An artificial intelligence-based approach for selecting the optimal day for triggering in antagonist protocol cycles.

Research Question: Can a machine-learning model suggest an optimal trigger day (or days), analysing three consecutive days, to maximize the number of total and mature (metaphase II [MII]) oocytes retrieved during an antagonist protocol cycle?

Design: This retrospective cohort study included 9622 antagonist cycles between 2018 and 2022. The dataset was divided into training, validation and test sets. An XGBoost machine-learning algorithm, based on the cycles' data, suggested optimal trigger days for maximizing the number of MII oocytes retrieved by considering the MII predictions, prediction errors and outlier detection results. Evaluation of the algorithm was conducted using a test dataset including three quality groups: 'Freeze-all oocytes', 'Fertilize-all' and 'ICSI-only' cycles. The model suggested 1, 2 or 3 days as trigger options, depending on the difference in potential outcomes. The suggested days were compared with the actual trigger day chosen by the physician and were labelled 'concordant' or 'discordant' in terms of agreement.

Results: In the 'freeze-all' test-set, the concordant group showed an average increase of 4.8 oocytes and 3.4 MII oocytes. In the 'ICSI-only' test set there was an average increase of 3.8 MII oocytes and 1.1 embryos, and in the 'fertilize-all' test set an average increase of 3.6 oocytes and 0.9 embryos was observed (P < 0.001 for all parameters in all groups).

Conclusions: Utilizing a machine-learning model for determining the optimal trigger days may improve antagonist protocol cycle outcomes across all age groups in freeze-all or fresh transfer cycles. Implementation of these models may more accurately predict the number of oocytes retrieved, thus optimizing physicians' decisions, balancing workloads and creating more standardized, yet patient-specific, protocols.