Explainable AI-Based Robust Prediction of Oocyte Retrieval Using AMH Data in ART Cycles
Abstract — Missing anti-Müllerian hormone (AMH) values limit the application of ovarian-response prediction models to real-world assisted reproductive technology data. This study evaluated whether artificial intelligence (AI)-based AMH imputation could preserve prediction of oocyte retrieval during controlled ovarian stimulation (COS). A retrospective multicenter dataset of 27,435 COS cycles collected from 2018 to 2023 was analyzed. Linear regression, XGBoost, LightGBM, and deep neural network models were compared for AMH imputation and downstream oocyte-yield prediction. In the measured-AMH setting, XGBoost achieved R² = 0.4063, MAE = 3.9988, and AUC = 0.8374. XGBoost also provided the highest AMH-imputation R² (0.2267). When missing AMH values were progressively simulated, AUC remained at or above 0.80 until missingness reached approximately 35.5% but decreased markedly at higher missing rates. SHAP analysis identified AMH and age as the principal predictors and showed compensatory contributions from stimulation protocol, hormone administration, and infertility cause. AI-based imputation therefore retained clinically useful information despite modest biomarker reconstruction accuracy and may support ovarian-response prediction when clinical records are incomplete.
Keywords — Anti-Müllerian Hormone; Artificial Intelligence; Imputation; Oocyte Retrieval; Ovarian Response
I. Introduction
The number of oocytes retrieved during COS is a central indicator of ovarian response and affects the number of fertilized oocytes, blastocysts, and embryos available for transfer. Consequently, reliable prediction of oocyte yield is valuable for selecting stimulation strategies and counseling patients [2], [3].
AMH is one of the most informative markers of ovarian reserve and is routinely combined with age, antral follicle count, and treatment-related factors in ovarian-response prediction [4]. However, AMH may be unavailable because testing was omitted, results were not transferred between institutions, or repeated treatment cycles used incomplete records.
Complete-case analysis discards these cycles and can reduce statistical power and clinical generalizability. Machine-learning imputation may recover part of the missing biomarker information, but the clinically important question is whether an imputed value preserves the downstream prediction task rather than exactly reproducing the laboratory measurement [5].
II. Methods
A. Study Population and Variables
The source database contained 69,255 ART cycles collected from eight infertility hospitals. Records were irreversibly de-identified. Cycles using frozen oocytes were excluded. Because validation of imputation required known ground-truth AMH values, analyses began with cycles having recorded AMH. Cases with missing values in other required variables were then excluded, yielding 27,435 COS cycles for final analysis.
The input variables were age, AMH, cause of infertility, ovarian stimulation protocol, FSH administration, and hCG administration. Infertility cause was grouped as female factor, male factor, combined factors, or unknown. Stimulation protocol was grouped as natural, GnRH antagonist, or GnRH agonist. FSH and hCG represented administration records rather than serum concentrations.
B. Outcomes and Preprocessing
The primary continuous outcome was the number of oocytes retrieved. For discrimination analysis, cycles were classified as normal responders (≥8 oocytes) or low responders (
C. Model Development
Linear regression was used as a conventional baseline. XGBoost and LightGBM were selected to capture nonlinear feature interactions, and a deep neural network (DNN) with hidden layers of 128, 64, and 32 neurons was used as the deep-learning comparator. Hyperparameters were chosen by grid search and cross-validation. The dataset was divided into training (80%) and validation (20%) sets.
The same four model families were first assessed for AMH imputation. XGBoost, which provided the highest R², was then used to generate imputed AMH values in the missingness experiments. Oocyte-yield models were compared under three conditions: actual AMH, partially missing and imputed AMH, and no AMH.
D. Implementation Details
XGBoost used 30 estimators, maximum depth 7, and a learning rate of 0.1. The estimator count was selected through grid search and cross-validation rather than a fixed default. Because XGBoost complexity is jointly determined by estimator count, tree depth, and learning rate, the compact ensemble was retained to limit overfitting under the available feature set. LightGBM used gradient-boosted decision trees with maximum depth 7, 31 leaves, and a learning rate of 0.1. The DNN was trained for 50 epochs with batch size 256, Adam optimization, and MAE loss. Analyses were conducted in Python using scikit-learn, XGBoost, LightGBM, TensorFlow, and Keras. These models represented a linear baseline, two tree-based ensembles, and a neural network under an identical clinical feature set.
E. Missingness Experiment and Evaluation
To create a known validation target, AMH values were randomly masked in complete records. Missingness was increased from 0% to 100% in 10% increments. At each level, missing AMH was estimated from the remaining clinical variables and then supplied to the downstream oocyte-retrieval model. AMH-imputation and continuous oocyte-yield performance were evaluated using mean absolute error (MAE) and coefficient of determination (R²). Classification of normal versus low response was evaluated using the area under the receiver operating characteristic curve (AUC). SHAP values were calculated to quantify the relative contribution of each variable to XGBoost predictions [6]. XGBoost achieved the highest explained variance for AMH recovery (R² = 0.2267), whereas the DNN achieved the lowest MAE (1.8279 ng/mL). XGBoost was selected for the subsequent missingness analysis because it provided the strongest overall balance of explained variance, error, and interpretability. All models achieved AUC values above 0.83. XGBoost yielded the highest R² and lowest MAE, while LightGBM showed a marginally higher AUC. XGBoost was retained as the principal prediction model because it performed consistently across regression and classification metrics.
F. Interpretation of Evaluation Metrics
R² and MAE assessed continuous oocyte-count prediction, whereas AUC assessed discrimination between low and normal responders. For the missingness experiment, the measured-AMH XGBoost result (R² = 0.4063) was defined as the study-specific empirical ceiling because it represented the best observed performance under the same cohort, feature set, and data split without artificial AMH missingness. This benchmark is distinct from the mathematical maximum of R² = 1.0. The percentages in Tables III and IV indicate R² retention relative to this empirical ceiling.
III. Result
A. Study Population and Benchmark Performance
The final cohort included 27,435 COS cycles. Median patient age was 37 years, median AMH was 2.41 ng/mL, and the median number of oocytes retrieved was 10. Most cycles used a GnRH-antagonist protocol. The measured-AMH XGBoost model achieved R² = 0.4063, MAE = 3.9988, and AUC = 0.8374, establishing the benchmark for evaluating imputed AMH.
B. Robustness under Simulated AMH Missingness
Prediction performance declined progressively as AMH missingness increased. AUC decreased from 0.8374 with measured AMH to 0.8277, 0.8168, and 0.8086 at 10%, 20%, and 30% missingness, respectively. At 40% missingness, AUC fell to 0.7928. Interpolation indicated that the AUC = 0.80 threshold was crossed at approximately 35.5% missingness.
Performance deteriorated more substantially beyond 60% missingness. At 100% missingness, equivalent to excluding AMH entirely, the model retained an AUC of 0.7189 but R² fell to 0.1687 and MAE increased to 4.8904. These findings demonstrate that imputed AMH provides measurable downstream value when a moderate proportion of biomarker records is absent.
C. SHAP-Based Feature Interpretation
SHAP analysis identified actual or imputed AMH as the most influential feature and age as the second most influential. When AMH was removed, the relative importance of age increased substantially. FSH and hCG administration, stimulation protocol, and infertility cause provided additional information and partially compensated for missing AMH. Across comparable AMH levels, predicted oocyte yield was generally lower with advancing age and higher in cycles using the antagonist protocol or hCG administration. Male-factor infertility was associated with comparatively higher predicted oocyte numbers. The predicted response plateaued or declined at very high AMH concentrations, highlighting the importance of multivariable interpretation rather than reliance on AMH alone. The feature hierarchy remained clinically coherent: the measured or imputed biomarker dominated when available, whereas age became the principal substitute when AMH was removed.
IV. Discussion
This study demonstrates that the clinical utility of an imputed biomarker cannot be judged solely by the accuracy of biomarker reconstruction. Although the best AMH-imputation model explained only 22.7% of the variance, the imputed values retained enough information to support robust downstream prediction. Incomplete AMH records could therefore be included without simply discarding entire treatment cycles.
The missingness experiment provides an operational threshold for data-quality management. Prediction remained above AUC 0.80 until approximately 35.5% of AMH values were missing. When missingness exceeded 60%, performance declined more rapidly, suggesting that imputation should complement rather than replace biomarker testing in severely incomplete datasets. Explainability results also support the biological plausibility of the model. AMH and age remained the dominant factors, while treatment context and infertility cause contributed to supplementary information. This compensation explains why a model with modest AMH-reconstruction R² could preserve useful classification performance. Such a distinction is relevant to medical AI systems deployed across institutions with heterogeneous documentation practices.
A. Comparison with Complete-Case Approaches
Conventional complete-case modelling excludes every cycle lacking AMH and therefore assumes that the missing records are dispensable. The present framework instead used the observable treatment context to estimate the missing biomarker and explicitly quantified the point at which this strategy ceased to preserve the target level of discrimination. This makes the model more compatible with routine multicenter data than approaches developed only from fully observed datasets.
B. Clinical Implications
• Imputation can increase the usable sample size of multicenter ART databases and reduce bias from complete-case exclusion.
• Predictions remain clinically informative when AMH missingness is moderate, but uncertainty should be communicated as missingness increases.
• SHAP-based interpretation can show which available clinical variables are supporting a prediction when AMH is unavailable.
C. Limitations
The analysis was retrospective and based on data collected in a specific regional health-care environment. Artificial random masking may not fully reproduce the mechanisms that cause AMH to be missing in clinical practice. Drug exposure was represented as binary administration variables rather than dose or serum concentration, and external prospective validation remains necessary. Although the 30-estimator configuration was selected through the prespecified tuning procedure, a broader estimator-range and early-stopping sensitivity analysis was not performed. Therefore, the modest R² should not be attributed solely to estimator count, and hyperparameters should be retuned in independent datasets.
D. Future Work
Future studies should validate the threshold in independent clinical populations and evaluate non-random missingness caused by patient selection, hospital workflow, or repeated cycles. Incorporating antral follicle count, medication dose, ultrasound images, and longitudinal measurements may improve both AMH reconstruction and prediction of downstream outcomes such as pregnancy and live birth.
V. Conclusion
AI-based imputation of missing AMH values preserved useful prediction of oocyte retrieval in 27,435 COS cycles. XGBoost achieved the highest AMH-imputation R² and strong downstream prediction using measured AMH. Predictive AUC remained at least 0.80 until AMH missingness reached approximately 35.5%, while SHAP analysis showed that age and treatment-related variables compensated for unavailable AMH. The framework offers a practical approach for maintaining ovarian-response decision support in incomplete real-world ART datasets.
Table 1. BASELINE CHARACTERISTICS
| Characteristic | Value |
|---|---|
| Cycles analyzed | 27,435 |
| Age, years | 37 [34–40] |
| AMH, ng/mL | 2.41 [1.26–4.39] |
| Oocytes retrieved | 10 [6–15] |
| Natural protocol | 80 |
| GnRH antagonist | 20,904 |
| GnRH agonist | 6,451 |
| Female-factor infertility | 10,704 |
| Male-factor infertility | 3,947 |
| Combined factors | 2,082 |
| Unknown cause | 10,702 |
Table 2. AMH IMPUTATION PERFORMANCE
| Model | R² | MAE |
|---|---|---|
| Linear regression | 0.2004 | 1.8720 |
| XGBoost | 0.2267 | 1.8331 |
| LightGBM | 0.2201 | 1.8378 |
| DNN | 0.2193 | 1.8279 |
Table 3. OOCYTE PREDICTION USING ACTUAL AMH
| Model | R² | R² retention | MAE | AUC |
|---|---|---|---|---|
| Linear regression | 0.3710 | 91.3% | 4.3174 | 0.8318 |
| XGBoost | 0.4063 | 100.0% | 3.9988 | 0.8374 |
| LightGBM | 0.4029 | 99.2% | 4.0150 | 0.8376 |
| DNN | 0.3847 | 94.7% | 4.0123 | 0.8347 |
Table 4. PERFORMANCE ACROSS AMH MISSINGNESS
| Missing AMH | R² | R² retention | MAE | AUC |
|---|---|---|---|---|
| 0% | 0.4063 | 100.0% | 3.9988 | 0.8374 |
| 10% | 0.3786 | 93.2% | 4.0835 | 0.8277 |
| 20% | 0.3561 | 87.6% | 4.1658 | 0.8168 |
| 30% | 0.3357 | 82.6% | 4.2602 | 0.8086 |
| 40% | 0.3038 | 74.8% | 4.3556 | 0.7928 |
| 50% | 0.2745 | 67.6% | 4.4906 | 0.7824 |
| 60% | 0.2572 | 63.3% | 4.5453 | 0.7692 |
| 70% | 0.2421 | 59.6% | 4.6697 | 0.7501 |
| 80% | 0.2196 | 54.0% | 4.7436 | 0.7450 |
| 90% | 0.1987 | 48.9% | 4.8246 | 0.7281 |
| 100% | 0.1687 | 41.5% | 4.8904 | 0.7189 |
Table 5. QUALITATIVE SHAP FEATURE HIERARCHY
| AMH condition | Leading SHAP variables |
|---|---|
| Measured AMH | AMH > age > FSH > infertility cause |
| Imputed AMH | Imputed AMH > age > FSH > infertility cause |
| Without AMH | Age > FSH > infertility cause > hCG/protocol |
References
[1] J. Lee, H.-J. Park, S.-G. Yang, and K.-Y. Kim, “AI-based imputation of anti-Müllerian hormone enables robust prediction of oocyte retrieval during controlled ovarian stimulation,” J. Assist. Reprod. Genet., vol. 43, pp. 1393–1402, 2026.
[2] M. Fanton, J. H. Cho, V. L. Baker, and K. Loewke, “A higher number of oocytes retrieved is associated with an increase in fertilized oocytes, blastocysts, and cumulative live birth rates,” Fertil. Steril., vol. 119, pp. 762–769, 2023.
[3] P. Wang et al., “The association between the number of oocytes retrieved and cumulative live birth rate in different female age strata,” Sci. Rep., vol. 13, Art. no. 14516, 2023.
[4] Y. Liu, Z. Pan, Y. Wu, J. Song, and J. Chen, “Comparison of anti-Müllerian hormone and antral follicle count in the prediction of ovarian response,” J. Ovarian Res., vol. 16, Art. no. 117, 2023.
[5] M. Kazijevs and M. D. Samad, “Deep imputation of missing values in time series health data: A review with benchmarking,” J. Biomed. Inform., vol. 144, Art. no. 104440, 2023.
[6] S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in Proc. 31st Conf. Neural Information Processing Systems (NIPS), Long Beach, CA, USA, 2017, pp. 4765–4774.
[7] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” in Proc. ACM SIGKDD, 2016, pp. 785–794.
[8] G. Ke et al., “LightGBM: A highly efficient gradient boosting decision tree,” in Proc. NIPS, 2017, pp. 3149–3157.