Consistent associations of BMI with ovarian response and pregnancy loss risk across PCOS, DOR, and tubal factor infertility: a complete-case mediation analysis of 8,923 IVF cycles

Baseline characteristics of the three cohorts are presented in Table S1. Given the distinct diagnostic criteria for PCOS, DOR, and tubal factor infertility, expected differences were observed across groups in age, hormonal profiles (FSH, AMH), ovarian response parameters, and reproductive outcomes. Before mediation analysis, we confirmed that BMI was significantly correlated with FSH ( r = -0.08 to -0.15, all P  < 0.001), and both FSH and BMI were significantly associated with oocyte yield in multivariable regression analyses (all P  < 0.01), satisfying the prerequisite conditions for mediation analysis (Table S2). Table  1 presents the cohort-specific path coefficients and effect decomposition from SEM. Indirect associations of BMI with oocyte yield via FSH were significant across all cohorts (β range: 0.018–0.025, all P  < 0.01), with the proportion mediated ranging from 5.31% to 8.21%. Direct effects (β range: 0.262–0.437) and total effects (β range: 0.286–0.462) were also positive and significant in all groups (all P  < 0.001). Table 1 Cohort-specific path coefficients and effect decomposition from SEM analysis Cohort Path/Effect Label Estimate 95% CI p -value Tubal ( n  = 3991) BMI → FSH a -0.168 (-0.248, -0.088) < 0.001 FSH → Oocyte b -0.146 (-0.184, -0.108) < 0.001 Indirect effect a×b 0.025 (0.011, 0.038) < 0.001 Direct effect c’ 0.437 (0.331, 0.544) < 0.001 Total effect total 0.462 (0.355, 0.569) < 0.001 Proportion mediated 5.31% (2.68%, 9.14%) < 0.001 PCOS ( n  = 2507) BMI → FSH a -0.109 (-0.172, -0.046) 0.001 FSH → Oocyte b -0.168 (-0.220, -0.117) < 0.001 Indirect effect a×b 0.018 (0.006, 0.030) 0.003 Direct effect c’ 0.299 (0.192, 0.406) < 0.001 Total effect total 0.318 (0.210, 0.425) < 0.001 Proportion mediated 5.77% (2.41%, 12.62%) < 0.001 DOR ( n  = 2425) BMI → FSH a -0.307 (-0.415, -0.200) < 0.001 FSH → Oocyte b -0.076 (-0.119, -0.034) < 0.001 Indirect effect a×b 0.023 (0.008, 0.039) 0.002 Direct effect c’ 0.262 (0.120, 0.405) < 0.001 Total effect total 0.286 (0.145, 0.426) < 0.001 Proportion mediated 8.21% (3.78%, 25.43%) < 0.001 SEM was performed in the complete-case mediation analysis set: PCOS n  = 2,507; DOR n  = 2,425; tubal factor n  = 3,991 (total n  = 8,923). All estimates are standardized regression coefficients. CI = confidence interval. Proportion mediated and its 95% CI and p -value were derived from counterfactual-based mediation analysis with 5,000 bootstrap resamples Cohort-specific path coefficients and effect decomposition from SEM analysis Tubal ( n  = 3991) PCOS ( n  = 2507) DOR ( n  = 2425) SEM was performed in the complete-case mediation analysis set: PCOS n  = 2,507; DOR n  = 2,425; tubal factor n  = 3,991 (total n  = 8,923). All estimates are standardized regression coefficients. CI = confidence interval. Proportion mediated and its 95% CI and p -value were derived from counterfactual-based mediation analysis with 5,000 bootstrap resamples Likelihood ratio tests (Table  2 ) showed no significant differences in indirect effects across cohorts ( p  = 0.771). Direct and total effects showed marginal overall differences ( p  = 0.079 and p  = 0.069, respectively), which became non-significant after Holm correction. Table 2 Overall comparisons of effects across cohorts using likelihood ratio tests Effect Type Model Comparison Δχ² Δdf Raw p -value Holm-adjusted p Indirect effect Constrained vs. Free 0.519 2 0.771 0.771 Direct effect Constrained vs. Free 5.086 2 0.079 0.207 Total effect Constrained vs. Free 5.345 2 0.069 0.207 Overall comparisons of effects across cohorts using likelihood ratio tests Bootstrap-based pairwise comparisons with Holm correction revealed no significant differences in indirect effects between any cohort pairs (all adjusted p  = 1.000). For direct and total effects, tubal factor showed higher effects than DOR (raw p   0.05; Table  3 ). Table 3 Pairwise comparisons of indirect, direct, and total effects using bootstrap tests from multi-group SEM Effect Type Comparison Effect Difference (95% CI) Bootstrap p Holm-adjusted p Cohen’s d Indirect effect PCOS vs. Tubal -0.0063 (-0.0246, 0.0115) 0.4996 1.0000 0.98 PCOS vs. DOR -0.0052 (-0.0257, 0.0140) 0.5984 1.0000 0.74 Tubal vs. DOR 0.0011 (-0.0206, 0.0212) 0.8848 1.0000 0.15 Direct effect PCOS vs. Tubal -0.1385 (-0.2881, 0.0092) 0.0684 0.1368 2.55 PCOS vs. DOR 0.0372 (-0.1264, 0.2089) 0.6780 0.6780 0.60 Tubal vs. DOR 0.1757 (0.0046, 0.3461) 0.0448 0.1344 2.84 Total effect PCOS vs. Tubal -0.1447 (-0.2940, 0.0032) 0.0544 0.1284 2.66 PCOS vs. DOR 0.0320 (-0.1303, 0.2037) 0.7236 0.7236 0.52 Tubal vs. DOR 0.1767 (0.0062, 0.3474) 0.0428 0.1284 2.86 Results are from multi-group SEM with 5,000 bootstrap iterations. Effect values are standardized regression coefficients; 95% CIs are bootstrap percentile intervals. Standardized differences were calculated using the pooled bootstrap SD of effect estimates, sqrt((SD₁² + SD₂²)/2), and are not equivalent to traditional Cohen’s d Pairwise comparisons of indirect, direct, and total effects using bootstrap tests from multi-group SEM Results are from multi-group SEM with 5,000 bootstrap iterations. Effect values are standardized regression coefficients; 95% CIs are bootstrap percentile intervals. Standardized differences were calculated using the pooled bootstrap SD of effect estimates, sqrt((SD₁² + SD₂²)/2), and are not equivalent to traditional Cohen’s d Cohort-specific mediation analyses using both the counterfactual framework (Table S3) and the product-of-coefficients method (Table S4) yielded results consistent with the SEM analysis. Wald tests for overall and pairwise comparisons (Table S5) aligned with likelihood ratio and bootstrap results. Cohort interaction tests (Table S6) further supported the absence of significant moderation by diagnostic group. Model fit indices (Table S7) confirmed excellent model performance across all cohorts (CFI: 0.963–0.998; TLI: 0.900–0.994; RMSEA: 0.018–0.061; SRMR: 0.005–0.015). In the BMI-restricted sensitivity analysis, the indirect effect of BMI on oocyte yield via FSH was similar across cohorts (tubal factor: 0.029, P  = 0.003; PCOS: 0.030, P  = 0.046; DOR: 0.030, P  = 0.035), with no significant between-cohort difference ( P  = 0.994; Supplementary Table S8). In a sensitivity analysis restricted to pretreatment confounders, the indirect effect via FSH remained significant in all three cohorts, whereas direct and total effects were significant only in PCOS (Supplementary Table S9). After confirming the absence of significant interactions between BMI and cohort for all outcomes (clinical pregnancy: P for interaction = 0.152; pregnancy loss: P  = 0.144; live birth: P  = 0.095), data from the three cohorts were pooled for subsequent analyses. The pooled dataset comprised 6,708 women for clinical pregnancy analysis, 4,182 for pregnancy loss analysis, and 6,523 for live birth analysis. As shown in Table  4 , after adjustment for potential confounders, obese women had significantly higher odds of pregnancy loss compared with normal-weight women (OR = 1.60, 95% CI: 1.02–2.53, P  = 0.041). Effect size analysis revealed an absolute risk difference of 5.4% for obese women, corresponding to a number needed to harm (NNH) of 19, indicating that for every 19 obese women, one additional pregnancy loss occurs compared to normal-weight women. Although not reaching statistical significance, a dose-response pattern was observed across BMI categories: underweight (OR = 0.91, 95% CI: 0.65–1.27, P  = 0.567), overweight (OR = 1.23, 95% CI: 0.91–1.66, P  = 0.183), and obese (OR = 1.60, 95% CI: 1.02–2.53, P  = 0.041), suggesting a progressive association with pregnancy loss risk across increasing BMI categories. For live birth, no statistically significant differences were observed across BMI categories, although a trend toward reduced live birth rate was noted in obese women (OR = 0.77, 95% CI: 0.57–1.03, P  = 0.078) (Table  4 ). Effect size analysis showed an absolute risk reduction of 2.8% for obese women (NNT = 36) and 3.4% for overweight women (NNT = 29), suggesting clinically meaningful but statistically non-significant effects. No significant associations were found between BMI and clinical pregnancy rate across any BMI category (all P  > 0.05, Table  4 ). Effect size analysis confirmed minimal absolute risk differences across groups. Table 4 Association between BMI categories and reproductive outcomes in the pooled cohort BMI Category Clinical Pregnancy ( n  = 6708) Pregnancy loss ( n  = 4182) Live Birth ( n  = 6523) Underweight (< 18.5 kg/m²)  n / total (%) 412/639 (64.5) 53/395 (13.4) 342/622 (55.0)  Adjusted OR (95% CI) 1.02 (0.84–1.24) 0.91 (0.65–1.27) 1.05 (0.87–1.26)   P value 0.822 0.567 0.616  Absolute risk difference (%, 95% CI) -0.83 (-4.54 to 2.88) -0.97 (-4.34 to 2.39) -0.16 (-4.07 to 3.75)  NNT/NNH NNT = 120 NNT = 103 NNT = 625 Normal weight (18.5–24.0 kg/m²)  n / total (%) 3230/4946 (65.3) 447/3106 (14.4) 2659/4822 (55.1)  Adjusted OR (95% CI) 1.00 (reference) 1.00 (reference) 1.00 (reference)  Absolute risk difference (%, 95% CI) 0 (reference) 0 (reference) 0 (reference) Overweight (24.0–28.0 kg/m²)  n / total (%) 537/839 (64.0) 89/508 (17.5) 419/810 (51.7)  Adjusted OR (95% CI) 1.05 (0.87–1.27) 1.23 (0.91–1.66) 0.95 (0.79–1.14)   P value 0.600 0.183 0.558  Absolute risk difference (%, 95% CI) -1.30 (-4.55 to 1.95) 3.13 (-0.18 to 6.44) -3.41 (-6.86 to 0.03)  NNT/NNH NNT = 77 NNH = 32 NNT = 29 Obese (≥ 28.0 kg/m²)  n / total (%) 181/270 (67.0) 33/167 (19.8) 134/256 (52.3)  Adjusted OR (95% CI) 0.97 (0.72–1.32) 1.60 (1.02–2.53) 0.77 (0.57–1.03)   P value 0.867 0.041 0.078  Absolute risk difference (%, 95% CI) 1.73 (-3.89 to 7.35) 5.37 (-0.69 to 11.43) -2.80 (-8.93 to 3.33)  NNT/NNH NNH = 58 NNH = 19 NNT = 36  P for interaction with cohort P  = 0.152 P  = 0.144 P  = 0.095 All models adjusted for age, infertility type, duration of infertility, treatment protocol, gonadotropin dose, estradiol and progesterone levels on hCG trigger day, insemination method, AMH, number of normal pronuclei, number of good-quality embryos, embryo type, transfer type, and number of good-quality embryos transferred. For pregnancy loss, NNH represents the number of obese women needed to observe one additional pregnancy loss compared to normal weight. For live birth and clinical pregnancy, NNT represents the number needed to treat to achieve one additional event. Bold values indicate statistical significance ( P  < 0.05). Abbreviations: NNT, number needed to treat; NNH, number needed to harm. For live birth, the analysis was underpowered (5.3%); the NNT is presented for descriptive purposes only and should not be used for clinical inference Association between BMI categories and reproductive outcomes in the pooled cohort All models adjusted for age, infertility type, duration of infertility, treatment protocol, gonadotropin dose, estradiol and progesterone levels on hCG trigger day, insemination method, AMH, number of normal pronuclei, number of good-quality embryos, embryo type, transfer type, and number of good-quality embryos transferred. For pregnancy loss, NNH represents the number of obese women needed to observe one additional pregnancy loss compared to normal weight. For live birth and clinical pregnancy, NNT represents the number needed to treat to achieve one additional event. Bold values indicate statistical significance ( P  < 0.05). Abbreviations: NNT, number needed to treat; NNH, number needed to harm. For live birth, the analysis was underpowered (5.3%); the NNT is presented for descriptive purposes only and should not be used for clinical inference In supplementary analyses treating BMI as a continuous variable, consistent patterns were observed: log-BMI was significantly associated with pregnancy loss (OR = 2.47, 95% CI: 1.07–5.70, P  = 0.034) but not with clinical pregnancy (OR = 1.58, 95% CI: 0.95–2.62, P  = 0.079) or live birth (OR = 0.97, 95% CI: 0.59–1.58, P  = 0.889) (Table S10). Post-hoc power analysis revealed that the statistical power to detect the observed associations was 70.2% for pregnancy loss, 57.2% for clinical pregnancy, and only 5.3% for live birth (Table S10). Sensitivity analyses across different R² assumptions showed that the minimum detectable OR for miscarriage ranged from 2.63 to 3.00 (Table S11). In a sensitivity analysis restricted to pretreatment confounders, the association between obesity and pregnancy loss remained in the same direction, with a borderline significant increase in risk (OR = 1.55, 95% CI 0.99–2.41, P  = 0.055; Table S12).

This retrospective cohort study included women who underwent their first in vitro fertilization/intracytoplasmic sperm injection (IVF/ICSI) cycle following controlled ovarian stimulation at Jiangxi Maternal and Child Health Hospital between January 2015 and December 2025. The study was approved by the reproductive Medicine Ethics Committee of Jiangxi Maternal and Child Health Hospital (Approval No. SZYY-202604), with informed consent waived due to its retrospective nature. All cycles meeting the following criteria were initially screened: first IVF/ICSI cycle; no use of donor sperm, donor oocytes, or preimplantation genetic testing; no azoospermia (obstructive or non-obstructive) in male partners; no chromosomal abnormalities in either partner; oocyte retrieval with yield > 0; and non-natural cycles. Cycles were excluded if any of the following conditions were present: Metabolic and endocrine disorders : type 1 or 2 diabetes; medications affecting metabolism (statins, fibrates, metformin, oral contraceptives within 3 months); severe hepatic or renal dysfunction; thyroid dysfunction; pituitary or adrenal abnormalities; autoimmune diseases; other endocrine disorders (congenital adrenal hyperplasia, Cushing’s syndrome, uncontrolled thyroid disorders, hyperprolactinemia, androgen-secreting tumors). Reproductive system abnormalities : severe male factor requiring donor sperm; uterine structural abnormalities; history of gestational trophoblastic disease; endometriosis stage III-IV; history of ovarian surgery; endometrioma; pelvic radiotherapy or chemotherapy. Pregnancy history : recurrent miscarriage. Metabolic and endocrine disorders : type 1 or 2 diabetes; medications affecting metabolism (statins, fibrates, metformin, oral contraceptives within 3 months); severe hepatic or renal dysfunction; thyroid dysfunction; pituitary or adrenal abnormalities; autoimmune diseases; other endocrine disorders (congenital adrenal hyperplasia, Cushing’s syndrome, uncontrolled thyroid disorders, hyperprolactinemia, androgen-secreting tumors). Reproductive system abnormalities : severe male factor requiring donor sperm; uterine structural abnormalities; history of gestational trophoblastic disease; endometriosis stage III-IV; history of ovarian surgery; endometrioma; pelvic radiotherapy or chemotherapy. Pregnancy history : recurrent miscarriage. Three cohorts were defined based on distinct ovarian phenotypes and metabolic profiles: ①  Tubal factor cohort (negative control) required: (1) tubal infertility confirmed by hysterosalpingography or laparoscopy; (2) normal ovarian reserve (AMH ≥ 1.1 ng/mL, FSH  7); and (3) strict metabolic normality (BMI < 24 kg/m², and all lipid parameters within normal range: TC < 5.2, LDL < 3.4, HDL ≥ 1.0, TG  37 years, AMH < 1.1, FSH ≥ 10, AFC ≤ 7), or any metabolic disturbance (BMI ≥ 24 or any lipid abnormality). ②  PCOS cohort required: (1) PCOS diagnosis by Rotterdam criteria; and (2) high ovarian reserve (AMH ≥ 1.1 ng/mL, FSH  7). Metabolic disturbances were permitted. Exclusion criteria included other ovarian disorders (endometriosis, diminished reserve) and conditions severely confounding ovarian response as listed in general exclusion criteria. ③  DOR cohort required at least two of: age ≥ 35 years, AMH < 1.1 ng/mL, AFC < 7, or FSH ≥ 10 IU/L, with age ≤ 40 years to exclude advanced age as the dominant driver of diminished reserve. Metabolic disturbances were permitted. Exclusion criteria included PCOS, premature ovarian insufficiency (POI), premature ovarian failure (POF), or poor ovarian response (POR) according to Bologna criteria if not meeting the above DOR criteria. ①  Tubal factor cohort (negative control) required: (1) tubal infertility confirmed by hysterosalpingography or laparoscopy; (2) normal ovarian reserve (AMH ≥ 1.1 ng/mL, FSH  7); and (3) strict metabolic normality (BMI < 24 kg/m², and all lipid parameters within normal range: TC < 5.2, LDL < 3.4, HDL ≥ 1.0, TG  37 years, AMH < 1.1, FSH ≥ 10, AFC ≤ 7), or any metabolic disturbance (BMI ≥ 24 or any lipid abnormality). ②  PCOS cohort required: (1) PCOS diagnosis by Rotterdam criteria; and (2) high ovarian reserve (AMH ≥ 1.1 ng/mL, FSH  7). Metabolic disturbances were permitted. Exclusion criteria included other ovarian disorders (endometriosis, diminished reserve) and conditions severely confounding ovarian response as listed in general exclusion criteria. ③  DOR cohort required at least two of: age ≥ 35 years, AMH < 1.1 ng/mL, AFC < 7, or FSH ≥ 10 IU/L, with age ≤ 40 years to exclude advanced age as the dominant driver of diminished reserve. Metabolic disturbances were permitted. Exclusion criteria included PCOS, premature ovarian insufficiency (POI), premature ovarian failure (POF), or poor ovarian response (POR) according to Bologna criteria if not meeting the above DOR criteria. Finally, 11,524 women were included: PCOS ( n  = 2,881), DOR ( n  = 4,354), and tubal factor infertility ( n  = 4,289). Mediation analysis included 8,923 women with complete data (PCOS n  = 2,507; DOR n  = 2,425; tubal factor n  = 3,991). Reproductive outcome analyses included clinical pregnancy ( n  = 6,708), pregnancy loss ( n  = 4,182), and live birth ( n  = 6,523) (Figure S1). Ovarian stimulation was individualized based on age, BMI, and ovarian reserve. Protocols included GnRH agonist (long, ultra-long, or short), GnRH antagonist, and other regimens (e.g., mild stimulation or progestin-primed ovarian stimulation, PPOS). Trigger criteria were at least 2–3 follicles ≥ 18 mm, with trigger methods including hCG, GnRH agonist, or both. Oocyte retrieval was performed 34–36 h post-trigger. IVF or ICSI was performed according to standard indications. Embryos were cultured to day 3 or day 5/6, and the number of embryos transferred was determined based on embryo quality, age, and previous cycle history. Primary outcome was oocyte yield. Secondary outcomes, analyzed only in fresh embryo transfer cycles, included: (1) clinical pregnancy, defined as the presence of a gestational sac on ultrasound 4–5 weeks after embryo transfer; (2) pregnancy loss rate = (Number of spontaneous or therapeutic abortions / Number of clinical pregnancies) × 100%; for this composite outcome, therapeutic abortion included medical or surgical treatment of ectopic pregnancy; and (3) live birth, defined as delivery of at least one live infant after 28 weeks of gestation. The mediating variable was basal FSH, measured on day 2–3 of the menstrual cycle. Missing data were not imputed because the extent and pattern of missingness differed across the three cohorts, and a uniform imputation approach was considered inappropriate. Complete-case analysis was used for mediation models, and available-case analysis was used for descriptive and outcome-specific analyses. Baseline variables across cohorts were compared using ANOVA or Kruskal-Wallis tests for continuous variables and chi-square or Fisher’s exact tests for categorical outcomes, as appropriate. Continuous data were presented as mean ± SD or median [IQR] based on normality (Shapiro-Wilk test). Pearson correlation coefficients were calculated to examine pairwise relationships among clinical variables within each cohort. Multivariable linear regression models were fitted within each cohort to assess associations of log-transformed BMI and FSH with log-transformed oocyte yield, adjusting for potential confounders (age, logAMH, infertility type, log infertility duration, treatment protocol, log Gn dose, log E 2 and log P on hCG day, log baseline E 2 , log HDL, and log baseline LH). Before conducting mediation analysis, restricted cubic spline regression was employed to examine the nonlinear relationship between BMI and oocyte yield. The results supported the linearity assumption, justifying the use of linear regression models in the mediation framework. A comprehensive mediation analysis was conducted to examine whether FSH mediated the association between BMI and oocyte yield across three groups. First, cohort-specific mediation analyses were performed using three complementary approaches to ensure robustness: (1) the product-of-coefficients method with bias-corrected bootstrap resampling (5,000 iterations); (2) structural equation modeling (SEM) to assess overall model fit; and (3) counterfactual-based mediation analysis to decompose the total association into indirect and direct components. Second, to test whether mediation pathways differed across cohorts, we employed two complementary approaches: (a) multi-group SEM with likelihood ratio tests and Wald tests for overall comparisons of indirect, direct, and total effects; and (b) interaction tests within the product-of-coefficients bootstrap framework to assess whether the mediated effect varied by diagnostic group. Third, pairwise comparisons between cohorts were conducted using both Wald tests and parametric bootstrap methods. All multiple comparisons were adjusted using the Holm procedure to control for Type I error. Given the retrospective observational design, mediation analysis was used to decompose statistical associations, and the results should not be interpreted as proof of causality. Residual confounding cannot be excluded. Given the broad BMI range across the three cohorts, an additional sensitivity analysis was performed restricting all cohorts to BMI 18.5–24 kg/m². In a sensitivity analysis, the mediation model was restricted to pretreatment confounders (age, infertility type, duration of infertility, baseline AMH, baseline E₂, baseline LH, and baseline HDL) to evaluate the potential impact of overadjustment by stimulation-related variables. The mediation analysis assumes no important unmeasured confounding of the BMI–FSH, FSH–oocyte yield, and BMI–oocyte yield relationships. This assumption cannot be fully verified in a retrospective observational study. To evaluate the association between body mass index (BMI) and reproductive outcomes (clinical pregnancy, miscarriage, and live birth), data from the three cohorts were pooled after confirming the absence of effect modification by cohort. Interaction between BMI and cohort was tested using likelihood ratio tests for both continuous and categorical specifications; non-significant interactions ( P  > 0.05) justified pooling. ① Primary analysis: BMI as a categorical variable . To enhance clinical interpretability, BMI was classified according to Chinese criteria: underweight (< 18.5 kg/m²), normal weight (18.5–24.0 kg/m², reference), overweight (24.0–28.0 kg/m²), and obese (≥ 28.0 kg/m²). Multivariable logistic regression models were fitted adjusting for potential confounders. Effect sizes were quantified using odds ratios (OR) with 95% confidence intervals (CI), absolute risk differences, and numbers needed to treat/harm ②  Supplementary analysis: BMI as a continuous variable . To assess linear dose-response relationships and ensure robustness of findings, log-transformed BMI was additionally analyzed as a continuous predictor in multivariable logistic regression models. For non-significant associations, post-hoc power analyses were performed to determine the minimum detectable OR at 80% power; for significant associations, sensitivity analyses were conducted to evaluate result stability. To address the potential for overadjustment by treatment-related variables, a sensitivity analysis restricted to pretreatment confounders ① Primary analysis: BMI as a categorical variable . To enhance clinical interpretability, BMI was classified according to Chinese criteria: underweight (< 18.5 kg/m²), normal weight (18.5–24.0 kg/m², reference), overweight (24.0–28.0 kg/m²), and obese (≥ 28.0 kg/m²). Multivariable logistic regression models were fitted adjusting for potential confounders. Effect sizes were quantified using odds ratios (OR) with 95% confidence intervals (CI), absolute risk differences, and numbers needed to treat/harm ②  Supplementary analysis: BMI as a continuous variable . To assess linear dose-response relationships and ensure robustness of findings, log-transformed BMI was additionally analyzed as a continuous predictor in multivariable logistic regression models. For non-significant associations, post-hoc power analyses were performed to determine the minimum detectable OR at 80% power; for significant associations, sensitivity analyses were conducted to evaluate result stability. To address the potential for overadjustment by treatment-related variables, a sensitivity analysis restricted to pretreatment confounders All statistical analyses were performed using R software (version 4.3.2). A two-tailed P  < 0.05 was considered statistically significant unless otherwise specified for multiple comparisons.

This study leveraged three large cohorts with distinct ovarian reserve (DOR to PCOS) and metabolic profiles to examine how BMI affects ART outcomes. A key finding is the remarkable consistency across cohorts: no significant cohort-by-BMI interactions were detected for any outcome, indicating that the associations of BMI with outcomes were consistent across infertility etiologies. Specifically, we found that: (1) BMI exerted a small but significant indirect effect on oocyte yield via FSH mediation (proportion mediated: 5.31–8.21%), which was consistent across all cohorts; and (2) regarding reproductive outcomes, obesity was significantly associated with increased pregnancy loss risk (OR = 1.60), showed a trend toward lower live birth rates, but did not affect clinical pregnancy. Previous studies have reported conflicting findings on the relationship between BMI and oocyte yield. One study of 22,392 patients undergoing their first PPOS cycle found no difference across BMI categories [ 12 ], but was confined to a single protocol. Another study of 3,852 patients undergoing 5,190 GnRH antagonist or PPOS cycles—which included repeated cycles from the same couples—also reported no association [ 13 ]. Smaller studies have yielded inconsistent results, with some showing no association and others reporting increased oocyte yield in higher BMI groups [ 14 – 17 ]. Comparison with previous large-scale studies highlights a key distinction: our study included diverse protocols (agonist, antagonist, PPOS, mild stimulation), whereas prior studies were largely confined to single protocols. This suggests that treatment protocols may modify the effect of BMI on oocyte yield, explaining the heterogeneity in conclusions. Moreover, those studies were observational and did not explore mechanisms. In contrast, our mediation analysis identified a small indirect association between BMI and oocyte yield via basal FSH. Because the proportion mediated was only 5.31–8.21% and confidence intervals were wide, this finding should be regarded as exploratory and does not establish a mechanism. Unlike conventional regression, which cannot assess mediation pathways, our SEM approach demonstrated excellent model fit, validating the correctness of our models. Additionally, the consistency of results across three mediation methods—counterfactual framework, product-of-coefficients, and SEM—further reinforces the robustness of our findings. However, these findings reflect statistical mediation rather than a causal biological pathway; a good SEM fit and consistency across methods support the robustness of the statistical association but cannot confirm an underlying mechanism, and basal FSH may also reflect ovarian reserve rather than solely mediating the BMI–oocyte yield association. The relationship between BMI and FSH remains controversial. Several studies have reported an inverse association between FSH levels and BMI in reproductive-aged women, while others have found positive correlations [ 18 ]. Data specifically from IVF cohorts are limited; one study of 326 cycles reported an inverse BMI-FSH association [ 18 ]. Our large cohort study supports an inverse relationship between BMI and basal FSH in reproductive-aged women. The underlying mechanisms warrant further investigation. The relationship between BMI and fresh embryo transfer outcomes remains controversial. A large study ( n = 30,746, PCOS excluded) found no differences across BMI categories, though group sizes varied considerably [ 19 ]. Another study of 5,981 low-prognosis women reported comparable pregnancy rates but increased miscarriage and decreased live birth in obese women [ 20 ]. A PCOS study ( n = 1,704) found no outcome differences [ 21 ], while smaller studies yielded inconsistent results [ 22 , 23 ]. Notably, none of these studies performed power or effect size analyses [ 24 ], leaving the reliability of their findings uncertain. In contrast, our study ( n = 6,708/4,182/6,523) combined logistic regression with power and effect size analyses. For clinical pregnancy, the effect size was minimal, suggesting that even statistical significance would lack clinical relevance. For live birth, the association was not statistically significant ( P = 0.078), and post-hoc power was only 5.3%. This analysis was therefore underpowered and inconclusive. The observed NNT should not be interpreted as evidence of clinical meaningfulness. It is important to note that many previous studies have used the standard definition of miscarriage, whereas our outcome was a composite pregnancy loss rate including spontaneous miscarriage, ectopic pregnancy, and therapeutic abortion. This definitional difference should be considered when comparing effect sizes. For pregnancy loss, BMI as a continuous variable was significantly associated with increased risk (OR = 2.47, 95% CI: 1.07–5.70, P  = 0.034). As a categorical variable, obese women showed significantly higher pregnancy loss risk (OR = 1.60, 95% CI: 1.02–2.53, P  = 0.041), with a progressive risk increase across BMI categories (underweight: OR = 0.91; overweight: OR = 1.23; obese: OR = 1.60). Taken together, these findings consistently support an association between obesity and pregnancy loss risk, and the integration of power and effect size analyses enables a more comprehensive interpretation of both statistical and clinical significance. Our findings have implications for clinical counseling. For obese patients, the focus should shift from BMI’s effect on oocyte yield to its impact on subsequent reproductive outcomes. For obese patients, counseling may focus on the consistent association with increased pregnancy loss risk (NNH = 19). Because the live birth analysis was underpowered, the trend toward lower live birth should not be used for clinical counseling. Weight management remains reasonable, but conclusions regarding live birth require confirmation in larger studies. Moreover, this increased risk was consistent across PCOS, DOR, and tubal factor cohorts, indicating that weight management is important not only for PCOS but also for other infertility etiologies. This study has several strengths. First, the large sample size and well-defined cohorts encompassing a spectrum of ovarian reserve (DOR to PCOS) and metabolic profiles enhance the representativeness of our findings. This design enabled us to demonstrate that associations of BMI with ART outcomes are remarkably conserved across distinct infertility etiologies, underscoring the fundamental role of metabolic status in reproductive function. Second, the combination of mediation analysis and outcome analysis provided a comprehensive evaluation of BMI’s impact on both oocyte yield (quantity) and reproductive outcomes (quality). Third, multiple statistical approaches consistently validated the robustness of our conclusions from different perspectives. Several limitations should be acknowledged. This was a retrospective, single-center study; residual confounding, selection bias from complete-case analysis, and the observational design mean that causality cannot be established. Cross-cohort comparability was limited because the tubal factor cohort was restricted to BMI < 24 kg/m² and basal FSH is an inherent component of the DOR definition. The pregnancy loss outcome was a composite of spontaneous miscarriage, ectopic pregnancy, and therapeutic abortion, which differs from the standard definition of spontaneous miscarriage. The live birth analysis was underpowered (5.3%), and the proportion mediated by FSH was small with wide confidence intervals; the mediation analysis should therefore be considered exploratory. Some variables in the mediation and reproductive outcome models may lie on the pathway between BMI and outcomes, so these models estimate conditional associations; a sensitivity analysis restricted to pretreatment confounders supported the robustness of the indirect association, but direct and total effects were significant only in PCOS. Reproductive outcomes were analyzed only in fresh embryo transfer cycles, and pregnancy loss was evaluated among women who achieved a clinical pregnancy, so estimates reflect risk among clinical pregnancies after fresh transfer. Finally, NNT/NNH values were based on crude absolute risk differences and should be interpreted cautiously. In conclusion, BMI was associated with oocyte yield through a small FSH-related indirect pathway (proportion mediated 5.31–8.21%, wide confidence intervals); this finding is exploratory. BMI was also associated with increased pregnancy loss risk in the three cohorts, although cross-cohort comparability was limited. The live birth analysis was underpowered and inconclusive. In pre-IVF counseling, the NNH of 19 for pregnancy loss may help set realistic expectations and encourage weight optimization. These findings do not establish causality and require confirmation in larger prospective studies.

Metabolic disturbances are closely associated with ovarian dysfunction [ 1 – 3 ], most prominently in women with polycystic ovary syndrome (PCOS) [ 4 , 5 ]. However, the underlying mechanisms by which metabolic factors influence ovarian function remain poorly understood [ 6 ]. More importantly, whether metabolic disorders modify or exacerbate pre-existing ovarian dysfunction—such as the hyperandrogenic high-reserve state in PCOS or the age-related diminished reserve in diminished ovarian reserve (DOR)—has not been systematically investigated. Recent genetic evidence, including our own Mendelian randomization study, suggests that body mass index (BMI) may influence ovarian function by downregulating basal follicle-stimulating hormone (FSH) levels. As FSH is a key driver of follicular recruitment and development [ 7 ], changes in its levels may mediate the effect of obesity on ovarian response [ 8 – 11 ]. However, this hypothesis has not been validated in clinical cohorts, nor has it been compared across populations with distinct ovarian functional statuses. To address this question, we designed a study encompassing three well-defined cohorts with distinct ovarian phenotypes: a PCOS cohort (high ovarian reserve with permissible metabolic disturbances), a DOR cohort (low ovarian reserve with permissible metabolic disturbances), and a tubal factor cohort (normal ovarian reserve with strictly defined metabolic normality, serving as a negative control). This design allows us to determine whether the influence of metabolic factors on ovarian function and reproductive outcomes varies by “disease context”—or conversely, whether it operates through conserved pathways independent of the underlying infertility etiology.

Supplementary Material 1. Supplementary Material 1.