Can machine learning models predict oocyte yield during assisted conception?: a systematic review

Accurately predicting oocyte yield is key to individualizing gonadotrophin dosing in assisted reproduction, where excessive responses carry significant risks and inadequate responses may compromise treatment success. Machine learning models are being developed to support this. This review evaluates the accuracy and clinical readiness of such models for predicting oocyte yield at transvaginal oocyte retrieval. A literature search of OVID MEDLINE, OVID EMBASE and the Cochrane Database identified nine relevant studies, eight retrospective and one prospective cohort, and encompassing 62,354 cycles. Study quality was assessed using the TRIPOD checklist and risk of bias was assessed with PROBAST. Accuracy was variably reported; the most frequent measure was mean absolute error, ranging from 0.62 to 4.13. A meta-analysis was not possible due to heterogeneity. Neural network models generally performed best. None of the studies were externally validated, limiting the generalizability. Most lacked transparency around data pre-processing, including handling missing data and variable transformation, reducing reproducibility. Although several models demonstrated promising internal performance, they remain at high risk of bias. Clinicians should be cautious in applying these models to practice until independently validated. Overall, machine learning-based prediction of oocyte yield remains in the early phase, and future development should prioritize external validation and transparent reporting.