EUVIMEDEUVIMED entdecken
Crossref

L26/O-263 Near–real-time adaptive machine learning for personalized continuous quality monitoring in IVF laboratories

E. Güell Penas, I. Vazquez Vazquez, I. Ochando Sánchez, J. Vázquez Vázquez, M. Fernandez Tarjuelo

Human Reproduction · 01.07.2026 · 41 (Supplement_1)

DOI: 10.1093/humrep/deag083.261 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von E. Güell Penas und 4 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Study question Can a continuously updated machine-learning (ML) model accurately predict key embryology laboratory process metrics and detect suboptimal performance in near-real-time (nRT)? Summary answer A data-driven, centre-specific predictive framework showed acceptable accuracy for core embryology metrics and allowed for the early detection of performance deviations at the laboratory level. What is known already Quality control in embryology laboratories is traditionally retrospective and based on key performance indicators (KPIs), such as maturation, fertilisation, and blastocyst development rates. This approach could delay problem detection and may limit early intervention. Traditional monitoring methods often do not fully account for variability between patients, and current laboratory software rarely incorporates AI modules capable of near real-time monitoring that is sensitive to individual patient characteristics. While machine learning has been applied to predict clinical outcomes, its use for continuous laboratory surveillance remains limited. Study design, size, duration Retrospective validation simulating nRT prediction, including a total of 606 treatment cycles from 495 patients and 111oocyte donors (Januart 2021–January 2026) from a single ART centre. The modelling pipeline, developed in the Ilitia software platform (Sentinel v1.0), used overlapping 30-cycle rolling windows. An initial 60-cycle training set generated the first prediction, after which each new cycle was added sequentially, producing 517 prediction windows of 30 consecutive cases with continuous model updating. Participants/materials, setting, methods Data included patient age, oocyte origin/type, and counts of retrieved, mature, inseminated, fertilised and usable embryos (D5/D6). Fertilization and usable-blastocyst rates were computed. A multivariate linear regression model was used for deploying the predictive model. Predictors included oocyte origin, patient age group and preceding laboratory variables. Agreement was assessed with Bland–Altman and concordance correlation coefficient (CCC). A triple alert system flagged underperformance (<5th percentile, P5), deviations from prediction and benchmark values from Vienna Consensus. Main results and the role of chance Window-level mean absolute error and standard deviation was calculated for fertilised and usable embryos across all rolling windows, yielding values of 1.07±0.27 and 1.17±0.26, respectively. Bland–Altman analysis showed acceptable agreement for both outcomes: fertilised embryos displayed a near-zero mean bias with 95% limits of agreement (LoA,−3.89 to 4.02), while usable embryos showed similar bias and LoA results (−3.83 to 3.42). These intervals indicated that prediction errors typically remained within 3–4 embryos, a clinically acceptable range given IVF cohort variability. Over 93% of paired differences fell within LoA, and concordance correlation coefficients confirmed good agreement for fertilised embryos (CCC=0.928) and moderate agreement for usable embryos (CCC=0.821).To reduce noise from overlapping windows, alerts were consolidated into non-overlapping 30-case blocks. Among 21 blocks analysed for fertilisation rate, 2 blocks (9.5%) triggered alerts: one fell below both the Vienna Consensus benchmark and the laboratory-specific P5 threshold, while the other remained benchmark-compliant but below cumulative P5, suggesting a subtle laboratory-level decline not captured by external standards. For usable blastocyst rate, 4 blocks (19%) triggered alerts: one detected by both benchmark and P5 violation, and three detected exclusively by P5 deviation. These sentinel alerts revealed early performance shifts that static external benchmarks alone failed to detect. Limitations, reasons for caution Single-centre retrospective design. Linear models may miss complex non-linear interactions. Performance thresholds depend on patient characteristics not evaluated here, such as hormone levels and semen quality, as well as culture media and incubators. Small clinics may struggle to achieve representative sample sizes. Local calibration and multicentre validation are needed. Wider implications of the findings This first design enables a shift from reactive to predictive–proactive laboratory quality control, allowing early intervention, continuous benchmarking, and easier regulatory validation through model interpretability. Leveraging big data with additional patient- and lab-level parameters is promising for more precise, personalized quality monitoring and timely detection of performance deviations. Trial registration number No

Zitieren

APA

Penas, E. G., Vazquez, I. V., Sánchez, I. O., Vázquez, J. V., & Tarjuelo, M. F. (2026). L26/O-263 Near–real-time adaptive machine learning for personalized continuous quality monitoring in IVF laboratories. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.261

Vancouver

Penas EG, Vazquez IV, Sánchez IO, Vázquez JV, Tarjuelo MF. L26/O-263 Near–real-time adaptive machine learning for personalized continuous quality monitoring in IVF laboratories. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.261

BibTeX
@article{Penas2026,
  author = {Penas, E. Güell and Vazquez, I. Vazquez and Sánchez, I. Ochando and Vázquez, J. Vázquez and Tarjuelo, M. Fernandez},
  title = {L26/O-263 Near–real-time adaptive machine learning for personalized continuous quality monitoring in IVF laboratories},
  journal = {Human Reproduction},
  year = {2026},
  volume = {41},
  number = {Supplement_1},
  doi = {10.1093/humrep/deag083.261}
}
Kostenlos starten →