A retrospective cohort analysis evaluating letrozole starting doses in women with polycystic ovary syndrome undergoing sequential ovulation induction found no statistically significant average difference in cumulative clinical pregnancy rates after statistical adjustment. The research examined whether an initial daily dose of 2.5 milligrams or 5.0 milligrams produced divergent pregnancy outcomes across patient subsets, testing whether baseline patient characteristics could guide individualized dosing strategies in routine clinical care.
The study addressed clinical decision-making during sequential ovulation induction combining letrozole with follicle-stimulating hormone, investigating whether patient-level heterogeneity influenced which starting dose conferred greater therapeutic benefit.
Conducted as a single-center retrospective study, the investigation included 1,917 women diagnosed with polycystic ovary syndrome according to the 2003 Rotterdam criteria who received treatment between 2020 and 2024. Among the cohort, 979 patients initiated treatment with 2.5 milligrams per day, while 938 patients began with 5.0 milligrams per day under routine care conditions for up to three treatment cycles. The primary study endpoint tracked across treatment was cumulative clinical pregnancy.
To account for potential confounding inherent in observational clinical records, the investigators employed stabilized inverse probability of treatment weighting alongside an augmented inverse-probability-weighted estimator to calculate adjusted risk differences. The analytical framework incorporated a causal forest model and pre-specified continuous interaction tests using patient-level bootstrap inference to explore treatment heterogeneity. The researchers additionally conducted overlap weighting as a sensitivity analysis and generated a doubly robust policy tree to model candidate dosing decisions, evaluating its predictive reliability on held-out data through five-fold internal cross-validation.
Unadjusted observational records indicated a cumulative clinical pregnancy rate of 52.3 percent in the 2.5-milligram starting group compared with 42.9 percent in the 5.0-milligram starting group. Following statistical adjustment with the augmented inverse-probability-weighted estimator, the adjusted risk difference was positive 4.00 percentage points, with a 95 percent confidence interval ranging from negative 2.34 to positive 10.35 percentage points. Because the confidence interval encompassed zero, the average outcome distinction between the two starting doses was not statistically significant. In the overlap weighting sensitivity analysis, the adjusted risk difference was estimated at positive 6.22 percentage points, with a 95 percent confidence interval spanning 1.42 to 11.03 percentage points.
Evaluation of treatment heterogeneity revealed that a body mass index by treatment interaction was the sole pre-specified interaction that remained statistically meaningful following Bonferroni correction, demonstrating an odds ratio ratio of 1.53 per one standard deviation increase in body mass index, with a 95 percent confidence interval of 1.16 to 2.03 and a Bonferroni-adjusted P-value of 0.01386. This interaction maintained a consistent direction in the overlap population. The constructed doubly robust policy tree assigned either 2.5 milligrams or 5.0 milligrams across seven terminal patient nodes, yielding an in-sample policy value of 57.84 percent. However, during five-fold internal cross-validation on held-out data, the policy tree demonstrated an estimated value of 51.18 percent, falling below the 51.43 percent policy value achieved by assigning a uniform 2.5-milligram starting dose to all patients.
The authors concluded that the observed body mass index interaction and the machine-learning policy tree represent exploratory, hypothesis-generating patterns rather than an established clinical treatment algorithm. Because the candidate decision model did not outperform uniform administration of the 2.5-milligram starting dose during internal cross-validation, the investigators emphasized that the candidate rule should not be used to direct clinical prescribing without prior evaluation in prospective multicenter clinical trials.
Subject of Research: Letrozole starting dose heterogeneity and pregnancy outcomes in polycystic ovary syndrome
Article Title: BMI-related heterogeneity in the association between the letrozole starting dose and pregnancy outcomes in PCOS: a machine-learning analysis of a retrospective cohort
Article References: Li, J.-T., Gao, Y., Lin, N., Li, R., Zhao, J., Huang, E., Wang, H., Xiong, W., & Liu, Y. (2026). BMI-related heterogeneity in the association between the letrozole starting dose and pregnancy outcomes in PCOS: a machine-learning analysis of a retrospective cohort. Journal of Ovarian Research. https://doi.org/10.1186/s13048-026-02327-z
Image Credits: AI Generated
DOI: 10.1186/s13048-026-02327-z
Keywords: Polycystic ovary syndrome, Letrozole starting dose, Ovulation induction, Follicle-stimulating hormone, Cumulative clinical pregnancy, Body mass index, Machine learning, Causal forest, Doubly robust policy tree, Augmented inverse-probability weighting, Stabilized inverse probability weighting, Overlap weighting, Rotterdam criteria, Treatment heterogeneity, Internal cross-validation, Retrospective cohort study, Reproductive medicine
News Source: Harold Sullivan. (November 15, 2026). Adjusted Letrozole Starting Doses Show No Average Pregnancy Difference in PCOS Study. Scienmag.



