Performance of prognostic modelling of high and low ovarian response to ovarian stimulation for IVF

Abstract

Study question: What is the performance of previously established regression models in predicting low and high ovarian response to 150 μg corifollitropin alfa/GnRH-antagonist ovarian stimulation in an independent dataset?

Summary answer: The outcome of ovarian stimulation with 150 μg corifollitropin alfa in a fixed, multiple dose GnRH-antagonist protocol can be validly predicted using logistic regression models with AMH being of paramount importance.

What is known already: Predictors of ovarian response have been identified in FSH/GnRH agonist protocols as well as ovarian stimulation with corifollitropin alfa/GnRH-antagonist. Multivariable response models have been established already, however, external validation of model performance has so far been lacking.

Study design, size, duration: Data from a prospective, multi-centre (n = 5), multi-national, investigator-initiated, observational cohort study were analysed. Infertile women (n = 211), body weight >60 kg, were undergoing ovarian stimulation with 150 μg corifollitropin alfa in a GnRH-antagonist multiple dose protocol for transvaginal oocyte retrieval for IVF. Demographic, sonographic and endocrine parameters were prospectively assessed on cycle Day 2 or 3 of spontaneous menstruation before ovarian stimulation. Main outcomes were low (<6 oocytes) and high (>18 oocytes) ovarian response.

Participants/materials, setting, methods: Firstly, previously established prediction models for low ovarian response (LOR) and high ovarian response (HOR) were tested using the original parameters. Secondly, re-estimated parameters generated from the present data were tested on the established models. Thirdly, for the development of new predictive models of both LOR and HOR, several logistic regression models were estimated. Resulting prediction models were compared by means of the area under the receiver operating characteristic curve (AUC) and bias-corrected Akaike's Information Criterion (AICc) to identify the most reasonable model for each scenario.

Main results and the role of chance: The previously established prediction models for low and high response performed remarkably well on this dataset (low response AUC 0.8879 (95% CI: 0.8185-0.9573) and high response AUC 0.8909 (95% CI: 0.8251-0.9568)). A newly developed simplified model for LOR with log-transformed AMH values and only age as another covariate showed an AUC of 0.8920 (95% CI: 0.8237-0.9603) with the lowest AICc of all models compared. For predicting HOR, we suggest a simplified model using AMH, FSH and AFC (AUC of 0.8976, 95% CI: 0.8206-0.9746).

Limitations, reasons for caution: All analyses were done on data from women with a body weight >60 kg. The newly developed simplified models may suffer from overfitting and need to be tested in further independent data sets.

Wider implications of the findings: Patient selection for ovarian stimulation with corifollitropin alfa should utilize established response prediction models. The clinical impact of this needs to be evaluated in future studies.

Study funding/competing interest(s): The study was funded by university funds. M.O.S., T.L. and I.R.K. have nothing to declare. G.G. has received personal fees and non-financial support from MSD, Ferring, Merck-Serono, Finox, TEVA, IBSA, Glycotope, Abbott, Marckryl Pharma, VitroLife, NMC Healthcare, ReprodWissen, ZIVA and BioSilu.

Trial registration number: Not applicable.

Keywords: GnRH antagonist; corifollitropin alfa; ovarian hyperstimulation syndrome; ovarian response; ovarian stimulation; poor response; response prediction.