EUVIMEDEUVIMED entdecken
Crossref

L26/O-317 Machine learning versus logistic regression for predicting cumulative live birth rates in young low-prognosis patients: A POSEIDON-stratified analysis and web-based tool

X. Zhen, W. Guo, L. Wang, R. Li, J. Qiao

Human Reproduction · 01.07.2026 · 41 (Supplement_1)

DOI: 10.1093/humrep/deag083.315 ↗

Kurz gefasst

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

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Study question Does machine learning offer superior predictive performance compared to conventional logistic regression for cumulative live birth rates in young, low-prognosis patients? Summary answer Machine learning demonstrated comparable performance to logistic regression. A derived nomogram and web-based calculator exhibited robust discrimination, particularly within POSEIDON Group 3 (AUC=0.817). What is known already Young patients (<35 years) with low oocyte yield (≤9) constitute a clinically heterogeneous cohort. While the POSEIDON criteria stratify them into Group 1 (normal reserve, unexpected low response) and Group 3 (diminished reserve), the distinct prognostic determinants of cumulative live birth rate (CLBR) within these strata remain ill-defined. Furthermore, the incremental value of complex machine learning (ML) algorithms over conventional statistical modeling in this specific population, and the feasibility of a practical clinical tool, remain to be elucidated. The ability to personalize care and manage patient expectations through an evidence-based tool is crucial. Study design, size, duration This retrospective cohort study analyzed the first IVF cycles of 4,768 patients aged under 35 years with an oocyte yield of ≤ 9, between 2020 and 2023. CLBR was defined as the delivery of ≥ 1 live birth following the utilization of all fresh and frozen embryos from the index cycle. The dataset was randomly partitioned into training (70%) and validation (30%) sets for model development and internal evaluation. Participants/materials, setting, methods Patients were stratified into POSEIDON Group 1 (n = 3,137) and Group 3 (n = 1,631). Five predictive algorithms were benchmarked: LASSO-penalized logistic regression, Random Forest, SVM, XGBoost, and LightGBM. Feature importance was interpreted via SHAP values. The most parsimonious yet accurate model was translated into a nomogram and web-based interface, designed for clinicians to provide real-time predictions and enhance patient management. Main results and the role of chance Group 1 displayed markedly superior baseline AMH (3.50±2.76 vs 0.69±0.31 ng/ml) and oocyte yield (6.37±2.21 vs 4.97±2.45) compared to Group 3, translating into a higher yield of transferable embryos (3.11±2.25 vs 2.26±1.96) and CLBR (41.6% vs 26.8%, P < 0.001). In validation, the LASSO-logistic regression yielded an AUC of 0.763 (95% CI 0.736–0.785), statistically comparable to XGBoost (0.760) and LightGBM (0.755). Consequently, the logistic model was selected. Crucially, subgroup analysis revealed superior discriminative power in the diminished reserve cohort (Group 3, AUC=0.817) compared to the more heterogeneous unexpected poor responder cohort (Group 1, AUC=0.744). SHAP analysis identified the number of transferable embryos as the predominant predictor (OR = 1.52, 95% CI 1.41–1.63), followed by endometrial thickness (OR = 1.13, 95% CI 1.08–1.19). Decision curve analysis confirmed that the nomogram conferred a higher net clinical benefit than default strategies across a broad threshold range. Limitations, reasons for caution Inherent limitations of the retrospective design include potential selection bias. While internally validated, external corroboration in multi-center cohorts is requisite to confirm generalizability and establish the clinical utility of the web-based tool. A prospective study will be necessary to evaluate its impact on clinical decision-making and patient outcomes. Wider implications of the findings Complex ML algorithms offered no prognostic advantage over logistic regression in this setting. The proposed tool facilitates evidence-based counseling, offering high-precision prognostication specifically for young women with diminished ovarian reserve (POSEIDON Group 3). It empowers clinicians with individualized, data-driven insights to enhance treatment planning and patient communication. Trial registration number Yes

Zitieren

APA

Zhen, X., Guo, W., Wang, L., Li, R., & Qiao, J. (2026). L26/O-317 Machine learning versus logistic regression for predicting cumulative live birth rates in young low-prognosis patients: A POSEIDON-stratified analysis and web-based tool. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.315

Vancouver

Zhen X, Guo W, Wang L, Li R, Qiao J. L26/O-317 Machine learning versus logistic regression for predicting cumulative live birth rates in young low-prognosis patients: A POSEIDON-stratified analysis and web-based tool. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.315

BibTeX
@article{Zhen2026,
  author = {Zhen, X. and Guo, W. and Wang, L. and Li, R. and Qiao, J.},
  title = {L26/O-317 Machine learning versus logistic regression for predicting cumulative live birth rates in young low-prognosis patients: A POSEIDON-stratified analysis and web-based tool},
  journal = {Human Reproduction},
  year = {2026},
  volume = {41},
  number = {Supplement_1},
  doi = {10.1093/humrep/deag083.315}
}
Kostenlos starten →