EUVIMEDEUVIMED entdecken
Crossref

L26/P-701 Center-Adapted Self-Learning Mature Oocyte Retrieval Prediction Model: Development, External Validation, and Clinical Utility

M. Kobanawa, J. Yoshida

Human Reproduction · 01.07.2026 · 41 (Supplement_1)

DOI: 10.1093/humrep/deag083.1030 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von M. Kobanawa und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Study question Can a center-adapted self-learning model improve MII prediction accuracy versus the original model and enable higher CLBR with earlier live birth attainment in routine practice? Summary answer Self-learning recalibration reduced MII prediction error versus the original model, potentially supporting improved CLBR and faster live birth attainment across embryo transfers. What is known already Mature oocyte (MII) yield prediction supports individualized ovarian stimulation and trigger decisions, yet models often degrade when applied across centers. Differences in stimulation regimens, monitoring intensity, trigger criteria, and oocyte assessment practices shift data distributions, causing miscalibration and systematic bias. Rebuilding models at every clinic is impractical. Self-learning approaches that update parameters using local real-world data may enable rapid center-specific recalibration, preserving the original structure while adapting to local practice. Evidence remains limited on whether improved calibration translates into better cumulative live birth rates (CLBR) and shorter time to live birth across embryo transfers, hindering broader implementation. Study design, size, duration Single-center retrospective observational study at an independent Japanese fertility center. Training comprised insurance-covered ovarian stimulation cycles using injectable gonadotropins from January–December 2023 (n = 152). Validation included patients undergoing their first ovarian stimulation between January–December 2024 (n = 96). Embryo transfer outcomes were available for n = 86 patients with at least one transfer. Model performance was evaluated in the 2024 cohort, and clinical signals were examined using transfer-based cumulative live-birth follow-up. Participants/materials, setting, methods We compared a self-learning recalibrated model with the original model for predicting number of MII. Accuracy was assessed using mean absolute error (MAE) and agreement within clinically meaningful thresholds (±1/±2 for MII). Paired differences in absolute MII error were compared using the Wilcoxon signed-rank test. Clinical utility evaluated CLBR and time-to-live-birth signals across transfers using discrete-time hazard models (cloglog) adjusted for age and body weight. Main results and the role of chance External validation (n = 96) showed that the self-learning model more closely matched observed MII than the original model. Agreement within ±2 MII was 76.0% (73/96) with self-learning versus 51.0% (49/96) with the original model (absolute difference +25.0 percentage points). Agreement within ±1 MII was 45.8% (44/96) versus 30.2% (29/96). Mean absolute error (MAE) decreased from 2.365 to 1.33 (−1.04 MII; 44% relative reduction). Paired absolute MII errors were significantly lower with self-learning (Wilcoxon signed-rank p = 3.99 × 10-6), and the improvement in ± 2 agreement was supported by McNemar’s test (p = 2.1 × 10−⁷). Clinical utility was explored in patients with embryo-transfer follow-up (n = 86). Defining agreement as |ΔMII|≤1, cumulative live birth (CLBR) was numerically higher in the agreement group: 71.1% (27/38) versus 62.5% (30/48). In logistic regression adjusted for age, body weight, and number of embryo transfers, agreement was not independently associated with live birth (adjusted OR 0.96, 95%CI 0.35–2.63; p = 0.929), but the absolute rates suggested a potential clinical signal warranting confirmation in larger cohorts. Discrete-time hazard models showed a lower live-birth hazard per +1 increase in |ΔMII| (adjusted HR 0.914), consistent with the hypothesis that reduced prediction error may support earlier attainment of live birth across transfers. Limitations, reasons for caution Retrospective single-center design limits causal inference and generalizability. Sample size was modest, and unmeasured confounding may remain despite adjustment. Trigger-timing benefit was inferred from prediction performance rather than prospectively tested. Clinical decisions were not model-guided. Selection bias is possible; replication in larger cohorts is needed. Wider implications of the findings Self-learning recalibration can adapt an MII prediction model to local practice, improving accuracy without rebuilding. If confirmed in larger studies, reduced prediction error may translate into higher CLBR and earlier live birth attainment, enabling scalable decision support and safer cross-center deployment for ovarian stimulation in routine care and research. Trial registration number No

Zitieren

APA

Kobanawa, M., & Yoshida, J. (2026). L26/P-701 Center-Adapted Self-Learning Mature Oocyte Retrieval Prediction Model: Development, External Validation, and Clinical Utility. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.1030

Vancouver

Kobanawa M, Yoshida J. L26/P-701 Center-Adapted Self-Learning Mature Oocyte Retrieval Prediction Model: Development, External Validation, and Clinical Utility. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.1030

BibTeX
@article{Kobanawa2026,
  author = {Kobanawa, M. and Yoshida, J.},
  title = {L26/P-701 Center-Adapted Self-Learning Mature Oocyte Retrieval Prediction Model: Development, External Validation, and Clinical Utility},
  journal = {Human Reproduction},
  year = {2026},
  volume = {41},
  number = {Supplement_1},
  doi = {10.1093/humrep/deag083.1030}
}
Kostenlos starten →