L26/P-387 Ai-driven trigger timing optimization in elective fertility preservation: comparing learning and non-learning models
Human Reproduction · 01.07.2026 · 41 (Supplement_1)
DOI: 10.1093/humrep/deag083.721 ↗Kurz gefasst
Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von R. Berkovitz shperling und 5 weiteren Autor:innen.
Abstract (Originalsprache)
Study question Can AI improve trigger timing in elective fertility preservation, and how does case-based learning affect AI performance compared to guideline-based reasoning? Summary answer A learning AI achieved 94.3% agreement with physicians, compared to 65.4% for non-learning models; cases where physicians diverged from AI recommendations showed higher oocyte yields. What is known already AI is increasingly used in reproductive medicine, with recent research showing its potential to identify the best time for IVF trigger. However, generative AI models like ChatGPT offer general advice that may not be customized for individual patients. The specific use of AI for trigger timing in elective fertility preservation—patients with unique characteristics—remains unstudied. Furthermore, no research has directly compared learning and non-learning AI methods using the same generative AI architecture. Study design, size, duration Retrospective cohort study of 563 elective fertility preservation cycles at a single tertiary center (December 2022–May 2025). Two AI models using GPT-5 architecture (OpenAI) were developed: a non-learning model simulating real-time chatbot use, and a supervised learning model (Random Forest) trained on institutional data. Participants/materials, setting, methods Women with normal ovarian reserve (AMH ≥1.0 ng/mL, AFC ≥5) undergoing elective fertility preservation at Tel Aviv Sourasky Medical Center. Trigger decisions were made by senior specialists with over 5 years of experience. AI performance was assessed using concordance rates (±2 days), mean absolute error (MAE), and prediction tendency. Learning curve analysis identified the minimum training data requirements. Main results and the role of chance The learning model achieved 94.3% agreement (531/563 cycles) and an MAE of 0.83 days. The non-learning model showed 60.4% agreement (340/563), an MAE of 1.86 days, and a consistent tendency toward earlier triggers (−1.59 days); 83.1% of its suggestions favored earlier triggering. The learning model demonstrated greater stability (SD 0.07 vs. 0.64 days across 1,000 iterations). Cases where physicians deviated from the learning model were associated with higher oocyte yields (median 21 vs. 13, P = 0.002), increased MII counts (16.5 vs. 10, P = 0.001), and better MII-to-follicle ratios (1.50 vs. 1.00, P < 0.001), suggesting these were mainly high responders requiring personalized timing. The learning curve analysis indicated that 120–150 cycles are necessary to reach over 90% agreement. Limitations, reasons for caution A single-center retrospective design limits how widely the results can be applied. AI was compared to physician decisions rather than an objectively optimal timing; therefore, we cannot determine if deviation led to better outcomes. Results might differ across institutions with different patient populations or protocols. Wider implications of the findings Non-learning AI provides general guidance that may not reflect individuals’ specific practice patterns. Learning-based AI showed high agreement and can serve as a decision-support tool. Physicians should prefer learning-based models over general-purpose models. Patients consulting with AI should be made aware of the limitations of general AI models. Trial registration number No
Zitieren
APA
shperling, R. B., Engelder, D., Khana, A., Kalma, Y., Feferkorn, I., & Cohen, Y. (2026). L26/P-387 Ai-driven trigger timing optimization in elective fertility preservation: comparing learning and non-learning models. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.721
Vancouver
shperling RB, Engelder D, Khana A, Kalma Y, Feferkorn I, Cohen Y. L26/P-387 Ai-driven trigger timing optimization in elective fertility preservation: comparing learning and non-learning models. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.721
BibTeX
@article{shperling2026,
author = {shperling, R. Berkovitz and Engelder, D. and Khana, A. and Kalma, Y. and Feferkorn, I. and Cohen, Y.},
title = {L26/P-387 Ai-driven trigger timing optimization in elective fertility preservation: comparing learning and non-learning models},
journal = {Human Reproduction},
year = {2026},
volume = {41},
number = {Supplement_1},
doi = {10.1093/humrep/deag083.721}
}
EUVIMED