Assessment and prediction models for the quantitative and qualitative reserve of the ovary using machine learning.

A total of 442 patients who underwent ART treatment from June 2021 to January 2023 were involved in this study. Table  1 summarizes the patient characteristics and feature values for machine learning. Medical records were collected, including age, BMI, gravidity, parity, menstrual cycle regularity, history of smoking, endometriosis, PCOS, history of ovarian cystectomy or oophorectomy, total number of oocyte retrieval cycles, AMH, FSH, E2, motile sperm count, and age of husband. Residual serum on the hCG trigger day was collected and levels of AGEs, sRAGE, GDF9, BMP15, d-ROMs, BAP, and zinc were measured in our laboratory. Binary classification models that assess the quantity of the ovarian reserve as the number of oocytes retrieved were developed with 19 features, of which 14 were retrieved from medical records and five were obtained by residual serum analysis, or with 14 features retrieved from medical records without data from residual serum analysis. Several models were created using 15 algorithms, and the four models with the highest AUC are shown in Table  2 . There was no significant difference between the models developed with and without data from residual serum analysis. Focusing on the random forest model, the feature values used for model creation were narrowed down to maximize the AUC and ACC (Fig.  2 ). The best-performing model was created with five features and had an AUC of 0.9101 and ACC of 0.8556. The values of feature importance and individual interpretations are shown in Fig.  3 . AMH had the highest feature importance, followed by FSH, age, history of ovarian cystectomy, and parity (Fig.  3 a). AMH was positively correlated with model output, while FSH, age, and history of ovarian cystectomy were negatively correlated with model output (Fig.  3 b). Table 2 AUC and ACC of models for assessing the quantity of the ovarian reserve Medical records Medical records with additional data P -value Random Forest Classifier AUC 0.8957 0.8889 0.713 ACC 0.8523 0.8427 0.814 Light Gradient Boosting Machine AUC 0.8903 0.8836 0.723 ACC 0.8421 0.8320 0.813 Gradient Boosting Classifier AUC 0.8862 0.8785 0.727 ACC 0.8393 0.8388 0.524 Logistic Regression AUC 0.8531 0.8437 0.712 ACC 0.7972 0.7770 0.900 To determine the necessity of data from residual serum analysis, in addition to data retrieved from medical records, several models were built with or without such additional data and compared. Models were created using 15 algorithms, and the four models ranked highest in terms of AUC are shown AUC; area under the receiver operating characteristic curve, ACC; accuracy AUC and ACC of models for assessing the quantity of the ovarian reserve To determine the necessity of data from residual serum analysis, in addition to data retrieved from medical records, several models were built with or without such additional data and compared. Models were created using 15 algorithms, and the four models ranked highest in terms of AUC are shown AUC; area under the receiver operating characteristic curve, ACC; accuracy Fig. 2 AUC and ACC of models for assessing the quantity of the ovarian reserve. The number of feature values used to create models was increased sequentially from one to 19, and the combination of feature values that maximized the AUC and ACC was extracted. The best-performing model was created with five features, which had an AUC of 0.9101 and ACC of 0.8556 AUC and ACC of models for assessing the quantity of the ovarian reserve. The number of feature values used to create models was increased sequentially from one to 19, and the combination of feature values that maximized the AUC and ACC was extracted. The best-performing model was created with five features, which had an AUC of 0.9101 and ACC of 0.8556 Fig. 3 ( a ) The values of feature importance of the best-performing model for the quantity of the ovarian reserve. The average values of feature importance for the random forest model created using five features are shown. ( b ) The SHAP values of the best-performing model for the quantity of the ovarian reserve. A representative bee swarm plot is shown in which the SHAP values of the dataset were calculated from the random forest model. The dot color indicates a feature value from low (blue) to high (red). The SHAP value demonstrates the impact on model output. History of ovarian cystectomy is a categorical feature, and the red and blue dots indicate a history and no history of ovarian cystectomy, respectively ( a ) The values of feature importance of the best-performing model for the quantity of the ovarian reserve. The average values of feature importance for the random forest model created using five features are shown. ( b ) The SHAP values of the best-performing model for the quantity of the ovarian reserve. A representative bee swarm plot is shown in which the SHAP values of the dataset were calculated from the random forest model. The dot color indicates a feature value from low (blue) to high (red). The SHAP value demonstrates the impact on model output. History of ovarian cystectomy is a categorical feature, and the red and blue dots indicate a history and no history of ovarian cystectomy, respectively Binary classification models that assess the quality of the ovarian reserve as the number of good-morphology embryos were developed with 21 features, of which 16 were retrieved from medical records and five were obtained by residual serum analysis, or with 16 features retrieved from medical records without data from residual serum analysis. Several models were created using 15 algorithms, and the four models with the highest AUC are shown in Table  3 . In contrast to the models assessing the quantity of the ovarian reserve, models developed with data from residual serum analysis had a significantly higher AUC and ACC than those developed without such data. Focusing on the random forest model, the feature values used for model creation were narrowed down to maximize the AUC and ACC (Fig.  4 ). The best-performing model was created with 14 features and had an AUC of 0.7983 and ACC of 0.7762. The values of feature importance and individual interpretations are shown in Fig.  5 . AMH had the highest feature importance, followed by BMP15, motile sperm count, FSH, age, OSI, zinc, age of husband, GDF9, AGEs/sRAGE, E2, gravidity, menstrual cycle regularity, and PCOS (Fig.  5 a). AMH, BMP15, zinc, and AGEs/sRAGE were positively correlated with model output, while FSH, age, OSI, GDF9, and E2 were negatively correlated with model output (Fig.  5 b). Table 3 AUC and ACC of models for assessing the quality of the ovarian reserve Medical records Medical records with additional data P -value Random Forest Classifier AUC 0.7448 0.7742 0.043 ACC 0.7244 0.7530 0.023 Light Gradient Boosting Machine AUC 0.7446 0.7668 0.077 ACC 0.7143 0.7411 0.030 Gradient Boosting Classifier AUC 0.7400 0.7798 0.009 ACC 0.7214 0.7512 0.010 Logistic Regression AUC 0.7219 0.7501 0.045 ACC 0.6774 0.7095 0.008 To determine the necessity of data from residual serum analysis, in addition to data retrieved from medical records, several models were built with or without such additional data and compared. Models were created using 15 algorithms, and the four models ranked highest in terms of AUC are shown AUC; area under the receiver operating characteristic curve, ACC; accuracy AUC and ACC of models for assessing the quality of the ovarian reserve To determine the necessity of data from residual serum analysis, in addition to data retrieved from medical records, several models were built with or without such additional data and compared. Models were created using 15 algorithms, and the four models ranked highest in terms of AUC are shown AUC; area under the receiver operating characteristic curve, ACC; accuracy Fig. 4 AUC and ACC of models for assessing the quality of the ovarian reserve. The number of feature values used to create models was increased sequentially from one to 21, and the combination of feature values that maximized the AUC and ACC was extracted. The best-performing model was created with 14 features, which had an AUC of 0.7983 and ACC of 0.7762 AUC and ACC of models for assessing the quality of the ovarian reserve. The number of feature values used to create models was increased sequentially from one to 21, and the combination of feature values that maximized the AUC and ACC was extracted. The best-performing model was created with 14 features, which had an AUC of 0.7983 and ACC of 0.7762 Fig. 5 ( a ) The values of feature importance of the best-performing model for the quality of the ovarian reserve. The average values of feature importance for the random forest model created using 14 features are shown. ( b ) The SHAP values of the best-performing model for the quality of the ovarian reserve. A representative bee swarm plot is shown in which the SHAP values of the dataset were calculated from the random forest model. The dot color indicates a feature value from low (blue) to high (red). The SHAP value demonstrates the impact on model output. Menstrual cycle and PCOS are categorical data. Red dots indicate a regular cycle or complicating PCOS, while blue dots indicate an irregular cycle or no complicating PCOS ( a ) The values of feature importance of the best-performing model for the quality of the ovarian reserve. The average values of feature importance for the random forest model created using 14 features are shown. ( b ) The SHAP values of the best-performing model for the quality of the ovarian reserve. A representative bee swarm plot is shown in which the SHAP values of the dataset were calculated from the random forest model. The dot color indicates a feature value from low (blue) to high (red). The SHAP value demonstrates the impact on model output. Menstrual cycle and PCOS are categorical data. Red dots indicate a regular cycle or complicating PCOS, while blue dots indicate an irregular cycle or no complicating PCOS

All procedures in this study were approved by the institutional review boards (authorization reference number, 3594-11) and signed informed consent was obtained from each patient. Medical records of patients who underwent ART treatment at the University of Tokyo Hospital and Phoenix ART Clinic from June 2021 to January 2023 were retrospectively collected. The inclusion criteria for candidates were as follows: (1) complete records shown in Tables  1 and (2) undergoing oocyte retrieval, and (3) a sufficient amount of residual serum. The exclusion criteria for candidates were as follows: (1) severe internal disease or malignancy, (2) received anticancer drugs or radiation therapy, and (3) only undergoing oocyte cryopreservation. A total of 442 patients were involved in this study. All patients had received controlled ovarian stimulation with clomiphene citrate, human menopausal gonadotropin (150–300 IU/day; HMG TEIZO or Gonapure; ASKA Pharmaceutical Co., Tokyo, Japan), or recombinant human follicle-stimulating hormone (FSH; 150–300 IU/day; Gonalef; Merck Biopharma Co., Tokyo, Japan) with downregulation using gonadotropin-releasing hormone (GnRH) agonist (Nasanyl; Pfizer Japan, Tokyo, Japan), GnRH antagonist (Ganirest; MSD K.K, Tokyo, Japan), or progestin. When leading follicles reached 18–20 mm in diameter, human chorionic gonadotropin (hCG; 10,000 IU; HCG Mochida; Mochida Pharmaceutical Co., Tokyo, Japan) was injected to trigger ovulation. Oocytes were retrieved 34 h after hCG administration. Embryologists conducted the fertilization procedure and monitored embryo development. Embryo quality was assessed morphologically according to the criteria established by the Istanbul Consensus Workshop on Embryo Assessment [ 27 ]. Good-quality embryos were defined as follows: (1) seven or more blastomeres and grade II or higher (Veeck criteria [ 28 ]) on day 3, or (2) grade III or higher and BB or higher (Gardner criteria [ 29 ]) on day 5/6. Patient characteristics and feature values for the machine learning dataset are presented in Table  1 . Table 1 Patient characteristics and feature values for the dataset Characteristic Median (range) N (%) Type Age (years) 38 (25–49) Numerical BMI (kg/m 2 ) 21 (15–36) Numerical Gravidity 0 (0–5) Numerical Parity 0 (0–3) Numerical Menstrual cycle Regular 358 (81.0%) Categorical Irregular 84 (19.0%) Categorical Current or past smoking Yes 55 (12.4%) Categorical No 387 (87.6%) Categorical Endometriosis Yes 91 (20.6%) Categorical No 351 (79.4%) Categorical PCOS Yes 34 (7.7%) Categorical No 408 (92.3%) Categorical Ovarian cystectomy Yes 41 (9.3%) Categorical No 401 (90.7%) Categorical Oophorectomy Yes 4 (0.9%) Categorical No 438 (99.1%) Categorical AMH (ng/mL) 2.15 (0.02–15.81) Numerical Baseline FSH (mIU/mL) 8.4 (0.3–27.4) Numerical Baseline estradiol (pg/mL) 39.6 (5–147.5) Numerical Total number of oocyte retrieval cycles 2 (1–24) Numerical Number of oocytes retrieved 7.5 (0–34) Numerical 0–3 100 (22.6%) Categorical 4≤ 342 (77.4%) Categorical Number of fertilized eggs 4 (0–27) Numerical Number of good embryos 1 (0–14) Numerical 0 145 (32.8%) Categorical 1≤ 297 (67.2%) Categorical Motile sperm count (×10 6 ) 75 (0.2–871) Numerical Age of husband (years) 40 (26–61) Numerical AGEs (ng/mL) 4420 (810–10887) Numerical sRAGE (pg/mL) 963 (239–2606) Numerical AGEs/sRAGE 444 (72–2661) Numerical GDF9 (pg/mL) 120 (1–844) Numerical BMP15 (pg/mL) 22.5 (0.3–501.6) Numerical d-ROMs (U.CARR) 386.5 (208–778) Numerical BAP (µmol/L) 2198 (1026–2679) Numerical OSI 18.2 (8.9–34.4) Numerical Zinc (µg/dL) 80 (47.1–201.6) Numerical Data are presented as median (range). BMI; body mass index, PCOS; polycystic ovary syndrome, AMH; anti-Müllerian hormone, FSH; follicle-stimulating hormone, AGEs; advanced glycation-end products, sRAGE; soluble receptor of AGEs, GDF9; growth differentiation factor 9, BMP15; bone morphogenetic protein 15, d-ROMs; Diacron reactive oxygen metabolites, BAP; biological antioxidant potential, OSI; oxidative stress index calculated as d-ROMs/BAP×100 Patient characteristics and feature values for the dataset Total number of oocyte retrieval cycles Data are presented as median (range). BMI; body mass index, PCOS; polycystic ovary syndrome, AMH; anti-Müllerian hormone, FSH; follicle-stimulating hormone, AGEs; advanced glycation-end products, sRAGE; soluble receptor of AGEs, GDF9; growth differentiation factor 9, BMP15; bone morphogenetic protein 15, d-ROMs; Diacron reactive oxygen metabolites, BAP; biological antioxidant potential, OSI; oxidative stress index calculated as d-ROMs/BAP×100 Residual serum on the hCG trigger day was collected, aliquoted, and stored at -80 °C for further analysis. Serum concentrations of advanced glycation-end products (AGEs), soluble receptor of AGEs (sRAGE), growth differentiation factor 9 (GDF9), and bone morphogenetic protein 15 (BMP15) were measured using an OxiSelect AGE Competitive ELISA Kit (Cell BioLabs, Inc., San Diego, CA, USA), human RAGE Quantikine ELISA Kit (R&D Systems, Minneapolis, MN, USA), Human GDF9 ELISA Kit (LifeSpan BioSciences, Inc., Lynnwood, WA, USA), and Human BMP15 ELISA Kit (CUSABIO, Houston, TX, USA), respectively. Oxidative stress and antioxidant potential were measured using a Diacron reactive oxygen metabolites (d-ROMs)/biological antioxidant potential (BAP) Kit and Free Radical Elective Evaluator (FREE Carrio Duo; Diacron International srl, Grosseto, Italy). The oxidative stress index (OSI) was calculated as d-ROMs/BAP×100 [ 30 ]. The zinc concentration was measured using a Metalloassay Zn LS Kit (Metallogenics Co., Chiba, Japan). All procedures were performed following the manufacturers’ instructions. The quantity of the ovarian reserve was classified by the number of oocytes retrieved as follows: (1) poor, 0–3 oocytes and (2) good, 4 or more oocytes, determined with reference to the Bologna criteria [ 31 ]. The quality of the ovarian reserve was classified by the number of good-morphology embryos as follows: (1) poor, no good-morphology embryo and (2) good, one or more good-morphology embryo. Models for binary classification of the ovarian reserve were developed using the PyCaret auto machine learning library ver. 2.3.10 in Python which automatically performed standardization, missing value imputation, data splitting, comparison of multiple models, and hyperparameter tuning [ 32 , 33 ]. The following 15 machine learning algorithms were trained and compared: Random Forest Classifier, Light Gradient Boosting Machine, Extreme Gradient Boosting, Gradient Boosting Classifier, Ada Boost Classifier, Logistic Regression, Extra Trees Classifier, Decision Tree Classifier, Ridge Classifier, Linear Discriminant Analysis, Naïve Bayes, Dummy Classifier, K Neighbors Classifier, Quadratic Discriminant Analysis, and Support Vector Machine. The dataset was randomly split into two groups, namely, the training (80%) and test (20%) datasets. Then, the training dataset was split 4:1 and cross-validated. Data imbalance was addressed by specifying the fix_imbalance parameter, which applies SMOTE. The area under the receiver operating characteristic curve (AUC) and accuracy (ACC) of the models were evaluated using the test dataset. The final outcomes of AUC and ACC were averages of 20 models created using differential seeds for generating random numbers (Fig.  1 a). Fig. 1 ( a ) Workflow for creating models. The dataset was randomly split into two groups, namely, the training (80%) and test (20%) datasets. Then, the training dataset was split 4:1 and cross-validated. The AUC and ACC of the models were evaluated using the test dataset. The final outcome of AUC and ACC were averages of 20 models created using differential seeds for generating random numbers. ( b ) Feature values used for supervised machine learning. The upper box presents feature values for assessment of the quantity of the ovarian reserve. The lower box presents feature values for assessment of the quality of the ovarian reserve. ( c ) Workflow for narrowing down the feature values. The number of feature values used to create models was increased sequentially from one to a certain number, and the combination of feature values that maximized the AUC and ACC was extracted ( a ) Workflow for creating models. The dataset was randomly split into two groups, namely, the training (80%) and test (20%) datasets. Then, the training dataset was split 4:1 and cross-validated. The AUC and ACC of the models were evaluated using the test dataset. The final outcome of AUC and ACC were averages of 20 models created using differential seeds for generating random numbers. ( b ) Feature values used for supervised machine learning. The upper box presents feature values for assessment of the quantity of the ovarian reserve. The lower box presents feature values for assessment of the quality of the ovarian reserve. ( c ) Workflow for narrowing down the feature values. The number of feature values used to create models was increased sequentially from one to a certain number, and the combination of feature values that maximized the AUC and ACC was extracted To assess the quantity of the ovarian reserve, binary classification models were built with 19 features, with 14 retrieved from medical records, including age, body mass index (BMI), gravidity, parity, menstrual cycle regularity, history of smoking, endometriosis, polycystic ovary syndrome (PCOS), history of ovarian cystectomy, history of oophorectomy, total number of oocyte retrieval cycles, AMH, FSH, and estradiol (E2), and five obtained by residual serum analysis, including ratio of AGEs to sRAGE (AGEs/sRAGE), GDF9, BMP15, OSI, and zinc. To assess the quality of the ovarian reserve, binary classification models were built with 21 features, with 16 retrieved from medical records, including age, BMI, gravidity, parity, menstrual cycle regularity, history of smoking, endometriosis, PCOS, history of ovarian cystectomy, history of oophorectomy, total number of oocyte retrieval cycles, AMH, FSH, E2, motile sperm count, and age of husband, and five obtained by residual serum analysis, including AGEs/sRAGE, GDF9, BMP15, OSI, and zinc. To determine the necessity of the data from residual serum analysis, in addition to those retrieved from medical records, several models were built with or without such additional data and compared. The feature values used to create the models are summarized in Fig.  1 b. The best-performing model was selected based on its AUC and ACC. The best-performing machine learning method was used hereinafter. To maximize the AUC and ACC of models, feature values used for model creation were narrowed down. The number of feature values used to create the models was increased sequentially from one to a certain number, and the combination of feature values that maximized the AUC and ACC was extracted (Fig.  1 c). The Shapley Additive exPlanations (SHAP) library was used to provide local interpretable explanations for the models [ 34 ]. All statistical analyses were performed using JMP Pro 15 software (RRID: SCR_022199; SAS Institute Inc., Cary, NC, USA). All numerical data were presented as medians and ranges, and categorical data were presented as counts and percentages (Table  1 ). In the experiment investigating whether the inclusion of the data from residual serum analysis influenced the performance of machine learning models, the Student’s t -test was performed on the mean AUC or ACC of the 20 developed models. P  < 0.05 was considered statistically significant.

We developed models to assess the quantity and quality of the ovarian reserve using machine learning. Our models are more accurate than currently popular methods for predicting the ovarian reserve. Furthermore, they can assess the ovarian reserve using only information obtained from a medical interview and single blood sampling. They can provide the future expectancy to patients seeking ART treatment and also provide information about the ovarian condition to women who may desire to have children in the future. Enabling easy measurement of the ovarian reserve with this model would allow a greater number of women to engage in preconception care and facilitate the delivery of personalized medical treatment for patients undergoing infertility therapy.

In this study, we created models for binary classification of the ovarian reserve using machine learning methods developed with many collected feature values, and selected the best-performing model based on its AUC and ACC. The best-performing model for the quantity of the ovarian reserve was the random forest model created with five features, which had an AUC of 0.9101 and ACC of 0.8556. Selected feature values consisted only of data from medical records, including AMH, FSH, age, history of ovarian cystectomy, and parity. In contrast to the model for the quantity of the ovarian reserve, the best-performing model for the quality of the ovarian reserve was created with data not only from medical records but also from residual serum analysis, which had an AUC of 0.7983 and ACC of 0.7762. 14 features were selected, including AMH, BMP15, motile sperm count, FSH, age, OSI, zinc, age of husband, GDF9, AGEs/sRAGE, E2, gravidity, menstrual cycle regularity, and PCOS. We expected to further improve the prediction accuracy of models by adding factors that were suggested to be associated with the ovarian reserve in previous research, including AGEs, sRAGE, GDF9, BMP15, oxidative stress, and zinc. AGEs are endogenously produced through the Maillard reaction between reducing carbohydrates and free amino groups of proteins or other substrates and are exogenously obtained from the diet and smoking [ 35 – 37 ]. They are involved in several diseases as proinflammatory molecules. AGEs bind to cell membrane receptors for AGEs (RAGE) and activate proinflammatory signaling pathways, resulting in ovarian follicular growth arrest and ovulatory dysfunction [ 38 – 40 ]. The extracellular form of RAGE, called sRAGE, is a decoy receptor for AGEs and prevents inflammatory signaling [ 41 ]. The levels of AGEs and sRAGE in FF are associated with ART treatment outcomes, such as retrieved oocyte number, mature oocyte number, fertilization rate, embryo number, embryo quality, and pregnancy rate [ 42 , 43 ]. GDF9 and BMP15 are members of the transforming growth factor-β superfamily and are predominantly secreted by oocytes. These paracrine factors interact with surrounding somatic cells via activation of the SMAD signaling pathway, which is related to follicular development, ovulation, and oocyte maturation [ 44 – 46 ]. Supplementation of in vitro maturation culture media with GDF9 and/or BMP15 enhances oocyte quality and embryo formation [ 47 , 48 ]. Many studies have suggested that GDF9 and BMP15 levels in FF, serum, or granulosa cells correlate with oocyte quality and ART treatment outcomes and they have been proposed as biomarkers for predicting oocyte development [ 16 , 49 – 52 ]. Oxidative stress refers to an imbalance of the redox system caused by generation of excessive amounts of reactive oxygen species (ROS) or depletion of the antioxidant system [ 53 ]. Intracellular ROS, which are mainly generated as byproducts of the mitochondrial respiratory chain, have important roles in physiological functions, including follicular growth, oocyte maturation, ovulation, fertilization, and implantation, while excessive levels of ROS are harmful for such processes [ 54 – 56 ]. Increased levels of ROS in FF adversely affect oocyte development and ART treatment outcomes [ 57 , 58 ]. Zinc is the second most abundant trace element in the body, and an estimated 3000 proteins bind zinc to maintain structural integrity and function [ 59 ]. Zinc plays an important role in physiological processes including oocyte development, ovulation, spermatogenesis, fertilization, and fetal development [ 60 – 63 ]. The amount of zinc in FF is associated with mature oocytes retrieved from patients undergoing ART treatment [ 64 ]. Although many previous studies examined the potential roles of AGEs, sRAGE, GDF9, BMP15, oxidative stress, and zinc in predicting the ovarian reserve or ART treatment outcomes, no study has examined the role of these factors in combination. Machine learning has an advantage for analyzing and integrating large amounts of data, in contrast to traditional statistical approaches, and therefore we used a machine learning approach to create models. The ovarian reserve, which was defined as the number and quality of oocytes in the ovaries in this study, is essential for female fertility. To evaluate and predict the quantity of the ovarian reserve, we first created binary classification models that assess the number of oocytes retrieved. The classification criteria to divide patients into two groups (poor, 0–3 oocytes retrieved and good, 4 or more oocytes retrieved) were determined with reference to the Bologna criteria [ 31 ]. We examined whether the AUC and ACC of the model was improved by adding data obtained from residual serum analysis to data obtained from medical records. There was no significant difference between the models developed with and without the additional serum data (Table  2 ). This suggests that these additional factors are not useful for predicting the quantity of the ovarian reserve in the models created. The best-performing model was the random forest model created with five features, which consisted of data from only medical records, including AMH, FSH, age, history of ovarian cystectomy, and parity. The SHAP values, which demonstrate the impact on model output, suggested that AMH had a positive effect on model output, while the other factors had a negative effect on model output. The SHAP values are entirely plausible interpretations of ovarian reserve prediction [ 65 , 66 ]. The AUC of this model was 0.9101, which was higher than that of the model developed with AMH alone (AUC of 0.8600) (Figs.  2 and 3 ). AMH is currently considered one of the most useful and consensual markers for predicting the quantity of the ovarian reserve in clinical situations, especially in oocyte retrieval after ovarian stimulation [ 12 , 67 – 69 ]. Although a number of studies have attempted to improve prediction of the number of oocytes retrieved by using multiple markers, these combination models are no better than those developed using each marker alone [ 12 , 70 , 71 ]. More recently, it was shown that machine learning algorithms have an advantage over classical logistic regression for prediction of the number of oocytes retrieved and perform better when created using multiple markers than simple markers, with features including age, BMI, gravidity, parity, baseline E2, duration of ovarian stimulation, and total amount of gonadotropin used for ovarian stimulation [ 72 ]. All five predictors selected in our model are determined during usual check-ups before ART treatment. Therefore, it might be more appropriate to adapt the model based on these five factors than the method using only AMH for assessing the number of oocytes retrieved during ART treatment. To evaluate the quality of the ovarian reserve, we created binary classification models that assess the number of good-morphology embryos. Classification criteria were used to divide patients into two groups (poor, no good-morphology embryos and good, one or more good-morphology embryo). A blastocyst development rate of 30–40% or higher is recommended by ESHRE ART laboratory performance indicators [ 73 ]. We adopted the number of good-morphology embryos, rather than the ratio of good-morphology to total embryos, to assess the quality of the ovarian reserve because the discrepancy between the true condition and the ratio of good-morphology to total embryos is large, especially in patients with a low number of oocytes retrieved, e.g., for a patient who has only one oocyte retrieved, the ratio of good-morphology to total embryos can be as extreme as 0% or 100%. Similar to analysis of the quantity of the ovarian reserve, we examined whether the AUC and ACC of the model was improved by adding data obtained from residual serum analysis to data obtained from medical records. In contrast to the models for assessing the quantity of the ovarian reserve, the models developed with data from residual serum analysis had a significantly higher AUC and ACC than those developed without these data (Table  3 ). The best-performing model was the random forest model created with 14 features, including data obtained not only from medical records but also from residual serum analysis (Figs.  4 and 5 ). We created a more accurate model for assessing the quality of the ovarian reserve by adding data obtained from residual serum analysis, including AGEs, sRAGE, GDF9, BMP15, OSI, and zinc, to data obtained from medical records. The bee swarm plots of SHAP values suggested that AMH, BMP15, zinc, AGEs/sRAGE, gravidity, and menstrual cycle positively affected model output, while FSH, age, OSI, GDF9, and E2 negatively affected model output. Most of these factors had generally plausible effects on oocyte quality. While the BMP15 and GDF9 levels in FF are positively correlated with oocyte quality [ 49 – 52 ], our results showed that GDF9 was negatively correlated with oocyte quality. A study reported that the blastocyst formation rate is higher in patients with lower GDF9 levels in FF than in controls [ 74 ]. GDF9 and BMP15 are predominantly secreted by oocytes and interact with surrounding somatic granulosa cells to control Kit ligand expression. GDF9 suppresses expression of Kit ligand, whereas BMP15 enhances it [ 75 ]. Our results are possibly explained by the finding that GDF9 and BMP15 have opposite effects on expression of Kit ligand, leading to modification of oocyte quality. AGEs/sRAGE were positively correlated with oocyte quality in this study, although many studies suggest that AGEs/sRAGE in FF are negatively associated with ART treatment outcomes [ 42 , 43 ]. This discrepancy might be due to the difference between serum and FF or unknown confounding factors. Meanwhile, a study reported that a higher AGE level in skin tends to be associated with better ART treatment outcomes [ 76 ]. Furthermore, our results suggest that the levels of these factors in serum are useful predictors in a clinical setting, although most previous studies examining their roles in prediction of the ovarian reserve used their concentrations in FF or granulosa cells and thus oocyte retrieval was required to obtain the data. This study has several limitations. Our models may overfit a specific population due to the small number of participating patients and facilities. All data used in this study were obtained from patients undergoing ART treatment, who might have different properties than non-infertile women. In addition, it is unclear whether ovarian stimulation for ART treatment impacts the levels of AGEs, sRAGE, GDF9, BMP15, OSI, and zinc. One study showed that the serum levels of GDF9 and BMP15 are not influenced by ovarian stimulation [ 49 ]. The ovarian stimulation protocol affects the level of oxidative stress, while the balance between ROS and antioxidants is important for evaluating the whole oxidative status [ 77 ]. Studies incorporating more participants and facilities are required to resolve issues such bias and improve the generalization of the results. In this study, we developed models to assess the quantity and quality of the ovarian reserve using machine learning. To assess the quantity of the ovarian reserve, our model is more accurate than the currently popular model that uses only AMH. To assess the quality of the ovarian reserve, no effective evaluation method is currently available, while our model has a sufficient accuracy for clinical application. Furthermore, our models can assess the ovarian reserve using only information obtained from a medical interview and a single blood sampling. They can provide the future expectancy to patients seeking ART treatment and also provide information about the ovarian condition to women who may desire to have children in the future. More research is needed to implement our ovarian reserve-predicting models in patients undergoing ART and the general population, which would realize precision medicine and preconception care.

The occurrence of age-related infertility has been increasing in modern society, where delayed childbearing means couples may require fertility treatment. Despite the growing number of patients treated with assisted reproductive technology (ART), it is difficult to compensate for decreased fertility due to aging [ 1 , 2 ]. The age-related decline of fertility is caused by a reduction of the ovarian reserve, which is represented by the number and quality of oocytes in the ovaries [ 3 – 5 ]. Females are born with approximately 1–2 million oocytes. Through continuous follicular development, atresia, and ovulation, the stock of oocytes is depleted over time, resulting in a reduction of the oocyte number from 300,000 at puberty to 1,000 at menopause [ 6 , 7 ]. Oocyte quality is deeply connected with fertilization competence and embryo development, and also declines progressively with age [ 8 ]. This decline is confirmed by the finding that oocyte donation from younger women overcomes age-related infertility [ 9 – 11 ]. The number of oocytes is predicted by some indicators, such as the serum concentration of anti-Müllerian hormone (AMH) and the antral follicle count measured by ultrasonography, which are the most sensitive and reliable markers in clinical situations [ 12 ]. There is still room for improvement in the accuracy of ovarian reserve prediction using AMH. Antral follicle count is subject to variability depending on the examiner’s level of expertise, therefore, a more accurate prediction method is required. Furthermore, there are no clinically available predictors, except female age, of the quality of oocytes [ 13 ]. Oocyte quality can only be assessed after patients undergo oocyte retrieval by determining the oocyte’s morphology, the follicular fluid (FF) content, or gene expression of granulosa cells [ 14 – 20 ]. While many studies have investigated oocyte quality and suggested candidate indicators, these cannot be easily used in clinical practice, especially before ART treatment. In recent years, artificial intelligence and machine learning have begun to be used in the field of medicine. Machine learning has an advantage for analyzing and integrating large amounts of data, in contrast to traditional statistical approaches [ 21 – 26 ]. Our final goal is to produce prediction models of the ovarian reserve, which is defined as the number and quality of oocytes in the ovaries, using multiple predictors that can be easily determined, such as by a questionnaire and single blood sampling. As a first step to achieve this goal, we created several machine learning models based on the patients undergoing oocyte retrieval because there is sufficient information about patients’ characteristics and oocyte properties. Enabling easy measurement of the ovarian reserve with this model would allow a greater number of women to engage in preconception care and facilitate the delivery of personalized medical treatment for patients undergoing infertility therapy.