Looking into the future: a machine learning powered prediction model for oocyte return rates after cryopreservation

Could a predictive model, using data from all US fertility clinics reporting to the Society for Assisted Reproductive Technology, estimate the likelihood of patients using their stored oocytes?

Multiple learner algorithms, including penalized regressions, random forests, gradient boosting machine, linear discriminant analysis and bootstrap aggregating decision trees were used. Data were split into training and test datasets. Patient demographics, medical and fertility diagnoses, partner information and geographic locations were analysed.

A total of 77,631 oocyte-cryopreservation cycles (2014-2020) were analysed. Patient age averaged 34.5 years. Treatment indications varied: planned (35.6%), gender-related (0.1%), medically indicated (15.5%), oncologic (5.7%) and unknown (42.3%). Infertility diagnoses were less common: unexplained infertility (1.8%), age-related infertility (3.2%), diminished ovarian reserve (9.9%) and endometriosis (1.6%). An ensemble model combining bootstrap aggregation classification and regression trees, stochastic gradient boosting and linear discriminant analysis yielded the highest predictive accuracy on test set (balanced accuracy: 0.83, sensitivity: 0.76, specificity: 0.91), with a receiver operating characteristic curve of 0.90 and precision-recall curve and area under the curve of 0.57. Key factors influencing the likelihood of returning for oocyte use included patient age, presence of a partner, race or ethnicity, the clinic's geographic region and oocyte cryopreservation indication.

This model demonstrated significant predictive accuracy, and is a valuable tool for patient counselling on oocyte cryopreservation. It helps to identify patients more likely to use stored oocytes, enhancing healthcare decision-making and the efficiency of gamete storage programmes. The model can be applied to self-financed and insurance-funded cycles.