This retrospective cohort study was conducted at the Reproductive Medicine Center of Shenzhen Hengsheng Hospital in accordance with the established guidelines for observational research in the field of reproductive medicine. The study population consisted of women diagnosed with OMAs who underwent ultrasound-guided EST before their IVF/ICSI cycle between January 2020 and December 2024. The study protocol was approved by the hospital’s Institutional Review Board (HSYY2022-12-22) and adhered to the principles outlined in the Declaration of Helsinki for medical research involving human subjects. Owing to the retrospective design of the study, the requirement for informed consent was waived, and appropriate data anonymization procedures were implemented. The reporting of this manuscript adheres to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement for prediction model development and validation (18).

The primary outcome was the CLBR, defined as the delivery of a live-born infant with a gestational age of at least 28 weeks resulting from all fresh and frozen-thawed embryo transfers following the first completed oocyte retrieval cycle after EST (19). The CLBR was determined for each patient, considering embryo transfers until a live birth was achieved or all embryos from the initial cycle were utilized. The follow-up period extended from oocyte retrieval to either live birth or final embryo transfer, with a minimum duration of 12 months, to adequately capture reproductive outcomes.

Sample size calculation followed the 10 events per variable (EPV) guideline for the logistic regression models (25). With an expected live birth rate of 50% and 5–8 predictor variables, 50–80 events (100–160 patients) were needed. Accounting for data exclusion and 20% follow-up loss, the target was set at 125–200 patients. According to the EPV principle, our sample size of 194 patients, which included 97 live-birth events, yielded an EPV of 19.4. This was sufficient for developing a prediction model that incorporated the five variables.

Continuous variables were described using the mean ± standard deviation (SD) for data with a normal distribution or median with interquartile range (IQR) for data not normally distributed, with normality checked via the Shapiro-Wilk test and visual inspection of histograms and Q-Q plots. Categorical variables were expressed as frequency and percentage. Comparisons between groups were conducted using Student’s t-test or the Mann-Whitney U test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables, with effect sizes reported as Cohen’s d for continuous variables and Cramér’s V for categorical variables. We used the R software (version 4.2.2) for all statistical analyses. This included packages such as rms for regression modeling and creating nomograms, glmnet for LASSO regression, pROC for ROC analysis, and mice for multiple imputation. SPSS version 26.0 (IBM Corp., Armonk, NY, USA) was used for the supplementary analyses. All analyses employed a two-sided significance level of P < 0.05, and random seeds were used to ensure the reproducibility of the results.

This study retrospectively analyzed 220 patients with OMAs who underwent ultrasound-guided EST followed by IVF/ICSI between January 2020 and December 2024 at our institution. In accordance with the stringent inclusion and exclusion criteria, 26 patients were excluded from the study: 11 were lost to follow-up, 5 had incomplete data, 7 presented with severe male factor infertility, and 3 did not undergo IVF/ICSI. Consequently, 194 patients were deemed eligible for the final analysis (Figure 1). Patients were randomly split 7:3 into a training cohort (n=135, 70%) for model development and a validation cohort (n=59, 30%) for internal validation, establishing a statistical foundation. The median age of the study population was 33.0 years, with a median AMH level of 2.83 ng/mL, eight oocytes retrieved, and four usable embryos produced. Among all baseline variables, only bT (training: 28.22 ng/dL vs. validation: 22.75 ng/dL, P = 0.045) and AFC (training: 11.00 vs. validation: 9.00, P = 0.030) showed statistically significant differences. All other variables, including age, BMI, infertility features, cyst parameters, COH protocols, and embryological outcomes, were comparable (P>0.05) (Table 1). This confirmed successful randomization and a good baseline balance, providing a reliable dataset for the subsequent model construction.

We conducted a univariate logistic regression analysis to evaluate the association between each candidate variable and CLBR. Seventy-one variables spanning six clinical domains were assessed. Using a liberal screening threshold of P < 0.10 to avoid prematurely discarding potentially informative predictors, 26 variables showed statistically significant associations with CLBR (Table 2). The complete univariate results for all 71 candidate variables are presented in Supplementary Table S1. From the 26 identified variables, the LASSO regression analysis, utilizing 10-fold cross-validation, determined an optimal λ value of 0.0807 (Figures 2A, B). Consequently, nine variables were retained: female age, male age, AFC, previous live birth history, COH protocol, number of oocytes retrieved, number of cryopreserved embryos, LH level on HCG trigger day, and cyst diameter. The nine variables were further refined using multivariate logistic regression with backward elimination. This process yielded a streamlined final model comprising four independent predictors: previous live birth history, COH protocol, total number of oocytes retrieved, and endometrioma diameter (Table 3).

We constructed a nomogram to predict the CLBR in women with OMAs after EST and IVF/ICSI treatment. The nomogram included previous live birth history, COH protocol, cyst diameter, and number of oocytes retrieved (Figure 3). To calculate individualized CLBR prediction probabilities using the nomogram, first, find values for the five predictors and read their scores on the “Points” scale. Then, sum these scores to get “Total Points.” Finally, the total points were projected downward to the “Diagnostic Possibility” scale to determine the predicted CLBR probability. Model calibration showed a clear link between total points and the predicted CLBR probability. A total score of 90 corresponded to a 10% CLBR, 115 to 20%, 131 to 30%, 145 to 40%, 157 to 50%, 170 to 60%, 183 to 70%, 199 to 80%, and 224 to 90%. This probability gradient helps doctors predict patient outcomes, provide advice, and set realistic treatment goals and decisions.

The discriminatory capacity of the nomogram was assessed using ROC curve analysis in both datasets to differentiate between patients who achieved and did not achieve cumulative live birth. The nomogram showed good discrimination, with an AUC of 0.849 (95% CI: 0.782-0.912) in the training dataset (Figure 4A) and 0.853 (95% CI: 0.754-0.952) in the validation dataset (Figure 4B). The AUC in the validation set was higher than that in the training set, suggesting minimal overfitting and good model generalizability. Overlapping CIs indicate the model’s stable discriminative ability. The DeLong test indicated no statistically significant difference between the AUCs of the training and validation sets (P = 0.9527), suggesting consistent discriminative performance across the datasets. Bootstrap resampling validation with 500 iterations was performed for both cohorts. The Bootstrap-validated AUC was 0.877 (95% CI: 0.817-0.937) in the training cohort and 0.879 (95% CI: 0.792-0.967) in the validation cohort, as shown in Figures 4C, D. Bootstrap CIs aligned with the original ROC analysis confirmed the robustness of the model discrimination. The Bootstrap AUC in the validation cohort exceeded that in the training cohort, demonstrating strong generalizability without overfitting.

Calibration analysis was performed to evaluate the agreement between the predicted probabilities and the observed outcomes. Calibration plots were constructed to visually assess the relationship between the predicted and actual CLBR. In the training set, the model demonstrated good calibration, with the calibration curve generally aligned with the ideal line. The key calibration metrics included a Brier score of 0.151, an intercept of 0.000, and a slope of 1.000, indicating accurate probability estimates without any systematic bias. Furthermore, the Hosmer-Lemeshow goodness-of-fit test yielded a non-significant result (P = 0.415), supporting the conclusion that there was no statistically significant difference between the predicted and observed CLBR. In the validation set, the model maintained a good calibration (Figure 5B). The Brier score was 0.158, which is comparable to that of the training set and indicates a low prediction error. The calibration intercept was -0.549 with a slope of 0.915, indicating a slight tendency to overestimate risk at lower probabilities and underestimate it at higher ones; nonetheless, the overall performance remained robust. Crucially, the Hosmer-Lemeshow test was also non-significant in the validation cohort (P = 0.083), confirming that the model’s predictions were well-calibrated and reliable for external data.

DCA was conducted to evaluate the clinical utility and net benefit of using the nomogram for decision-making in both the training and validation sets (Figure 6). DCA determines the range of threshold probabilities at which the model provides a greater net benefit than the default strategies of treating all or none of the patients. In the training set (Figure 6A), the nomogram demonstrated a significant net benefit across a wide range of threshold probabilities, from approximately 5% to 90%. This indicates that if a clinician’s threshold for intervention lies within this range, using the nomogram to guide decisions is superior to “treat all” and “treat none” patients. The model achieved a peak net benefit of approximately 0.48 at a 10% threshold. Even at more moderate thresholds, the benefit remained substantial; for instance, at a threshold probability of 25%, the net benefit was approximately 0.40, and at 50%, it was approximately 0.30. This means that at a 25% threshold, using the model is equivalent to a strategy that correctly identifies 40 additional CLBR cases per 100 patients without increasing unnecessary interventions compared with the default strategies. In the validation set (Figure 6B), the clinical value of the model was confirmed. The nomogram provided a positive net benefit for threshold probabilities ranging from approximately 5%–85%. The net benefit curve for the nomogram was consistently higher than that of both the “treat all” and “treat none” strategies within this range. The peak net benefit in the validation cohort was approximately 0.38, observed at a low threshold of approximately 5-10%. At the 25% threshold, the net benefit was approximately 0.25, and at the 40% threshold, it was approximately 0.20. These results robustly demonstrate that the nomogram is not only statistically accurate but also clinically practical, offering tangible benefits for guiding personalized treatment decisions in women with OMAs undergoing IVF/ICSI.

In this study, we developed and validated a comprehensive nomogram to predict CLBR in women with OMAs undergoing EST and IVF/ICSI. Our model incorporated four clinical variables: previous live birth history, cyst diameter, COH protocol, and number of oocytes retrieved, showing good discriminative ability with AUCs of 0.849 and 0.853 in the training and validation sets, respectively. It demonstrated good calibration and clinical utility for individualized fertility counseling in the study. Bootstrap resampling validation (500 iterations) further confirmed the robustness of the model, with Bootstrap AUCs of 0.877 and 0.879 in the training and validation sets, respectively. The Bootstrap CIs were consistent with the original analysis, demonstrating reliable model performance estimates. The development of this nomogram addresses a critical gap, as women with OMAs undergoing ovary-sparing EST represent a unique patient subset. Lavadia et al. (9) confirmed that EST is a viable treatment option that preserves the ovarian reserve and improves IVF/ICSI outcomes compared to cystectomy. Olawade et al. (26) demonstrated that AI-driven personalized treatment protocols for IVF could optimize ovarian stimulation and improve clinical outcomes using individualized approaches. Our model builds on this by providing a tailored prognostic tool that quantifies individual success probabilities, moving beyond generalized treatment recommendations to facilitate precise, patient-centered, and clinical decision-making. Moreover, our model can be integrated into clinical workflows at various decision points, ranging from the initial consultation to treatment monitoring, thereby facilitating the advancement of precision in reproductive medicine.

This nomogram serves as a practical tool for translating complex statistical data into individualized clinical guidance, fostering shared decision-making for women with OMAs undergoing IVF/ICSI after EST. Its primary utility lies in pre-treatment counseling, where it provides a quantitative estimate of the CLBR to manage patient expectations. Based on our DCA results, a threshold probability of 30% was a reasonable benchmark. For patients with a predicted CLBR above this threshold, clinicians can offer reassurance and encourage proceeding with IVF/ICSI treatment. Conversely, for those with a prognosis below 30%, the model facilitates crucial discussions on alternative strategies. This may include optimizing the COH protocol, considering a repeat EST for larger cysts, or realistically counseling on the potential need for multiple retrieval cycles. It is imperative to underscore that this nomogram was designed as a decision-support tool to augment, not supplant, clinical judgment. The predictions provided by the model should be integrated with the clinician’s experience, the patient’s individual circumstances, and shared decision-making. It serves as an objective reference to facilitate more personalized counseling and expectation management, helping both clinicians and patients to better understand the likelihood of success before embarking on the IVF/ICSI journey.

This study had several notable strengths that enhanced the validity and clinical applicability of our findings. To our knowledge, this is the first study to develop and validate a prediction model specifically for the CLBR in women with OMAs following EST and subsequent IVF/ICSI treatment, addressing a critical gap in reproductive medicine. Second, we employed a rigorous statistical methodology adhering to the TRIPOD guidelines, including a comprehensive three-stage variable selection process (univariate regression, LASSO regression, and multivariate logistic regression) to minimize overfitting and ensure the parsimony of the model. Third, Bootstrap resampling validation with 500 iterations provided a robust assessment of model stability and reproducibility, confirming the reliability of our predictions. Fourth, we used CLBR as the primary outcome, which is more clinically meaningful than the single-cycle pregnancy rate, as it captures the full reproductive potential of all embryos from a single oocyte retrieval cycle. Finally, the model demonstrated consistent performance across both the training and validation cohorts, with good discrimination (AUC 0.849-0.853), acceptable calibration, and favorable clinical utility, as evidenced by decision curve analysis, supporting its potential for clinical implementation.

This study has some limitations that should be considered when interpreting its findings. First, the retrospective, single-center design may limit its generalizability. Although our sample size of 194 patients was adequate, larger multicenter cohorts are required for robustness. Second, our model excluded variables such as genetic polymorphisms, inflammatory biomarkers, and surgical history due to data constraints. Future studies should include these factors to improve performance. Third, the finding that a history of previous live birth is a negative predictor requires cautious interpretation, possibly reflecting endometriosis progression, selection bias, or unmeasured confounders such as the time interval since the last live birth. External validation is essential for confirming this association. Another limitation of this study is the statistically significant difference in two key ovarian reserve markers—basal testosterone and AFC—between the training and validation sets, despite random allocation. This baseline imbalance could potentially challenge the assessment of the model’s generalizability. Ideally, methods such as Propensity Score Matching or rigorous covariate adjustment should be employed to eliminate such confounding bias. Due to the retrospective nature and limited sample size of our study, we did not implement these more complex statistical corrections. Although our model demonstrated good predictive performance on the validation set despite the baseline differences, which to some extent proves its robustness, we must acknowledge that this methodological shortcoming may limit the direct applicability of our findings. Future research, particularly in larger prospective cohorts, should prioritize ensuring comparability between groups through strict matching or stratified designs to obtain more reliable validation results.

While this study establishes a foundation for individualized CLBR prediction in women with OMAs after EST and IVF/ICSI, several directions warrant further investigation. First, external validation in independent multicenter cohorts across diverse geographic regions and clinical settings is essential to confirm the generalizability and robustness of the nomogram. Second, prospective studies are needed to evaluate the model’s performance in real-world clinical practice and assess its impact on clinical decision-making, patient satisfaction, and reproductive outcomes. Third, the incorporation of emerging biomarkers (e.g., AMH kinetics, inflammatory markers, and genetic polymorphisms) and advanced analytical techniques (e.g., machine learning algorithms and artificial intelligence) may further enhance predictive accuracy and enable dynamic risk stratification. Finally, extending this prediction framework to other endometrioma management strategies (e.g., laparoscopic cystectomy and expectant management) and investigating the cost-effectiveness of nomogram-guided treatment selection represent important areas for future research to optimize fertility care in women with OMAs.