combination (Y) was transformed into a probability. The discriminatory capacity was assessed using the c-statistics. Calibration was made using a logistic regression that included Y as the unique variable and live birth as the outcome. Data are presented using both the original and the calibrated models. Performance was evaluated correlating the mean predicted chances of live births in the five quintiles and the observed rates.

Two-hundred-eleven live births (27%) were obtained. The c-statistic was 0.64 (95% CI: 0.61-0.67, P < 0.001). The slope of the linear predictor (calibration slope) expressed as an Odds Ratio was 1.81 (95% CI: 1.46-2.24, P < 0.001), corresponding to a beta of 0.630. The calibration intercept was +0.349 (P = 0.13). While a clear discrepancy exists using the original model, data appear properly distributed with the calibrated model. The Pearson coefficient of the correlation between the mean predicted chances of live births in the five quintiles and the observed rates was 0.99 (P = 0.002).

Data were collected retrospectively, thus exposing them to potential inaccuracies. The selection criteria for access to IVF adopted in our center might be too stringent, leading to the exclusion of women with a poor, yet acceptable chance of live birth. Therefore, the validity of the model in women with a very low chance of live birth could not be tested.

The van Loendersloot's model can be used in other contexts but it is important that it has local calibration. It may help in counseling couples about their chance of success but it cannot be used to exclude treatments. Further research is needed to improve the discriminatory performance of IVF predictive models.

None.

Not applicable.