EUVIMEDEUVIMED entdecken
Crossref

L26/P-768 Personalized oocyte yield prediction: A causal machine learning model estimates outcomes for specific r-hFSH starting doses by correcting for dose selection bias

R. Hourvitz, A. Luz, N. Itzhak, S. Reuvenny, E. Moran, M. Baum, E. Maman, A. Hochberg, E. Hariton, N. Polyzos, W. Zheng, T. D’hooghe und 1 weitere

Human Reproduction · 01.07.2026 · 41 (Supplement_1)

DOI: 10.1093/humrep/deag083.1095 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von R. Hourvitz und 12 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Study question Can a machine learning algorithm accurately predict oocyte yield forspecific r-hFSH starting doses by disentangling true drug effects from doseselection bias? Summary answer A causal-inference machine learning model correcting for dose selectionbias accurately predicts oocyte yield for individual r-hFSH starting doses. What is known already Initial FSH dosing relies heavily on clinician experience or generalizednomograms. A major challenge in analyzing historic FSH dosing data isdose selection bias, or more precisely confounding by indication: cliniciansprescribe higher FSH doses to patients with poorer prognosis.Consequently, previous studies often confused the patient’s poor prognosiswith the drug’s dosing effect, incorrectly inferring that higher doses causelower oocyte yields. To date, no tool has reliably separated the expectedeffect of FSH dose from the patient’s underlying ovarian potential to providepatient-specific predictions under different FSH starting doses. Study design, size, duration This retrospective, multi-center cohort study collected data from 10 IVFclinics across Europe, North America, and Asia (2014–2024). The datasetconsisted of 22,686 mixed-protocol cycles used for model development and778 r-hFSH-only cycles for testing.The study focused on developing and validating a patient-specificapproach, using age, BMI, AMH, and AFC, to predict expected oocyte yieldfor different initial r-hFSH doses, while accounting for dose selection bias. Participants/materials, setting, methods A causal-inference framework (orthogonalization) utilizing three XGBoostmodels was employed.The first two models estimated the patient’s expected average oocyte yieldand the typical FSH starting dose that clinicians prescribe, given thepatient’s profile. A third model then quantified the true drug dosing effect byanalyzing how deviations from this typical FSH starting dose resulted indeviations from the expected average oocyte yield. This approach isolatestreatment effects from prognosis, minimizing dose selection bias. Main results and the role of chance The model showed strong predictive performance on the r-hFSH-only testset (R²=0.66; MAE=3.65 oocytes, n = 778), exceeding performancepreviously reported in the literature.To assess clinical plausibility, oocyte yields were simulated for each patientacross r-hFSH starting doses of 75–450 IU. To ensure reliability, an outlierdetection model utilizing AMH and AFC defined a plausible dose range foreach patient by removing extreme starting doses where predictionaccuracy was uncertain due to lack of reliable data. Within this plausibledose range, the dose–yield slopes were analyzed, classifying patterns asflat (variation ≤2%), increasing (positive slope), or decreasing (negativeslope).This approach was compared with a standard (“naïve”) XGBoost model.The naïve model frequently produced biologically implausible patterns,predicting lower yield with higher dose in 41.7% of patients. In contrast, ourbias-corrected approach predicted increasing yields in 99.4% of patients inthe test dataset.Pattern | Our Approach | Naive ApproachDecreasing Yield | 0% | 41.7%Increasing Yield | 99.4% | 58.3%Flat | 0.6% | 0%Compared with the naïve approach, our bias-corrected approach markedlyreduced clinically implausible decreasing dose–yield patterns, supportingmore credible predictions on oocyte yield based on patient profile and r-hFSH starting doses. Limitations, reasons for caution The primary limitation is the study's retrospective design. The model alsodoes not account for key clinical factors such as previous cycle ovarianresponse, reasons for specific treatment, or infertility causes. Thesevariables were deliberately omitted to prioritize the tool's ease of use. Wider implications of the findings By accounting for causal effects, this framework provides a robust, data-driven tool for r-hFSH dose planning. Separating true drug effects fromunderlying prognosis and dose selection bias is essential for accuratepersonalization, enabling reliable outcome estimation under differentdosing scenarios. Trial registration number No

Zitieren

APA

Hourvitz, R., Luz, A., Itzhak, N., Reuvenny, S., Moran, E., Baum, M., Maman, E., Hochberg, A., Hariton, E., Polyzos, N., Zheng, W., D’hooghe, T., & Hourvitz, A. (2026). L26/P-768 Personalized oocyte yield prediction: A causal machine learning model estimates outcomes for specific r-hFSH starting doses by correcting for dose selection bias. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.1095

Vancouver

Hourvitz R, Luz A, Itzhak N, Reuvenny S, Moran E, Baum M, et al. L26/P-768 Personalized oocyte yield prediction: A causal machine learning model estimates outcomes for specific r-hFSH starting doses by correcting for dose selection bias. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.1095

BibTeX
@article{Hourvitz2026,
  author = {Hourvitz, R. and Luz, A. and Itzhak, N. and Reuvenny, S. and Moran, E. and Baum, M. and Maman, E. and Hochberg, A. and Hariton, E. and Polyzos, N. and Zheng, W. and D’hooghe, T. and Hourvitz, A.},
  title = {L26/P-768 Personalized oocyte yield prediction: A causal machine learning model estimates outcomes for specific r-hFSH starting doses by correcting for dose selection bias},
  journal = {Human Reproduction},
  year = {2026},
  volume = {41},
  number = {Supplement_1},
  doi = {10.1093/humrep/deag083.1095}
}
Kostenlos starten →